final report residential nonparticipant customer profile study

146
FINAL REPORT Residential Nonparticipant Customer Profile Study MA19X06-B-RESNONPART Date: February 6, 2020

Upload: others

Post on 10-Dec-2021

2 views

Category:

Documents


0 download

TRANSCRIPT

FINAL REPORT

Residential Nonparticipant

Customer Profile Study MA19X06-B-RESNONPART

Date: February 6, 2020

DNV GL – www.dnvgl.com February 6, 2020 Page i

Table of contents

1 EXECUTIVE SUMMARY ..................................................................................................... 1

1.1 Study purpose and objectives 1

1.2 Key findings, implications, recommendations, and considerations 3 1.2.1 Comparison to the Market Barriers Study 6

1.3 Methodology overview 8 1.3.1 Provisions 9

2 INTRODUCTION ............................................................................................................ 12

2.1 Study purpose, objectives, and research questions 12

2.2 Organization of report 14

3 METHODOLOGY AND APPROACH ..................................................................................... 15

3.1 Participation by block group list preparation 15

3.2 Participant-Nonparticipant list preparation 15 3.2.1 Renter flag 17 3.2.2 Modeling 21

3.3 Hot spot analysis 27

4 ANALYSIS AND RESULTS ............................................................................................... 28

4.1 ACS variable correlations 28 4.1.1 Hot spot analysis 30

4.2 Electric location participation findings 31 4.2.1 ACS variable correlations with electric participation 31 4.2.2 Electric participation block group-level models 34 4.2.3 Electric participation individual-level models 36

4.3 Electric savings/consumption findings 41 4.3.1 Block group electric analysis 41 4.3.2 Individual-level electric analysis – Statewide 45 4.3.3 Individual-level electric analysis – Enhanced 46 4.3.4 Location participation and savings/consumption inconsistencies 47

4.4 Gas location participation findings 50 4.4.1 ACS variable correlations with gas participation 50 4.4.2 Gas participation block group-level models 51 4.4.3 Gas participation individual-level models 56

4.5 Gas savings/consumption findings 60 4.5.1 Block group gas analysis 60 4.5.2 Individual-level gas analysis – Statewide 63 4.5.3 Individual-level gas analysis – Enhanced 64 4.5.4 Location participation and savings/consumption inconsistencies 66

5 CONCLUSIONS, RECOMMENDATIONS, AND CONSIDERATIONS ........................................... 69

5.1 Conclusions 69 5.1.1 Location participation rates for 2013-2017 are negatively associated with all three

term sheet variables: moderate-income households, renter households, and limited

English-speaking households. 69 5.1.2 PA efforts to obtain participation from large multifamily locations have resulted in

some inclusion of the populations outlined in the term sheet. 69 5.1.3 When participation is measured using 2013-2017 savings/consumption, it is positively

correlated with low income and multifamily at the block group level. 69 5.1.4 Location participation rates for 2013-2017 are affected by multiple factors. 69

DNV GL – www.dnvgl.com February 6, 2020 Page ii

5.1.5 Most of the variables investigated are correlated with each other, especially in the ACS data. 69

5.1.6 Term sheet-related populations are geographically clustered in urban areas. 70 5.1.7 Limited-English speakers are more likely to rent, and renters are less likely to

participate. 70 5.1.8 The effects of the examined variables on participation are similar in both electric and

gas markets. 70 5.1.9 The individual-level models are mostly consistent with the block group-level models. 70 5.1.10 Comparison to the Market Barriers Study 70

5.2 Recommendations 73

5.3 Considerations 74

6 APPENDIX A: PARTICIPANT/NONPARTICIPANT LIST PREPARATION ..................................... 75

7 APPENDIX B: LOW PARTICIPATION TABLE ....................................................................... 86

8 APPENDIX C: BIVARIATE MAPS .................................................................................... 122

9 APPENDIX D: CLOSE-UP HOTSPOT MAPS ....................................................................... 135

List of figures Figure 3-1. Renter flag definition ...................................................................................................... 19 Figure 4-1. Statewide ACS variable hot spot map, subset to urbanized land areas only ............................ 30 Figure 4-2. Block group-level electric participation correlations (term sheet variables) ............................. 32 Figure 4-3. Block group-level electric participation correlations (extra ACS variables) .............................. 33 Figure 4-4. Renter*Multifamily interaction on electric account participation probability – statewide ........... 38 Figure 4-5. Renter*Limited English interaction on electric account participation probability – enhanced data .................................................................................................................................................... 40 Figure 4-6. Renter*Multifamily interaction on electric account participation probability – enhanced data .... 41 Figure 4-7. Block group electric correlations ....................................................................................... 42 Figure 4-8. Multifamily interactions with low income and renter ............................................................ 44 Figure 4-9. Individual-level correlations with savings over consumption, statewide electric ...................... 45 Figure 4-10. Individual-level correlations with savings over consumption, enhanced electric .................... 46 Figure 4-11. Block group correlations, electric, no LIMF savings ............................................................ 49 Figure 4-12. Block group-level gas participation correlations (term sheet variables) ................................ 50 Figure 4-13. Block group-level gas participation correlations (extra ACS variables) ................................. 51 Figure 4-14. Expected gas location participation rate (renter * multifamily) ........................................... 55 Figure 4-15. Renter* multifamily interaction on gas account participation probability – statewide ............. 58 Figure 4-16. Renter* multifamily interaction on gas account participation probability – enhanced data ...... 60 Figure 4-17. Block group gas correlations .......................................................................................... 61 Figure 4-18. Interaction of low income and multifamily on gas savings/consumption ............................... 63 Figure 4-19. Interaction of rent and multifamily on gas savings/consumption ......................................... 63 Figure 4-20. Individual-level correlations with savings over consumption, statewide gas .......................... 64 Figure 4-21. Individual-level correlations with savings over consumption, enhanced gas .......................... 65 Figure 4-22. Block group correlations, gas, no LIMF savings ................................................................. 68 Figure 7-1. Community Outreach Metric formula and features .............................................................. 86 Figure 9-1. ACS variable hot spot map, closeup of Worcester and surrounding areas ............................. 136 Figure 9-2. ACS variable hot spot map, closeup of Springfield, Holyoke, and surrounding areas .............. 137 Figure 9-3. ACS variable hot spot map, closeup of Lawrence and surrounding areas .............................. 138 Figure 9-4. ACS variable hot spot map, closeup of Fitchburg and surrounding areas .............................. 139 Figure 9-5. ACS variable hot spot map, closeup of Fall River and surrounding areas .............................. 140 Figure 9-6. ACS variable hot spot map, closeup of Boston and surrounding areas ................................. 141

DNV GL – www.dnvgl.com February 6, 2020 Page iii

List of tables Table 1-1. Key findings of residential nonparticipant studies ................................................................... 6 Table 3-1. Final record counts by fuel and PA ..................................................................................... 16 Table 3-2. Number of electric accounts per location, 2014-2017 ........................................................... 18 Table 3-3. Number of gas accounts per location, 2014-2017 ................................................................ 18 Table 3-4. Number of accounts per location for Experian’s likely renters not flagged as a condo location.... 19 Table 3-5. Experian renter flag location comparison ............................................................................ 19 Table 3-6. Electric PA renter summary ............................................................................................... 20 Table 3-7. Gas PA renter summary ................................................................................................... 20 Table 3-8. PNP renter flag block group comparison with ACS block group percent renter.......................... 21 Table 4-1. ACS variable correlations .................................................................................................. 28 Table 4-2. Initial electric block group models ...................................................................................... 34 Table 4-3. Electric location participation models with some variables removed ........................................ 35 Table 4-4. Electric consumption weighted location participation models renter variable present/absent ..... 35 Table 4-5. Electric block group model results with all variables ............................................................. 36 Table 4-6. Individual-level correlations – statewide electric .................................................................. 37 Table 4-7. Electric account participation models coefficients – statewide ................................................ 38 Table 4-8. Individual-level correlations – Enhanced electric data ........................................................... 39 Table 4-9. Electric account participation models coefficients – enhanced data ......................................... 40 Table 4-10. Electric savings/consumption model results, block group-level ............................................. 43 Table 4-11. Individual-level savings over consumption models, statewide electric ................................... 45 Table 4-12. Individual-level savings over consumption models, enhanced electric ................................... 47 Table 4-13. Block group participation models comparison, electric ........................................................ 47 Table 4-14. Savings/consumption models, aggregated account-level data, electric .................................. 48 Table 4-15. Initial gas block group models ......................................................................................... 52 Table 4-16. Gas location participation models; renter variable present/absent ........................................ 53 Table 4-17. Gas consumption weighted location participation models renter variable present/absent ......... 53 Table 4-18. Gas block group model results with all variables ................................................................ 55 Table 4-19. Individual-level correlations – Statewide Gas..................................................................... 56 Table 4-20. Gas account participation models coefficients – statewide ................................................... 57 Table 4-21. Individual-level correlations – Enhanced Gas data .............................................................. 58 Table 4-22. Gas account participation models coefficients – enhanced data ............................................ 59 Table 4-23. Gas savings/consumption model results ........................................................................... 62 Table 4-24. Individual-level savings over consumption models, statewide gas ........................................ 64 Table 4-25. Individual-level savings over consumption models, enhanced gas ........................................ 66 Table 4-26. Block group participation models comparison, electric ........................................................ 66 Table 4-27. Savings/consumption models, aggregated account-level data, electric .................................. 67 Table 5-1. Key findings of residential nonparticipant studies ................................................................. 71 Table 6-1. Unique accounts in tracking data (Fuel*PA*year) ................................................................ 75 Table 6-2. Residential unique accounts by fuel and PA ......................................................................... 76 Table 6-3. C&I Unique accounts by fuel and PA (before removing nonresidential) ................................... 77 Table 6-4. Final record counts by fuel and PA ..................................................................................... 79 Table 7-1. Highest and lowest scoring town analysis ........................................................................... 88 Table 7-2. Community outreach metric table: Dual-fuel ....................................................................... 89 Table 7-3. Community outreach metric table: Gas .............................................................................. 99 Table 7-4. Community outreach metric table: Electric........................................................................ 109

DNV GL – www.dnvgl.com February 6, 2020 Page 1

1 EXECUTIVE SUMMARY

1.1 Study purpose and objectives

DNV GL conducted the Residential Nonparticipant Analysis Study (MA19X06-B-RESNONPART; “NPA”) for the

Massachusetts Program Administrators (PAs) and Energy Efficiency Advisory Council (EEAC) Consultants

from February 1, 2019, to December 20, 2019. The study’s overall purpose is to assess relationships

between residential participation rates and the variables specified in the PAs’ October 19, 2018 term sheet,

which stipulates:

“Special Focus on Renters, Moderate Income, Non-English Speaking, and Small Business Customers:

The Program Administrators will conduct tailored evaluations in 2019 that address participation

levels and potential unaddressed barriers for (a) businesses (small, medium and large) and (b)

residential customers by income levels and by non-English speaking populations (utilizing proxy

methods that do not rely on specific income or demographic information from Mass Save®

participants). The Program Administrators will leverage the existing EM&V framework, and present

full results of the studies to the EEAC.”1

This study was developed to provide information relevant to this term sheet stipulation. The study

objectives, as established by the working group, are as follows:

1. Quantify recent (2013-2017) levels of participation in PA programs for renters, moderate-income

customers, and non-English-speaking customers

2. Quantify how various factors (including but not limited to income level, language barriers, building

ownership, and single or multifamily) are associated with participation in Mass Save residential programs

3. Address the possibility of ecological fallacy2 when using block group-level variables

4. Establish a baseline level of participation that can be used to assess the effectiveness of PA efforts to

increase outreach

The main body of this report addresses the first 3 objectives. It presents the results of the tasks involving

statistical modeling to characterize the relationships between the term sheet variables, participation, and

other available demographic variables. The multiple levels of modeling also helped determine the extent of

ecological fallacy in the block group-level analyses. We also provide implications and recommendations

derived from these findings. Two additional, separate deliverables added to the information in this report:

• A “low participation” table that shows levels of the term sheet variables and location participation rates

by town and block group level. This deliverable addresses objectives 1 and 4. A copy of the town-level

data provided through this deliverable is included in Appendix B (Section 7) of this document.

• A series of bivariate maps that provide visualizations of areas with high concentrations of the term sheet

characteristics that have low participation rates. A copy of this deliverable is included in Appendix C of

this document.

This study provided a foundational analysis of participation patterns to identify underserved customers. This

analysis fed into a subsequent study called the “Residential Nonparticipant Market Characterization and

Barriers Study” (Market Barriers Study). The Market Barriers Study was a market assessment that

1 http://ma-eeac.org/wordpress/wp-content/uploads/Term-Sheet-10-19-18-Final.pdf

2 Ecological fallacy occurs when one makes conclusions about an individual based on group-level variables.

DNV GL – www.dnvgl.com February 6, 2020 Page 2

conducted primary research (including surveys and in-depth interviews) to explore market barriers for the

underserved customer segments, and to develop recommendations to more effectively reach those

segments.

This study is also related to the 2013-2017 Residential Customer Profile Study. Both reports have generally

consistent findings.

DNV GL – www.dnvgl.com February 6, 2020 Page 3

1.2 Key findings, implications, recommendations, and considerations

Study

Objective

Key finding Implication Consideration/Recommendations

1,2

Location participation rates for

2013-2017 are negatively

associated with all three term

sheet variables: moderate-

income households, renter

households, and limited English-

speaking households.

The populations represented by the

term sheet participate at lower levels

than other populations. This is not

surprising based on known barriers with

these populations.

Continue to explore program designs that will

help these populations overcome the

participation barriers they face.

1,2

Especially when participation is

measured as the ratio of 2013-

2017 savings to consumption,

there is evidence that the PAs

have had some success in

getting low-income and

multifamily (MF) locations to

participate.

The PAs’ success with low-income,

multifamily locations has enabled some

of the term sheet-related populations to

benefit from the programs because

multifamily is correlated with renters,

low-income, and limited English

speaking.

All PAs tracking participation at the individual

unit level would increase the accuracy of

evaluation results for MF buildings. For

implementers, knowing at a finer granularity

what they did in a building would help identify

opportunities to revisit buildings that have

participated in the past.

1,2

When participation is measured

using 2013-2017

savings/consumption, it is

positively correlated with low

income and multifamily at the

block group level. Additional

analysis revealed that savings

from the low income multifamily

programs account for these

positive correlations.

The PAs’ low income multifamily

programs are succeeding at achieving

depth of savings in areas with high

concentrations of low income

households and multifamily housing.

Future participation analyses should remain

aware of the strengths and limitations of each

way of measuring participation, and carefully

choose which definition to use in each analysis

based on the conceptual question being asked.

Consider whether the recently completed

multifamily study can answer any of the

outstanding questions, and consider

conducting a follow-up multifamily study to

further explore the outstanding questions.

DNV GL – www.dnvgl.com February 6, 2020 Page 4

Study

Objective

Key finding Implication Consideration/Recommendations

2

Location participation for 2013-

2017 is individually, negatively

correlated with the following

characteristics: low income and

pre-1950 construction and

individually; positively

correlated with average or

higher income and post-1950

construction.

Participation is affected by multiple

factors.

PAs could use this information for program

targeting.

2

Most of the variables

investigated are correlated with

each other, especially in the

ACS block group data.

Increasing participation among one of

the term sheet-related populations is

likely to also increase participation

among the other term sheet-related

populations.

Use the bivariate maps provided in this study

and other methods of identifying geospatial

areas to target program outreach efforts. The

bivariate maps provide a way to identify the

block groups that are the greatest outliers in

terms of the high incidence of term sheet

variables and low participation rates.

2

Term sheet-related populations

are geographically clustered in

urban areas.

The PAs’ community outreach

partnerships and continued

engagement with local community

stakeholders including the action

agencies and civic groups are likely to

have a positive impact on the term

sheet populations and benefit from

positive word of mouth from trusted

voices at the local level.

The PAs may benefit from examining their

time series billing and tracking data to identify

individuals who both participated in the

programs, and are in areas of high interest.

Alternatively or in addition, PAs could use

these geographic clusters to inform future

implementation of the Municipal and

Community Partnership Strategy.

DNV GL – www.dnvgl.com February 6, 2020 Page 5

Study

Objective

Key finding Implication Consideration/Recommendations

2

Limited English speakers are

more likely to rent, and renters

are less likely to participate.

Program designs that increase the

participation of renters also are likely to

increase participation of limited-English

speakers. However, additional outreach

or enhanced program design would be

required to fully address the population

of limited-English speakers.

Continue to explore program designs to

increase renter participation. Also, continue to

explore program designs that specifically

address limited-English speakers above and

beyond the renter programs.

2

The effects of the examined

variables on participation are

similar in both electric and gas

markets.

Participation barriers faced by the

populations identified by the term sheet

variables are probably similar for both

electric and gas participation.

Continue to communicate program successes

and setbacks between electric and gas

implementers to try to learn from each other’s

experiences.

3

The individual-level models are

mostly consistent with the block

group-level models.

Conclusions based on the block group-

level information are unlikely to suffer

from ecological fallacy for the analyses

examined in this study. However, one

should always be wary of the possibility

of ecological fallacy in the future.

Future analyses of the types of questions

examined in this study should utilize more

readily accessible block-group-level data.

4

In a separate deliverable, DNV

GL provided a metric that

reflects the ratio of the presence

of term sheet characteristics and

previous participation rates.

This metric was intended to help the

implementation team select

communities in which to focus

additional efforts.

Use the metric provided for the Mass Save

Municipal and Community Partnership Strategy

as the baseline for assessing the effectiveness

of future PA efforts to address the term sheet

populations.

DNV GL – www.dnvgl.com February 6, 2020 Page 6

1.2.1 Comparison to the Market Barriers Study

The Market Barriers Study conducted primary research (including over 1,700 surveys and in-depth

interviews) to explore nonparticipant characteristics, investigate participation barriers, and identify PA

engagement opportunities. Key findings from that study, and their relationship to this study’s findings, are

shown in Table 1-1. Findings from both studies were generally consistent, and there were no cases of

completely contradictory findings.

Table 1-1. Key findings of residential nonparticipant studies

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Self-reported program participation agreed 68% with participation

as determined by tracking data.

This is a high level of agreement

between self-reported participation

and the tracking data. In addition,

most (88%) of the self-reported

participants that were not identified

via the tracking data indicated their

participation took place outside the

5-year window of tracking data.

Participants identified via the

tracking data but not self-reporting

tended to be renters, live in

multifamily buildings, and/or live in

their current home for 4 or fewer

years.

Nonparticipants were more likely

to live in:

• Rental units

• Low to moderate income

households

• Households that speak a non-

English language or report

having limited English

proficiency

Renters, customers who live

in moderate income

households, and customers

living in limited-English

speaking households are less

likely to participate than

customers without these

characteristics.

The findings are consistent across

both studies.

DNV GL – www.dnvgl.com February 6, 2020 Page 7

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Nonparticipants also tend to:

• Have lower education levels

• Be less aware of Mass Save

• Do not fully trust government,

landlords, or free offers

• Prioritize time and resources

for other issues such as food

and well-being

• Need additional information or

understanding of Mass Save

offerings, processes, and

benefits

• Perceive energy efficiency as

irrelevant or not applicable to

them

No related findings

These findings highlight some of the

increased depth and additional topics

the Market Barriers study was able

to address beyond the data that

were available to the Nonparticipant

Customer Profile study.

Nonparticipant renters are more

likely to:

• Reside in lowest participating

Census tracts

• Self-identify as low income

Most of the variables

investigated are correlated

with each other, especially in

the ACS block group data.

Term sheet-related

populations are

geographically clustered in

urban areas.

The findings are consistent across

both studies.

DNV GL – www.dnvgl.com February 6, 2020 Page 8

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Moderate income nonparticipants

are more likely to reside in higher

participating Census tracts.

Most of the variables

investigated are correlated

with each other, especially in

the ACS block group data.

Term sheet-related

populations are

geographically clustered in

urban areas.

On the surface, these findings seem

contradictory. However, it is

important to note that the two

studies used different definitions of

moderate income. In the Market

Barriers study, the definition of

moderate income was $50,000 to

$75,000. In the Customer Profile

study, the definition of moderate

income was $40,000 to $59,999.

This means that part of the

moderate income group in the

Market Barriers study falls into the

“average or higher income” category

for the Customer Profile study. When

this cross-categorization is taken

into account, the findings are largely

consistent across both studies.

1.3 Methodology overview

DNV GL conducted the following tasks:

1. Prepare data. These analyses utilized three different data sets: block group data, individual-level data,

and enhanced individual-level data.

a. DNV GL developed the block group-level data as part of the overall project’s first deliverable (Low

Participation Table). We enhanced these data with ACS variables covering additional income

categories, construction date, heating fuel type, and urban/rural status.

b. We prepared the individual-level data as part of the Nonparticipant Profile Study’s second Deliverable

(Participant-Nonparticipant List; “PNP list”). This deliverable was primarily intended to serve as a

sample frame for surveys conducted as part of a parallel project. We enhanced this data with a

calculated renter flag.

c. We added the enhanced individual-level Experian data to the individual-level data set. However,

because these data are available for Eversource customers only, this data set has more variables but

fewer customers than the individual-level data.

2. Conduct block group-level analyses. These analyses included correlations and ordinary least squares

(OLS) regression models. Dependent variables for these models were location participation rate and the

ratio of savings to consumption (savings/consumption); independent variables were proportions of

DNV GL – www.dnvgl.com February 6, 2020 Page 9

households in block groups with certain characteristics such as moderate income. The independent

variables were based on ACS data.

3. Conduct individual-level analyses. We conducted analyses parallel to the block group-level analyses

using the individual-level data. The dependent variables for these models were account participation and

savings/consumption. Independent variables were individual-level variables reflecting demographic

characteristics available statewide through sources such as the Tax Assessor data (multifamily, year of

construction) and those available only for Eversource through Experian data (income, language).

Throughout the analysis phase, DNV GL referenced the 2013-2017 Residential Customer Profile Study report

which contains similar analyses. When discrepancies were found, we conducted extra analysis to resolve the

differences.

1.3.1 Provisions

The following provisions and limitations pertain to this study.

1. The core data used for this study only covers the five-year period from 2013 – 2017, and

participation in PA programs. The populations under consideration might have participated prior to

and/or after this window or may have participated in programs sponsored by organizations other than

the PAs.

2. This study excludes savings from several sources: upstream programs, behavioral programs, and

delivered fuel projects (the latter for some analyses). The upstream and behavioral programs account

for approximately 8% of the total residential electric and 14% of the total residential gas savings from

2013-2017. Delivered fuel savings were excluded because we do not have reliable information for the

total consumption of delivered fuels. Therefore, the ratios in this report that rely on savings (e.g.,

savings/consumption) are impossible to calculate for delivered fuels. Participation in delivered fuel

projects is included in our (binary) participation metrics.

3. There is no perfect way to measure or define “participation.” This study utilized four different

definitions, each of which has strengths and weaknesses.

a. Location participation measures whether a building participated. Location participation is similar to

coverage rate. It is best used to answer questions such as, “How many buildings did the program

touch, regardless of size?” A strength of location participation is that it is stable over time, so it is

useful for time-series analyses. Weaknesses of location participation are: It overstates participation

of multifamily (MF) buildings because a location is considered a participant if any sub-unit within the

building has participated; single-family (SF) homes count the same amount as MF buildings; and

building-level data grain does not match the ACS data grain, which is at the household level.

b. Consumption-weighted location participation uses location participation multiplied by the total

consumption associated with each location. This metric gives a sense of how much of the underlying

consumption the programs addressed. It is best used to account for the impact differences between

SF and MF locations and to answer questions such as, “How well has the program addressed large

customers?” The strengths of this variable are that it is stable over time, like location participation,

and that it differentiates between SF and MF buildings. Weaknesses are that it overstates

participation at MF locations to an even greater extent than location participation due to the effects

of very large MF buildings; it is not at the same grain as ACS variables; and it can be difficult to

interpret.

DNV GL – www.dnvgl.com February 6, 2020 Page 10

c. Account participation measures whether a unique account ID has participated at least once over a

certain timeframe. This variable is best used when one wants to answer questions about individuals

who participated, such as, “What income levels did past participants have?” Strengths of this metric

are that it is closest to the same grain as the ACS variables; it is directly accessible from PA tracking

data; and it is relatively easy to interpret. The main weakness of this variable is that it is not stable

over time. Account IDs change when new residents move in or out of a specific dwelling, so a high

proportion of dwellings with frequent resident turnover (e.g., rented apartments) will tend to lower

account participation rates by increasing the denominator while not changing the numerator.

Account participation also does not accurately capture participation at the individual housing unit

level in master-metered MF buildings.

d. Savings/consumption measures the ratio of savings achieved in a block group or for a specific

account between 2013-2017 and the average consumption for the block group or account from

2013-2017. The strengths of this metric are that it shows depth of participation and locations with

greater consumption (i.e., large MF buildings) have greater relative weight in the analyses. This also

creates a limitation: those large consumers can dominate block-group sums and averages while they

do not have any exceptional weighting in individual account-level analyses. Other limitations of this

metric are that it does not fully account for consumption added for new construction, and at the

account level, it does not fully capture savings that occurred at the same location under a different

account-holder.

4. Several limitations apply to the account-level analyses:

a. The individual and enhanced individual-level analyses are limited to unique 2017 accounts.

The analyses cover participation for these accounts at any time between 2013-2017. However, the

analyses do not cover 2013-2017 participation for accounts that were not active in 2017.

b. Any participation that occurred by accounts that did not exist in 2017 is not represented.

This particularly affects renters, who are a key population of this study. For example, if

someone living at a home participated in 2015, then moved out in 2016, and a new person moved in

in 2017, the 2015 participation would not be modeled.

c. Accurately accounting for participation in master metered multifamily buildings, especially

low-income multifamily buildings, has proven difficult with the data available from some

PAs. Some PAs track participation of individual units within multifamily buildings, but other PAs track

participation only at the whole building level. The evaluation team cannot account for data that does

not exist. We have used location and consumption-weighted location participation to partially account

for the whole-building approach of some PAs, but these participation variables have their own

limitations. Individual household characteristics also cannot be modeled at the building level.

d. The enhanced individual-level analyses are limited to Eversource accounts. The Experian

data with which we enhanced the data were only available for Eversource customers. We attempted

to integrate audit data from the other PAs that contains some similar data. However, the audit data

did not cover other PAs’ customer base sufficiently to aid the analyses in this study.

e. The method used to identify renters in the statewide data set is not completely accurate.

While this method does highly correlate with more rigorous ACS and Experian methods of

determining renter status, it is not a perfect correlation. The method will undercount long-tenured

renters and overcount changes to account status for other reasons (e.g., name change due to

marriage/divorce).

DNV GL – www.dnvgl.com February 6, 2020 Page 11

f. The ACS definition of multifamily (5+ units) does not fully align with Tax Assessor data.

The Tax Assessor data has categories for 4-8 units and over 8 units.

5. The block group and individual analyses utilize different data grains. The block group-level

analyses use location and consumption-weighted location participation. The individual-level analyses use

account participation. The difference in data grain means that caution should be exercised when directly

comparing the results across the levels of analysis. However, the fact that the results of all three

analyses are similar is positive. This indicates a level of consistency in the real-life situation that the

models are trying to capture. Furthermore, the previous memo that reported on block group-level

results included account participation, and the outcomes for account and location participation were

highly congruent in those findings.

6. The analyses presented in this study are not adversely affected by small populations meeting

the term sheet characteristics. The way to interpret the block group level correlations is “as the

proportion of households that meet term sheet characteristics increases, the rate of participation tends

to decrease.” That the term sheet households are relatively rare does not drive the relationship; as you

get a higher concentration of those households in a block group, we expect the (total) participation rates

in that block group to decrease. For the individual-level models, we are modeling the probability that a

particular house will participate based on the term sheet (and other) characteristics. The individual

model looks at the participation rate of the accounts with that characteristic, rather than the rate of the

characteristic in the participant population.

7. The study did not examine differential funding levels. PA funding for low income programs is not

necessarily uniform throughout the state. This study did not factor any such differences into the

analyses.

DNV GL – www.dnvgl.com February 6, 2020 Page 12

2 INTRODUCTION

2.1 Study purpose, objectives, and research questions

DNV GL carried out the Residential Nonparticipant Analysis Study (MA19X06-B-RESNONPART; “NPA”) for the

Massachusetts Program Administrators (PAs) and Energy Efficiency Advisory Council (EEAC) Consultants

from February 1, 2019, to September 30, 2019. The study’s overall purpose was to assess relationships

between participation rates and the variables specified in the PAs’ October 19, 2018 term sheet, which

stipulates:

“Special Focus on Renters, Moderate Income, Non-English Speaking, and Small Business Customers:

The Program Administrators will conduct tailored evaluations in 2019 that address participation

levels and potential unaddressed barriers for (a) businesses (small, medium and large) and (b)

residential customers by income levels and by non-English speaking populations (utilizing proxy

methods that do not rely on specific income or demographic information from Mass Save®

participants). The Program Administrators will leverage the existing EM&V framework, and present

full results of the studies to the EEAC.” 3

This study was developed to provide information relevant to the term sheet stipulation. The study

objectives, as established by the working group, are as follows:

1. Quantify recent levels of participation in PA programs for moderate-income customers, renters, and non-

English-speaking customers

2. Quantify how various factors (including but not limited to income level, language barriers, building

ownership, and single or multifamily) are associated with participation in Mass Save residential

programs

3. Address the possibility of ecological fallacy4 when using block group-level variables

4. Establish a baseline level of participation that can be used to assess the effectiveness of PA efforts to

increase outreach

This report addresses the first 3 objectives. It presents the results of the tasks involving statistical modeling

to characterize the relationships between the term sheet variables, participation, and other available

demographic variables. The multiple levels of modeling also helped determine the extent of ecological fallacy

in the block group-level analyses. We also provide implications and recommendations derived from those

findings. There are two additional, separate deliverables that add to the information presented in this report:

• A “low participation” table that shows levels of the term sheet variables and location participation rates

by town and block group level. This deliverable addresses objectives 1 and 4. A copy of the town-level

data provided by this deliverable is included in Appendix B (Section 7) of this document.

