From Health Research to Social Research: Privacy, Methods, Approaches

Leslie L Roos

Distinguished ProfessorDepartment of Community Health

SciencesManitoba Centre For Health Policy

For Comparison: Panel Study of Income Dynamics (PSID)

• Compare administrative data with the only social science study noted on “the NSF’s list of its fifty most significant projects in its fifty year history”

• Highlights issues in extending population-based (health) administrative data to facilitate social research

Information-Rich Environment: Using Administrative Data

ResearchRegistry

Medical

VitalStatistics

Home Care NursingHome

Hospital

ProviderPharmaceutical

Immunization Monitoring

Inflammatory Bowel Disease Database

Sleep Lab Clinical Data

Alcoholism Panel Surveys

Cancer Registry Educational and Social Data

Heart Health Survey

Aging in Manitoba Study

National Population Health

Survey

Table 1. Comparing Longitudinal Primary Data and Population-Based Administrative Data

CharacteristicsLongitudinal Primary Data (Panel Study in Income Dynamics)

Population-based Administrative Data (Manitoba and other sites)

Number of cases Several thousand or smaller (5,000 households in PSID)

Often over a million

Cost High on a per person basis Very low on a per person basis

Representativeness Often national Often from a province or state

Population Studied Subjects sampled and tracked Built on a registry of an entire population

Research Design Often complex designs need to increase power and control costs

Given a population, complex designs can be imposed retrospectively as needed

Record Linkage Useful in some contexts Critical to check data quality and expand scope of information sources

Individual Follow-up Before and after an event Before and after an event

Coverage and Loss to Follow-up

Nonresponse and differential attrition possible Differential attrition possible

Updating New data must be collected and merged with existing data

Multifile information must be cleaned and merged with existing data. Cleaning relies on record linkage

Time Information must be collected (typically annually or at longer intervals)

Information provided at relatively short intervals (from daily to annually)

Place Information at time of study (historical reconstruction possible)

Detailed information usually provided close to date of move

Longitudinal Goes back many years/Corrections for immigration in PSID)

Goes back many years

Neighborhoods Flexible construction from postal code or census area

Flexible construction from postal code or census area; large N may permit flexible assignment to generate nearest 'neighbors'

Life Events Collected as part of design Possibly available from registry or other sources

Family and Intergenerational Data

Collected as part of design (PSID); sibling and intergenerational studies facilitated

Assessing a family composition at any point in time possible; sibling and twin studies facilitated

Limits Important information likely to be collectable for entire sample

Important information may be missing or available only for a subpopulation

Measures Defined by researchers; scaling possible Defined by others for administrative purposes; creating meaningful variables may be very time-consuming; scaling possible

Intellectual History Scope of data collection often expanded to provide a rich data set

Scope expanded beyond initial health care data by receiving files from other agencies

Expanding Capabilities

a) Link files to incorporate new data sets while preserving privacy and confidentiality

b) Measure such outcomes as educational achievement at population level

Moving from health research to social research involves preparatory work

with databases, which must be organized to:

Expanding Capabilities (cont’d)

c) Use place of residence data (for any point in time) to calculate the number of moves, number of years in relatively poor neighborhoods, etc.

d) Estimating family composition at any point in time (tracking marital status, family size, ages of siblings and twins). Social variables used for more powerful research designs.

• Canada Foundation for Innovation and provincially funded Data Laboratory

• Highest standards of security, privacy & confidentiality of data

• No names, no addresses

• Probabilistic record linkages across files as needed

• Data for research, not for administrative use

• Provincial privacy offices kept fully informed

Respect for Privacy In Building Manitoba

Database

Linking while Preserving Privacy

• Multi stage de-identification process:– Trustee preparation– Manitoba health links Trustee’s file to

the encrypted PHIN– Crosswalk file provided to MCHP – Trustee provides data file– MCHP stores data separately and

unlinked

Currently Available: Measures to Understand Well

BeingAvailable for population for one or more years:

• Age, grade level, school attendance, marks, achievement tests for Grade 3, Grade 9, Grade 12

• Well-identified health conditions such as asthma and diabetes plus a measure of general health status can be studied through childhood and adolescence

• Healthcare utilization and costs• Receipt of social assistance

Not Available

• Survey-based measure of household income and education are not available for entire population

• Manitoba comparisons of income measure found risk estimates for health status measures derived from neighborhood income not attenuated relative to those from household income

• May not be true everywhere

Small Area Measures from Statistics Canada

• Publicly available• Use dissemination/enumeration areas• Often encompass several six digit postal

code areas• Indicators such as mean household

income and education, unemployment rates, etc.