• A series of bivariate maps that provide visualizations of areas with high concentrations of the term sheet

characteristics and low participation rates. A copy of this deliverable is included in Appendix C of this

document.

This study provided a foundational analysis of participation patterns to identify underserved customers. This

analysis fed into a subsequent study called the “Residential Nonparticipant Market Characterization and

Barriers Study” (Market Barriers Study). The Market Barriers Study was a market assessment that

3 http://ma-eeac.org/wordpress/wp-content/uploads/Term-Sheet-10-19-18-Final.pdf

4 Ecological fallacy occurs when one makes conclusions about an individual based on group-level variables.

DNV GL – www.dnvgl.com February 6, 2020 Page 13

conducted primary research (including surveys and in-depth interviews) to explore market barriers for the

underserved customer segments, and to develop recommendations to more effectively reach those

segments.

DNV GL – www.dnvgl.com February 6, 2020 Page 14

2.2 Organization of report

The remainder of this report focuses on the first two objectives listed above, and is structured as follows:

• Section 3 – Methodology and approach

• Section 4 – Analysis and results

• Section 4.1 – ACS variable correlations

• Section 4.2 – Electric location participation findings

• Section 4.3 – Electric savings/consumption findings

• Section 4.4 – Gas location participation findings

• Section 4.5 – Gas savings/consumption findings

• Section 4.5.4 – Conclusions, recommendations, and considerations

• Section 6 – Appendix A – Participant/nonparticipant list preparation

• Section 7 – Appendix B – Low participation table

• Section 8 – Appendix C – Bivariate maps

• Section 9 – Appendix D - Close-up hotspot maps

DNV GL – www.dnvgl.com February 6, 2020 Page 15

3 METHODOLOGY AND APPROACH

This section provides more detail on the study tasks. In general, this study uses the definition of participant

from the MA Technical Reference Manual (TRM): a participant is “a customer who installs a measure through

regular program channels and receives any benefit (i.e., incentive) that is available through the program

because of their participation.”

3.1 Participation by block group list preparation

The study’s first objective was to quantify recent levels of participation of the populations defined by the

term sheet variables. This objective was achieved through a separate deliverable called the Low Participation

Table, also referred to as Deliverable 1. This is an Excel workbook that presents, by census block group, the

proportion of households that meet each term sheet variable criteria and the consumption-weighted location

participation of that block group. DNV GL also aggregates the block group data at the town5 level. Controls

allow the PAs to filter the table by electric PA and gas PA, and sort by town name and participation rate. The

workbook has separate worksheets for electric-only, gas-only, and dual-fuel participation. A copy of this

Excel book is embedded in Appendix B of this report.

The data used for the Low Participation Table includes PA billing and tracking records for 2013-2017,

processed by DNV GL’s MA data management team. After combining the individual years of data, we made

four changes to the data to prepare it for the Low Participation Table:

1. Using building type information from tax data and a string search on customer names (e.g. names that

included various permutations of “HOUSING AUTHORITY”), we identified housing authorities and

subsidized multifamily housing in the C&I billing data. These records were added to the residential billing

data to account for the fact that buildings with C&I accounts housing low-income tenants participate

through the PAs’ low-income initiatives. Without including the consumption of these buildings, the

denominator for any savings achieved calculations would be incorrect for low-income participants.

2. For accounts that had bad address information for one or more years but currently have good address

information, we back-filled that information to make sure that each account is assigned to the correct

block group for as many years as possible.

3. We removed a single outlier/error in the National Grid data that had 1-month consumption of greater

than 1 million kWh.

4. With guidance from the PAs, we removed non-gas MMBTU savings (i.e., MMBTU savings in the electric

tracking data for delivered fuel projects) from the tracking data due to inconsistencies. We did not,

however, remove delivered fuel participants from our participant counts.

3.2 Participant-Nonparticipant list preparation

The primary purpose of the Participant-Nonparticipant (PNP) list was to provide a population for a related

study to draw samples of current residents to survey. Key characteristics needed by the survey study were

addresses, block groups, and 2013-2017 participation status (as a binary variable). Because some

residentials buildings, particularly housing authorities, can be on C&I rates, DNV GL needed to merge C&I

and residential billing data to provide the survey team with a complete population. DNV GL was also able to

leverage this data for individual-level analyses as a secondary use. DNV GL prepared the Participant-Non-

5 Towns as defined by the Tax Assessor.

DNV GL – www.dnvgl.com February 6, 2020 Page 16

participant (PNP) list using the following process. This was a completely separate effort from the block group

list preparation.

1. Combined 2013 – 2017 residential tracking data6

2. Merged 2017 Residential and C&I billing data

a. Added geocode information

b. Added tax assessor information

c. Filtered out C&I accounts (e.g., streetlights and nonresidential locations)

3. Identified participants from the tracking data

4. Identified most recent account holder based on mid-2018 lists provided by PAs

The result of this process was a data file at the account data grain of unique residential and multifamily

accounts billed in 2017, with account and location participation variables based on the 2013-2017

(residential only) tracking data and with additional location-level data such as Tax Assessor information

appended. The final number of unique account IDs by fuel and PA are shown in Table 3-1. Detailed data

preparation methods are described in Appendix A.

Table 3-1. Final record counts by fuel and PA

Fuel

PA From Residential Billing

From C&I Billing1 Total

E Cape Light Compact 190,212 1,703 191,915

E Eversource 1,245,185 34,552 1,279,737

E National Grid 1,168,581 38,961 1,207,542

E Unitil 29,103 354 29,457

Total Electric 2,633,081 75,570 2,708,651

G Berkshire 35,223 621 35,844

G Columbia 348,951 3,262 352,213

G Eversource 315,619 3,375 318,994

G Liberty 51,867 147 52,014

G National Grid 845,449 8,989 854,438

G Unitil 16,641 134 16,775

Total Gas 1,613,750 16,528 1,630,278

Grand Total 4,246,831 92,098 4,338,929

1 C&I records were included to ensure we accounted for multifamily buildings; excluding accounts positively identified as non-residential such as

businesses, streetlights, and cellphone towers.

Besides the difference in data grain (block group versus account), there are three key differences between

the block group list and the PNP list. First, the block group list did not attempt to match specific participating

accounts with specific billed accounts; it summed all participation and all consumption. In contrast, the PNP

list does match specific account ids across billing and tracking databases. Second, the PNP list contains only

information from accounts billed in 2017. In contrast, the block group list contains information for accounts

6 While we included C&I billing data to ensure we had all potential residential buildings, we includined only residential tracking data.

DNV GL – www.dnvgl.com February 6, 2020 Page 17

billed anytime between 2013 and 2017. Finally, the PNP list contains more C&I records (billed in 2017) than

the block group list. The block group list added only C&I records that could be positively identified as

housing authorities or low-income housing. In contrast, the PNP list included all C&I records and only

excluded those that could be positively identified as not residential. Thus, the PNP list errs on the side of

including C&I account that actually are not residential. This was partly because the primary purpose of the

list was for survey sampling purposes and the surveys would have the ability to screen out truly non-

residential locations.

3.2.1 Renter flag

This section outlines the steps we used to prepare the data before identifying likely renters. The goal of this

data processing was to identify all accounts in the PNP list that are likely renters. The PAs did not have

direct information on renter status. Therefore, DNV GL had to implement a logical process to determine if an

account was a likely renter. This determination had to be based on data available for all PAs, which was

limited. We identified high turnover rate (i.e., a higher number of accounts associated with a location from

2014-2017) as a proxy for renter status.

This method has limitations and relies on several assumptions. We assumed that residents in rented

properties will move in and out more quickly than in non-rented properties. We set a threshold on the

number of account changes to represent this tendency. However, an account change does not necessarily

mean a renter turnover or even a sale or change of ownership, and the threshold approach underrepresents

long-tenured renters. Without additional information to draw on, we also assumed condos are owned rather

than rented, so rented condos are also underrepresented. Unfortunately, there is no good way to resolve

these limitations without conducting surveys.

However, our approach provides a means of assigning likely renter status using data readily available from

all PAs. Experian uses a much more robust approach to assigning likely renter status.7 This approach begins

with self-reported data from surveys, then includes a comparison between the names in the Tax Assessor

records and the current resident, and finally includes statistical interpolation to fill in otherwise missing data.

Despite the limitations of our approach, we agree with Experian’s assignments 90% of the time.8

Renter flag definition

The data processing began with the Participant-Nonparticipant (PNP) list produced as part of Deliverable 2.

This list was based on all accounts billed in 2017. Using the locations (latitude and longitude pairs) in the

PNP list, we identified all accounts associated with those locations from the combination of the 2015-2017

C&I billing data and the 2014-2017 residential billing data.

We then counted the number of 2014-2017 accounts associated with each location between 2014 and 2017

by PA and fuel type. We counted accounts by PA and fuel type at a location, as some locations will be served

by both a gas and an electric PA but others will not, resulting in variations in the number of expected

accounts at a given location.

Finally, lacking any other information, we assumed condos are owned and flagged condo locations by

looking for permutations of “CONDO” in the customer name and using the tax assessor building types that

correspond to condos. If any account at a given location was flagged as a condo, we classified the location

(latitude and longitude) as a condo.

7 Based on the data dictionary provided with Experian’s data.

8 For Eversource, which is the only PA for which we have Experian data.

DNV GL – www.dnvgl.com February 6, 2020 Page 18

Table 3-2 and Table 3-3 present the number of accounts per location for electric and gas accounts

respectively. For electric, most locations had two or fewer accounts within the four-year timespan of 2014-

2017. For gas, most locations had three or fewer accounts within the four years.

Table 3-2. Number of electric accounts per location, 2014-2017

Number of Accounts Percentage of Locations Cumulative Percentage

1 27% 27%

2 39% 66%

3 15% 82%

4 5% 87%

5 2% 89%

6 3% 92%

7 2% 93%

8 1% 94%

9 1% 95%

10+ 5% 100%

Table 3-3. Number of gas accounts per location, 2014-2017

Number of Accounts Percentage of Locations Cumulative Percentage

1 25% 25%

2 14% 39%

3 28% 67%

4 16% 83%

5 3% 85%

6 3% 89%

7 2% 91%

8 3% 93%

9 1% 95%

10+ 5% 100%

Next, we compared Experian data available for Evesrource, which has a flag for likely renters, to our data on

the number of accounts per location to determine the number of accounts at a location over a four year

period that indicates a likely renter. Table 3-4 presents the number of accounts per location for (Eversource)

records that Experian flagged as likely homeowners and likely renters, respectively. Ninety percent of

locations that Experian indicated were likely owned had 3 or fewer accounts between 2014 and 2017, while

56% of locations Experian indicated were likely rented had 4 or more accounts between 2014 and 2017. We

selected 4 or more accounts as our rental threshold. At this threshold, the majority of locations are rented

according to Experian and almost all of the locations with 3 or fewer accounts are reported as owned by

Experian.9

9 It is difficult to identify single-family homes that are renter-occupied.

DNV GL – www.dnvgl.com February 6, 2020 Page 19

Table 3-4. Number of accounts per location for Experian’s likely renters not flagged as a condo

location

Number of accounts

Percentage of Locations

Cumulative Percentage

1 14% 14%

2 20% 34%

3 10% 44%

4 10% 54%

5 7% 61%

6 7% 68%

7 5% 73%

8 5% 78%

9 3% 81%

10+ 19% 100%

Figure 3-1 presents how the initial renter flag was assigned. We processed electric and gas accounts

separately, and set the location to “rented” if either analysis indicated it. Finally, we checked our assignment

against the Experian “likely renter” assignment. Both assignments agreed 90% of the time (see Table 3-5).

Figure 3-1. Renter flag definition

Table 3-5. Experian renter flag location comparison

Category Percentage

Match 90%

Don't Match 10%

Other methods investigated

We investigated the feasibility of using two other methods to determine renters, both of which relied on

text-based matches on account owner names. The first alternative method compared the PA’s account

owner name to the tax assessor’s plot owner name. The second method attempted to track changes in the

account owner name(s) associated with a location across the years of available billing data. We determined

that both methods were impractical because of the limitations of matching text strings for the volume of

data involved in this project. The method we described using account ID is simpler, more reliable, and has a

90% agreement with Experian’s assignments. We determined that achieving an agreement rate better than

90% using the name matches is unlikely and even if possible, would have used a large portion of the

remaining project budget.

DNV GL – www.dnvgl.com February 6, 2020 Page 20

Renter flag summaries

Table 3-6 and Table 3-7 summarize the renter flag for the electric and gas PAs, respectively. The percentage

of accounts renting includes all the locations in the PNP list.

Table 3-6. Electric PA renter summary

PA Percentage of

locations renting

CLC 12%

Eversource 17%

National Grid 19%

Unitil 36%

Table 3-7. Gas PA renter summary

PA Percentage of

locations renting

Berkshire 8%

Columbia 13%

Eversource 20%

Liberty 30%

National Grid10 48%

Unitil 43%

As an additional diagnostic for the accuracy of our renter flag, DNV GL aggregated the renter flags up to the

block group level and computed the proportion of each block group our method assigned as renters. We

then compared those proportions to those reported in the ACS. Table 3-8 presents a distribution of how well

our renter flag corresponds to the ACS based on the absolute value of the difference between our block

group proportion and the ACS block group proportion. A plurality of the block groups is within 10% of each

other and the majority are within 20%.

Even if the flags were identical at the level of an individual, we expect some variation between the two

metrics due to differences in reporting grain and the included population. The ACS metric is based on

households, whereas the DNV GL metric is based on accounts, some of which are master metered.

Additionally, the DNV GL metric includes only PA accounts of a given fuel type and does not include accounts

from other service territories that may exist within a block group.

10 This method overestimates renters for National Grid. National Grid estimates that between 25% and 30% of their residential gas customers are

renters.

DNV GL – www.dnvgl.com February 6, 2020 Page 21

Table 3-8. PNP renter flag block group comparison with ACS block group percent renter

Difference in % renter1

% of Electric PA Block Groups

% of Gas PA Block Groups

0% - 10% 43% 33%

11% - 20% 29% 28%

21% - 30% 15% 18%

31% - 40% 7% 9%

41% - 50% 3% 4%

51% - 60% 1% 2%

61% - 70% 1% 2%

71% or more 0% 3%

1 Calculated as the absolute value of the difference between the ACS percent renter and the PNP percent renter by block group.

3.2.2 Modeling

DNV GL analyzed three levels of models: block group, individual, and enhanced individual. We modeled

electric and gas participation separately, using the same methods in both cases.

1. Block group-level models. For these models, the dependent variable is location participation,

consumption-weighted location participation, or savings/consumption within a block group. Predictor

variables are block group-level American Community Survey (ACS) variables expressed as proportions

of locations or households that match the term sheet variables. These models utilize the most readily

available data and do not contain personal identifying information. However, because they operate on

proportions of block groups, they can be difficult to interpret and are susceptible to ecological fallacy.

2. Individual-level models. These models use the binary definition of participation at the account level and

savings/consumption computed at the account level. They predict participation using individual-level

variables available from the unified Massachusetts residential database prepared for the Residential

Customer Profile Study (RCPS). This database includes information from the PAs’ billing and tracking

data as well as information appended from the Massachusetts Tax Assessors data.

3. Enhanced individual-level models (available for Eversource only). These models are similar to the

standard individual-level models, but they include additional location-level variables available through

third-party data sources, specifically Eversource’s Experian data. DNV GL attempted to integrate audit

data available for the other PAs. Due to a high percentage of missing values in the variables of interest,

pertinent data was available for less than 100 accounts. Therefore, we left the audit data out of this

analysis to help reduce evaluation costs and used only the Experian data for Eversource.

Block group-level models

The initial models predict block group-level participation rates using variables based on block group-level

ACS data and US Census Urban and Rural Classification data. Participation rates are informed by program

participation data from 2013 to 2017 and 2013-2017 billing data. For the electric analyses, we included only

block groups where one of the electric PAs was listed as a utility of service by the Department of Public

Utilities (DPU). For the gas analyses, we included block groups where one of the gas PAs was listed as the

utility of service by the DPU.

We used three different definitions of participation, each calculated at the block group level. The

participation variables are defined as:

DNV GL – www.dnvgl.com February 6, 2020 Page 22

• Location-level participation: Location-level participation is calculated as the number of unique

locations that participated at least once between 2013 and 2017, divided by the number of unique

locations in the 2013-2017 billing data.

• Consumption-weighted participation: is defined as the sum of 2017 consumption for any locations

that participated at least once between 2013-2017 divided by the sum of 2017 consumption for all

locations.

• Savings/Consumption: is defined as the sum of 2013-2017 savings divided by the average 2013-

2017 consumption, both taken at the block group level. We drop block groups for whom the metric

exceeds 1.11 While a value greater than 1 is theoretically possible, it is extremely rare, and these

instances most likely represent substantial new construction occurring in the block group.

����� ��� ����� ������

���������=

��2013 − 2017 ��������� !"

#�������2013 − 2017 ������������ !"

The predictor variables include ACS variables associated with the term sheet:

• Moderate income: We used the proportion of households with incomes between 56 and 85% of

statewide median income12 (based on ACS variables B19001e9, B19001e10, B19001e11, and S1903e2).

Note, this is a proxy for true moderate income in two different ways; moderate income is defined as

61% to 80% of state median income, factoring in household size. We had only block-group level data.

Thus, we chose income ranges available in the ACS data that most closely approximated the eligibility

ranges set by Mass Save for a household size of 2.

• Renter-occupied housing: The proportion of occupied households in the block group that are renters

(based on ACS variables B25003e, B25003e3)

• Multifamily housing: The proportion of occupied households in the block group living in buildings with

5 or more units (based on ACS variables in table B25024)

• Limited English-speaking households: The proportion of households in the block group reporting as

speaking limited English at home (based on ACS variables in table C16002)

DNV GL also investigated the relationships between participation and several other ACS variables, including:

• Low income: Defined as the proportion of households with less than 56% statewide median income13

(based on ACS variables B19001e1, B19001e2, B19001e3, B19001e4, B19001e5, B19001e6, B19001e7,

and B19001e8). This variable is a proxy for low income, which is officially defined as less than or equal

to 60% of state median income.

• Average or higher income: Defined as the proportion of households with greater than moderate

income (based on ACS variables B19001e1, B19001e12, B19001e13, B19001e14, B19001e15,

B19001e16, B19001e17). Note that this definition may not be considered truly high income by other

standards. We include the variable as a comparison point for moderate income.

• Construction year: Defined as the proportion of housing units built within the following ranges of

years: before 1949, 1950-1969, 1970-1989, 1990-2009, and 2010 or later (based on ACS variables in

table B25034)

• Heating fuel type: Defined as the proportion of occupied housing units with primary heating fuels of

natural gas, electric, or non-utility fuels. Non-utility fuels include all fuel types other than natural gas or

11 0.3% of block groups in the electric analysis were affected. 0.1% of block groups were affected in the gas analysis.

12 Based on the grand median statewide income, independent of household size.

13 Based on the grand median statewide income, independent of household size.

DNV GL – www.dnvgl.com February 6, 2020 Page 23

electric, such as fuel oil, propane, kerosene, wood, and solar (based on ACS variables from table

B25040).

• Urban/rural: Defined using the 2010 U.S. Census definition of urban and rural at the block level. The

block level is smaller than the block-group level. All homes in urban blocks were coded as urban, and

then we computed the proportion of urban homes within a block group based on the number of homes

in the urban blocks divided by the total number of homes in the block group. Rural was calculated in the

same way. It should be noted that the majority of blocks in MA are coded as urban (there are

approximately 4 times as many urban as rural blocks). Furthermore, according to the 2010 Census,14

92% of the state population lives in urban areas, which account for 38% of the land area.

The first step in our analysis was to look at zero-order correlations between each of the participation

variables and each of the ACS variables. These results are discussed in detail in Section 4.1. These

correlations estimate the direct relationship between participation and the ACS variables of interest. They

represent the simplest and most direct relationship between the variables of interest. Because they

represent the relationships between only two variables at a time, they do not account for nuances or the

effects of multiple variables.

For example, both renting and limited English are positively correlated with each other and negatively

correlated with participation rates. The correlations themselves do not provide any insight into how the

three variables relate. One possibility is that the three correlations are completely independent. Another

possibility is that the relationship of one variable is explained by another. For example, if limited English

speakers are more likely to rent, and renters are less likely to participate, does renter status explain why

limited English speakers are less likely to participate? A third possibility is that there is another variable at

work. In this case, perhaps renters and limited English speakers both tend to have lower incomes, and it is

lower income that explains decreased participation. Regression models allow one to account for these

interrelated effects.

The next analytic step was to use regression modeling to estimate the independent and interrelated effects

of the ACS variables on participation. We used ordinary least squares (OLS) to estimate the following

general block group participation model. We repeated this model for location participation and consumption-

weighted location participation. The primary model15 estimated participation rate as a function of the

proportion of moderate-income homes, the proportion of renter-occupied housing units, the proportion of

multifamily housing units, and the proportion of limited English speaking households. We also ran additional

models using different permutations of the ACS variables to better understand how the term sheet variables

interacted with each other and the non-term sheet variables. These results are discussed in Section 4.2.

For variables that are perfectly collinear, meaning that there is a mutually exclusive relationship between

them (e.g. urban and rural), one of the variables must be dropped from models as there is no variation for

the model to pick up. As an example, the proportion of rural PA households is 1 minus the proportion of

14 https://www2.census.gov/geo/docs/reference/ua/PctUrbanRural_State.xls

15 $� = % + Β()�*� + Β+,���� + Β-).� + Β/0��� + 1

Where:

- $� is the participation rate

- )�*� is the proportion of households at 56 - 85% of statewide median income

- ,���� is the proportion of renter occupied housing units

- ).� is the proportion of housing units in 5+ unit buildings

- 0��� is the proportion of limited English-speaking households

DNV GL – www.dnvgl.com February 6, 2020 Page 24

urban PA households. Because these two variables contain the same information and are linearly related to

one another, they cannot both be included in the same model. The variable dropped from the model can be

interpreted as the reference case, meaning that the coefficient for urban in this example is the impact of

urban vs. rural. For most models, we dropped the proportion of households above 85% of the statewide

median income, the proportion of households built in 1950 or later, the proportion of households with non-

utility heating, and the proportion of rural households.

Individual-level modeling

The primary purpose of the individual modeling was to assess whether the findings based on the block

group-level models were consistent when using individual-level models. Because the block group-level

models are based on proportions of households within a geographic area, they are susceptible to the

ecological fallacy. That is, an association between two variables in the block group-level models does not

necessarily link the two variables at an individual household level. For example, moderate-income and low

participation were negatively correlated in the block group models. This means that as the proportion of

moderate-income households increases, the participation rate of the block group decreases. However, it

should not be interpreted as “moderate-income households are less likely to participate.” Such an

interpretation requires individual-level modeling.

The drawback of the individual-level modeling is that it is more difficult to reliably determine demographic

characteristics at this level. The ACS variables are readily accessible, but the PAs do not track most of those

demographic variables at the account level in either their billing or program tracking databases. Obtaining

variables at this data grain often requires costly data purchases. In some cases, the individual data provided

are derived from the same aggregate data that we are trying to test, and therefore are subject to the same

ecological fallacy concerns. We ran models that paralleled the block group models as closely as possible

using the account-level information available from the PA billing and tracking databases. Experian

demographics data was available for Eversource customers, but not other PAs.

• Account participation: For the individual-level models, we used account participation as the

dependent variable. Account participation is closest in data grain to the individual-level variables we

used as independent variables in these models. The two primary drawbacks of using account

participation are that it can fail to represent participation in master-metered multifamily buildings and in

cases where the tracking system does not have unit-level measure records, and it underrepresents

participation in time-series analyses.

The first limitation of failing to represent all master-metered multifamily participation is impossible to

completely overcome. The best option is to use location participation. However, using location

participation as a dependent variable would put the dependent and independent variables at different

data grains. This introduces complications such as how to calculate individual-level variables such as

income or primary language at a location level. Location-level participation is also susceptible to a bias

for master-metered multifamily participation in the opposite direction as account-level participation.

Because location-level participation was set to 1 if any of the units at a location participated, it tends to

overrepresent participation because larger buildings (those with more units) have more chances for at

least one unit to participate. Thus, location-level participation assessed at the account level results in

DNV GL – www.dnvgl.com February 6, 2020 Page 25

every account within a participating location being set as a participant, which definitely over-represents

participation.16

The time-series limitations for account-level participation were mostly absent in the individual-level

modeling because of the way we prepared the data sets. The basis of the data were active accounts in

2017. Because we limited the basis to a single year and associated any participation that occurred from

2013-2017 to that single year of billed accounts, the data set was able to essentially avoid any time

series variability.

• Savings/consumption: For the individual-level metric, we start with the accounts in the PNP list. This

list is based on accounts active in 2017 because its initial purpose was to provide a sample for the

qualitative surveys. Using this starting point allows the study to add in a savings/consumption metric at

the individual level and still meet the established reporting timeline.

Calculating the individual-level metric presents some challenges. First, not all accounts active in 2017

were also active all the way back to 2013, and the number of active accounts gets smaller the further

we go back in time. Furthermore, some accounts’ consumption in the first active year is extremely low

(<1,000 kWh). To derive a typical consumption level for the account against which to normalize the

savings, we remove the years with <1,000 kWh of consumption before averaging the annual

consumption values we have for each account. For the numerator, we include any savings we have

records of. Similar to the block-group-level metric, we drop any account whose metric exceeds 1.17 We

calculate the formulation as follows:

2�*���*�� ����� ������

���������=

��2013 − 2017 ������+3(45 �678"

#������ ���������+3(-9+3(4

The strength of this metric is that it presents a more scalable metric of participation than account- or

location-level participation rates.18 Whereas single-family and large multifamily sites count equally in

location participation, the savings/consumption metric scales with the energy-saving opportunities

inherent in larger sites like high-rise multifamily buildings.

This metric also has several limitations. As noted above, the individual-level formulation fails to account

for the participation of previous residents (accounts) that the current resident may be benefiting from.

This limitation is more relevant to more transient populations such as renters, which are one of this

study’s key populations of interest. We also cannot identify situations where savings were recorded in

the tracking database associated with a single site but were in reality distributed across multiple sites.

DNV GL has seen an example of this in another study when all the fire stations in a city participated in a

program. There were savings across approximately 10 stations, but all of the savings were tracked at a

single station. When these situations occur, it can cause savings to exceed consumption. Because we

excluded cases where the metric exceeded 1, any cases such as these are not in the present analysis.

We prepared the independent variables slightly differently for the standard (all PA) models and the

enhanced (Eversource-only) models. For the enhanced models, we assigned the variables based only on

Experian data whenever possible.

16 The situation where a building owner retrofits lighting in common areas is an example of when this could happen. In this case, one could argue

that all tenants in the building benefit from that participation. However, that level of indirect benefit was not the focus and is not captured in

this study. 17 This affected 4% of electric accounts and 9% of gas accounts.

18 This should not be taken or construed as a better or more accurate representation of participation than metrics used in other studies, and is not a

subjective statement on the quality, accuracy, or appropriateness of those other metrics to address different research questions.

DNV GL – www.dnvgl.com February 6, 2020 Page 26

• Moderate income:

- (statewide models) We attempted to use building value and value per square foot data available

from the Tax Assessor’s data as a proxy for income in the all PA data set. However, all such

variables we were able to create using the Tax Assessor data failed to correlate with income in the

Experian data set, so we decided it was not a viable proxy. Thus, we did not have a viable income

variable in the all-PA dataset.

- (enhanced models only) The Experian data includes a variable that listed income ranges. The ranges

are in $15,000 or greater increments. As we did for the ACS data, we selected the two ranges that

most closely corresponded to the definition of moderate income. Based on grand statewide median

income, the moderate-income range in Massachusetts is approximately $41,000 to $63,000. The

two ranges available from Experian data that we coded as moderate-income were $35,000 to

$49,999 and $50,000 to $74,999.

• Renter:

- (statewide models) We coded renter as a dummy variable in the all PA dataset based on the

description in section 0.

- (enhanced models) We coded a dummy variable as 1 if Experian listed the record as a renter or

likely renter.

• Multifamily housing:

- (statewide models) We coded multifamily housing as a dummy variable set to 1 if the Tax Assessor

building type information indicated a multifamily type.19

- (enhanced models) We coded multifamily housing as a dummy variable if Experian listed the record

as multifamily, likely multifamily, condo, or likely condo.

• Limited English-speaking households (enhanced models only): We coded this dummy variable as 1

if Experian listed the record as likely limited-English speaking.

Other variables we included as independent variables in these models include the following:

• Construction year:

- (statewide models) We coded a dummy variable for pre-1950 construction based on construction

year data from the Tax Assessor data.

- (enhanced models) We coded a dummy variable for pre-1950 construction based on the construction

year in the Experian data.

• Urban/Rural (both data sets): We set an urban dummy variable at the account level based on

geocoding of the account’s location. When the account was geocoded to an urban block according to the

U.S. Census, we coded the account as urban.

These analyses followed the same procedure as for the block group level. First, we analyzed correlations of

the independent and dependent variables. Then we fit regression models using a variety of combinations of

the independent variables. A key difference between these models and the block group-level models is that

the individual-level models utilize binary logistic regression. Like OLS models, binary logistic models still

19 The Multifamily Census project has additional methods of cross-checking for multifamily status, but those data were not available at the time we

prepared the data for this report.

DNV GL – www.dnvgl.com February 6, 2020 Page 27

provide information about the individual and interrelated effects of the independent variables on the

dependent variables. The main difference between the two types of models is the unit of the regression

coefficients. However, for the purposes of this report, the distinctions have little practical implication.

3.3 Hot spot analysis

To illustrate how key ACS block group characteristics cluster spatially, DNV GL generated a composite

number for each block group and then input these composite numbers to generate a geographic hot spot

map based on areas where there are concentrations of consistently high (or low) block group scores.20 We

defined an area as the block groups adjacent to a given block group. 21 The composite number is calculated

by taking the sum of the ratios of the following attributes:

1. Percentage of households that are moderate income22

2. Percentage of households that are renter-occupied

3. Percentage of households with a primary language other than English

4. Percentage of households that are in structures of 5 or more units (i.e., a proxy for multifamily)

5. Percentage of structures that were built prior to 1940

The hot spot analysis does not factor in participation status.

As an example, a block group with 25% of each of the above 5 attributes would have a composite number of

(.25+.25+.25+.25+.25) = 1.25. The lowest score possible for a block group is zero; the highest score

possible is 5.

There are benefits and drawbacks to this scoring approach. The primary benefit is that using a ratio as the

input for the composite number gives DNV GL a scale-less number as the input to the hot spot algorithm.

The primary drawback is that the ratio does not take scale into account. For example, a 60-household block

group that is 100% renter-occupied and generates a composite score of 1 appears the same as a 600-

household block group that is 100% renter-occupied (also a composite score of 1).

The scaling drawback is the main reason why DNV GL chose to take the extra processing step of mapping

the statistical hot spots based on adjacent block groups rather than just mapping each block group’s

composite score. Taking this extra step means that a small block group that has a high score in isolation but

is surrounded by low-scoring block groups will not be identified as a hot spot, and may even show up as a

cold spot.23

20 For technical details on ESRI’s implementation of hot spot analysis and additional context, please see https://pro.arcgis.com/en/pro-app/tool-

reference/spatial-statistics/h-how-hot-spot-analysis-getis-ord-gi-spatial-stati.htm. 21 We chose to look at adjacent block group areas in consideration of differences in block group area between densely populated cities like Boston and

rural towns like Mt. Washington, and of our understanding that stakeholders are more interested in local geographic differences than large scale regional differences. Other representations we considered were a Euclidean distance function (e.g., look at all block groups within 5 miles) and

an inverse distance weight (e.g., look at all block groups within 5 miles, but as we get further away, assign a lower hot spot impact weight to

the block groups). In both cases, DNV GL felt that the regional scale differences between the rural western and central parts of the state versus

the densely populated eastern part of the state precluded using distance as our weighting variable. 22 As defined earlier in this report

23 A different but conceptually similar type of analysis looking at clusters and outliers identifies these types of block groups; more can be read about

this analysis at https://pro.arcgis.com/en/pro-app/tool-reference/spatial-statistics/h-how-cluster-and-outlier-analysis-anselin-local-m.htm.

DNV GL – www.dnvgl.com February 6, 2020 Page 28

4 ANALYSIS AND RESULTS

4.1 ACS variable correlations

Across all the block groups in the state that receive PA electric or gas service, most of the term-sheet

variables are positively correlated with each other (Table 4-1). This means that as the proportion of the

block group matching one variable increases, so do the proportions of the other variables. The only variable

pairs that are not significantly correlated are moderate-income and multifamily, moderate income and

natural gas heating, and moderate income and urban. All other correlations are positive and statistically

significant. The relationship between the proportion of renters and the proportion of households in

multifamily buildings is particularly strong (r=0.67), as is the relationship between renters and households

who speak limited English (r=0.58).24

Table 4-1. ACS variable correlations

Proportion of homes that are…

Lo

w I

nco

me

Mo

derate

In

com

e

Avg

or H

igh

er

In

com

e

Ren

ters

Mu

ltif

am

ily

Lim

ited

En

gli

sh

Pre-1

95

0

Co

nst.

Po

st-

19

50

Co

nst.

Ele

ctr

ic

Heati

ng

Gas H

eati

ng

No

n-E

lec.

Or

Gas H

eati

ng

Urb

an

Low Income - 0.05 -0.93 0.70 0.45 0.53 0.25 -0.25 0.42 0.08 -0.32 0.14

Moderate Income

0.05 - -0.43 0.16 0.00 0.10 0.12 -0.12 0.04 0.03 -0.05 0.00

Average or Higher Income

-0.93 -0.43 - -0.69 -0.41 -0.51 -0.27 0.27 -0.40 -0.08 0.31 -0.12

Renters 0.70 0.16 -0.69 - 0.67 0.58 0.45 -0.45 0.54 0.21 -0.52 0.28

Multifamily 0.45 0.00 -0.41 0.67 - 0.37 0.07 -0.07 0.68 0.00 -0.40 0.21

Limited English 0.53 0.10 -0.51 0.58 0.37 - 0.27 -0.27 0.34 0.19 -0.38 0.19

Pre-1950 Construction

0.25 0.12 -0.27 0.45 0.07 0.27 - -1.00 -0.04 0.36 -0.33 0.23

Post-1950 Construction

-0.25 -0.12 0.27 -0.45 -0.07 -0.27 -1.00 - 0.04 -0.36 0.33 -0.23

Electric Heating 0.42 0.04 -0.40 0.54 0.68 0.34 -0.04 0.04 - -0.27 -0.33 0.14

Gas Heating 0.08 0.03 -0.08 0.21 0.00 0.19 0.36 -0.36 -0.27 - -0.82 0.49

Non-electric or Gas Heating

-0.32 -0.05 0.31 -0.52 -0.40 -0.38 -0.33 0.33 -0.33 -0.82 - -0.56

Urban 0.14 0.00 -0.12 0.28 0.21 0.19 0.23 -0.23 0.14 0.49 -0.56 -

All correlations are statistically significant at p<.01 except Moderate Income and Multifamily, Moderate Income and Gas Heating, and Moderate

Income and Urban.

For the additional ACS variable correlations, low and moderate-income, renters, multifamily, limited English

language, natural gas heating, and urban status all tend to cluster together. By definition when looking at

block group statistics, this clustering indicates an underlying geographical commonality. There are

essentially areas around the state, generally urban ones, where people who tend to fall into all of these

categories live, and the housing stock there tends to be rented, multifamily housing constructed before

1950. Statistically significant, positive correlations include:

24 Higher numbers indicate stronger correlations. When the number is positive, it means that as one variable increases, the other also tends to

increase. When the number is negative, it means that as one variable increase, the other tends to decrease.

DNV GL – www.dnvgl.com February 6, 2020 Page 29

• Low income is positively correlated with: moderate-income (r=0.05), renter (r=0.70), multifamily

(r=0.45), limited English (r=0.53), pre-1950 construction (r=0.25), natural gas heating (r=0.08),

electric heating (r=0.42), urban (r=0.14).

• Moderate income is positively correlated with: low income (r=0.05), renter (r=0.16), limited English

(r=0.10), pre-1950 construction (r=0.12), electric heating (r=0.04).

• Average or higher income is positively correlated with: post-1950 construction (r=0.27), non-utility

heating (r=0.33), rural (r=0.13).

• Renter is positively correlated with: low income (r=0.70), moderate income (r=0.16), multifamily

(r=0.67), limited English (r=0.58), pre-1950 construction (r=0.45), natural gas heating (r=0.21),

electric heating (r=0.54), and urban (r=0.28).

• Multifamily housing is positively correlated with: low income (r=0.45), renter (r=0.67), limited English

(r=0.37), pre-1950 construction (r=0.07), electric heating (r=0.68), and urban (r=0.21).

• Limited English is positively correlated with: low income (r=0.53), moderate income (r=0.10), renter

(r=0.58), multifamily (r=0.37), pre-1950 construction (r=0.27), natural gas heating (r=0.19), electric

heating (r=0.34), and urban (r=0.19).

DNV GL – www.dnvgl.com February 6, 2020 Page 30

4.1.1 Hot spot analysis

The result of the composite number hot spot analysis is presented in Figure 4-1. DNV GL applied an overlay

to the map to shade out areas where households are not likely to be present (e.g. lakes, wetlands, forests).

Detailed regional views are also provided in Appendix B. Finally, because this analysis looked at the ACS

variables rather than PA-supplied data, DNV GL included municipal-served towns so that stakeholders in

these areas would also be able to leverage the spatial analyses.

Figure 4-1. Statewide ACS variable hot spot map, subset to urbanized land areas only

As Figure 4-1 shows, the spatial clustering of the composite score correlates with the highly populated urban

cores across the state. We believe one of the benefits of the hot spot map, particularly using the adjacent

block group spatial representation for the statistical model, is that local stakeholders can identify and

engage the areas that contain greater proportions of the key ACS variables while still preserving the

statewide view of where hot spots occur. For example, a stakeholder in a less populated area like Greenfield

(the small red cluster to the northwest-central part of the state, north of the more intense Springfield

cluster) has the same opportunity to identify target clusters of block groups that stakeholders in larger cities

like Boston or Cambridge have.

There are some areas, such as the Cape region, where most of the block groups comprise cold spots. This

should not be interpreted to mean that there are no individual priority households in this area or that there

This figure does not include participation status.

DNV GL – www.dnvgl.com February 6, 2020 Page 31

is no program opportunity to serve customers in this region. A critical distinction of this analysis is that we

are looking at these block groups as a single composite number of the multiple inputs and not as individual

inputs. If we were to focus on just one of the variables (for example, renter-occupied homes) the results on

the Cape would likely be different, partly due to seasonal and rental households.25

4.2 Electric location participation findings

4.2.1 ACS variable correlations with electric participation

Figure 4-2 presents the correlations between the three different participation variables and the ACS

variables. Correlations between the participation rates and ACS variables were usually negative and

statistically significant. This indicates that block groups with higher concentrations of the term sheet

variables have lower participation rates. Key findings include:

• Renter status is most strongly (negatively) correlated with (unweighted) location participation. This

means that as a general trend, as the proportion of households in the block group that rent increases,

the proportion of locations in the block group that participate decreases.

• Limited English speaking is the next most strongly correlated variable with location participation. As for

renter status, the negative correlation means that as a general trend, as the proportion of households in

the block group that have limited English speaking increases, participation rates decrease.

• Moderate income and multifamily have the third and fourth correlation strengths. The negative

correlation is interpreted the same way as the other two variables.

• Except for multifamily, the correlations for location participation and consumption-weighted location

participation are similar. In the case of multifamily, it is significantly negatively correlated with location

participation but was significantly positively correlated with consumption-weighted location participation.

This is most likely caused by very high weights being assigned to large participating multifamily

buildings because they have high consumption. The weighting causes the participation in those buildings

to count very heavily.

25 The 2013-2017 Residential Customer Profile Study report explores these types of spatial nuances and interactions in greater detail in its section on

exploratory factor analysis.

DNV GL – www.dnvgl.com February 6, 2020 Page 32

Figure 4-2. Block group-level electric participation correlations (term sheet variables)

Figure 4-3 shows the correlations between participation and the additional ACS variables. Key findings

include:

• The proportion of low-income households is negatively correlated with participation, while the proportion

of average or higher-income households is positively correlated with participation.26

• The proportion of homes built before 1949 is negatively correlated with participation. All other home

vintage variables are positively correlated with participation. 1949 is the threshold where knob and tube

wiring and asbestos insulation become much less common. Other studies have shown that these older

building technologies are barriers to efficiency program participation because they have to be

remediated to bring homes up to code before other efficiency measures can be installed.27

• The proportion of homes with natural gas heat is negatively correlated with electric participation.

• The proportion of homes with electric heat is negatively correlated with location participation but

positively correlated with consumption-weighted location participation. This dichotomy is probably

caused by the high weights given to the very large multifamily buildings. This implies that the large

multifamily buildings are more likely to have electric heat.

• The proportion of homes with non-utility heating is positively correlated with participation.

• The proportion of urban or rural households is not significantly correlated with (unweighted)

participation. This is likely an artifact of low variance in the proportion of urban/rural households per

block group because the urban/rural designation is recorded at the block level. Also, approximately 82%

of block groups are 100% urban, so there are few that have values other than 1. This could make a

correlation difficult to detect.

26 The 2017 Residential Profile report had different results for low-income participation. As we will show in Section 4.3, low-income is positively

correlated with block group level participation measured as the ratio of savings to consumption. The 2017 Residential Profile report used a

similar metric. 27 DNV GL reported in a previous Eversource-only deliverable that knob and tube wiring and the asbestos are barriers in older homes because of the

extra cost required to bring those systems up to code.

-0.20

-0.39

-0.14

-0.25

-0.18

-0.15

0.19

-0.09

-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5

Proportion of Moderate Income Housing Units

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ Unit Buildings

Proportion of Limited English-speaking Households

Correlation

Location Participation, 2013-2017 (n=4,345)

Consumption Weighted Location Participation, 2013-2017 (n=4,325)

all correlations statistically

significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 33

Figure 4-3. Block group-level electric participation correlations (extra ACS variables)

-0.36

-0.20

0.40

-0.34

0.34

-0.09

-0.13

0.18

0.02

-0.18

-0.18

0.23

-0.30

0.30

0.21

-0.16

0.03

0.05

-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5

Proportion of Low Income Housing Units

Proportion of Moderate Income Housing Units

Proportion of Average or Higher Income

Housing Units

Proportion of Housing Units built before 1950

Proportion of Housing Units built in 1950 or

Later

Proportion of Households with Electric Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PA Heating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,345)

Consumption Weighted Location Participation, 2013-2017 (n=4,325)

§ correlation NOT statistically significant at p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 34

4.2.2 Electric participation block group-level models

Table 4-2 presents the electric block group model results for location and consumption-weighted location

participation. In both models, the relationships between most of the ACS variables and participation rate

remain similar to the zero-order correlations. This indicates that they each have independent effects on

participation. The exceptions are multifamily housing and limited English; the effects of which are different

in the models than in the zero-order correlations. This indicates that the effects of these two variables are

interacting with at least one of the other variables in the model.

Table 4-2. Initial electric block group models

Variable

Location Participation

Consumption Weighted

Participation

Coefficient Coefficient

Intercept 0.389 *** 0.393 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.173 *** -0.164 ***

Proportion of Renter Occupied Housing Units -0.178 *** -0.221 ***

Proportion of Housing Units in 5+ Unit Buildings 0.081 *** 0.280 ***

Proportion of Limited English-speaking Households -0.027 0.027

*** p <.001; ** p < .01; * p < .1

There are three effects that change from the zero-order correlations when all the variables are included in

the models:

1. The strength of the effect of renting decreases.

2. The effect of limited English-speaking households becomes non-significant.28

3. The effect of multifamily housing becomes much more positive; so much so that it changes signs in the

location participation model.

Additional models (location participation: Table 4-3; consumption-weighted location participation: Table 4-4)

revealed that the renter variable is interacting with the multifamily and limited English variables. For both

types of participation, when the renter variable is absent from the model, the effects of multifamily housing

and limited English are similar to their zero-order correlations.

28 There is a sign change in the estimate for this variable for unweighted and location weighted participation. However, because neither value is

statistically significantly different than zero, DNV GL did not attempt to interpret the sign change.

DNV GL – www.dnvgl.com February 6, 2020 Page 35

Table 4-3. Electric location participation models with some variables removed

Variable

Without % Renter

Full Model

Coefficient Coefficient

Intercept 0.363 *** 0.389 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.246 *** -0.173 ***

Proportion of Renter Occupied Housing Units -0.178 ***

Proportion of Housing Units in 5+ Unit Buildings -0.032 *** 0.081 ***

Proportion of Limited English-speaking Households -0.232 *** -0.027

*** p <.001; ** p < .01; * p < .1

Table 4-4. Electric consumption weighted location participation models renter variable

present/absent

Variable

Without % Renter

Full Model

Coefficient Coefficient

Intercept 0.362 *** 0.393 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.255 *** -0.164 ***

Proportion of Renter Occupied Housing Units -0.221 ***

Proportion of Housing Units in 5+ Unit Buildings 0.139 *** 0.280 ***

Proportion of Limited English-speaking Households -0.228 *** 0.027

*** p <.001; ** p < .01; * p < .1

The interactions between renter status and the other two variables suggest that the effect of the other

variable (limited English or multifamily) is largely explained by the tendency of those groups to rent.

• Limited English is a barrier to participation mostly because of the tendency of that population to rent.

Limited English speakers tend to rent and not participate. Renters also tend to not participate. When the

tendency of renters to not participate is accounted for, we no longer see a significant effect of Limited

English. This is illustrated in the following pattern:

- Renter and limited English are positively correlated with each other (r=0.58) and both negatively

correlated with location participation (Renter r = -0.41; Limited English r= -0.26)

- In the model without Renter, Limited English remains a significant, negative predictor of

participation (Location coefficient=-0.232)

- When Renter is added to the model, renter is a significant, negative predictor of participation

(Location coefficient=-0.178) and limited English becomes non-significant (Location coefficient=-

0.027).

- Consumption-weighted participation has the same pattern, but with slightly different numeric

values.

Adding the other variables to the model, such as home construction vintage and heating fuel, did not cause

substantial changes to the results reported in this section (Table 4-5). The effects of multifamily and limited

English decreased slightly, but that is not unexpected when adding variables to the model that are known to

correlate with other variables.

DNV GL – www.dnvgl.com February 6, 2020 Page 36

Table 4-5. Electric block group model results with all variables29

Variable

Location Participation

Consumption-weighted Location

Participation

Coefficient Coefficient

Intercept 0.363 *** 0.372 ***

Proportion of Households at less than 56% of Statewide

Median Income -0.110 *** -0.120 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.183 ***

-0.179 ***

Proportion of Renter Occupied Housing Units -0.089 *** -0.133 ***

Proportion of Housing Units in 5+ Unit Buildings 0.019 * 0.171 ***

Proportion of Limited English-speaking Households 0.009 0.053 *

Proportion of Households Built in 1949 or Earlier -0.070 *** -0.071 ***

Proportion of Households with Natural Gas Heating -0.036 *** -0.026 *

Proportion of Households with Electric Heating 0.027 * 0.165 ***

Proportion of Urban PA Households 0.089 *** 0.072 ***

4.2.3 Electric participation individual-level models

Data available at the individual account level (from all PAs) included renting, multifamily, construction date,

and urban/rural. In the enhanced individual-level models (Eversource only), we were able to also include the

effects of income and limited English.

As stated in the method section, we used account-level participation for the individual models because it is

at a grain similar to most of the independent variables. This type of participation reflects participation while

the resident lived in a particular house or apartment.

Correlations for individual-level data

Participation is negatively correlated with all four demographic variables available statewide30 (Table 4-6).

This means that people with moderate incomes, who rent, who live in multifamily housing, or who have

limited English skills are less likely to participate. Unlike the block group level models, these findings are not

subject to ecological fallacy and can be interpreted to indicate effects at an individual level. The individual-

level correlations are consistent with the block-group-level correlations in direction, though generally

weaker. Income and language information is not available at this data grain.

29 For variables that are perfectly collinear, meaning that there is a mutually exclusive relationship between them (e.g. urban and rural), one of the

variables must be dropped as there is no variation for the model to pick up. As an example, the proportion of rural PA households is 1 minus the

proportion of urban PA households. Because these two variables contain the same information and are linearly related to one another, they cannot

both be included in the same model. The variable dropped from the model can be interpreted as the reference case, meaning that the coefficient for

urban in this example is the impact of urban vs. rural. For these models, we dropped the proportion of households above 85% of the statewide

median income, the proportion of households built in 1950 or later, the proportion of households with non-utility heating, and the proportion of rural

households.

30 These results are for the combined statewide data set. We did not test results for individual PAs.

DNV GL – www.dnvgl.com February 6, 2020 Page 37

Table 4-6. Individual-level correlations – statewide electric

Variable Account

Participation, 2013-2017

Multifamily Renter Urban Pre-1950

Construction

Account Participation, 2013-2017

- -0.12 -0.20 -0.05 -0.11

Multifamily -0.12 - 0.32 0.09 0.09

Renter -0.20 0.32 - 0.18 0.30

Urban -0.05 0.09 0.18 - 0.15

Pre-1950 Construction

-0.11 0.09 0.30 0.15 -

Over 2 million observations per correlation. All correlations statistically significant p<0.001

Models for individual-level data

The individual-level models use binary logistic regression to predict a dependent variable that is set to 0 for

nonparticipating accounts and 1 for participating accounts. The coefficients of these models are logs of odd

ratios, which are difficult to interpret directly. The magnitude of the coefficient still reflects the relative

strength of the relationship. The sign indicates the direction of the relationship. Positive signs mean the

probability of participation increases as the independent variable increases and vice versa for negative

values.

Table 4-7 shows several model results predicting account participation for the variables that are available

statewide. These model results are consistent with the block group-level models. All of the independent

variables are negative predictors of participation. That is, renters are less likely to participate than non-

renters; people living in multifamily buildings are less likely to participate than those in other buildings, etc.

Because these models utilize individual-level variables, they can be interpreted as individual-level effects.

One effect is inconsistent with the block group level models. The main effect of multifamily residency in the

individual level models is the opposite of the effect in the block group models. In the block group models,

multifamily was a positive predictor of participation, especially when controlling for the effect of renting. The

difference in results is probably caused by the differences in the dependent variables in each model. In the

block group models, the dependent variable was location (or consumption weighted location) participation.

In the individual level models, the dependent variable is account participation. Location participation is

probably skewing the model outcomes. Because participation by any single unit in a multifamily building

creates a location participation value of 1, the more units in the building, the greater the probability that at

least one of them participated. In contrast, account participation is not affected by the number of units in a

building, and the model results indicate that individual units in multifamily buildings have a lower probability

of participating than those in non-multifamily buildings.

DNV GL – www.dnvgl.com February 6, 2020 Page 38

Table 4-7. Electric account participation models coefficients – statewide

Variable Model 1 Model 2 Model 3

Intercept -0.901 *** -0.900 *** -0.748 ***

Renter -0.850 *** -0.853 *** -0.753 ***

Resides in multifamily building

-1.018 *** -1.305 *** -1.204 ***

Renter * Multifamily1 0.303 *** 0.152 **

Resides in Urban block -0.051 ***

Pre-1950 construction -0.305 ***

*** p <.001; ** p < .01; * p < .1 1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

Similar to the block group-level model results, there is also a significant (although small) interaction

between renter and multifamily status in these models. Figure 4-4 illustrates the nature of that interaction

by graphing the expected probability of participation for the four combinations of renter and multifamily

status. For non-renters (Rent=0) and renters (Rent=1), multifamily has a negative effect on the probability

of participation. However, the negative effect of multifamily is not as strong for renters, which is indicated

by the smaller distance between the two columns in the renter category. It should be noted this relationship

is symmetrical: the negative effect of renting is less in multifamily buildings than it is in non-multifamily

buildings.

Figure 4-4. Renter*Multifamily interaction on electric account participation probability –

statewide

Correlations for enhanced individual-level data

In the enhanced individual-level data set, which primarily uses the Experian demographic data available only

for Eversource accounts, account participation is negatively correlated with multifamily, limited English,

moderate-income, pre-1950 construction, and urban (Table 4-8). These correlations are in the same

directions as the block-group and non-enhanced individual-level correlations. The effect for multifamily is

similar in magnitude, but the other effects are weaker than in the other data sets.

29%

15%

10%

6%

0%

5%

10%

15%

20%

25%

30%

35%

Rent 0 Rent 1

Pro

bability o

f Part

icip

ation

Multifamily 0 Multifamily 1

DNV GL – www.dnvgl.com February 6, 2020 Page 39

Table 4-8. Individual-level correlations – Enhanced electric data

Variable

Accou

nt

Parti

cip

ati

on

,

20

13

-2

01

7

Ren

ter

Mu

ltif

am

ily

Lim

ited

En

gli

sh

Low

In

co

me

Mod

erate

In

com

e

Avg

or H

igh

er

In

com

e

Pre-1

95

0

Con

str

ucti

on

Urb

an

Account Participation, 2013-2017

-

-0.14 -0.19 -0.04 -0.04 -0.04 0.06 -0.11 -0.05

Renter -0.14

- 0.41 0.18 0.30 -0.01 -0.23 0.20 0.11

Multifamily -0.19 0.41

- 0.14 0.19 0.01 -0.16 0.26 0.17

Limited English -0.04 0.18 0.14

- 0.09 -0.01 -0.06 0.08 0.09

Low Income -0.04 0.30 0.19 0.09

- -0.32 -0.50 0.08 0.05

Moderate Income -0.04 -0.01 0.01 -0.01 -0.32

- -0.66 0.02 -0.05

Average or Higher Income

0.06 -0.23 -0.16 -0.06 -0.50 -0.66

- -0.08 0.00

Pre-1950 Construction

-0.11 0.20 0.26 0.08 0.08 0.02 -0.08

- 0.15

Urban -0.05 0.11 0.17 0.09 0.05 -0.05 0.00 0.15

- Approximately 1 million observations per correlation. All correlations statistically significant p<.001 except for Average or higher income and urban.

Models for enhanced individual-level data

The results of the enhanced models are very similar to those for the individual-level data. Renter and

multifamily are both negatively associated with participation (Table 4-9).

The enhanced models include a variable for moderate-income that the individual-level models cannot. As

observed in the correlations and in the block group-level models, moderate-income is negatively associated

with participation. The enhanced models also include a variable for limited English that is negatively

associated with participation.

DNV GL – www.dnvgl.com February 6, 2020 Page 40

Table 4-9. Electric account participation models coefficients – enhanced data

Variable Model 1 Model 2 Model 3 Model 4 Model 5

Intercept -0.723 *** -0.658 *** -0.649 *** -0.650 *** -0.590 ***

Experian Low Income

-0.059 ***

Experian Moderate

Income -0.149 *** -0.158 *** -0.155 *** -0.156 *** -0.166 ***

Experian Renter -0.524 *** -0.646 *** -0.596 *** -0.552 ***

Experian resides in

multifamily building

-1.017 *** -0.781 *** -0.843 *** -0.782 *** -0.770 ***

Experian limited English

-0.096 *** -0.057 *** -0.058 *** -0.137 *** -0.083 ***

Resides in Urban Block

0.030 **

Pre-1950

construction -0.159 ***

Experian Renter * Multifamily1

0.259 ***

Experian Renter * Limited English1

0.317 ***

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

The interaction between renting and limited English is significant in the enhanced individual-level models.

Figure 4-5 illustrates the nature of that interaction by graphing the expected probability of participation for

the four combinations of renter and limited English status. For non-renters (Rent=0), limited English has a

positive effect on the probability of participation. However, the relationship switches for renters (Rent=1),

such that limited English decreases the probability of participation. Despite being statistically significant, it is

difficult to imagine a real-life scenario that can explain this relationship. It is most likely a statistical artifact.

Figure 4-5. Renter*Limited English interaction on electric account participation probability –

enhanced data

31%

15%

26%

16%

0%

5%

10%

15%

20%

25%

30%

35%

Rent 0 Rent 1

Pro

bability o

f Part

icip

ation

Limited English 0 Limited English 1

DNV GL – www.dnvgl.com February 6, 2020 Page 41

There is also a statistically significant interaction between renter and multifamily in the enhanced models.

Figure 4-6 illustrates the nature of that relationship. It is very similar to the relationship observed in the

individual-level models.

Figure 4-6. Renter*Multifamily interaction on electric account participation probability –

enhanced data

4.3 Electric savings/consumption findings

4.3.1 Block group electric analysis

Figure 4-7 shows the zero-order correlations between the electric participation metrics, income, the term

sheet variables, and several other ACS variables at the block group level. Location participation correlations

are repeated here for the sake of contrast. As Figure 4-7 shows, correlations change for four key variables

depending on how participation is measured. Low income, renter, multifamily, and limited English correlate

negatively with location participation. By contrast, these variables correlate positively with the

savings/consumption metric. This pattern suggests that while the electric PAs might have lower participation

in block groups with more low income, renters, multifamily, and limited English households in terms of

number of participants, those participants have a relatively high depth of savings (and accordingly, a high

savings-to-consumption ratio).

33%

21%

18%

13%

0%

5%

10%

15%

20%

25%

30%

35%

Rent 0 Rent 1

Pro

ba

bil

ity

of

Pa

rtic

ipa

tio

n

Multifamily 0 Multifamily 1

DNV GL – www.dnvgl.com February 6, 2020 Page 42

Figure 4-7. Block group electric correlations

-0.36

-0.20

0.40

-0.39

-0.14

-0.25

-0.34

0.34

-0.09

-0.13

0.18

0.02

-0.18

-0.18

0.23

-0.15

0.19

-0.09

-0.30

0.30

0.21

-0.16

0.03

0.05

0.07

-0.11

-0.02

0.01

0.14

0.08

-0.17

0.17

0.12

-0.07

0.00

0.01

-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5

Proportion of Low Income Housing Units

Proportion of Moderate Income Housing Units

Proportion of Average or Higher Income HousingUnits

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ Unit Buildings

Proportion of Limited English-speakingHouseholds

Proportion of Housing Units built before 1950

Proportion of Housing Units built in 1950 or Later

Proportion of Households with Electric Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PA Heating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,345)

Consumption Weighted Location Participation, 2013-2017 (n=4,325)

Savings over Average Consumption, 2013-2017 (n=4,338)

§ correlation NOT statistically significant at p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 43

The large multifamily locations, where PAs are able to treat many sub-units within a building at once, are

most likely driving this pattern. Such large multifamily buildings count as only a single location in the

location participation metric, and have the potential to have a large savings/consumption value.

An ordinary least squares (OLS) model that predicts savings/participation with low and moderate income,

renter, multifamily, and limited English terms supports the hypothesis that multifamily locations are partially

driving the correlation patterns. This model is shown in Table 4-10.

Table 4-10. Electric savings/consumption model results, block group-level

Variable Model 1 Model 2

Coefficient Coefficient

Intercept 0.062 *** 0.060 ***

Proportion of Households at less than 56% of Statewide Median Income

0.024 *** 0.029 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.051 *** -0.049 ***

Proportion of Renter Occupied Housing Units -0.044 *** -0.036 ***

Proportion of Housing Units in 5+ Unit Buildings 0.048 *** 0.047 ***

Proportion of Limited English-speaking Households 0.054 ***

*** p <.001; ** p < .01; * p < .1

As Table 4-10 shows:

• In both models, the multifamily effect is reduced relative to the zero-order correlation (0.05 vs. 0.18)

but is still significantly greater than zero. This suggests that multifamily still has a meaningful effect

even when controlling for all the other variables.

• The effect for renters goes from not significantly greater than zero correlation in the zero-order

correlation to a significantly less than zero coefficient in the regression model. This is most likely caused

by an interaction with multifamily. Many renters live in multifamily housing, so when the effect of

multifamily is taken into account, the effect of renters actually reverses. In other words, renters in

multifamily housing are seeing relatively deep savings, but renters in non-multifamily housing are

seeing relatively shallow savings (possibly due solely to low participation numbers).

• Effects of moderate income and limited English are close to the same as their correlations, so the

interactions between those variables and multifamily appear to be minimal.

• The low-income effect becomes more strongly positive. This suggests that there is covariation with

renting for the low-income group that weakens the relationship with low-income at the correlational

level. When that effect is controlled for in the models, the independent relationship with low-income

emerges and is positive. In other words, independent of renting, multifamily, and limited English (which

are highly correlated with low income), low income is positively associated with participation when

measured as the ratio of savings to consumption. This is likely due to PA efforts in their low-income

programs.

As Figure 4-8 shows, there are statistically significant interactions between multifamily and low income, and

multifamily and renter. The blue lines on the graphs indicate that low income and renter status have

different effects when the proportion of multifamily housing is high or low. In block groups with little

DNV GL – www.dnvgl.com February 6, 2020 Page 44

multifamily housing (dark blue line), we expect participation to decrease when there are more renters

and/or more low income households. However, in block groups with a lot of multifamily housing (light blue

lines), we expect participation to increase with when there are more renters and/or more low income

households. As multifamily increases, so do the expected participation rates for low income or renters

increases. These interactions provide additional evidence that something is happening in multifamily

buildings or areas that changes the effect of the other variables. PAs achieving deeper savings in multifamily

buildings, particularly low income multifamily buildings, could explain these interactions.

Figure 4-8. Multifamily interactions with low income and renter

0%

5%

10%

15%

20%

LowIncome

0%

LowIncome50%

LowIncome100%E

xpecte

d P

art

icip

ation

Rate

s

MF 0% MF 50% MF 100%

0%

5%

10%

15%

20%

Rent 0% Rent 50% Rent 100%Expecte

d P

art

icip

ation

Rate

s

MF 0% MF 50% MF 100%

DNV GL – www.dnvgl.com February 6, 2020 Page 45

4.3.2 Individual-level electric analysis – Statewide

Figure 4-9 shows the statewide, zero-order, account-level correlations between the electric participation

metrics, multifamily, renter, urban, and pre-1950 construction (other characteristics are not available at the

account level). Account participation correlations are repeated here for the sake of contrast. Unlike the

block-group level correlations, the account-level correlations with savings/consumption go in the same

direction as for account participation.

Figure 4-9. Individual-level correlations with savings over consumption, statewide electric

Table 4-11 shows results for models that DNV GL used to isolate the effects of the different variables

independently of the other variables. In all three models, the coefficients go in the same direction and are

approximately the same magnitude as the zero-order correlations. This means that the covariation between

the various characteristics (e.g., multifamily residents tend to rent) does not mask or substantially change

the effects of any singular characteristic. There is a statistically significant interaction between renter and

multifamily that indicates there is a significant overlap in the effects of each variable. However, that overlap

is not enough to change overall (correlation) versus independent (model coefficient) effect of either variable.

Table 4-11. Individual-level savings over consumption models, statewide electric

Variable Model 1 Model 2 Model 3

Intercept 0.048 *** 0.048 *** 0.053 ***

Renter -0.020 *** -0.020 *** -0.017 ***

Resides in multifamily building -0.015 *** -0.027 *** -0.024 ***

Renter * Multifamily 1 0.012 *** 0.008 ***

Resides in Urban block -0.002 ***

Pre-1950 construction -0.008 ***

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

-0.12

-0.20

-0.05

-0.11

-0.07

-0.12

-0.03

-0.07

-0.25 -0.15 -0.05 0.05 0.15 0.25

Multifamily

Renter

Urban

Pre-1950 Construction

Correlation

Account Participation, 2013-2017

Savings/ Avg. Consumption, 2013-2017

all correlations statistically significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 46

4.3.3 Individual-level electric analysis – Enhanced

Figure 4-10 shows the Eversource-only, enhanced data, zero-order account-level correlations between the

electric participation metrics, income, renter, multifamily, limited English, construction vintage, and urban.

Account participation correlations are repeated here for the sake of contrast. Unlike the block-group level

correlations and similar to the statewide account-level correlations, the enhanced account-level correlations

with savings/consumption go in the same direction as for enhanced account participation.

Figure 4-10. Individual-level correlations with savings over consumption, enhanced electric

Table 4-12 shows results for models that DNV GL used to isolate the effects of the different variables

independently of the other variables. Except for low and moderate income, the coefficients for each of the

variables are similar to the zero-order correlations. This means the overall effect of the variable (correlation)

is similar to the independent effect when controlling for the covariation with the other characteristics

(coefficients). For low and moderate income, the coefficients in the models shift, which indicates that the

overall effect of those variables is partially affected by overlaps with the other characteristics.

-0.04

-0.04

0.06

-0.14

-0.19

-0.04

-0.11

-0.05

-0.01

-0.01

0.02

-0.09

-0.12

-0.03

-0.07

-0.03

-0.20 -0.10 0.00 0.10 0.20

Low Income

Moderate Income

Average or Higher Income

Renter

Multifamily

Limited English

Pre-1950 Construction

Urban

Correlation

Account Participation, 2013-2017

Savings/ Avg. Consumption, 2013-2017

all correlations statistically significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 47

Table 4-12. Individual-level savings over consumption models, enhanced electric

Variable Model 1 Model 2 Model 3

Intercept 0.058 *** 0.059 *** 0.058 ***

Experian Low Income 0.005 *** 0.001 ** 0.005 ***

Experian Moderate Income 0.000 0.000 0.000

Experian Renter -0.016 *** -0.017 *** -0.016 ***

Experian resides in multifamily building -0.023 *** -0.025 *** -0.024 ***

Experian limited English -0.002 *** -0.002 *** -0.004 ***

Multifamily * Limited English1 0.004 ***

Experian Low Income * Multifamily1 0.009 ***

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

4.3.4 Location participation and savings/consumption inconsistencies

There are inconsistent findings when participation is measured as location participation and when it is

measured as savings/consumption. In particular, at the block group level, low income and multifamily are

negatively correlated with location participation, but positively correlated with savings/consumption (Table

4-13).

Table 4-13. Block group participation models comparison, electric

Variable

Location Participation

Savings /Consumption

Correlation Correlation

Proportion of Households at less than 56% of Statewide Median Income

-0.36 ** 0.07 **

Proportion of Housing Units in 5+ Unit Buildings -0.14 ** 0.14 ** *** p <.001; ** p < .01; * p < .1

The relationships between low income and savings/consumption and multifamily and savings/consumption

remain positive and statistically significant in the block group savings/consumption models (Table 4-10).

However, at the individual account level, low income has a near-zero effect and multifamily has a negative

effect (Table 4-12).

We explored three potential explanations for these inconsistencies:

1. The data set used for the individual models has data for fewer accounts than the data set used for the

block group models. The missing data could account for the differences. To test this explanation, DNV

GL, aggregated the account-level data set up to the block group level and reran the correlations. These

correlations look very similar to the ACS block group model results with positive effects for low income

and multifamily (Table 4-14). Because the aggregated model appears the same as the ACS block group

models, we do not think the differences in the number of accounts covered by each data set explains the

discrepancies.

2. Aggregating the data up to the block group level somehow changes the relationships of the

variables. Our specific hypothesis is that the very large accounts dominate the block group statistics

because of their very high consumption (and potentially savings) values. However, in the account-level

analyses, these accounts have no greater weight in the statistics than any other (potentially tiny)

DNV GL – www.dnvgl.com February 6, 2020 Page 48

account. To test this explanation, we removed the largest 1% of accounts from each block group based

on annual consumption from the aggregated accounts data set and reran the correlations. With the

largest 1% of accounts removed, the effect of multifamily becomes not statistically significant. However,

the low-income correlation remains statistically greater than zero (Table 4-14). This pattern partially

supports this explanation. The positive correlation between the proportion of the block group that lives

in multifamily buildings and block group level savings/consumption is at least partially driven by the

large multifamily buildings. However, this pattern does not explain the low income effects.

Table 4-14. Savings/consumption models, aggregated account-level data, electric

Variable

Aggregated Accounts

Savings/ Avg.

Consumption

Aggregated Accounts,

<99 percentile by consumption,

Savings/ Avg. Consumption

Intercept 0.060 *** 0.062 ***

Proportion of Households at less than 56% of Statewide Median Income

0.018 ** 0.015 **

Proportion of Households at 56 - 85% of Statewide

Median Income -0.042 *** -0.027 **

Proportion of Renter Occupied Housing Units -0.033 *** -0.037 ***

Proportion of Housing Units in 5+ Unit Buildings 0.025 *** 0.006

Proportion of Limited English-speaking Households -0.012 -0.021 *

*** p <.001; ** p < .01; * p < .1

3. The positive savings/consumption correlations are caused by high levels of savings achieved through the

PAs’ low income multifamily programs. To test this explanation, we removed all savings associated with

low income multifamily programs from the block group data set and reran the correlations. In these

correlations (Figure 4-11), savings/consumption is negatively correlated with low income and

multifamily, at almost the same magnitude as location participation rates. This pattern supports this

explanation that the PAs’ low income multifamily programs achieve deep enough savings to counteract

the generally lower participation rates in areas with high concentrations of low income residents and

multifamily housing.

DNV GL – www.dnvgl.com February 6, 2020 Page 49

Figure 4-11. Block group correlations, electric, no LIMF savings

-0.46

-0.19

0.50

-0.49

-0.23

-0.35

-0.35

0.35

-0.17

-0.13

0.23

0.00

-0.16

-0.18

0.22

-0.14

0.20

-0.08

-0.30

0.30

0.22

-0.16

0.03

0.05

-0.37

-0.12

0.38

-0.41

-0.21

-0.30

-0.32

0.32

-0.16

-0.12

0.21

-0.09

-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5

Proportion of Low Income Housing Units

Proportion of Moderate Income Housing Units

Proportion of Average or Higher Income Housing

Units

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ Unit Buildings

Proportion of Limited English-speaking

Households

Proportion of Housing Units built before 1950

Proportion of Housing Units built in 1950 or Later

Proportion of Households with Electric Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PA Heating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,337)

Consumption Weighted Location Participation, 2013-2017 (n=4,329)

Savings over Average Consumption, 2013-2017 (n=4,348)

§ correlation NOT statistically

significant at p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 50

4.4 Gas location participation findings

4.4.1 ACS variable correlations with gas participation

Figure 4-12 presents the correlations between gas location and gas consumption-weighted location

participation and the primary ACS variables. Correlations between these variables are negative and

statistically significant. This indicates that block groups with higher concentrations of the term sheet

variables have lower participation rates.

Figure 4-12. Block group-level gas participation correlations (term sheet variables)

Figure 4-13 shows the correlations between gas participation and the additional ACS variables. Key findings

include:

• The proportion of low-income households is negatively correlated with participation, while the proportion

of high-income households is positively correlated with participation.

• The proportion of homes built before 1949 is negatively correlated with participation. All other home

vintage variables are positively correlated with participation. 1949 is the threshold where knob and tube

wiring and asbestos insulation become much less common. Other studies have shown that these older

building technologies are barriers to efficiency program participation because they have to remediated to

bring homes up to code before other efficiency measures can be installed.31

• The proportion of homes with natural gas heat or electric heat is negatively correlated with gas program

participation. In contrast, the proportion of homes with non-utility heat is positively correlated with gas

program participation. DNV GL cannot account for this pattern at this time.

31 DNV GL reported in a previous Eversource-only deliverable that knob and tube wiring and the asbestos are barriers in older homes because of the

extra cost required to bring those systems up to code.

-0.24

-0.46

-0.20

-0.33

-0.23

-0.41

-0.08

-0.34

-0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5

Proportion of Moderate Income Housing Units

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ Unit Buildings

Proportion of Limited English-speakingHouseholds

Correlation

Location Participation, 2013-2017 (n=4,127) all correlations statistically

significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 51

• The proportion of urban and rural households is significantly correlated with participation. Residents in

block groups with a greater proportion of urban blocks are less likely to participate in gas programs than

residents in block groups with a lower proportion of urban blocks. This finding should be interpreted with

caution because of the skew towards urban blocks in Massachusetts.

Figure 4-13. Block group-level gas participation correlations (extra ACS variables)

4.4.2 Gas participation block group-level models

Table 4-15 presents the gas block group model results for location and gas consumption-weighted

participation variables. In both models, the relationships between the ACS variables and participation rate

remain similar to the zero-order correlations. This indicates that they each have independent effects on

participation. The exceptions are multifamily housing and limited English; the effects of which are different

in the models than in the zero-order correlations. This indicates these variables are interacting with at least

one of the other variables in the model.

-0.48

-0.24

0.52

-0.31

0.31

-0.18

-0.11

0.24

-0.10

-0.46

-0.23

0.51

-0.33

0.33

-0.09

-0.16

0.22

-0.11

-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6

Proportion of Low Income Housing Units

Proportion of Moderate Income Housing Units

Proportion of Average or Higher Income

Housing Units

Proportion of Housing Units built before 1950

Proportion of Housing Units built in 1950 orLater

Proportion of Households with Electric Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PA Heating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,133)

Consumption Weighted Location Participation, 2013-2017 (n=4,053)

§ correlation NOT statistically significant at p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 52

Table 4-15. Initial gas block group models

Variable

Location Participation

Consumption Weighted

Participation

Estimate Estimate

Intercept 0.339*** 0.358***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.227*** -0.217***

Proportion of Renter Occupied Housing Units -0.198*** -0.215***

Proportion of Housing Units in 5+ Unit Buildings 0.082*** 0.163***

Proportion of Limited English-speaking Households -0.105*** -0.197*** *** p <.001; ** p < .01; * p < .1

DNV GL – www.dnvgl.com February 6, 2020 Page 53

There are three effects that change from the zero-order correlations when all the variables are added

together into the models:

1. The effect of renting remains statistically significant but is much weaker than the zero-order correlation.

2. The effect of limited English-speaking households remains statistically significant but is much weaker

than the zero-order correlation.

3. The effect of multifamily housing changes sign from a negative correlation to a positive regression

coefficient.

Additional models (location participation: Table 4-16; consumption-weighted location participation: Table

4-17) revealed that the renter variable is interacting with the multifamily and limited English variables. For

both types of participation, when the renter variable is absent from the model, the effects of multifamily

housing and limited English are similar to their zero-order correlations.

Table 4-16. Gas location participation models; renter variable present/absent

Variable

Without % Rental

Full Model

Estimate Estimate

Intercept 0.313*** 0.339***

Proportion of Households at 56 - 85% of Statewide Median

Income -0.315*** -0.227***

Proportion of Renter Occupied Housing Units -0.198***

Proportion of Housing Units in 5+ Unit Buildings -0.046*** 0.082***

Proportion of Limited English-speaking Households -0.334*** -0.105*** *** p <.001; ** p < .01; * p < .1

Table 4-17. Gas consumption weighted location participation models renter variable

present/absent

Variable

Without % Renter

Full Model

Coefficient Coefficient

Intercept 0.331 *** 0.358 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.317 *** -0.217 ***

Proportion of Renter Occupied Housing Units -0.215 ***

Proportion of Housing Units in 5+ Unit Buildings 0.024 ** 0.163 ***

Proportion of Limited English-speaking Households -0.448 *** -0.197 ***

*** p <.001; ** p < .01; * p < .1

The interpretation of these interactions is similar to that for electric participation. The interactions between

renter and the other two variables suggest that the effect of the other variable (limited English or

multifamily) is partially explained by those groups’ tendency to also rent.

• Limited English is a barrier to participation partially because of the tendency of that population to rent.

Limited English speakers tend to rent and not participate. Renters also tend to not participate. When the

tendency of renters to not participate is accounted for, the strength of the effect of Limited English

decreases by about half. The fact that Limited English remains a significant predictor indicates that

DNV GL – www.dnvgl.com February 6, 2020 Page 54

renting only partially explains the effect of Limited English. Even after the tendency of limited English

speakers to rent is accounted for, there still appears to be a statistically significant language barrier.

This is illustrated in the following pattern:

- Renter and limited English are positively correlated with each other (r=0.58) and both negatively

correlated with location participation (Renter r = -0.46; Limited English r= -0.33)

- In the model without Renter, Limited English remains a significant, negative predictor of

participation (Location coefficient=-0.448)

- When Renter is added to the model, renter is a significant, negative predictor of participation

(Location coefficient=-0.215) and limited English remains significant, but has a lesser magnitude

(Location coefficient=-0.197).

- Consumption-weighted participation has the same pattern, but with slightly different numeric

values.

• Multifamily housing is a barrier to participation mostly because of the tendency of that population to

rent; absent the association with renting, multifamily locations are positively associated with location

participation. A similar pattern applies to multifamily housing and renting. People living in multifamily

housing tend to rent and not participate. Renters tend to not participate. When we account for the

nonparticipation tendency of renters, the effect of Multifamily actually becomes slightly (but statistically

significantly) positive. In the case of consumption-weighted participation, Multifamily starts out

positively associated with participation and renting. Renting is still negatively associated with

participation, so it decreases the Multifamily effect. When the effect of Renting is removed from

Multifamily, the positive effect of Multifamily becomes stronger.32

A graph of the expected values from the regression model illustrates this relationship (Figure 4-14). The

positive effect of multifamily is evident in the vertical separation between the lines. That vertical

separation decreases as the proportion of renters increases.

32 The amplification of the positive effect of Multifamily in the location weighted model is caused by the weighting. Large multifamily buildings have a

higher chance of participation both because they are targeted more aggressively by the PAs and because there are more residents there and

therefore more chance for at least one of them to participate. The weighting makes it so that the whole building gets credit even if only one unit

participates.

DNV GL – www.dnvgl.com February 6, 2020 Page 55

Figure 4-14. Expected gas location participation rate (renter * multifamily)

Adding the other variables to the model, such as home construction vintage and heating fuel, did not cause

significant changes to the results reported in this section (Table 4-18). The direction of all coefficients stayed

the same. The strengths of the coefficients decreased to some degree, but that is expected when adding

variables with covariance to a model.

Table 4-18. Gas block group model results with all variables

Variable

Location

Participation

Consumption-weighted

Location Participation

Coefficient Coefficient

Intercept 0.363 *** 0.411 ***

Proportion of Households at less than 56% of Statewide Median Income

-0.260 *** -0.259 ***

Proportion of Households at 56 - 85% of Statewide Median Income

-0.199 *** -0.243 ***

Proportion of Renter Occupied Housing Units -0.084 *** -0.064 ***

Proportion of Housing Units in 5+ Unit Buildings 0.047 *** 0.116 ***

Proportion of Limited English-speaking Households -0.032 -0.112 ***

Proportion of Households Built in 1949 or Earlier -0.043 *** -0.060 ***

Proportion of Households with Natural Gas Heating 0.006 -0.001

Proportion of Households with Electric Heating 0.017 0.017

Proportion of Urban PA Households 0.014 -0.003

0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

Rent 0% Rent 50% Rent 100%

Expecte

d P

art

icip

ation R

ate

s

MF 0% MF 50% MF 100%

DNV GL – www.dnvgl.com February 6, 2020 Page 56

4.4.3 Gas participation individual-level models

Data available at the individual account level (statewide) included renting, multifamily, construction date,

and urban/rural. In the enhanced individual-level models (Eversource only), we were able to also include the

effects of income and limited English.

As stated in the method section, we used account-level participation for the individual models because it is

at a grain similar to most of the independent variables. This type of participation reflects participation while

the resident lived in a particular house or apartment.

Correlations for statewide data

Participation is negatively correlated with all four demographic variables available statewide (Table 4-19).

Unlike the block group level models, these findings are not subject to ecological fallacy and can be

interpreted to indicate effects at an individual level. The individual-level correlations are consistent with the

block-group-level correlations in direction. The renter and pre-1950 construction correlations are not as

strong; the multifamily and urban correlations are of similar magnitude. The comparisons between block

group and individual-level correlations for gas are very similar to those for electric. Income and language

data is not available at this data grain.

Table 4-19. Individual-level correlations – Statewide Gas

Variable Account

Participation, 2013-2017

Multifamily Renter Urban Pre-1950

Construction

Account Participation, 2013-

2017 - -0.09 -0.13 -0.02 -0.09

Multifamily -0.09

- 0.25 0.04 0.13

Renter -0.13 0.25

- 0.06 0.27

Urban -0.02 0.04 0.06

- 0.11

Pre-1950 Construction -0.09 0.13 0.27 0.11 - Over 1 million observations per correlation. All correlations statistically significant p<0.001

Models for individual-level data

The individual-level models use binary logistic regression to predict a dependent variable that is set to 0 for

nonparticipating accounts and 1 for participating accounts. The coefficients of these models are logs of odd

ratios, which are difficult to interpret directly. The magnitude of the coefficient still reflects the relative

strength of the relationship. The sign indicates the direction of the relationship. Positive signs mean the

probability of participation increases as the independent variable increases and vice versa for negative

values.

Table 4-20 shows several model results predicting account participation for the variables that are available

statewide. These model results are consistent with the block group-level models. All of the independent

variables are negative predictors of participation. That is, renters are less likely to participate than non-

renters; people living in multifamily buildings are less likely to participate than those in other buildings, etc.

Because these models utilize individual-level variables, they can be interpreted as individual-level effects.

One effect is inconsistent with the block group level models. The main effect of multifamily residency in the

individual level models is the opposite of the effect in the block group models. In the block group models,

DNV GL – www.dnvgl.com February 6, 2020 Page 57

multifamily was a positive predictor of participation, especially when controlling for the effect of renting. The

difference in results is probably caused by the differences in the dependent variables in each model. In the

block group models, the dependent variable was location (or consumption weighted location) participation.

In the individual level models, the dependent variable is account participation. Location participation is

probably skewing the model outcomes. Because participation by any single unit in a multifamily building

creates a location participation value of 1, the more units in the building, the greater the probability that at

least one of them participated. In contrast, account participation is not affected by the number of units in a

building, and the model results indicate that individual units in multifamily buildings have a lower probability

of participating than those in non-multifamily buildings.

Table 4-20. Gas account participation models coefficients – statewide

Variable Model 1 Model 2 Model 3

Intercept -1.489 *** -1.489 *** -1.256 ***

Renter -0.569 *** -0.569 *** -0.461 ***

Resides in multifamily building -1.699 *** -1.575 *** -1.535 ***

Renter * Multifamily 1 -0.134 * -0.183 *

Resides in Urban block -0.098 ***

Pre-1950 construction -0.309 ***

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

Similar to the block group-level model results, there is also a significant interaction between renter and

multifamily status in these models. Figure 4-15 illustrates the nature of that interaction by graphing the

expected probability of participation for the four combinations of renter and multifamily status. For non-

renters (Rent=0) and renters (Rent=1), multifamily has a negative effect on the probability of participation.

However, the negative effect of multifamily is weaker for renters, which is indicated by the shorter distance

between the two columns in the renter category. It should be noted this relationship is symmetrical: the

negative effect of renting is lesser in multifamily buildings than it is in non-multifamily buildings. This is the

same in the electric participation models.

DNV GL – www.dnvgl.com February 6, 2020 Page 58

Figure 4-15. Renter* multifamily interaction on gas account participation probability – statewide

Correlations for enhanced individual-level data

In the enhanced individual-level data set, which primarily uses the Experian demographic data available only

for Eversource accounts, account participation is negatively correlated with multifamily, limited English,

moderate-income, pre-1950 construction, and urban (Table 4-21). These correlations are in the same

directions as the block-group and non-enhanced individual-level correlations. The effect for multifamily is

similar in magnitude, but the other effects are weaker than in the other data sets.

Table 4-21. Individual-level correlations – Enhanced Gas data

Variable

Accou

nt

Parti

cip

ati

on

,

20

13

-2

01

7

Low

In

co

me

Mod

erate

In

com

e

Avg

or H

igh

er

In

com

e

Ren

ter

Mu

ltif

am

ily

Lim

ited

En

gli

sh

Pre-1

95

0

Con

str

ucti

on

Urb

an

Account Participation, 2013-2017

- -0.09 -0.04 0.11 -0.18 -0.20 -0.04 -0.09 -0.02

Low Income -0.09 - -0.32 -0.53 0.30 0.21 0.07 0.15 0.03

Moderate Income -0.04 -0.32 - -0.64 -0.01 0.03 0.02 0.10 0.03

Average or Higher Income

0.11 -0.53 -0.64 - -0.23 -0.20 -0.07 -0.21 -0.05

Renter -0.18 0.30 -0.01 -0.23 - 0.40 0.11 0.29 0.06

Multifamily -0.20 0.21 0.03 -0.20 0.40 - 0.09 0.43 0.08

Limited English -0.04 0.07 0.02 -0.07 0.11 0.09 - 0.05 0.02

Pre-1950 Construction -0.09 0.15 0.10 -0.21 0.29 0.43 0.05 - 0.11

Urban -0.02 0.03 0.03 -0.05 0.06 0.08 0.02 0.11 -

Approximately 1 million observations per correlation. All correlations statistically significant p<.01.

18%

11%

4%2%

0%

5%

10%

15%

20%

25%

30%

35%

Rent 0 Rent 1

Pro

bability o

f Part

icip

ation

MF 0 MF 1

DNV GL – www.dnvgl.com February 6, 2020 Page 59

Models for enhanced individual-level data

The results of the enhanced models are very similar to those for the all-PA data models. Renter and

multifamily are both negatively associated with participation (Table 4-22).

The enhanced models include a variable for moderate-income that the all-PA data models could not. As

observed in the correlations and in the block group-level models, moderate-income is negatively associated

with participation. The enhanced models also include a variable for limited English that is negatively

associated with participation.

Table 4-22. Gas account participation models coefficients – enhanced data

Variable Model 1 Model 2 Model 3 Model 4 Model 5

Intercept -0.982 *** -0.872 *** -0.884 *** -0.873 *** -0.708 ***

Experian Low

Income -0.320 ***

Experian Moderate Income

-0.212 *** -0.233 *** -0.239 *** -0.233 *** -0.262 ***

Experian Renter -1.080 *** -0.847 *** -1.069 *** -0.953 ***

Experian Resides in Multifamily Building

-1.221 *** -0.870 *** -0.780 *** -0.870 *** -0.693 ***

Experian Limited English

-0.145 *** -0.110 *** -0.108 *** -0.103 *** -0.136 ***

Resides in Urban

Block 0.068 *

Pre-1950 Construction

-0.449 ***

Experian Renter * Multifamily1

-0.561 ***

Experian Renter * Limited English1

-0.052

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other variable.

The interaction between renting and limited English is of similar magnitude as in the statewide individual-

level model, but not statistically significant in the enhanced models.

There is a statistically significant interaction between renter and multifamily in the enhanced models. Figure

4-16 illustrates the nature of that relationship. It is similar to the relationship observed in the individual-

level models. In the enhanced individual-level models, the negative effect of multifamily on the probability

of participation is lesser for renters than it is for non-renters (and vice versa).

DNV GL – www.dnvgl.com February 6, 2020 Page 60

Figure 4-16. Renter* multifamily interaction on gas account participation probability – enhanced

data

4.5 Gas savings/consumption findings

4.5.1 Block group gas analysis

Figure 4-17 shows the zero-order correlations between the gas participation metrics, income, the term sheet

variables and several other ACS variables at the block group level. Similar to electric, there is a positive

correlation between multifamily housing and savings/consumption, and a negative correlation between

multifamily housing and location participation.

For gas, most of the characteristics that were negatively correlated with location participation have

correlations with savings/consumption that are not significantly different than zero. A notable exception is a

correlation between multifamily and gas savings/consumption, which is positive (compared to a negative

correlation with location participation.) This pattern suggests that gas PAs are achieving deeper

savings/consumption in high-multifamily areas. It also suggests they are achieving enough gas savings per

participating household to counteract the lower participation rates for renters, moderate income, limited

English, and low income customers.

28%

14%15%

4%

0%

5%

10%

15%

20%

25%

30%

35%

Rent 0 Rent 1

Pro

bability o

f Part

icip

ation

MF 0 MF 1

DNV GL – www.dnvgl.com February 6, 2020 Page 61

Figure 4-17. Block group gas correlations

-0.48

-0.24

0.52

-0.46

-0.20

-0.33

-0.31

0.31

-0.18

-0.11

0.24

-0.10

-0.46

-0.23

0.51

-0.41

-0.08

-0.34

-0.33

0.33

-0.09

-0.16

0.22

-0.11

0.00

-0.02

0.00

0.02

0.11

0.01

-0.08

0.08

0.12

-0.07

-0.02

0.01

-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6

Proportion of Low Income Housing Units

Proportion of Moderate Income HousingUnits

Proportion of Average or Higher IncomeHousing Units

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ UnitBuildings

Proportion of Limited English-speakingHouseholds

Proportion of Housing Units built before1950

Proportion of Housing Units built in 1950 orLater

Proportion of Households with Electric

Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PAHeating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,133)

Consumption Weighted Location Participation, 2013-2017 (n=4,053)

Savings over Average Consumption, 2013-2017 (n=4,498)

§ correlation NOT statistically significant at p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 62

Ordinary least squares (OLS) models revealed interesting dynamics between low and moderate income,

renter, multifamily, and limited English regarding savings/consumption. When low income, moderate

income, renter, multifamily, and limited English are in the model (model 1 in Table 4-23), moderate income,

renter, and limited English all become statistically significant predictors of savings/consumption. The

strength of the effect of multifamily (relative to the zero-order correlation) is reduced in model 1 while the

effect of moderate income is increased. This suggests that the covariation between moderate income and

multifamily status reduces the overall effect (correlation) of moderate income. When that effect is controlled

for in the model, moderate income becomes a significant predictor. The interaction between multifamily and

renter (model 4) helps explain why the renter effects in the models are stronger than the correlations.

Table 4-23. Gas savings/consumption model results

Variable Model 1 Model 2 Model 3 Model 4

Coefficient Coefficient Coefficient Coefficient

Intercept 0.057 *** 0.056 *** 0.061 *** 0.059 ***

Proportion of Households at less than 56% of Statewide Median Income

-0.003 0.000 -0.027 *** -0.005

Proportion of Households at 56 - 85% of Statewide Median Income

-0.044 *** -0.043 *** -0.036 *** -0.041 ***

Proportion of Renter Occupied Housing Units

-0.036 *** -0.031 *** -0.030 *** -0.039 ***

Proportion of Housing Units in 5+ Unit Buildings

0.046 *** 0.045 *** 0.013 * 0.022 *

Proportion of Limited English-speaking Households

0.030 ** 0.020 * 0.027 **

Proportion of Households at less than 56% of Statewide Median Income * Proportion of Housing Units in 5+ Unit Buildings

0.082 ***

Proportion of Renter Occupied Housing Units * Proportion of Housing Units in 5+ Unit Buildings

0.037 **

* Statistically significantly different from 0 at p<.05 or stronger

There are several significant interactions occurring in the gas models. First, as shown in Model 3 above,

multifamily and low income interact. The effect of multifamily inverts the low income effect such that at high

levels of low income and multifamily, we expect higher participation rates, but at high levels of low income

and low levels of multifamily, we expect very low participation rates (as measured by savings/consumption;

Figure 4-18). This is the same pattern observed in the electric analysis. Even though the main effect (zero-

order correlation) of low income is lowered participation rates, the significant interaction supports the

hypothesis that multifamily programs, particularly low income multifamily programs, are achieving relatively

deep savings.

DNV GL – www.dnvgl.com February 6, 2020 Page 63

Figure 4-18. Interaction of low income and multifamily on gas savings/consumption

As shown in Model 4 above, there is also a significant interaction between multifamily and renting. Renting

has a stronger effect on non-multifamily buildings than in multifamily buildings, as indicated by the distance

between the lines at MF=0% and MF=100% in Figure 4-19. In multifamily buildings, there is almost no

distinction in expected savings/consumption ratios based on renting status. However, in non-multifamily

buildings, renters can be expected to have lower savings/consumption ratios than non-renters.

Figure 4-19. Interaction of rent and multifamily on gas savings/consumption

4.5.2 Individual-level gas analysis – Statewide

Figure 4-20 shows the statewide, zero-order, account-level correlations between the gas participation

metrics, multifamily, renter, urban, and pre-1950 construction (other characteristics are not available at the

account level). Account participation correlations are repeated here for the sake of contrast. Unlike the

0%

2%

4%

6%

8%

10%

12%

Low Income 0% Low Income 50% Low Income 100%

Expecte

d P

art

icip

ation R

ate

s

MF 0% MF 50% MF 100%

0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

MF 0% MF 50% MF 100%

Expecte

d P

art

icip

ation R

ate

s

Rent 0% Rent 50% Rent 100%

DNV GL – www.dnvgl.com February 6, 2020 Page 64

block-group level correlations, the account-level correlations with savings/consumption go in the same

direction as for account participation.

Figure 4-20. Individual-level correlations with savings over consumption, statewide gas

Table 4-24 shows results for models that DNV GL used to isolate the effects of the different variables

independently of the other variables. In all three models, the coefficients go in the same direction and are

approximately the same magnitude as the zero-order correlations. This means that the covariation between

the various characteristics (e.g., multifamily residents tend to rent) does not mask or substantially change

the effects of any singular characteristic.

Table 4-24. Individual-level savings over consumption models, statewide gas

Variable Model 1 Model 2 Model 3

Intercept 0.028 *** 0.028 *** 0.032 ***

Renter -0.009 *** -0.009 *** -0.006 ***

Resides in multifamily

building -0.015 *** -0.017 *** -0.015 ***

Renter * Multifamily 1 0.002 0.000

Resides in Urban block 0.001

Pre-1950 construction -0.009 ***

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

4.5.3 Individual-level gas analysis – Enhanced

Figure 4-21 shows the Eversource-only, enhanced data, zero-order account-level correlations between the

electric participation metrics, income, renter, multifamily, limited English, construction vintage, and urban.

-0.09

-0.13

-0.02

-0.09

-0.06

-0.07

-0.01

-0.07

-0.15 -0.10 -0.05 0.00 0.05 0.10 0.15

Multifamily

Renter

Urban

Pre-1950 Construction

Correlation

Account Participation, 2013-2017

Savings/ Avg. Consumption, 2013-2017all correlations statistically significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 65

Account participation correlations are repeated here for the sake of contrast. Unlike the block-group level

correlations and similar to the statewide account-level correlations, the enhanced account-level correlations

with savings/consumption go in the same direction as for enhanced account participation.

Figure 4-21. Individual-level correlations with savings over consumption, enhanced gas

Table 4-25 shows results for models that DNV GL used to isolate the effects of the different variables

independently of the other variables. Except for low and renter, the coefficients for each of the variables are

similar to the zero-order correlations. This means the overall effect of the variable (correlation) is similar to

the independent effect when controlling for the covariation with the other characteristics (coefficients).

The strength of the effects for low income and renter was reduced in the models relative to their

correlations. This suggests that some of the overall effect (correlation) of these variables is because of their

overlap with other variables in the model. For example, renter and multifamily are correlated, so the overall

negative effect of renters is partially due to the overlapping negative effect of multifamily.

-0.09

-0.04

0.11

-0.18

-0.20

-0.04

-0.09

-0.02

-0.05

-0.02

0.06

-0.12

-0.12

-0.03

-0.07

-0.01

-0.30 -0.20 -0.10 0.00 0.10 0.20 0.30

Low Income

Moderate Income

Average or Higher Income

Renter

Multifamily

Limited English

Pre-1950 Construction

Urban

Correlation

Account Participation, 2013-2017

Savings/ Avg. Consumption, 2013-2017

all correlations statistically significant p<.01

DNV GL – www.dnvgl.com February 6, 2020 Page 66

Table 4-25. Individual-level savings over consumption models, enhanced gas

Variable Model 1 Model 2 Model 3

Intercept 0.048 *** 0.048 *** 0.048 ***

Experian Low Income -0.005 *** -0.004 *** -0.005 ***

Experian Moderate Income -0.005 *** -0.005 *** -0.005 ***

Experian Renter -0.020 *** -0.020 *** -0.020 ***

Experian resides in multifamily building -0.017 *** -0.017 *** -0.017 ***

Experian limited English -0.004 *** -0.004 *** -0.003 ***

Experian Renter * Limited English1 -0.001

Experian Low Income * Multifamily1 -0.002

*** p <.001; ** p < .01; * p < .1

1 Notation of variable a*b indicates an interaction term. A significant interaction means the effect of one variable depends on the level of the other

variable.

4.5.4 Location participation and savings/consumption inconsistencies

There are inconsistent findings when participation is measured as location participation and when it is

measured as savings/consumption. In particular, at the block group level, low income and multifamily are

negatively correlated with location participation, but non-correlated or positively correlated with

savings/consumption (Table 4-26).

Table 4-26. Block group participation models comparison, electric

Variable

Location Participation

Savings /Consumption

Correlation Correlation

Proportion of Households at less than 56% of Statewide Median Income

-0.48 ** 0.00

Proportion of Housing Units in 5+ Unit Buildings -0.20 ** 0.11 ** *** p <.001; ** p < .01; * p < .1

The relationships between low income and savings/consumption remain non-significant in most of the

regression models, and the relationship between multifamily and savings/consumption remains positive and

statistically significant in all of the regression models (Table 4-23). However, at the individual account level,

low income and multifamily both have negative effects (Table 4-25).

We explored three potential explanations for these inconsistencies:

1. The data set used for the individual models has data for fewer accounts than the data set used for the

block group models. The missing data could account for the differences. To test this explanation, DNV

GL, aggregated the account-level data set up to the block group level and reran the correlations. These

correlations look very similar to the ACS block group model results with non-significant effects for low

income and a positive (but non-significant) effect for multifamily (Table 4-14). The lack of a statistically

significant multifamily effect suggests that the individual account data set could be missing some

multifamily building savings that are otherwise represented in the block group models.

2. Aggregating the data up to the block group level somehow changes the relationships of the

variables. Our specific hypothesis is that the very large accounts dominate the block group statistics

because of their very high consumption (and potentially savings) values. However, in the account-level

analyses, these accounts have no greater weight in the statistics than any other (potentially tiny)

DNV GL – www.dnvgl.com February 6, 2020 Page 67

account. To test this explanation, we removed the largest 1% of accounts from each block group based

on annual consumption from the aggregated accounts data set and reran the correlations. With the

largest 1% of accounts removed, the effect of low income becomes significantly negative and the

multifamily effect remains not statistically significant (Table 4-14). This pattern partially supports this

explanation through the changes in the low income coefficients. However, multifamily is still not

explained.

Table 4-27. Savings/consumption models, aggregated account-level data, electric

Variable

Aggregated

Accounts Savings/ Avg. Consumption

Aggregated Accounts,

<99 percentile by consumption,

Savings/ Avg. Consumption

Intercept 0.040 *** 0.041 ***

Proportion of Households at less than 56% of Statewide Median Income

-0.004 -0.013 **

Proportion of Households at 56 - 85% of Statewide Median Income

-0.014 * -0.006

Proportion of Renter Occupied Housing Units -0.011 ** -0.012 **

Proportion of Housing Units in 5+ Unit Buildings 0.004 0.005

Proportion of Limited English-speaking Households -0.032 *** -0.029 ***

*** p <.001; ** p < .01; * p < .1

3. The positive savings/consumption correlations are caused by high levels of savings achieved through the

PAs’ low income multifamily programs. To test this explanation, we removed all savings associated with

low income multifamily programs from the block group data set and reran the correlations. In these

correlations (Figure 4-22), savings/consumption is negatively correlated with low income, and not

significantly associated with multifamily. This pattern supports the explanation that the PAs’ low income

multifamily programs achieve deep enough savings to counteract the generally lower participation rates

in areas with high concentrations of low income residents and multifamily housing. The magnitude of the

remaining effect also suggests that savings outside the low income multifamily program remain deep

enough to partially counteract the effect of low participation.

DNV GL – www.dnvgl.com February 6, 2020 Page 68

Figure 4-22. Block group correlations, gas, no LIMF savings

-0.50

-0.23

0.55

-0.48

-0.22

-0.36

-0.31

0.31

-0.20

-0.11

0.25

-0.10

-0.46

-0.23

0.50

-0.40

-0.07

-0.34

-0.33

0.33

-0.09

-0.16

0.22

-0.11

-0.10

0.00

0.09

-0.08

0.01

-0.07

-0.11

0.11

0.04

-0.05

0.03

-0.01

-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6

Proportion of Low Income Housing Units

Proportion of Moderate Income Housing Units

Proportion of Average or Higher Income Housing Units

Proportion of Renter Occupied Housing Units

Proportion of Housing Units in 5+ Unit Buildings

Proportion of Limited English-speaking Households

Proportion of Housing Units built before 1950

Proportion of Housing Units built in 1950 or Later

Proportion of Households with Electric Heating

Proportion of Households with Gas Heating

Proportion of Households with Non-PA Heating

Proportion of Urban PA Households

Correlation

Location Participation, 2013-2017 (n=4,126)

Consumption Weighted Location Participation, 2013-2017 (n=4,055)

Savings over Average Consumption, 2013-2017 (n=4,521)

§ correlation NOT

statistically significant at

p<.05 or stronger

DNV GL – www.dnvgl.com February 6, 2020 Page 69

5 CONCLUSIONS, RECOMMENDATIONS, AND CONSIDERATIONS

5.1 Conclusions

5.1.1 Location participation rates for 2013-2017 are negatively associated with all three term sheet variables: moderate-income

households, renter households, and limited English-speaking

households.

These relationships hold across the block group-level correlations and models as well as the individual-level

correlations and models. This indicates that populations represented by the term sheet participate at lower

levels than other populations.

5.1.2 PA efforts to obtain participation from large multifamily locations

have resulted in some inclusion of the populations outlined in the

term sheet.

There is evidence that the PAs have had some success in getting multifamily locations to participate,

especially when savings/consumption is considered. Because multifamily is correlated with renters, low-

income, and limited English speaking, the PAs’ success with multifamily locations has enabled some of the

term sheet-related populations to obtain benefits from the programs.

5.1.3 When participation is measured using 2013-2017

savings/consumption, it is positively correlated with low income and multifamily at the block group level.

In contrast to location participation rates, participation measured as the ratio of savings to consumption is

positively correlated with: low income, renter, multifamily, and limited English. However, these positive

correlations hold only at the block group level of analysis. In the individual-level analyses, the low income

and the term sheet variables are still negatively associated with participation measured as

savings/consumption. We suspect but have not been able to fully confirm that the differences between the

block group-level and individual-level findings are caused by the participation of the largest MF buildings.

The very large MF buildings tend to dominate the block group totals and averages. However, those buildings

have greater weight than any other account in the individual-level analyses. The inconsistency between the

block group and individual level models for this metric of participation indicates that the risk of ecological

fallacy is greater in these analyses than in the location participation analyses.

5.1.4 Location participation rates for 2013-2017 are affected by multiple factors.

Location participation is negatively correlated with: low income, pre-1950 construction, and PA-delivered

heating fuel (electric or gas). Location participation is positively correlated with average or higher income,

post-1950 construction, and receiving heating fuel from a non-PA source.

5.1.5 Most of the variables investigated are correlated with each other,

especially in the ACS data.

This suggests that increasing participation among one of the term sheet-related populations will also

increase participation among the other term sheet-related populations. For example, increasing participation

DNV GL – www.dnvgl.com February 6, 2020 Page 70

among renters will also create some increased participation among limited English speakers (and vice

versa).

5.1.6 Term sheet-related populations are geographically clustered in

urban areas.

In addition to being correlated with each other, the term sheet variables also exhibit spatial clustering in

urban areas. This suggests that the PAs’ community outreach partnerships and continued engagement with

local community stakeholders including the action agencies and civic groups are likely to have a positive

impact on the term sheet populations and benefit from positive word of mouth from trusted voices at the

local level.

5.1.7 Limited-English speakers are more likely to rent, and renters are

less likely to participate.

The tendency to rent explains some but not all of the reduced participation rates of the limited-English

population. Program designs that increase the participation of renters will also increase participation of

limited-English speakers as a “bonus.” However, additional outreach or program design would be required to

fully address the population of limited-English speakers.

5.1.8 The effects of the examined variables on participation are similar

in both electric and gas markets.

This indicates that the participation barriers faced by the populations identified by the term sheet variables

are probably similar for both electric and gas participation.

5.1.9 The individual-level models are mostly consistent with the block group-level models.

This indicates that conclusions based on the block group-level information are unlikely to suffer from

ecological fallacy for the analyses examined in this study. However, one should always be wary of the

possibility of ecological fallacy in the future. In particular, the effects of multifamily tended to be inconsistent

between the block group level and individual level models.

5.1.10 Comparison to the Market Barriers Study

The Market Barriers Study conducted primary research (including over 1,700 surveys and in-depth

interviews) to explore nonparticipant characteristics, investigate participation barriers, and identify PA

engagement opportunities. Key findings from that study, and their relationship to this study’s findings, are

shown in Table 1-1. Findings from both studies were generally consistent, and there were no cases of

completely contradictory findings.

DNV GL – www.dnvgl.com February 6, 2020 Page 71

Table 5-1. Key findings of residential nonparticipant studies

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Self-reported program participation agreed 68% with participation

as determined by tracking data.

This is a high level of agreement

between self-reported participation

and the tracking data. In addition,

most (88%) of the self-reported

participants that were not identified

via the tracking data indicated their

participation took place outside the

5-year window of tracking data.

Participants identified via the

tracking data but not self-reporting

tended to be renters, live in

multifamily buildings, and/or live in

their current home for 4 or fewer

years.

Nonparticipants were more likely

to live in:

• Rental units

• Low to moderate income

households

• Households that speak a non-

English language or report

having limited English

proficiency

Renters, customers who live

in moderate income

households, and customers

living in limited-English

speaking households are less

likely to participate than

customers without these

characteristics.

The findings are consistent across

both studies.

DNV GL – www.dnvgl.com February 6, 2020 Page 72

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Nonparticipants also tend to:

• Have lower education levels

• Be less aware of Mass Save

• Do not fully trust government,

landlords, or free offers

• Prioritize time and resources

for other issues such as food

and well-being

• Need additional information or

understanding of Mass Save

offerings, processes, and

benefits

• Perceive energy efficiency as

irrelevant or not applicable to

them

No related findings

These findings highlight some of the

increased depth and additional topics

the Market Barriers study was able

to address beyond the data that

were available to the Nonparticipant

Customer Profile study.

Nonparticipant renters are more

likely to:

• Reside in lowest participating

Census tracts

• Self-identify as low income

Most of the variables

investigated are correlated

with each other, especially in

the ACS block group data.

Term sheet-related

populations are

geographically clustered in

urban areas.

The findings are consistent across

both studies.

DNV GL – www.dnvgl.com February 6, 2020 Page 73

Residential Nonparticipant

Market Barriers

Residential

Nonparticipant

Customer Profile

Discussion

Moderate income nonparticipants

are more likely to reside in higher

participating Census tracts.

Most of the variables

investigated are correlated

with each other, especially in

the ACS block group data.

Term sheet-related

populations are

geographically clustered in

urban areas.

On the surface, these findings seem

contradictory. However, it is

important to note that the two

studies used different definitions of

moderate income. In the Market

Barriers study, the definition of

moderate income was $50,000 to

$75,000. In the Customer Profile

study, the definition of moderate

income was $40,000 to $59,999.

This means that part of the

moderate income group in the

Market Barriers study falls into the

“average or higher income” category

for the Customer Profile study. When

this cross-categorization is taken

into account, the findings are largely

consistent across both studies.

5.2 Recommendations

Based on the results of this study, DNV GL offers the following recommendations:

• Continue working closely with LEAN to incorporate their extensive data resources into this analysis.

• Future analyses of the types of questions examined in this study should utilize more readily accessible

ACS block-group-level data. The low participation table completed as deliverable 1 (and included in

Appendix B) is an example of such an analysis. Future analyses should remain aware of the strengths

and limitations of each way of measuring participation, and carefully choose which definition to use in

each analysis based on the conceptual question being asked.

• All PAs tracking participation at the individual unit level would increase the accuracy of evaluation results

for MF buildings. For implementers, knowing at a finer granularity what they did in a building would help

identify opportunities to revisit buildings that participated in the past.

• The PAs engage in various activities to communicate program successes and setbacks with each other.

We did not find any evidence that this activity is not effective. We recommend the PAs continue these

communication efforts so that electric and gas implementers continue to learn from each other’s

experiences.

• Use the metric provided for the Mass Save Municipal and Community Partnership Strategy as the

baseline for assessing the effectiveness of future PA efforts to address the term sheet populations. This

metric is provided in the separate Low Participation Table at both the block-group and town-level grains,

DNV GL – www.dnvgl.com February 6, 2020 Page 74

and is included in this report in Appendix B. In general, we recommend examining multiyear trends

rather than focusing on singular years.

5.3 Considerations

DNV GL also offers the following considerations:

• Future analyses should continue to assess and leverage data like Tax Assessor data to refine models.

For example, home type (e.g. Cape Cod, Victorian, Ranch) and construction date can help indicate

potential barriers. Home value might also be usable as a proxy for income.

• Given the overlap and geographic clustering of term sheet variables, the PAs may benefit from

examining their time series billing and tracking data to identify individuals who both participated in the

programs, and are in areas of high interest. Alternatively or in addition, PAs could use these geographic

clusters to inform future implementation of the Municipal and Community Partnership Strategy.

• Continue to explore program designs that will help these populations overcome the participation barriers

that they face. From our work nationally, DNV GL recognizes that these types of questions and

populations are not a unique consideration to MA PAs. The PAs should consider a meta-review of how

their ACEEE ranked nationally leading programs both set the stage for other states, as well as where

there are opportunities to quickly assimilate new program designs and ideas identified in the literature

review.

• Use the bivariate maps provided in this study and other methods of identifying geospatial areas to target

program outreach efforts. The bivariate maps provide a way to identify the block groups that are the

greatest outliers in terms of the high incidence of term sheet variables and low participation rates.

• Continue to explore program designs to increase renter participation. Also, continue to explore program

designs that specifically address limited-English speakers above and beyond the renter programs.

• The present study found evidence that there are significant interactions between multifamily buildings

and the term sheet variables that affect participation rates. The PAs treat multifamily locations

differently than single-family locations, and in some cases even record billing data differently. The

current study was not able to fully explore or explain all of these relationships due to budget and

schedule constraints. Carefully examine whether the recently completed Navigant MF Census study can

shed additional light on unanswered questions resulting from this study. If important questions remain,

consider conducting a follow-up multifamily study that attempts to fill in those gaps. A specific question

that should be considered in the scope of such a follow-up study is whether MF locations should use a

different participation metric than SF locations, or if using the same metric, whether it should be

considered separately for MF and SF locations.

• For the findings related to non-term sheet variables (construction year, urban/rural, and non-PA heating

fuel), limitations such as the skew in the distribution of urban versus rural block groups and the handling

of data related to delivered fuel projects should be considered more closely before taking action on

these findings.

DNV GL – www.dnvgl.com February 6, 2020 Page 75

6 APPENDIX A: PARTICIPANT/NONPARTICIPANT LIST PREPARATION

This appendix contains additional details of how we prepared the PNP list. These steps were previously

documented in a memo delivered to the working group on May 17, 2019.

Tracking data preparation

After combining the 2013 through 2017 residential tracking data, we had a total of 7,569,297 records across

all PAs, fuels, and year. These represent 1,087,153 unique account ids. Approximately 25% of the records

have a missing account id and are associated with the residential lighting program or residential products

program. Another 25% of the records are associated with account ids that appear at least 11 times per

year.

The following table shows how the number of unique accounts is distributed among fuels, PAs, and years.

These counts are all within 8% (most within 5%) of previous counts provided by DNV GL. The discrepancies

are caused by continuous updating of the databases since the previous counts were provided based on

ongoing clarification from the PAs.

Table 6-1. Unique accounts in tracking data (Fuel*PA*year)

Fuel PA 2013 2014 2015 2016 2017 Unique

Accounts

2013-2017

E Cape Light Compact

10,165 13,580 13,725 11,462 14,722 50,290

E Eversource 63,526 80,956 78,387 74,094 73,615 305,285

E National Grid 96,893 81,549 70,199 67,785 73,071 362,700

E Unitil 497 741 1,096 1,427 1,463 4,556

Electric Total 173,094 178,840 165,422 156,784 164,888 722,831

G Berkshire 2,073 2,132 1,773 1,529 1,762 8,064

G Columbia 12,668 12,291 13,472 16,726 16,030 66,962

G Eversource 14,545 12,602 14,262 14,331 18,326 62,203

G Liberty 1,194 1,669 1,545 1,554 1,625 7,146

G National Grid 55,333 50,996 48,330 36,752 43,611 218,501

G Unitil 127 282 330 458 367 1,446

Gas Total 85,940 79,972 79,712 71,350 81,721 364,322

Grand Total 259,034 258,812 245,134 228,134 246,609 1,087,153

Billing data preparation

The residential and C&I billing data were prepared separately but in similar ways.

Residential

The billing data processing started with the merged 2017 residential billing data from all of the PAs. The

grain of these data is multiple records per account id by month. We started with 52,925,322 records across

the state.

DNV GL – www.dnvgl.com February 6, 2020 Page 76

Step 1: Deduplicate down to unique accounts by month. Some accounts have multiple lines per

month, such as when they receive a bill correction. This step reduced the list down to a single record per

month per account. The final record count statewide was 47,511,738.

Step 2: Deduplicate down to unique accounts. In the output from step 1, we would see 12 records per

account that were active the entire year. This step reduces the data down to a single record per account

that was active sometime in 2017. We ended with 4,256,178 records after this step, for a reduction of

approximately 90%. Reducing the count by a factor of approximately 10 makes sense because some

accounts would not be active the entire year or receive bills every month. Each PA saw similar levels of

reduction. These counts are generally within 1% or less of the counts in the summary of data completeness

from the 2017 data intake process. The one exception is Columbia, for which we have approximately 13,000

additional unique account ids. All of the Columbia records appear to be legitimate locations.

Table 6-2. Residential unique accounts by fuel and PA

Fuel PA Input

Residential Billing Data

Step 1: Accounts *

Month

Step 2: Unique

Accounts

E Cape Light Compact

3,917,354 2,101,570 191,267

E Eversource 16,139,224 12,619,015 1,248,640

E National Grid 13,962,591 13,950,887 1,170,963

E Unitil 288,309 285,792 29,140

Electric Total 34,307,478 28,957,264 2,640,010

G Berkshire 415,850 415,418 35,310

G Columbia 4,386,899 4,362,685 349,432

G Eversource 3,217,134 3,179,842 316,173

G Liberty 572,665 572,441 51,917

G National Grid 9,865,803 9,865,799 846,678

G Unitil 159,493 158,289 16,658

Gas Total 18,617,844 18,554,474 1,616,168

Grand Total 52,925,322 47,511,738 4,256,178

C&I

The processing of the C&I data followed the same first two steps as residential. We then added an additional

filter to remove C&I records that were not residential. The billing data processing started with the merged

2017 C&I billing data from all of the PAs. The grain of these data is approximately account id by month. We

started with 6,315,653 records across the state.

Step 1: Deduplicate down to unique accounts by month. Some accounts have multiple lines per

month, such as when they receive a bill correction. This step reduced the list down to a single record per

month per account. The final statewide record count was 6,204, 575.

Step 2: Deduplicate down to unique accounts. In the output from step 1, we would see 12 records per

account that were active the entire year. This step reduces the data down to a single record per account

DNV GL – www.dnvgl.com February 6, 2020 Page 77

that was active sometime in 2017. We ended with 538,598 records after this step, for a reduction of

approximately 90%. Reducing the count by a factor of approximately 10 makes sense because some

accounts would not be active the entire year or receive bills every month. Each PA saw similar levels of

reduction. These counts are within 1% of the 2017 summary of data completeness. Columbia was an

exception here as well, with an additional 1,500 records on our list. Again, those all appear to be legitimate

locations.

Table 6-3. C&I Unique accounts by fuel and PA (before removing nonresidential)

Fuel PA Input

Residential Billing Data

Step 1: Accounts *

Month

Step 2: Unique

Accounts

E Cape Light

Compact 283,947 279,234 25,535

E Eversource 2,017,772 1,925,861 171,420

E National Grid 2,124,855 2,123,052 181,765

E Unitil 42,479 40,694 3,909

Electric Total 4,469,053 4,368,841 382,629

G Berkshire 63,632 63,566 5,380

G Columbia 468,886 468,505 35,502

G Eversource 339,869 329,636 30,187

G Liberty 48,757 48,668 4,334

G National Grid 907,108 907,107 78,730

G Unitil 18,348 18,252 1,836

Gas Total 1,846,600 1,835,734 155,969

Grand Total 6,315,653 6,204,575 538,598

Final billing file

The final step for preparing the billing file was to merge the residential and C&I data sets and then remove

records that do not appear to be residential.

Remove nonresidential records. The final step was to remove records that could not be identified as

single-family or multi-family through the tax assessor data and any that were clearly streetlights, cell phone

boosters, and wells. We also checked account names for variations on “housing authority” to make sure we

included as many of these records as we could. We applied the following filters during this step (see

Appendix A for documentation of how we classified each rate code):

4. Remove records according to rate codes. This resulted in a reduction of 37,319 records.

5. Remove records that were coded as neither residential nor multifamily based on tax assessor use codes.

This resulted in a reduction of 413,477 records.

6. Remove records with the following text strings in the customer name. This resulted in a reduction of

2,727 records.

'VERIZON' 'CINGULAR' 'T-MOBILE'

DNV GL – www.dnvgl.com February 6, 2020 Page 78

'AT&T' 'COMCAST' 'CELLULAR'

'SPRINT' 'LIGHT' 'PACKAGING' 'CVS' 'LIBRARY' 'WATERDEPT' 'TRAFFIC'

'FLASH' 'GOLFCOURSE' 'CHURCH' 'FLORIST' 'FITNESSCOMPANY' 'TAVERN' 'YOGA'

'BUSINESSCENTR' 'BUSINESSCENTER' 'AUTOMOTIVECENTER' 'PRINTING' 'YACHTCLUB' 'OFFICEPRODUCTS'

'ISLANDRESORTS' 'AIRPOR' 'AIRPORT' 'AUTORENTAL' 'PHYSICALTHER' 'AUTOTECH' 'WAREHOUS'

'HARBORPRESERV' 'CHIROPRACTIC' 'STOPANDSHOP' 'MEDICAL' 'POWEREQUIPMENT' ‘BARBERS EDGE’ ‘BARGAIN DISCOUNT MARKETS’

‘BEEF’ ‘SEAFOOD’ ‘SELF STORAGE’ ‘HSE OF PIZZA’

7. Add back in any records removed during steps 1-3 where the customer name contained the following

text strings. This added back in 7,021 records.

'HSG' 'HSNG' 'HOUISNG' 'HSING'

'HAS AUTH' '*FRHA' 'HSE' 'HOUSING' ' APTS' 'APARTMENTS'

' APT '

DNV GL – www.dnvgl.com February 6, 2020 Page 79

The final record counts by fuel and PA follow.

Table 6-4. Final record counts by fuel and PA

Fuel PA From Residential

Billing From C&I Billing Total

E Cape Light Compact 190,212 1,703 191,915

E Eversource 1,245,185 34,552 1,279,737

E National Grid 1,168,581 38,961 1,207,542

E Unitil 29,103 354 29,457

Total Electric 2,633,081 75,570 2,708,651

G Berkshire 35,223 621 35,844

G Columbia 348,951 3,262 352,213

G Eversource 315,619 3,375 318,994

G Liberty 51,867 147 52,014

G National Grid 845,449 8,989 854,438

G Unitil 16,641 134 16,775

Total Gas 1,613,750 16,528 1,630,278

Grand Total 4,246,831 92,098 4,338,929

Set participation flags

Account participation

We determined account participation by identifying any account ids in the final billing file that also appeared

in the final tracking file.

Address participation

We determined to address participation by identifying any addresses in the final billing file that also

appeared in the final tracking file. Addresses were matched based on the combination of Address1 (street

number and name), Address2 (supplemental address information), standardized city name, and zip code.

Because these were matches made on text strings, literal differences such as spelling out “street” versus

abbreviating it as “st” would fail to match.

Location participation

Most of the records (98% of residential, 90% of C&I) had a point-level geocoded latitude and longitude

location. For those records, if any records at a particular latitude*longitude had either account participation

or address participation, then location participation was set to 1. For records that lacked point-level

geocoded locations, location-level participation is undefined (coded as -99).

DNV GL – www.dnvgl.com February 6, 2020 Page 80

We also performed a second match directly between geocoded locations in the 2013-2017 tracking data and

the 2017 billing data to identify locations where participation occurred but the current account id was not in

the 2017 billing data and the address text strings did not match.

Other variables

ID low-income participants

Low-income participants were identified based on a crosswalk of specific rate codes previously provided by

the PAs. The following rate codes by PA were identified as low income. See Appendix A for documentation of

how we classified each rate code.

Add current account holder

The PAs provided current account holder lists in mid-2018. We flagged any account id in these lists as a

current account holder.

Experian variables

Below are relevant excerpts from the Experian data dictionary that provide additional details about how

Experian created some of the data fields that DNV GL used.

0113A - Combined Homeowner – DNV GL used Combined Homeowner as one of our checks of our Renter

flag. We coded R and T as renters and H as owners. Other values were coded as neither.

Combined homeowner is a mixture of several data elements / fields. This element provides these separate

data components in a single position. Homeowner information indicates the likelihood of a consumer owning

a home, and is received from tax assessor and deed information. For records where exact Homeownership

information is not available, homeownership propensity is calculated using a proprietary statistical model

which predicts the likelihood of homeownership. Renter status is derived from self reported data. Unit

numbers are not used to infer rented status because units may be owner condominium/coop. Probable

Renter is calculated using an algorithm based on lack of Homeowner, the Address Type, and Census Percent

Renter.

Valid Values:

H = Homeowner

9 = Extremely Likely

8 = Highly Likely

7 = Likely

R = Renter

T = Probable Renter

U = Unknown

0118 - Dwelling Type – DNV GL used Dwelling Type for the Multifamily variable. We coded M and P as

multifamily and S and P as not multifamily.

DNV GL – www.dnvgl.com February 6, 2020 Page 81

Each household is assigned a dwelling type code based on United States Postal Service (USPS) information.

Single Family Dwelling Units are residences for one family or living unit (S). If the address contains an

apartment number or has a small dwelling size (5 units or less), the code is set to Multi-Family (A). Marginal

Multi Family Dwelling Units lack an apartment number and are considered of questionable deliverability (M).

Values also include P.O. Boxes (P) and Unknown dwelling types (U).

Values:

S = Single Family

A = Multi-Family & Condominiums

M = Marginal Multi-Family

P = Post Office Box

U = Unknown

0741 - YEAR BUILT - (ENHANCED) – DNV GL used year built for the pre-1950 construction variable after

stripping the first digit off the variable.

Year built is based on country assessor's records, the year the residence was built or through the application

of a predictive model.

Values:

First position contains the model confidence flag with the following values:

1 = Extremely Likely

5 = Likely

Position 2-5 contains the year built values:

0000-9999

D106N - Est Household Income V5 – DNV GL used this variable to set the Low, Moderate, and Average

or Higher income variables. Categories A, B, C were set to Low Income; Moderate income was categories D,

E; Average or Higher income was F through L.

Estimated Household Income Code V5 (in thousands) is the total estimated income for a living unit, and

incorporates several highly predictive individual and household level variables. The income estimation is

determined using multiple statistical methodologies to predict the income estimate for the living unit. Note

for Enrichment: When there is insufficient data to match a customer's record to our enrichment master for

estimated income in thousands, a median estimated income in thousands based on the Experian modeled

incomes assigned to other living units in the same ZIP+4 area is used. In the rare case that the ZIP+4 is not

on the record, median income in thousands is based on the incomes assigned to other records in that ZIP

region. The median level data applied to records for this element can be identified through the Enrichment

Mandatory Append - Total Enrichment Match Type indicator (G). VALUES:

A= $1,000-$14,999

B= $15,000-$24,999

DNV GL – www.dnvgl.com February 6, 2020 Page 82

C= $25,000-$34,999

D= $35,000-$49,999

E= $50,000-$74,999

F= $75,000-$99,999

G= $100,000-$124,999

H= $125,000-$149,999

I= $150,000-$174,999

J= $175,000-$199,999

K= $200,000-$249,999

L= $250,000+

U= Unknown

0108L – Language – DNV GL used this variable to set the Limited English variable. Values not equal to 01

were considered limited English.

Ethnic Insight is a comprehensive predictive name analysis process which identifies ethnic origin, probable

religion, and the language preference of individuals.

Valid values:

00=Unknown

01=English

03=Danish

04=Swedish

05=Norwegian

06=Finnish

07=Icelandic

08=Dutch

10=German

12=Hungarian

13=Czech

14=Slovakian

17=French

19=Italian

20=Spanish

DNV GL – www.dnvgl.com February 6, 2020 Page 83

21=Portuguese

22=Polish

23=Estonian

24=Latvian

25=Lithuanian

27=Georgian

29=Armenian

30=Russian

31=Turkish

33=Greek

34=Farsi

35=Moldavian

36=Bulgarian

37=Romanian

38=Albanian

40=Slovenian

41=Serbo-Croatian

44=Azeri

45=Kazakh

46=Pashto

47=Urdu

48=Bengali

49=Indonesian

51=Burmese

52=Mongolian

53=Mandarin

54=Cantonese

56=Korean

57=Japanese

58=Thai

DNV GL – www.dnvgl.com February 6, 2020 Page 84

59=Malay

60=Laotian (Include Hmong)

61=Khmer

62=Vietnamese

63=Sinhalese

64=Uzbeki

65=Hmong

68=Hebrew

70=Arabic

72=Turkmeni

73=Tajik

74=Kirghiz

7E=Nepali

7F=Samoan

80=Tongan

86=Oromo

88=Gha

8G=Tibetan

8I=Swazi

8J=Zulu

8K=Xhosa

8M=Afrikaans

8O=Comorian

8S=Ashanti

8T=Swahili

8X=Hausa

92=Bantu

94=Dzongha

95=Amharic

97=Twswan

DNV GL – www.dnvgl.com February 6, 2020 Page 85

9E=Somali

9F=Macedonian

9N=Tagalog

9O=Sotho

9R=Malagasy

9S=Basque

DNV GL – www.dnvgl.com February 6, 2020 Page 86

7 APPENDIX B: LOW PARTICIPATION TABLE

Overview of Community Outreach Metric for 2017 Residential Nonparticipant Customer Profile

Study (MA19X06-B-RESNONPART)

This appendix provides a brief description of the Community Outreach Metric added to the Low Participation

Table for the 2017 Residential Nonparticipant Customer Profile Study (RNCPS). An embedded copy of the

Excel file with the table appears at the end of the appendix.

Community Outreach Metric formulation

Figure 7-1 shows how to calculate the Community Outreach Metric, along with its key features:

Figure 7-1. Community Outreach Metric formula and features3334

We calculate the Community Outreach Metric by taking the sum of the proportion of households matching

each term sheet variable, and dividing it by the account participation rate. We compute and report this

metric at both the block group level and the town level (as categorized by the Massachusetts State Tax

Assessor).

Key features of the Community Outreach Metric include:

• It looks at each block group as an independent point within an absolute data range of 0% to 100% for

each term sheet characteristic. This makes it easy to identify block groups with high proportions of one

or more term sheet characteristics. However, because the block groups are independent, this metric

does not show how block groups compare with one another in terms of high or low proportions of term

sheet characteristics.

• Its numerator is based on the term sheet characteristics. This means that the metric’s numeric value

increases as the prevalence of the term sheet characteristics increases. Households get counted once for

each characteristic they exhibit. For example, a renting, moderate-income household with limited

English (matching 3 of the term sheet characteristics) would be counted 3 times in the numerator.

• Its denominator is based on historic participation rates. By capturing historic participation in the

denominator, the metric’s numeric value is lower for communities that already have high participation

levels.

• The metric does not depend on community size. This puts small and large communities on equal footing

and gives program implementers the option to consider community population size separately from

other factors.

33 Census block groups are discreet, usually small, geographic areas. Since they are discreet and non-overlapping, data counts (e.g., number of low

income homes) can be aggregated and summed to less detailed geographic areas (e.g., a city, MSA, or county) and can be recombined to reflect non-standard areas (e.g., a CAP’s service zip codes, or informal neighborhoods) without having to restructure the data.

34 Deliverable 1 was an Excel workbook organized by block group and town. It listed the proportion of households in each block group (town) that

meet the three term sheet characteristics and that live in multifamily buildings. It also listed consumption-weighted location participation for

each block group. The tables are dynamic, allowing users to restrict the table to any permutation of PA territories.

��%�������, %��*�����. ������, %������*. ������ℎ"

#����� ����������� ����

Considers each block group2 individually

Households matching multiple term sheet characteristics count multiple times

Can be weighted to reflect different priorities by community or program year

Can be calculated directly from Deliverable 13

DNV GL – www.dnvgl.com February 6, 2020 Page 87

• The PAs can easily update the metric by updating the ACS data and the PAs’ data in future years.

Key limitations of the metric include:

• In communities with smaller populations, the measurements of the term sheet characteristics and

participation rates are more sensitive to small changes. The denominators in these measurements

represent the entire population. When those denominators are small, small changes in the numerators

have stronger effects on the resulting measurement. For example, a difference of 10 renters would

cause a 1% change in the renter proportion for a block group with 1,000 homes, but it would cause a

10% change in the renter proportion for a block group with 100 homes.

• Account participation rate is calculated by dividing the total number of unique account IDs that

participated from 2013 to 2017 (unique across the entire period) by the total number of unique (across

the entire period) billed account ids from 2013 to 2017. The two primary drawbacks of using account

participation are that it can fail to fully represent participation in master-metered multifamily buildings,

and it underrepresents participation in time-series analyses. Master-metered multifamily buildings count

as a single account in the metric, even if multiple residences were treated by the programs. Account ids

change every time someone moves in or out of a premise. This means the denominator of account

participation rate will tend to increase faster than the numerator because new residents have a less than

100% probability to participate in a program.

Data Quality Checks

DNV GL verified that the Community Outreach Metric correlates positively with the proportions of each block

group that matches the term sheet variables. The Community Outreach Metric also correlates negatively

with the historic account participation rate.

We examined the 5 highest and 5 lowest scoring towns in greater depth. We verified that these towns match

areas that the bivariate maps (Appendix C) identify as areas of low participation and high term sheet

characteristics. Furthermore, we examined why each of the 5 highest and 5 lowest scoring towns scored

how it did. For the towns with the highest scores, the deciding factors were low participation (small

denominator) and high renter population (large numerator). For the towns with the lowest scores, the

deciding factors were moderate participation rate (moderately large denominator) and low renter

population, low moderate-income population, and low limited-English-speaking population (extremely small

numerator). Table 7-1 provides more detail.

DNV GL – www.dnvgl.com February 6, 2020 Page 88

Table 7-1. Highest and lowest scoring town analysis

High metric

towns

Numerator Denominator

Town Renter

%

Moderate

income %

Limited

English %

Dual Fuel Participation

rate

LAWRENCE 70% 15% 25% 6%

FALL RIVER 65% 17% 9% 6%

FITCHBURG 48% 14% 5% 6%

NEW BEDFORD 58% 16% 13% 8%

CHELSEA 74% 17% 27% 11%

Low metric

towns Numerator Denominator

Town Renter

%

Moderate

income %

Limited

English %

Dual Fuel Participation

rate

BOLTON 4% 1% 0% 37%

CARLISLE 4% 3% 1% 36%

HINGHAM 3% 3% 0% 22%

NORFOLK 5% 6% 0% 39%

BOXFORD 2% 8% 1% 35%

DNV GL provided the community outreach metric deliverable to the PAs in a dynamic Excel file. The

following tables provide the top-level content included in that deliverable. The Excel file provided additional

functionality to sort, slice, and drill down into the block group level under each town. This functionality is not

possible to reproduce in a static document.

DNV GL – www.dnvgl.com February 6, 2020 Page 89

Table 7-2. Community outreach metric table: Dual-fuel

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

ABINGTON 21% 32% 17% 1% 21% 39% 41% 37% 375 2031 2.39

ACTON 23% 23% 7% 3% 27% 46% 47% 46% 1074 3484 1.24

ACUSHNET 3% 12% 13% 2% 22% 25% 31% 21% 252 1829 1.22

ADAMS 16% 33% 21% 1% 19% 24% 33% 20% 287 1710 2.81

AGAWAM 22% 26% 17% 2% 18% 30% 35% 28% 659 5962 2.45

AMESBURY 23% 31% 13% 0% 17% 33% 35% 32% 620 3139 2.58

AMHERST 35% 55% 11% 3% 13% 36% 38% 33% 207 1245 5.42

ANDOVER 18% 20% 9% 4% 21% 34% 41% 30% 771 6013 1.54

ARLINGTON 22% 39% 9% 3% 22% 38% 42% 36% 2259 9704 2.37

ASHBY 0% 12% 9% 0% 18% 22% 21% 25% 18 123 1.12

ASHLAND 8% 20% 10% 3% 28% 48% 50% 47% 816 3146 1.18

ATTLEBORO 14% 35% 14% 3% 15% 25% 30% 20% 432 5584 3.48

AUBURN 7% 19% 14% 1% 26% 38% 44% 30% 143 805 1.25

AVON 5% 26% 19% 2% 18% 21% 31% 15% 39 691 2.61

AYER 18% 35% 18% 2% 19% 33% 39% 29% 207 1205 2.94

BARNSTABLE 8% 25% 16% 3% 20% 25% 34% 21% 1921 17133 2.26

BEDFORD 20% 29% 9% 4% 34% 46% 46% 46% 835 2947 1.22

BELLINGHAM 13% 21% 10% 3% 22% 29% 35% 25% 179 1520 1.59

BERLIN 1% 16% 8% 0% 25% 41% 33% 62% 9 77 0.98

BEVERLY 26% 42% 12% 1% 19% 32% 34% 31% 1336 6222 2.88

BILLERICA 14% 18% 11% 2% 24% 34% 32% 36% 1842 8648 1.30

BOLTON 0% 4% 1% 0% 37% 44% 45% 32% 6 29 0.15

BOSTON 40% 63% 13% 12% 12% 29% 38% 26% 12700 77783 7.43

DNV GL – www.dnvgl.com February 6, 2020 Page 90

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

BOURNE 6% 23% 12% 1% 18% 25% 32% 21% 503 5248 2.01

BOXFORD 2% 2% 8% 1% 35% 42% 40% 43% 372 1209 0.29

BOYLSTON 0% 13% 7% 0% 12% 55% 100% 41% 0 0 1.71

BREWSTER 7% 16% 16% 0% 23% 31% 40% 25% 274 2584 1.37

BRIDGEWATER 14% 25% 11% 2% 18% 26% 35% 19% 218 2819 2.07

BROCKTON 20% 45% 16% 11% 11% 18% 26% 14% 655 13700 6.74

BROOKFIELD 4% 44% 30% 0% 18% 23% 31% 17% 13 94 4.15

BROOKLINE 52% 51% 8% 7% 10% 31% 32% 31% 1294 7239 6.70

BURLINGTON 29% 30% 12% 5% 27% 44% 48% 41% 880 3475 1.74

CAMBRIDGE 53% 62% 10% 6% 8% 30% 39% 27% 1975 14138 9.63

CANTON 23% 22% 11% 4% 21% 34% 42% 30% 568 4162 1.74

CARLISLE 2% 4% 3% 1% 36% 41% 40% 41% 207 709 0.23

CARVER 1% 10% 11% 0% 25% 36% 40% 33% 207 1229 0.85

CHATHAM 3% 20% 12% 0% 16% 19% 24% 17% 304 4677 2.06

CHELMSFORD 18% 17% 11% 2% 30% 44% 40% 46% 2704 8701 0.99

CHELSEA 40% 74% 17% 27% 11% 22% 36% 17% 288 3614 10.34

CHESHIRE 1% 22% 25% 0% 20% 32% 31% 33% 38 320 2.34

CHICOPEE 0% 14% 6% 6% 9% 16% 62% 15% 0 0 2.82

CLARKSBURG 2% 15% 23% 0% 22% 22% 22% 22% 10 186 1.70

CLINTON 21% 46% 18% 3% 16% 32% 34% 31% 274 1696 4.30

COHASSET 8% 20% 9% 0% 27% 37% 42% 34% 350 1478 1.09

CONCORD 1% 6% 4% 0% 31% 36% 67% 36% 5 12 0.30

DALTON 5% 24% 18% 1% 22% 30% 37% 27% 226 1380 1.97

DARTMOUTH 11% 24% 15% 3% 21% 26% 32% 24% 824 6129 2.02

DNV GL – www.dnvgl.com February 6, 2020 Page 91

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

DEDHAM 17% 32% 12% 2% 23% 32% 37% 30% 1268 6520 2.07

DEERFIELD 3% 26% 16% 0% 24% 30% 33% 26% 42 349 1.74

DENNIS 8% 21% 18% 2% 18% 24% 33% 20% 867 9092 2.29

DIGHTON 6% 12% 13% 1% 17% 27% 41% 15% 25 373 1.49

DOVER 0% 4% 7% 1% 33% 44% 43% 45% 6 26 0.34

DRACUT 20% 23% 13% 1% 19% 36% 38% 36% 1297 6515 1.95

DUDLEY 9% 24% 17% 2% 20% 28% 34% 14% 57 531 2.16

DUNSTABLE 0% 3% 10% 1% 29% 35% 37% 32% 71 310 0.45

DUXBURY 6% 11% 7% 0% 25% 31% 39% 27% 346 2886 0.72

EAST BRIDGEWATER 7% 17% 15% 1% 21% 19% 16% 22% 168 1885 1.57

EAST BROOKFIELD 2% 16% 14% 0% 21% 13% 16% 9% 16 110 1.39

EAST LONGMEADOW 9% 16% 11% 2% 25% 22% 12% 27% 585 4189 1.16

EASTHAM 0% 15% 15% 0% 20% 24% 29% 21% 180 2173 1.51

EASTHAMPTON 20% 44% 16% 1% 17% 31% 41% 23% 207 2282 3.53

EASTON 11% 18% 11% 1% 22% 31% 39% 25% 299 2942 1.41

ESSEX 9% 22% 15% 1% 20% 32% 31% 33% 87 487 1.88

EVERETT 16% 61% 20% 16% 9% 18% 27% 16% 715 8276 10.33

FAIRHAVEN 11% 26% 13% 2% 18% 26% 29% 25% 659 4896 2.33

FALL RIVER 30% 65% 17% 9% 6% 13% 20% 11% 878 16484 15.07

FALMOUTH 6% 22% 13% 1% 18% 24% 29% 22% 1161 11988 1.98

FITCHBURG 17% 48% 14% 5% 6% 19% 27% 16% 342 6877 11.96

FOXBOROUGH 23% 34% 9% 1% 20% 30% 33% 29% 401 3243 2.16

FRAMINGHAM 27% 43% 13% 12% 21% 36% 41% 34% 2101 10188 3.15

FRANKLIN 16% 19% 10% 1% 22% 34% 41% 31% 884 5971 1.35

DNV GL – www.dnvgl.com February 6, 2020 Page 92

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

FREETOWN 0% 15% 12% 1% 26% 32% 33% 28% 27 233 1.04

GARDNER 28% 53% 18% 2% 13% 29% 33% 22% 40 660 5.63

GEORGETOWN 13% 20% 25% 0% 22% 22% 70% 21% 3 1 2.10

GLOUCESTER 12% 37% 14% 3% 19% 31% 35% 29% 844 4679 2.88

GRAFTON 12% 28% 13% 1% 23% 38% 34% 41% 552 2604 1.81

GRANBY 11% 19% 5% 0% 27% 29% 32% 23% 17 204 0.87

GREAT BARRINGTON 18% 35% 22% 5% 17% 20% 11% 28% 58 560 3.51

GREENFIELD 20% 46% 16% 2% 18% 33% 42% 26% 324 2405 3.54

GROTON 0% 5% 15% 0% 24% 33% 8% 33% 0 0 0.80

HADLEY 13% 27% 23% 2% 24% 40% 48% 32% 52 355 2.18

HALIFAX 3% 10% 15% 2% 27% 32% 36% 27% 42 633 0.99

HAMILTON 10% 17% 9% 3% 26% 24% 15% 35% 198 853 1.11

HAMPDEN 2% 8% 14% 1% 30% 33% 37% 29% 179 749 0.76

HANOVER 7% 14% 11% 2% 24% 31% 36% 27% 263 2343 1.11

HANSON 2% 9% 8% 1% 24% 29% 34% 25% 169 1783 0.75

HARVARD 3% 7% 5% 0% 35% 46% 45% 47% 48 179 0.35

HARWICH 7% 15% 15% 1% 20% 24% 33% 21% 576 6071 1.53

HATFIELD 8% 25% 19% 0% 22% 36% 42% 32% 71 507 2.03

HAVERHILL 23% 41% 15% 4% 14% 27% 28% 26% 1915 12228 4.10

HINGHAM 0% 3% 3% 0% 22% 33% 21% 33% 1 26 0.27

HOLBROOK 11% 20% 15% 2% 18% 26% 36% 19% 125 1640 2.04

HOLLISTON 10% 15% 9% 1% 35% 41% 46% 38% 694 2913 0.74

HOPEDALE 5% 21% 10% 3% 28% 39% 44% 30% 46 321 1.23

HOPKINTON 5% 15% 7% 2% 29% 47% 50% 45% 696 2407 0.83

DNV GL – www.dnvgl.com February 6, 2020 Page 93

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

HUDSON 0% 0% 13% 3% 29% 33% 100% 33% 3 0 0.58

IPSWICH 23% 30% 16% 2% 17% 33% 20% 33% 1 0 2.71

KINGSTON 11% 23% 10% 0% 26% 41% 44% 38% 268 1343 1.30

LANCASTER 3% 19% 15% 3% 26% 36% 36% 39% 23 160 1.37

LANESBOROUGH 12% 15% 17% 0% 14% 19% 17% 21% 4 164 2.17

LAWRENCE 25% 70% 15% 25% 6% 14% 26% 11% 414 10031 18.57

LEE 17% 30% 16% 3% 17% 27% 29% 26% 127 1225 2.79

LEICESTER 10% 32% 17% 4% 26% 38% 39% 27% 15 99 2.02

LENOX 30% 40% 19% 2% 18% 33% 32% 34% 145 1050 3.38

LEOMINSTER 26% 45% 14% 6% 19% 35% 43% 27% 710 4178 3.40

LEXINGTON 13% 19% 7% 4% 34% 45% 47% 43% 1268 4783 0.88

LINCOLN 13% 19% 7% 2% 36% 41% 44% 39% 187 734 0.76

LONGMEADOW 6% 9% 9% 4% 28% 33% 38% 31% 832 4863 0.79

LOWELL 36% 58% 16% 14% 11% 23% 29% 21% 2410 16682 8.12

LUDLOW 10% 25% 15% 7% 18% 30% 36% 25% 362 3452 2.59

LUNENBURG 7% 22% 15% 1% 18% 33% 34% 32% 41 465 2.07

LYNN 28% 56% 14% 15% 12% 22% 28% 19% 1855 14888 7.39

MALDEN 33% 59% 13% 16% 14% 28% 32% 26% 1453 9147 6.23

MANCHESTER 12% 31% 10% 1% 22% 31% 31% 32% 245 1196 1.90

MARION 6% 23% 17% 0% 21% 26% 31% 24% 107 770 1.86

MARLBOROUGH 30% 45% 14% 11% 17% 34% 37% 32% 1185 6727 4.20

MARSHFIELD 12% 22% 11% 1% 19% 27% 34% 25% 605 6800 1.72

MASHPEE 8% 13% 14% 1% 19% 29% 39% 25% 682 6436 1.44

MATTAPOISETT 1% 19% 18% 0% 22% 30% 37% 26% 183 1181 1.67

DNV GL – www.dnvgl.com February 6, 2020 Page 94

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

MAYNARD 11% 28% 10% 0% 22% 37% 39% 36% 421 2177 1.79

MEDFIELD 9% 12% 6% 0% 30% 36% 46% 31% 405 2475 0.60

MEDFORD 21% 44% 13% 7% 16% 31% 38% 29% 2147 13563 3.96

MEDWAY 9% 15% 6% 1% 26% 32% 42% 26% 265 2143 0.80

MELROSE 26% 33% 14% 3% 24% 38% 43% 37% 1420 5965 2.11

MENDON 0% 11% 3% 0% 35% 41% 42% 35% 13 58 0.42

METHUEN 18% 29% 14% 6% 13% 26% 30% 24% 711 9715 3.70

MILFORD 16% 32% 16% 5% 20% 32% 37% 29% 525 3521 2.66

MILLBURY 12% 27% 12% 2% 25% 40% 46% 32% 221 1177 1.59

MILLIS 9% 19% 13% 0% 25% 35% 41% 29% 144 983 1.28

MILTON 10% 17% 8% 1% 30% 37% 41% 36% 1453 6376 0.89

MONSON 5% 24% 22% 0% 27% 33% 33% 38% 106 61 1.71

MONTAGUE 16% 48% 17% 2% 17% 28% 36% 19% 56 617 3.99

NAHANT 16% 32% 15% 1% 20% 28% 26% 29% 205 1216 2.33

NATICK 24% 27% 13% 3% 25% 43% 47% 40% 1371 6049 1.71

NEEDHAM 12% 17% 8% 3% 29% 40% 41% 39% 1174 5380 0.94

NEW BEDFORD 20% 58% 16% 13% 8% 16% 24% 15% 1893 23265 10.77

NEW MARLBOROUGH 1% 12% 15% 0% 14% 19% 18% 100% 1 1 1.91

NEWBURY 5% 15% 10% 1% 23% 27% 26% 29% 97 497 1.12

NEWBURYPORT 18% 26% 12% 0% 21% 40% 43% 39% 1151 4417 1.83

NEWTON 15% 29% 8% 5% 24% 38% 41% 37% 4804 21917 1.74

NORFOLK 2% 5% 6% 0% 39% 44% 48% 32% 63 380 0.28

NORTH ADAMS 18% 46% 15% 2% 17% 27% 35% 24% 393 2372 3.74

NORTH ANDOVER 23% 27% 9% 1% 21% 33% 40% 29% 549 4731 1.73

DNV GL – www.dnvgl.com February 6, 2020 Page 95

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

NORTH BROOKFIELD 8% 30% 18% 0% 19% 17% 20% 14% 58 631 2.58

NORTHAMPTON 25% 44% 13% 2% 14% 30% 42% 24% 572 5617 4.12

NORTHBOROUGH 7% 17% 10% 1% 28% 41% 41% 42% 402 1710 1.01

NORTHBRIDGE 17% 41% 15% 1% 16% 28% 28% 28% 167 1266 3.46

NORTON 12% 18% 11% 1% 19% 34% 34% 34% 270 2825 1.59

NORWELL 3% 7% 10% 0% 29% 33% 42% 27% 153 1295 0.61

ORLEANS 8% 23% 14% 1% 21% 29% 35% 23% 131 1331 1.79

OXFORD 4% 29% 16% 0% 27% 38% 40% 22% 6 37 1.67

PALMER 15% 34% 17% 0% 19% 30% 26% 52% 114 82 2.62

PEMBROKE 9% 13% 12% 1% 22% 31% 34% 28% 296 2985 1.16

PEPPERELL 7% 22% 10% 0% 24% 35% 36% 34% 270 1351 1.35

PITTSFIELD 14% 40% 16% 2% 16% 24% 33% 21% 1025 10536 3.69

PLAINVILLE 15% 26% 11% 0% 14% 31% 31% 30% 54 615 2.74

PLYMOUTH 12% 24% 15% 2% 22% 35% 39% 31% 1222 7086 1.90

PLYMPTON 4% 13% 11% 0% 30% 34% 40% 21% 11 131 0.79

QUINCY 35% 53% 14% 12% 15% 31% 38% 28% 2478 18357 5.44

RANDOLPH 21% 31% 16% 8% 17% 25% 35% 21% 500 5508 3.25

READING 25% 30% 8% 0% 22% 39% 82% 39% 2 2 1.77

REHOBOTH 0% 10% 13% 0% 27% 33% 33% 21% 6 59 0.89

REVERE 25% 52% 18% 13% 13% 23% 33% 19% 958 8693 6.62

ROCHESTER 0% 12% 15% 1% 32% 35% 35% 35% 66 258 0.87

ROCKLAND 19% 28% 15% 1% 21% 37% 41% 35% 551 3072 2.10

ROCKPORT 3% 19% 4% 0% 19% 29% 32% 2% 2 113 1.23

SALEM 29% 52% 14% 5% 15% 36% 44% 33% 1256 6675 4.64

DNV GL – www.dnvgl.com February 6, 2020 Page 96

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

SALISBURY 8% 25% 15% 0% 15% 28% 29% 28% 363 1939 2.59

SANDWICH 3% 15% 14% 0% 23% 28% 36% 23% 525 4609 1.23

SAUGUS 14% 21% 12% 2% 20% 30% 36% 27% 925 5373 1.80

SCITUATE 5% 14% 7% 0% 21% 29% 38% 26% 421 4423 1.03

SEEKONK 1% 13% 13% 1% 22% 26% 31% 22% 171 2206 1.28

SHARON 8% 14% 8% 4% 29% 38% 43% 36% 720 4067 0.87

SHERBORN 4% 7% 5% 1% 36% 43% 45% 41% 100 401 0.36

SHIRLEY 10% 37% 16% 0% 20% 43% 49% 36% 98 396 2.57

SHREWSBURY 18% 14% 8% 5% 11% 36% 52% 36% 4 0 2.44

SOMERSET 3% 19% 13% 3% 15% 21% 25% 19% 538 5557 2.37

SOMERVILLE 25% 65% 14% 7% 9% 25% 29% 24% 2029 15808 9.63

SOUTHBOROUGH 1% 11% 7% 1% 31% 40% 43% 37% 241 1012 0.61

SOUTHBRIDGE 17% 57% 18% 10% 12% 22% 31% 15% 164 1766 6.96

SOUTHWICK 10% 22% 10% 0% 21% 29% 30% 28% 49 471 1.55

SPENCER 14% 38% 13% 2% 16% 29% 37% 21% 110 743 3.24

SPRINGFIELD 19% 52% 16% 11% 11% 18% 29% 13% 2196 27996 7.23

STOCKBRIDGE 11% 32% 18% 0% 16% 27% 22% 32% 29 259 3.16

STONEHAM 30% 36% 14% 3% 23% 38% 45% 33% 633 3074 2.30

STOUGHTON 18% 28% 13% 5% 16% 30% 34% 28% 498 5108 2.85

STOW 14% 16% 5% 0% 25% 39% 54% 38% 0 0 0.82

SUDBURY 4% 9% 6% 1% 40% 51% 49% 51% 1178 3396 0.39

SUNDERLAND 27% 51% 14% 0% 41% 55% 55% 62% 0 2 1.57

SUTTON 5% 12% 12% 0% 31% 53% 45% 78% 17 56 0.76

SWAMPSCOTT 14% 23% 8% 3% 26% 39% 41% 38% 773 3355 1.30

DNV GL – www.dnvgl.com February 6, 2020 Page 97

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

SWANSEA 2% 13% 14% 3% 17% 22% 27% 20% 429 4398 1.72

TEWKSBURY 12% 14% 10% 1% 22% 43% 41% 44% 1177 5486 1.13

TOPSFIELD 2% 8% 10% 0% 31% 38% 44% 35% 217 879 0.60

TOWNSEND 9% 20% 13% 0% 14% 19% 23% 17% 137 1553 2.35

TYNGSBOROUGH 10% 14% 13% 2% 25% 37% 35% 39% 623 2538 1.12

UPTON 7% 23% 7% 1% 26% 37% 39% 35% 126 524 1.16

UXBRIDGE 6% 22% 12% 1% 23% 40% 40% 40% 182 1005 1.54

WAKEFIELD 0% 8% 10% 3% 11% 25% 17% 52% 0 2 1.81

WALPOLE 11% 18% 11% 1% 23% 32% 40% 28% 569 4075 1.29

WALTHAM 30% 50% 13% 6% 14% 33% 35% 32% 1612 8724 5.09

WAREHAM 6% 23% 15% 1% 17% 31% 35% 29% 744 5370 2.29

WARREN 12% 37% 14% 0% 14% 19% 22% 16% 36 291 3.54

WATERTOWN 24% 49% 12% 4% 15% 35% 41% 33% 1323 7533 4.29

WAYLAND 6% 12% 4% 2% 34% 43% 44% 42% 787 2797 0.53

WEBSTER 17% 51% 19% 4% 13% 24% 31% 16% 118 1227 5.61

WENHAM 10% 12% 13% 0% 28% 38% 37% 39% 125 493 0.91

WEST BRIDGEWATER 3% 14% 16% 0% 21% 22% 19% 24% 79 1051 1.40

WEST BROOKFIELD 13% 29% 15% 0% 20% 18% 20% 16% 52 277 2.17

WEST NEWBURY 2% 4% 8% 0% 33% 42% 40% 47% 62 167 0.36

WEST SPRINGFIELD 23% 40% 18% 5% 15% 24% 33% 21% 688 6265 4.14

WESTBOROUGH 33% 39% 8% 3% 21% 44% 49% 42% 566 2446 2.39

WESTFORD 5% 11% 9% 2% 29% 43% 41% 44% 1645 5951 0.74

WESTHAMPTON 0% 9% 13% 0% 29% 32% 33% 22% 1 4 0.76

WESTMINSTER 8% 15% 5% 0% 22% 34% 34% 34% 48 244 0.91

DNV GL – www.dnvgl.com February 6, 2020 Page 98

Town

Ratio of

MF/Total

Units

Ratio

Renter

Occupied

Units

(ACS)

Ratio of

Households

at 56 - 85%

of

Statewide

Median

Income

(ACS)

Ratio

Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas &

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Mmbtu)

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Gas

Combined

2013-2017

Consumption

Weighted

Participation

Rate

(Therms)

Dual Fuel

Participants

Combined

2013-2017

Dual Fuel

Locations

Combined

2013-

2017

Community

Outreach

Metric

WESTON 7% 15% 5% 3% 30% 34% 32% 35% 680 2860 0.78

WESTPORT 4% 16% 10% 2% 16% 20% 24% 18% 270 3113 1.79

WESTWOOD 17% 13% 9% 2% 28% 41% 44% 39% 474 2065 0.85

WEYMOUTH 29% 36% 15% 2% 21% 31% 36% 28% 1571 9271 2.57

WHATELY 0% 15% 19% 2% 26% 30% 31% 25% 7 43 1.37

WHITMAN 12% 28% 14% 0% 21% 36% 36% 37% 454 2255 1.99

WILBRAHAM 4% 11% 8% 1% 29% 35% 41% 31% 544 2408 0.68

WILLIAMSTOWN 11% 27% 16% 2% 19% 28% 30% 27% 145 1083 2.37

WINCHESTER 11% 15% 8% 3% 31% 42% 43% 41% 1138 4426 0.81

WINTHROP 21% 43% 15% 4% 15% 32% 38% 30% 602 3989 4.15

WOBURN 24% 38% 14% 4% 23% 36% 41% 33% 953 4894 2.44

WORCESTER 26% 57% 16% 11% 12% 24% 33% 21% 3301 26158 7.24

WRENTHAM 7% 20% 10% 0% 23% 28% 37% 22% 116 1195 1.28

YARMOUTH 11% 21% 21% 3% 20% 26% 36% 23% 1302 10874 2.26

Grand Total 22% 40% 13% 6% 16% 30% 36% 27% 136746 924392 3.67

DNV GL – www.dnvgl.com February 6, 2020 Page 99

Table 7-3. Community outreach metric table: Gas

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

ABINGTON 21% 32% 17% 1% 21% 37% 2.39

ACTON 23% 23% 7% 3% 27% 46% 1.24

ACUSHNET 3% 12% 13% 2% 22% 21% 1.22

ADAMS 16% 33% 21% 1% 19% 20% 2.81

AGAWAM 22% 26% 17% 2% 18% 28% 2.45

AMESBURY 23% 31% 13% 0% 17% 32% 2.58

AMHERST 35% 55% 11% 3% 13% 33% 5.42

ANDOVER 18% 20% 9% 4% 21% 30% 1.54

ARLINGTON 22% 39% 9% 3% 22% 36% 2.37

ASHBY 0% 12% 9% 0% 18% 25% 1.12

ASHLAND 8% 20% 10% 3% 28% 47% 1.18

ATTLEBORO 14% 35% 14% 3% 15% 20% 3.48

AUBURN 7% 19% 14% 1% 26% 30% 1.25

AVON 5% 26% 19% 2% 18% 15% 2.61

AYER 18% 35% 18% 2% 19% 29% 2.94

BARNSTABLE 8% 25% 16% 3% 20% 21% 2.26

BEDFORD 20% 29% 9% 4% 34% 46% 1.22

BELLINGHAM 13% 21% 10% 3% 22% 25% 1.59

BELMONT 7% 37% 11% 4% 20% 33% 2.62

BERKLEY 0% 0% 13% 0% 9% 16% 1.37

BERLIN 1% 16% 8% 0% 25% 62% 0.98

BEVERLY 26% 42% 12% 1% 19% 31% 2.88

BILLERICA 14% 18% 11% 2% 24% 36% 1.30

BOLTON 0% 4% 1% 0% 37% 32% 0.15

BOSTON 40% 63% 13% 12% 12% 26% 7.44

DNV GL – www.dnvgl.com February 6, 2020 Page 100

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

BOURNE 6% 23% 12% 1% 18% 21% 2.01

BOXBOROUGH 38% 27% 13% 1% 34% 41% 1.19

BOXFORD 2% 2% 8% 1% 35% 43% 0.29

BOYLSTON 3% 14% 14% 1% 20% 33% 1.48

BRAINTREE 21% 30% 11% 3% 32% 37% 1.36

BREWSTER 7% 16% 16% 0% 23% 25% 1.37

BRIDGEWATER 14% 25% 11% 2% 18% 19% 2.07

BROCKTON 20% 45% 16% 11% 11% 14% 6.74

BROOKFIELD 4% 44% 30% 0% 18% 17% 4.15

BROOKLINE 52% 51% 8% 7% 10% 31% 6.70

BURLINGTON 29% 30% 12% 5% 27% 41% 1.74

CAMBRIDGE 53% 62% 10% 6% 8% 27% 9.63

CANTON 23% 22% 11% 4% 21% 30% 1.74

CARLISLE 2% 4% 3% 1% 36% 41% 0.23

CARVER 1% 10% 11% 0% 25% 33% 0.85

CHATHAM 3% 20% 12% 0% 16% 17% 2.06

CHELMSFORD 18% 17% 11% 2% 30% 46% 0.99

CHELSEA 40% 74% 17% 27% 11% 17% 10.34

CHESHIRE 1% 22% 25% 0% 20% 33% 2.34

CHICOPEE 19% 42% 18% 5% 8% 18% 7.87

CLARKSBURG 2% 15% 23% 0% 22% 22% 1.70

CLINTON 21% 46% 18% 3% 16% 31% 4.30

COHASSET 8% 20% 9% 0% 27% 34% 1.09

CONCORD 18% 23% 6% 1% 24% 37% 1.24

DALTON 5% 24% 18% 1% 22% 27% 1.97

DANVERS 22% 31% 12% 1% 15% 32% 2.97

DNV GL – www.dnvgl.com February 6, 2020 Page 101

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

DARTMOUTH 11% 24% 15% 3% 21% 24% 2.02

DEDHAM 17% 32% 12% 2% 23% 30% 2.07

DEERFIELD 3% 26% 16% 0% 24% 26% 1.74

DENNIS 8% 21% 18% 2% 18% 20% 2.29

DIGHTON 6% 12% 13% 1% 17% 15% 1.49

DOVER 0% 4% 7% 1% 33% 45% 0.34

DRACUT 20% 23% 13% 1% 19% 36% 1.95

DUDLEY 9% 24% 17% 2% 20% 14% 2.16

DUNSTABLE 0% 3% 10% 1% 29% 32% 0.45

DUXBURY 6% 11% 7% 0% 25% 27% 0.72

EAST BRIDGEWATER 7% 17% 15% 1% 21% 22% 1.57

EAST BROOKFIELD 2% 16% 14% 0% 21% 9% 1.39

EAST LONGMEADOW 9% 16% 11% 2% 25% 27% 1.16

EASTHAM 0% 15% 15% 0% 20% 21% 1.51

EASTHAMPTON 20% 44% 16% 1% 17% 23% 3.53

EASTON 11% 18% 11% 1% 22% 25% 1.41

ESSEX 9% 22% 15% 1% 20% 33% 1.88

EVERETT 16% 61% 20% 16% 9% 16% 10.33

FAIRHAVEN 11% 26% 13% 2% 18% 25% 2.33

FALL RIVER 30% 65% 17% 9% 6% 11% 15.07

FALMOUTH 6% 22% 13% 1% 18% 22% 1.98

FITCHBURG 17% 48% 14% 5% 6% 16% 11.96

FOXBOROUGH 23% 34% 9% 1% 20% 29% 2.16

FRAMINGHAM 27% 43% 13% 12% 21% 34% 3.15

FRANKLIN 16% 19% 10% 1% 22% 31% 1.35

FREETOWN 0% 15% 12% 1% 26% 28% 1.04

DNV GL – www.dnvgl.com February 6, 2020 Page 102

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

GARDNER 28% 53% 18% 2% 13% 22% 5.63

GEORGETOWN 9% 16% 12% 0% 27% 32% 1.07

GLOUCESTER 12% 37% 14% 3% 19% 29% 2.88

GRAFTON 12% 28% 13% 1% 23% 41% 1.81

GRANBY 11% 19% 5% 0% 27% 23% 0.87

GREAT BARRINGTON 18% 35% 22% 5% 17% 28% 3.51

GREENFIELD 20% 46% 16% 2% 18% 26% 3.54

GROTON 5% 14% 7% 1% 22% 31% 0.96

GROVELAND 6% 18% 13% 1% 22% 27% 1.50

HADLEY 13% 27% 23% 2% 24% 32% 2.18

HALIFAX 3% 10% 15% 2% 27% 27% 0.99

HAMILTON 10% 17% 9% 3% 26% 35% 1.11

HAMPDEN 2% 8% 14% 1% 30% 29% 0.76

HANOVER 7% 14% 11% 2% 24% 27% 1.11

HANSON 2% 9% 8% 1% 24% 25% 0.75

HARVARD 3% 7% 5% 0% 35% 47% 0.35

HARWICH 7% 15% 15% 1% 20% 21% 1.53

HATFIELD 8% 25% 19% 0% 22% 32% 2.03

HAVERHILL 23% 40% 15% 3% 14% 27% 4.09

HINGHAM 20% 17% 8% 0% 18% 34% 1.40

HOLBROOK 11% 20% 15% 2% 18% 19% 2.04

HOLDEN 7% 13% 11% 2% 20% 28% 1.24

HOLLISTON 10% 15% 9% 1% 35% 38% 0.74

HOPEDALE 5% 21% 10% 3% 28% 30% 1.23

HOPKINTON 5% 15% 7% 2% 29% 45% 0.83

HUDSON 18% 25% 14% 5% 15% 28% 3.01

DNV GL – www.dnvgl.com February 6, 2020 Page 103

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

HULL 19% 32% 12% 1% 16% 29% 2.76

IPSWICH 14% 27% 11% 2% 18% 31% 2.17

KINGSTON 11% 23% 10% 0% 26% 38% 1.30

LAKEVILLE 13% 23% 19% 0% 3% 10% 15.56

LANCASTER 3% 19% 15% 3% 26% 39% 1.37

LANESBOROUGH 12% 15% 17% 0% 14% 21% 2.17

LAWRENCE 25% 70% 15% 25% 6% 11% 18.57

LEE 17% 30% 16% 3% 17% 26% 2.79

LEICESTER 10% 32% 17% 4% 26% 27% 2.02

LENOX 30% 40% 19% 2% 18% 34% 3.38

LEOMINSTER 26% 45% 14% 6% 19% 27% 3.40

LEXINGTON 13% 19% 7% 4% 34% 43% 0.88

LINCOLN 13% 19% 7% 2% 36% 39% 0.76

LITTLETON 7% 16% 6% 1% 22% 29% 1.06

LONGMEADOW 6% 9% 9% 4% 28% 31% 0.79

LOWELL 36% 58% 16% 14% 11% 21% 8.12

LUDLOW 10% 25% 15% 7% 18% 25% 2.59

LUNENBURG 7% 22% 15% 1% 18% 32% 2.07

LYNN 28% 56% 14% 15% 12% 19% 7.39

LYNNFIELD 13% 15% 10% 2% 19% 31% 1.43

MALDEN 33% 59% 13% 16% 14% 26% 6.23

MANCHESTER 12% 31% 10% 1% 22% 32% 1.90

MANSFIELD 22% 28% 11% 1% 15% 25% 2.76

MARBLEHEAD 6% 21% 11% 2% 21% 30% 1.60

MARION 6% 23% 17% 0% 21% 24% 1.86

MARLBOROUGH 30% 45% 14% 11% 17% 32% 4.20

DNV GL – www.dnvgl.com February 6, 2020 Page 104

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

MARSHFIELD 12% 22% 11% 1% 19% 25% 1.72

MASHPEE 8% 13% 14% 1% 19% 25% 1.44

MATTAPOISETT 1% 19% 18% 0% 22% 26% 1.67

MAYNARD 11% 28% 10% 0% 22% 36% 1.79

MEDFIELD 9% 12% 6% 0% 30% 31% 0.60

MEDFORD 21% 44% 13% 7% 16% 29% 3.96

MEDWAY 9% 15% 6% 1% 26% 26% 0.80

MELROSE 26% 33% 14% 3% 24% 37% 2.11

MENDON 0% 11% 3% 0% 35% 35% 0.42

MERRIMAC 5% 13% 10% 0% 17% 25% 1.33

METHUEN 18% 29% 14% 6% 13% 24% 3.70

MIDDLEBOROUGH 6% 11% 19% 2% 5% 6% 6.28

MIDDLETON 15% 15% 6% 0% 17% 38% 1.28

MILFORD 16% 32% 16% 5% 20% 29% 2.66

MILLBURY 12% 27% 12% 2% 25% 32% 1.59

MILLIS 9% 19% 13% 0% 25% 29% 1.28

MILTON 10% 17% 8% 1% 30% 36% 0.89

MONSON 5% 24% 22% 0% 27% 38% 1.71

MONTAGUE 16% 48% 17% 2% 17% 19% 3.99

NAHANT 16% 32% 15% 1% 20% 29% 2.33

NATICK 24% 27% 13% 3% 25% 40% 1.71

NEEDHAM 12% 17% 8% 3% 29% 39% 0.94

NEW BEDFORD 20% 58% 16% 13% 8% 15% 10.77

NEW MARLBOROUGH 1% 12% 15% 0% 14% 100% 1.91

NEWBURY 5% 15% 10% 1% 23% 29% 1.12

NEWBURYPORT 18% 26% 12% 0% 21% 39% 1.83

DNV GL – www.dnvgl.com February 6, 2020 Page 105

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

NEWTON 15% 29% 8% 5% 24% 37% 1.74

NORFOLK 2% 5% 6% 0% 39% 32% 0.28

NORTH ADAMS 18% 46% 15% 2% 17% 24% 3.74

NORTH ANDOVER 23% 27% 9% 1% 21% 29% 1.73

NORTH ATTLEBOROUGH 18% 30% 14% 1% 8% 18% 5.85

NORTH BROOKFIELD 8% 30% 18% 0% 19% 14% 2.58

NORTH READING 12% 13% 12% 1% 18% 32% 1.46

NORTHAMPTON 25% 44% 13% 2% 14% 24% 4.12

NORTHBOROUGH 7% 17% 10% 1% 28% 42% 1.01

NORTHBRIDGE 17% 41% 15% 1% 16% 28% 3.46

NORTON 12% 18% 11% 1% 19% 34% 1.59

NORWELL 3% 7% 10% 0% 29% 27% 0.61

NORWOOD 26% 42% 11% 4% 13% 29% 4.39

ORLEANS 8% 23% 14% 1% 21% 23% 1.79

OXFORD 4% 29% 16% 0% 27% 22% 1.67

PALMER 15% 34% 17% 0% 19% 52% 2.62

PEABODY 26% 35% 17% 4% 14% 26% 4.15

PEMBROKE 9% 13% 12% 1% 22% 28% 1.16

PEPPERELL 7% 22% 10% 0% 24% 34% 1.35

PITTSFIELD 14% 39% 16% 2% 16% 22% 3.65

PLAINVILLE 15% 26% 11% 0% 14% 30% 2.74

PLYMOUTH 12% 24% 15% 2% 22% 31% 1.90

PLYMPTON 4% 13% 11% 0% 30% 21% 0.79

QUINCY 35% 53% 14% 12% 15% 28% 5.44

RANDOLPH 21% 31% 16% 8% 17% 21% 3.25

RAYNHAM 19% 23% 12% 1% 13% 21% 2.73

DNV GL – www.dnvgl.com February 6, 2020 Page 106

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

READING 17% 21% 9% 1% 22% 38% 1.36

REHOBOTH 0% 10% 13% 0% 27% 21% 0.89

REVERE 25% 52% 18% 13% 13% 19% 6.62

ROCHESTER 0% 12% 15% 1% 32% 35% 0.87

ROCKLAND 19% 28% 15% 1% 21% 35% 2.10

ROCKPORT 3% 19% 4% 0% 19% 2% 1.23

ROWLEY 13% 17% 10% 0% 20% 32% 1.34

SALEM 29% 52% 14% 5% 15% 33% 4.64

SALISBURY 8% 25% 15% 0% 15% 28% 2.59

SANDWICH 3% 15% 14% 0% 23% 23% 1.23

SAUGUS 14% 21% 12% 2% 20% 27% 1.80

SCITUATE 5% 14% 7% 0% 21% 26% 1.03

SEEKONK 1% 13% 13% 1% 22% 22% 1.28

SHARON 8% 14% 8% 4% 29% 36% 0.87

SHERBORN 4% 7% 5% 1% 36% 41% 0.36

SHIRLEY 10% 37% 16% 0% 20% 36% 2.57

SHREWSBURY 21% 26% 11% 7% 18% 30% 2.48

SOMERSET 3% 19% 13% 3% 15% 19% 2.37

SOMERVILLE 25% 65% 14% 7% 9% 24% 9.63

SOUTH HADLEY 14% 26% 16% 1% 14% 28% 3.04

SOUTHBOROUGH 1% 11% 7% 1% 31% 37% 0.61

SOUTHBRIDGE 17% 57% 18% 10% 12% 15% 6.96

SOUTHWICK 10% 22% 10% 0% 21% 28% 1.55

SPENCER 14% 38% 13% 2% 16% 21% 3.24

SPRINGFIELD 19% 52% 16% 11% 11% 13% 7.23

STERLING 3% 15% 13% 0% 57% 59% 0.49

DNV GL – www.dnvgl.com February 6, 2020 Page 107

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

STOCKBRIDGE 11% 32% 18% 0% 16% 32% 3.16

STONEHAM 30% 36% 14% 3% 23% 33% 2.30

STOUGHTON 18% 28% 13% 5% 16% 28% 2.85

STOW 4% 10% 8% 0% 24% 30% 0.73

SUDBURY 4% 9% 6% 1% 40% 51% 0.39

SUNDERLAND 27% 51% 14% 0% 41% 62% 1.57

SUTTON 5% 12% 12% 0% 31% 78% 0.76

SWAMPSCOTT 14% 23% 8% 3% 26% 38% 1.30

SWANSEA 2% 13% 14% 3% 17% 20% 1.72

TAUNTON 16% 38% 17% 5% 5% 14% 11.26

TEWKSBURY 12% 14% 10% 1% 22% 44% 1.13

TOPSFIELD 2% 8% 10% 0% 31% 35% 0.60

TOWNSEND 9% 20% 13% 0% 14% 17% 2.35

TYNGSBOROUGH 10% 14% 13% 2% 25% 39% 1.12

UPTON 7% 23% 7% 1% 26% 35% 1.16

UXBRIDGE 6% 22% 12% 1% 23% 40% 1.54

WAKEFIELD 10% 15% 15% 2% 18% 10% 1.81

WALPOLE 11% 18% 11% 1% 23% 28% 1.29

WALTHAM 30% 50% 13% 6% 14% 32% 5.09

WAREHAM 6% 23% 15% 1% 17% 29% 2.29

WARREN 12% 37% 14% 0% 14% 16% 3.54

WATERTOWN 24% 49% 12% 4% 15% 33% 4.29

WAYLAND 6% 12% 4% 2% 34% 42% 0.53

WEBSTER 17% 51% 19% 4% 13% 16% 5.61

WELLESLEY 11% 17% 5% 2% 31% 33% 0.79

WENHAM 10% 12% 13% 0% 28% 39% 0.91

DNV GL – www.dnvgl.com February 6, 2020 Page 108

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Gas Combined

2013-2017

Consumption

Weighted

Participation

Rate (Therms)

Community

Outreach

Metric

WEST BOYLSTON 6% 16% 14% 2% 22% 31% 1.43

WEST BRIDGEWATER 3% 14% 16% 0% 21% 24% 1.40

WEST BROOKFIELD 13% 29% 15% 0% 20% 16% 2.17

WEST NEWBURY 2% 4% 8% 0% 33% 47% 0.36

WEST SPRINGFIELD 23% 40% 18% 5% 15% 21% 4.14

WESTBOROUGH 33% 39% 8% 3% 21% 42% 2.39

WESTFORD 5% 11% 9% 2% 29% 44% 0.74

WESTHAMPTON 0% 9% 13% 0% 29% 22% 0.76

WESTMINSTER 8% 15% 5% 0% 22% 34% 0.91

WESTON 7% 15% 5% 3% 30% 35% 0.78

WESTPORT 4% 17% 10% 2% 16% 18% 1.86

WESTWOOD 17% 13% 9% 2% 28% 39% 0.85

WEYMOUTH 29% 36% 15% 2% 21% 28% 2.57

WHATELY 0% 15% 19% 2% 26% 25% 1.37

WHITMAN 12% 28% 14% 0% 21% 37% 1.99

WILBRAHAM 4% 11% 8% 1% 29% 31% 0.68

WILLIAMSTOWN 11% 27% 16% 2% 19% 27% 2.37

WILMINGTON 9% 15% 10% 2% 17% 32% 1.56

WINCHESTER 11% 15% 8% 3% 31% 41% 0.81

WINTHROP 21% 43% 15% 4% 15% 30% 4.15

WOBURN 24% 38% 14% 4% 23% 33% 2.44

WORCESTER 26% 57% 16% 11% 12% 21% 7.24

WRENTHAM 7% 20% 10% 0% 23% 22% 1.28

YARMOUTH 11% 21% 21% 3% 20% 23% 2.26

Grand Total 21% 39% 13% 6% 16% 27% 3.56

DNV GL – www.dnvgl.com February 6, 2020 Page 109

Table 7-4. Community outreach metric table: Electric

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

ABINGTON 21% 32% 17% 1% 21% 41% 2.39

ACTON 23% 23% 7% 3% 27% 47% 1.24

ACUSHNET 3% 12% 13% 2% 22% 31% 1.22

ADAMS 16% 33% 21% 1% 19% 33% 2.81

AGAWAM 22% 26% 17% 2% 18% 35% 2.45

ALFORD 0% 10% 13% 0% 18% 18% 1.30

AMESBURY 23% 31% 13% 0% 17% 35% 2.58

AMHERST 36% 56% 11% 3% 13% 37% 5.53

ANDOVER 18% 20% 9% 4% 21% 41% 1.54

ARLINGTON 22% 39% 9% 3% 22% 42% 2.37

ASHBURNHAM 0% 5% 17% 0% 35% 27% 0.64

ASHBY 0% 12% 9% 0% 18% 21% 1.12

ASHFIELD 4% 18% 14% 1% 31% 34% 1.05

ASHLAND 8% 20% 10% 3% 28% 50% 1.18

ATHOL 10% 30% 19% 1% 21% 34% 2.42

ATTLEBORO 14% 35% 14% 3% 15% 30% 3.48

AUBURN 7% 18% 14% 1% 28% 46% 1.13

AVON 5% 26% 19% 2% 18% 31% 2.61

AYER 18% 35% 18% 2% 19% 39% 2.94

BARNSTABLE 8% 25% 16% 3% 20% 34% 2.26

BARRE 5% 25% 18% 0% 22% 29% 1.93

BECKET 0% 9% 17% 1% 17% 23% 1.60

BEDFORD 20% 29% 9% 4% 34% 46% 1.22

BELCHERTOWN 6% 21% 13% 2% 36% 42% 1.01

DNV GL – www.dnvgl.com February 6, 2020 Page 110

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

BELLINGHAM 9% 20% 11% 3% 23% 34% 1.45

BERLIN 1% 15% 10% 0% 25% 35% 0.96

BERNARDSTON 0% 15% 18% 0% 29% 37% 1.13

BEVERLY 25% 40% 12% 1% 19% 35% 2.73

BILLERICA 14% 18% 11% 2% 24% 32% 1.30

BLACKSTONE 9% 30% 17% 1% 19% 29% 2.48

BLANDFORD 0% 6% 19% 0% 17% 19% 1.48

BOLTON 2% 6% 4% 0% 38% 44% 0.27

BOSTON 43% 65% 13% 12% 12% 39% 7.77

BOURNE 6% 23% 12% 1% 18% 32% 2.01

BOXFORD 2% 2% 8% 1% 35% 40% 0.29

BOYLSTON 0% 13% 7% 0% 12% 100% 1.71

BREWSTER 7% 16% 16% 0% 23% 40% 1.37

BRIDGEWATER 14% 25% 11% 2% 18% 35% 2.07

BRIMFIELD 1% 19% 10% 1% 24% 30% 1.26

BROCKTON 22% 46% 16% 12% 10% 27% 6.97

BROOKFIELD 3% 20% 16% 0% 22% 30% 1.62

BROOKLINE 52% 51% 8% 7% 10% 32% 6.70

BUCKLAND 3% 23% 18% 0% 26% 34% 1.54

BURLINGTON 29% 30% 12% 5% 27% 48% 1.74

CAMBRIDGE 53% 63% 11% 6% 8% 40% 9.93

CANTON 23% 22% 11% 4% 21% 42% 1.74

CARLISLE 2% 4% 3% 1% 36% 40% 0.23

CARVER 1% 7% 14% 0% 25% 45% 0.85

CHARLTON 10% 19% 11% 1% 27% 45% 1.14

DNV GL – www.dnvgl.com February 6, 2020 Page 111

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

CHATHAM 3% 20% 12% 0% 16% 24% 2.06

CHELMSFORD 18% 17% 11% 2% 30% 40% 0.99

CHELSEA 40% 74% 17% 27% 11% 36% 10.34

CHESHIRE 1% 22% 25% 0% 20% 31% 2.34

CHESTER 1% 15% 12% 2% 12% 12% 2.41

CHESTERFIELD 2% 10% 18% 0% 30% 37% 0.93

CHICOPEE 0% 14% 6% 6% 9% 62% 2.82

CHILMARK 0% 20% 9% 1% 8% 12% 3.52

CLARKSBURG 3% 10% 21% 0% 24% 25% 1.30

CLINTON 21% 46% 18% 3% 16% 34% 4.30

COHASSET 8% 20% 9% 0% 27% 42% 1.09

COLRAIN 0% 21% 26% 0% 27% 31% 1.72

CONCORD 1% 6% 4% 0% 31% 67% 0.30

CONWAY 1% 11% 17% 0% 28% 34% 0.98

CUMMINGTON 2% 19% 18% 1% 17% 27% 2.26

DALTON 5% 24% 18% 1% 22% 37% 1.97

DARTMOUTH 10% 22% 14% 3% 21% 31% 1.85

DEDHAM 17% 32% 12% 2% 23% 37% 2.07

DEERFIELD 3% 26% 16% 0% 24% 33% 1.74

DENNIS 8% 21% 18% 2% 18% 33% 2.29

DIGHTON 5% 10% 13% 0% 18% 37% 1.29

DOUGLAS 10% 19% 12% 1% 24% 34% 1.28

DOVER 1% 6% 7% 0% 32% 41% 0.40

DRACUT 20% 23% 13% 1% 19% 38% 1.95

DUDLEY 9% 24% 17% 2% 20% 34% 2.16

DNV GL – www.dnvgl.com February 6, 2020 Page 112

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

DUNSTABLE 0% 3% 10% 1% 29% 37% 0.45

DUXBURY 6% 11% 7% 0% 25% 39% 0.72

EAST BRIDGEWATER 7% 17% 15% 1% 21% 16% 1.57

EAST BROOKFIELD 2% 14% 11% 0% 22% 15% 1.12

EAST LONGMEADOW 9% 16% 11% 2% 25% 12% 1.16

EASTHAM 0% 15% 15% 0% 20% 29% 1.51

EASTHAMPTON 20% 44% 16% 1% 17% 41% 3.53

EASTON 11% 18% 11% 1% 22% 39% 1.41

EDGARTOWN 0% 22% 16% 3% 10% 14% 4.04

ERVING 1% 16% 18% 0% 26% 36% 1.30

ESSEX 9% 22% 15% 1% 20% 31% 1.88

EVERETT 16% 61% 20% 16% 9% 27% 10.33

FAIRHAVEN 11% 26% 13% 2% 18% 29% 2.33

FALL RIVER 30% 65% 17% 9% 6% 20% 15.07

FALMOUTH 6% 22% 13% 1% 18% 29% 1.98

FITCHBURG 17% 48% 14% 5% 6% 27% 11.96

FLORIDA 2% 11% 18% 0% 20% 23% 1.46

FOXBOROUGH 23% 34% 9% 1% 20% 33% 2.16

FRAMINGHAM 31% 46% 13% 12% 20% 42% 3.47

FRANKLIN 16% 19% 10% 1% 22% 41% 1.35

FREETOWN 0% 12% 11% 1% 27% 32% 0.90

GARDNER 26% 49% 19% 2% 14% 32% 5.03

AQUINNAH 0% 40% 9% 0% 7% 11% 6.86

GEORGETOWN 13% 20% 25% 0% 22% 70% 2.10

GILL 3% 16% 15% 2% 27% 35% 1.21

DNV GL – www.dnvgl.com February 6, 2020 Page 113

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

GLOUCESTER 13% 36% 14% 2% 19% 35% 2.69

GOSHEN 1% 10% 18% 0% 25% 33% 1.12

GRAFTON 12% 28% 13% 1% 23% 34% 1.81

GRANBY 6% 15% 7% 1% 28% 38% 0.79

GRANVILLE 0% 12% 15% 0% 16% 20% 1.73

GREAT BARRINGTON 15% 33% 21% 4% 18% 14% 3.21

GREENFIELD 18% 45% 16% 2% 19% 42% 3.34

GROTON 0% 5% 15% 0% 24% 8% 0.80

HADLEY 11% 26% 21% 1% 24% 45% 2.03

HALIFAX 3% 10% 15% 2% 27% 36% 0.99

HAMILTON 10% 17% 9% 3% 26% 15% 1.11

HAMPDEN 2% 8% 14% 1% 30% 37% 0.76

HANCOCK 33% 19% 16% 0% 10% 13% 3.56

HANOVER 7% 14% 11% 2% 24% 36% 1.11

HANSON 2% 9% 8% 1% 24% 34% 0.75

HARDWICK 9% 31% 16% 0% 21% 24% 2.18

HARVARD 3% 8% 6% 0% 35% 45% 0.41

HARWICH 7% 15% 15% 1% 20% 33% 1.53

HATFIELD 8% 25% 19% 0% 22% 42% 2.03

HAVERHILL 23% 41% 15% 4% 14% 28% 4.10

HAWLEY 4% 18% 21% 0% 21% 22% 1.91

HEATH 0% 12% 19% 1% 21% 29% 1.54

HINGHAM 0% 3% 3% 0% 22% 21% 0.27

HINSDALE 3% 19% 21% 1% 19% 32% 2.22

HOLBROOK 11% 20% 15% 2% 18% 36% 2.04

DNV GL – www.dnvgl.com February 6, 2020 Page 114

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

HOLLAND 0% 14% 20% 0% 18% 23% 1.94

HOLLISTON 10% 15% 9% 1% 35% 46% 0.74

HOPEDALE 5% 21% 10% 3% 28% 44% 1.23

HOPKINTON 5% 15% 7% 2% 29% 50% 0.83

HUBBARDSTON 2% 10% 16% 0% 25% 31% 1.05

HUDSON 0% 0% 13% 3% 29% 100% 0.58

HUNTINGTON 4% 15% 12% 0% 19% 26% 1.43

IPSWICH 23% 30% 16% 2% 17% 20% 2.71

KINGSTON 10% 21% 11% 0% 27% 44% 1.16

LAKEVILLE 0% 9% 9% 0% 23% 28% 0.79

LANCASTER 3% 19% 15% 3% 26% 36% 1.37

LANESBOROUGH 10% 17% 12% 0% 14% 16% 2.11

LAWRENCE 28% 72% 15% 26% 6% 28% 19.46

LEE 17% 30% 16% 3% 17% 29% 2.79

LEICESTER 12% 29% 17% 3% 25% 38% 1.94

LENOX 30% 40% 19% 2% 18% 32% 3.38

LEOMINSTER 26% 45% 14% 6% 19% 43% 3.40

LEVERETT 3% 11% 11% 0% 26% 34% 0.87

LEXINGTON 13% 19% 7% 4% 34% 47% 0.88

LEYDEN 4% 15% 16% 1% 26% 35% 1.22

LINCOLN 13% 19% 7% 2% 36% 44% 0.76

LONGMEADOW 6% 9% 9% 4% 28% 38% 0.79

LOWELL 36% 58% 16% 14% 11% 29% 8.12

LUDLOW 10% 25% 15% 7% 18% 36% 2.59

LUNENBURG 6% 20% 14% 1% 19% 34% 1.82

DNV GL – www.dnvgl.com February 6, 2020 Page 115

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

LYNN 28% 56% 14% 15% 12% 28% 7.39

MALDEN 33% 59% 13% 16% 14% 32% 6.23

MANCHESTER 12% 31% 10% 1% 22% 31% 1.90

MARION 4% 22% 15% 0% 23% 32% 1.59

MARLBOROUGH 30% 45% 14% 11% 17% 37% 4.20

MARSHFIELD 12% 22% 11% 1% 19% 34% 1.72

MASHPEE 8% 13% 14% 1% 19% 39% 1.44

MATTAPOISETT 1% 19% 18% 0% 22% 37% 1.67

MAYNARD 11% 28% 10% 0% 22% 39% 1.79

MEDFIELD 9% 12% 6% 0% 30% 46% 0.60

MEDFORD 21% 44% 13% 7% 16% 38% 3.96

MEDWAY 9% 15% 6% 1% 26% 42% 0.80

MELROSE 26% 33% 14% 3% 24% 43% 2.11

MENDON 0% 12% 3% 0% 33% 39% 0.46

METHUEN 18% 29% 14% 6% 13% 30% 3.70

WORTHINGTON 2% 13% 13% 0% 23% 29% 1.13

MILFORD 16% 32% 16% 5% 20% 37% 2.66

MILLBURY 12% 27% 12% 2% 25% 46% 1.59

MILLIS 9% 19% 13% 0% 25% 41% 1.28

MILLVILLE 5% 23% 12% 0% 24% 31% 1.41

MILTON 10% 17% 8% 1% 30% 41% 0.89

MONSON 4% 17% 20% 0% 29% 33% 1.31

MONTAGUE 12% 41% 16% 1% 18% 35% 3.32

MONTEREY 1% 20% 17% 0% 13% 17% 2.83

MONTGOMERY 0% 2% 16% 1% 28% 31% 0.68

DNV GL – www.dnvgl.com February 6, 2020 Page 116

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

MOUNT WASHINGTON 1% 15% 18% 1% 15% 9% 2.26

NAHANT 16% 32% 15% 1% 20% 26% 2.33

NANTUCKET 1% 36% 16% 3% 17% 22% 3.24

NATICK 24% 27% 13% 3% 25% 47% 1.71

NEEDHAM 12% 17% 8% 3% 29% 41% 0.94

NEW ASHFORD 4% 17% 11% 1% 13% 7% 2.26

NEW BEDFORD 20% 58% 16% 13% 8% 24% 10.77

NEW BRAINTREE 0% 11% 11% 0% 26% 29% 0.85

NEW MARLBOROUGH 0% 13% 20% 0% 14% 13% 2.36

NEW SALEM 0% 11% 17% 0% 31% 37% 0.87

NEWBURY 5% 15% 10% 1% 23% 26% 1.12

NEWBURYPORT 17% 26% 12% 0% 21% 42% 1.81

NEWTON 15% 29% 8% 5% 24% 41% 1.74

NORFOLK 2% 5% 6% 0% 39% 48% 0.28

NORTH ADAMS 18% 46% 15% 2% 17% 35% 3.74

NORTH ANDOVER 23% 27% 9% 1% 21% 40% 1.73

NORTH ATTLEBOROUGH 0% 7% 12% 0% 50% 25% 0.38

NORTH BROOKFIELD 8% 30% 18% 0% 19% 20% 2.58

NORTHAMPTON 25% 44% 13% 2% 14% 42% 4.12

NORTHBOROUGH 7% 17% 10% 1% 28% 41% 1.01

NORTHBRIDGE 14% 35% 14% 1% 18% 30% 2.69

NORTHFIELD 5% 23% 23% 0% 23% 33% 2.00

NORTON 12% 18% 11% 1% 19% 34% 1.59

NORWELL 3% 7% 10% 0% 29% 42% 0.61

OAK BLUFFS 5% 22% 16% 2% 11% 19% 3.67

DNV GL – www.dnvgl.com February 6, 2020 Page 117

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

OAKHAM 0% 6% 16% 0% 24% 31% 0.90

ORANGE 16% 31% 15% 0% 21% 38% 2.25

ORLEANS 8% 23% 14% 1% 21% 35% 1.79

OTIS 1% 12% 16% 0% 10% 19% 2.69

OXFORD 12% 28% 11% 1% 24% 37% 1.64

PALMER 10% 26% 14% 0% 21% 31% 1.91

PAXTON 2% 12% 19% 0% 25% 79% 1.28

PELHAM 1% 17% 15% 1% 27% 31% 1.20

PEMBROKE 9% 13% 12% 1% 22% 34% 1.16

PEPPERELL 7% 22% 10% 0% 24% 36% 1.35

PERU 0% 8% 17% 0% 19% 22% 1.33

PETERSHAM 0% 16% 13% 0% 23% 27% 1.29

PHILLIPSTON 0% 6% 13% 1% 21% 26% 0.97

PITTSFIELD 14% 40% 16% 2% 16% 33% 3.69

PLAINFIELD 0% 11% 21% 0% 23% 27% 1.38

PLAINVILLE 15% 26% 11% 0% 14% 31% 2.74

PLYMOUTH 9% 22% 15% 2% 22% 39% 1.72

PLYMPTON 4% 13% 11% 0% 30% 40% 0.79

PROVINCETOWN 21% 33% 14% 2% 13% 40% 3.74

QUINCY 35% 53% 14% 12% 15% 38% 5.44

RANDOLPH 21% 31% 16% 8% 17% 35% 3.25

READING 25% 30% 8% 0% 22% 82% 1.77

REHOBOTH 0% 9% 11% 2% 26% 32% 0.81

REVERE 25% 52% 18% 13% 13% 33% 6.62

RICHMOND 2% 10% 14% 0% 20% 26% 1.18

DNV GL – www.dnvgl.com February 6, 2020 Page 118

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

ROCHESTER 0% 10% 13% 1% 32% 36% 0.76

ROCKLAND 19% 28% 15% 1% 21% 41% 2.10

ROCKPORT 8% 31% 17% 0% 21% 35% 2.22

ROWE 2% 22% 19% 0% 20% 28% 2.04

ROYALSTON 0% 17% 24% 0% 17% 16% 2.47

RUSSELL 1% 10% 21% 0% 18% 28% 1.68

RUTLAND 4% 13% 12% 0% 24% 36% 1.06

SALEM 29% 52% 14% 5% 15% 44% 4.64

SALISBURY 8% 25% 15% 0% 15% 29% 2.59

SANDISFIELD 0% 8% 23% 1% 14% 18% 2.27

SANDWICH 3% 15% 14% 0% 23% 36% 1.23

SAUGUS 14% 21% 12% 2% 20% 36% 1.80

SAVOY 1% 11% 17% 0% 23% 27% 1.22

SCITUATE 5% 14% 7% 0% 21% 38% 1.03

SEEKONK 1% 13% 13% 1% 22% 31% 1.28

SHARON 8% 14% 8% 4% 29% 43% 0.87

SHEFFIELD 3% 19% 15% 0% 17% 21% 1.98

SHELBURNE 11% 36% 19% 1% 22% 31% 2.54

SHERBORN 4% 7% 5% 1% 36% 45% 0.36

SHIRLEY 8% 30% 15% 0% 21% 45% 2.16

SHREWSBURY 18% 14% 8% 5% 11% 52% 2.44

SHUTESBURY 2% 11% 11% 0% 25% 34% 0.86

SOMERSET 3% 19% 13% 3% 15% 25% 2.37

SOMERVILLE 26% 65% 14% 8% 9% 29% 9.69

SOUTHAMPTON 5% 10% 18% 1% 23% 30% 1.27

DNV GL – www.dnvgl.com February 6, 2020 Page 119

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

SOUTHBOROUGH 1% 11% 7% 1% 31% 43% 0.61

SOUTHBRIDGE 17% 57% 18% 10% 12% 31% 6.96

SOUTHWICK 9% 19% 13% 0% 21% 29% 1.50

SPENCER 12% 36% 12% 1% 17% 37% 2.89

SPRINGFIELD 21% 53% 16% 12% 11% 30% 7.38

STOCKBRIDGE 11% 32% 18% 0% 16% 22% 3.16

STONEHAM 30% 36% 14% 3% 23% 45% 2.30

STOUGHTON 18% 28% 13% 5% 16% 34% 2.85

STOW 14% 16% 5% 0% 25% 54% 0.82

STURBRIDGE 12% 18% 14% 3% 26% 41% 1.35

SUDBURY 4% 9% 6% 1% 40% 49% 0.39

SUNDERLAND 36% 57% 14% 3% 23% 48% 3.13

SUTTON 3% 9% 10% 0% 32% 42% 0.61

SWAMPSCOTT 14% 23% 8% 3% 26% 41% 1.30

SWANSEA 2% 13% 14% 3% 17% 27% 1.72

TEMPLETON 5% 10% 19% 0% 50% 38% 0.57

TEWKSBURY 12% 14% 10% 1% 22% 41% 1.13

TISBURY 2% 32% 18% 5% 14% 20% 4.00

TOLLAND 0% 15% 12% 0% 13% 13% 2.01

TOPSFIELD 2% 8% 10% 0% 31% 44% 0.60

TOWNSEND 9% 20% 13% 0% 14% 23% 2.35

TRURO 4% 17% 10% 0% 18% 26% 1.54

TYNGSBOROUGH 10% 14% 13% 2% 25% 35% 1.12

TYRINGHAM 1% 9% 8% 1% 14% 14% 1.29

UPTON 5% 18% 6% 1% 28% 40% 0.84

DNV GL – www.dnvgl.com February 6, 2020 Page 120

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

UXBRIDGE 5% 21% 12% 1% 24% 40% 1.46

WAKEFIELD 0% 8% 10% 3% 11% 17% 1.81

WALES 2% 20% 16% 2% 17% 21% 2.17

WALPOLE 11% 18% 11% 1% 23% 40% 1.29

WALTHAM 30% 50% 13% 6% 14% 35% 5.09

WARE 10% 29% 20% 1% 21% 34% 2.34

WAREHAM 6% 23% 15% 1% 17% 35% 2.29

WARREN 10% 30% 17% 0% 16% 23% 2.92

WARWICK 1% 10% 14% 0% 25% 30% 0.96

WASHINGTON 0% 1% 15% 0% 15% 21% 1.07

WATERTOWN 24% 49% 12% 4% 15% 41% 4.29

WAYLAND 6% 12% 4% 2% 34% 44% 0.53

WEBSTER 15% 46% 18% 4% 14% 31% 4.85

WELLFLEET 1% 16% 18% 0% 19% 29% 1.79

WENDELL 0% 14% 17% 2% 37% 33% 0.92

WENHAM 10% 12% 13% 0% 28% 37% 0.91

WEST BRIDGEWATER 3% 14% 16% 0% 21% 19% 1.40

WEST BROOKFIELD 13% 29% 15% 0% 20% 20% 2.17

WEST NEWBURY 1% 5% 7% 0% 32% 39% 0.36

WEST SPRINGFIELD 23% 40% 18% 5% 15% 33% 4.14

WEST STOCKBRIDGE 1% 10% 16% 0% 18% 12% 1.43

WEST TISBURY 0% 15% 10% 0% 14% 19% 1.76

WESTBOROUGH 33% 39% 8% 3% 21% 49% 2.39

WESTFIELD 8% 10% 10% 2% 33% 10% 0.65

WESTFORD 5% 11% 9% 2% 29% 41% 0.74

DNV GL – www.dnvgl.com February 6, 2020 Page 121

Town

Ratio of MF/Total

Units

Ratio Renter

Occupied

Units (ACS)

Ratio of Households at

56 - 85% of Statewide

Median Income (ACS)

Ratio Limited

English

Speaking

Households

Unweighted

Participation

Rate

Electric

Combined

2013-2017

Consumption

Weighted

Participation

Rate (KWh)

Community

Outreach Metric

WESTHAMPTON 0% 9% 13% 0% 29% 33% 0.76

WESTMINSTER 7% 13% 5% 0% 23% 35% 0.81

WESTON 7% 15% 5% 3% 30% 32% 0.78

WESTPORT 4% 16% 10% 2% 16% 24% 1.75

WESTWOOD 17% 13% 9% 2% 28% 44% 0.85

WEYMOUTH 30% 36% 15% 2% 21% 38% 2.52

WHATELY 0% 15% 19% 2% 26% 31% 1.37

WHITMAN 12% 28% 14% 0% 21% 36% 1.99

WILBRAHAM 4% 11% 8% 1% 29% 41% 0.68

WILLIAMSBURG 5% 22% 11% 0% 27% 36% 1.21

WILLIAMSTOWN 11% 27% 16% 2% 19% 30% 2.37

WINCHENDON 9% 22% 16% 1% 15% 24% 2.57

WINCHESTER 11% 15% 8% 3% 31% 43% 0.81

WINDSOR 0% 1% 15% 0% 19% 22% 0.84

WINTHROP 21% 43% 15% 4% 15% 38% 4.15

WOBURN 24% 38% 14% 4% 23% 41% 2.44

WORCESTER 27% 57% 16% 11% 12% 32% 7.30

WRENTHAM 5% 16% 9% 0% 25% 37% 1.00

YARMOUTH 11% 21% 21% 3% 20% 36% 2.26

Grand Total 21% 39% 13% 6% 16% 35% 3.57

DNV GL – www.dnvgl.com February 6, 2020 Page 122

8 APPENDIX C: BIVARIATE MAPS

DNV GL originally provided the bivariate maps to the PAs in a powerpoint file. The slides of that file are

reproduced as images in this appendix.

DNV GL – www.dnvgl.com February 6, 2020 Page 123

DNV GL – www.dnvgl.com February 6, 2020 Page 124

DNV GL – www.dnvgl.com February 6, 2020 Page 125

Combined gas and electric consumption weighted participation * Renter-occupied households

DNV GL – www.dnvgl.com February 6, 2020 Page 126

Combined gas and electric consumption weighted participation * Limited-English households

DNV GL – www.dnvgl.com February 6, 2020 Page 127

Combined gas and electric consumption weighted participation * Moderate income (40k-60k)

households

DNV GL – www.dnvgl.com February 6, 2020 Page 128

Electric consumption weighted participation * Renter-occupied households

DNV GL – www.dnvgl.com February 6, 2020 Page 129

Electric consumption weighted participation * Limited-English households

DNV GL – www.dnvgl.com February 6, 2020 Page 130

Electric consumption weighted participation * Moderate income (40k-60k) households

DNV GL – www.dnvgl.com February 6, 2020 Page 131

Gas consumption weighted participation * Renter-occupied households

DNV GL – www.dnvgl.com February 6, 2020 Page 132

Gas consumption weighted participation * Limited-English households

DNV GL – www.dnvgl.com February 6, 2020 Page 133

Gas consumption weighted participation * Moderate income (40k-60k) households

DNV GL – www.dnvgl.com February 6, 2020 Page 134

DNV GL – www.dnvgl.com February 6, 2020 Page 135

9 APPENDIX D: CLOSE-UP HOTSPOT MAPS

The close-up hot spot maps provide the same information as the statewide hotspot map, but covering

smaller areas. DNV GL generated a composite number for each block group and then input these composite

numbers to generate a geographic hot spot map based on areas where there are concentrations of

consistently high (or low) block group scores.35 We defined an area as the block groups adjacent to a given

block group. 36 The composite number is calculated by taking the sum of the ratios of the following

attributes:

1. Percentage of households that are moderate income37

2. Percentage of households that are renter-occupied

3. Percentage of households with a primary language other than English

4. Percentage of households that are in structures of 5 or more units (i.e., a proxy for multifamily)

5. Percentage of structures that were built prior to 1940

The hot spot maps do not factor in participation status.

35 For technical details on ESRI’s implementation of hot spot analysis and additional context, please see https://pro.arcgis.com/en/pro-app/tool-

reference/spatial-statistics/h-how-hot-spot-analysis-getis-ord-gi-spatial-stati.htm. 36 We chose to look at adjacent block group areas in consideration of differences in block group area between densely populated cities like Boston and

rural towns like Mt. Washington, and of our understanding that stakeholders are more interested in local geographic differences than large scale regional differences. Other representations we considered were a Euclidean distance function (e.g., look at all block groups within 5 miles) and

an inverse distance weight (e.g., look at all block groups within 5 miles, but as we get further away, assign a lower hot spot impact weight to

the block groups). In both cases, DNV GL felt that the regional scale differences between the rural western and central parts of the state versus

the densely populated eastern part of the state precluded using distance as our weighting variable. 37 As defined earlier in this report

DNV GL – www.dnvgl.com February 6, 2020 Page 136

Figure 9-1. ACS variable hot spot map, closeup of Worcester and surrounding areas

DNV GL – www.dnvgl.com February 6, 2020 Page 137

Figure 9-2. ACS variable hot spot map, closeup of Springfield, Holyoke, and surrounding areas

DNV GL – www.dnvgl.com February 6, 2020 Page 138

Figure 9-3. ACS variable hot spot map, closeup of Lawrence and surrounding areas

DNV GL – www.dnvgl.com February 6, 2020 Page 139

Figure 9-4. ACS variable hot spot map, closeup of Fitchburg and surrounding areas

DNV GL – www.dnvgl.com February 6, 2020 Page 140

Figure 9-5. ACS variable hot spot map, closeup of Fall River and surrounding areas

DNV GL – www.dnvgl.com February 6, 2020 Page 141

Figure 9-6. ACS variable hot spot map, closeup of Boston and surrounding areas

ABOUT DNV GL Driven by our purpose of safeguarding life, property and the environment, DNV GL enables organizations to advance the safety and sustainability of their business. We provide classification and technical assurance along with software and independent expert advisory services to the maritime, oil and gas, and energy industries. We also provide certification services to customers across a wide range of industries. Operating in more than 100 countries, our 16,000 professionals are dedicated to helping our customers make the world safer, smarter, and greener.