• Canadian studies of the socioeconomic gradient in utilization often use small area markers

Steps for comparing educational achievement across geographical

areas parallel those developed for health research

1) Obtaining indicators of socioeconomic risk for each small area

2) Developing individual data on educational achievement

3) Calculate rates of achievement based on small areas

4) Track the entire population5) Combine census enumeration areas as

appropriate (for quartiles or quintiles based on socioeconomic status of area of residence)

Data Used

• Up to nine years of Manitoba birth cohorts (1978-1987, excluding the 1983 cohort)

• Takes advantage of available provincial Grade 12 test scores

• About 13,000 in each cohort surviving and remaining in the province for the first 18 years of life

• Large sample for same-sex pairs (33,000) and multiple births (2,000)

Population-Based Data Provide a

Different Perspective

Grade 12 (S4) Performance, by Winnipeg SES Group, Language Arts Standards

Test, 2001/02

75%83%

92%87%

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Low Low-Mid Middle High

SES

Pass/Fail rates of test writers 18 year olds who should have written

27%

52%

65%

77%

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Low Low-Mid Middle High

SES

Withdrawn

In Grade 11(S3) or lower

In Grade 12(S4), but noLA Test Mark

Drop Course,Absent,Exempt, Incomplete

Fail

Pass

Estimating Family Circumstances

• Number of children in the family• Birth order• Mother’s marital status at birth of first

child• Age of mother at birth of first child• Whether or not family was receiving

income assistance• Number of changes of residence

Other Social Variables

• Number of years in different types of families

• Number of “family structure” changes (parental separations, remarriages)

• Number of years living with disabled parent

• NOTE: Almost all variables can be measured at different points in time

Importance of Social Variables

• The first six social variables (i.e. number of children, birth order, etc.) were better predictors of educational achievement than survey model (includes household income and parental education) from PSID

• Predictive power varies with outcome selected (much lower for health outcomes)

• Used as control variables in sibling and sibling/neighbourhood research

Emergence of the Socioeconomic Gradient

Association between socioeconomic status and health may vary at different life stages

Questions include:

1) How do socioeconomic gradients evolve for both males and females as children grow older?2) What can we learn about gradient development by looking at the trajectories of individuals over time?3) Are socioeconomic gradients in child health changing over time?

Problems with Non-Experimental Design

• Analyzing casual relationship is difficult

• Omitted variables and measurement error are likely to bias the coefficients attached to measured variables

• Standard statistical analyses suffer from this problem (a growing literature in economics on this issue)

• Whole population, sibling and twin studies each have different strengths and weaknesses

• Siblings and twin designs ‘control’ differently often using ‘family fixed effects’ statistical models

Problems with Non-Experimental Design

Literature Review

• Relationship between birth weight and infant mortality decreases when differences between twins examined

• Twin samples help eliminate unobserved heterogeneity across families BUT twins generally of lower birth weight than singletons

• Gestational length not examined in twin studies

• Canadian data provide uniform access to health insurance whereas coverage in U.S. may vary even among siblings

Approach One: Favourite Economist Tactic

• Summarizing cross-family variation (each family as a separate condition) provides estimate of effects of presumably causal individual variables (such as birth weight) by differencing out family-specific characteristics affecting all children.

• Comparisons of siblings within families help make up for lack of control over variables measuring household income and parental education

Comparing Twin and Sibling Estimates

• Exploits benefits of both identification strategies

• Siblings provide a more representative sample but strategy risks biased estimates:A) Potential change in parental investment following the birth of first childB) Potential change in socioeconomic status

between births

• Using twins eliminates these potential biases but sample is limited and unrepresentative

• Patterns of postnatal development (as a function of infant health) may differ for siblings and for twins h could affect results

Results

• Infant health strongly predicted both high school completion and income assistance take-up and length, controlling for a number of possible cofounders

• Long-term consequences of infant health were found across families, within siblings, and within twin pairs.

Measuring Infant Health: Five Models*

1) Ordinary Least Squares (OLS) using entire sample

2) OLS using sample of children with siblings3) Sibling sample including family fixed

effects4) OLS using sample of twins5) Twin sample including family fixed effects

*Looked at longer-term measures of child health and social outcomes

Approach Two:Sibling Neighbourhood Design

A) Compare correlations for siblings and unrelated neighbours of similar ages

B) Use to identify an “upper bound on the influence of neighbourhoods, because neighbour correlations reflect similar family backgrounds as well as shared community backgrounds”

C) Extend this approach to consider schoolmates and peers

Study Opportunity Structures

• Urban-rural comparisons

• Socio-economic gradient (comparing means and correlations by income)

• Comparisons across dimensions of well being

Sibling Neighbourhood Design

• Uses location of each family in a 6 digit postal code area

Matches a) with neighbour in postal code area, then b) those ‘leftover’, with closest ‘leftover’ in

census enumeration area (neat linear program)

Figure 1. Sibling - Neighbourhood Designs

Neighbourhood A

Family 1Sibling 1aSibling 1b

Family 2Sibling 2aSibling 2b

Neighbourhood B

Family 3Sibling 3aSibling 3b

Family 4Sibling 4aSibling 4b

Sibling and Neighbour Correlations Girls Outside Winnipeg:

1978-1987 Cohorts (excluding 1983)Pregnancy Before Age 19 Sibling (Unadjusted) Neighbour (Unadjusted)Neighbour (Adjusted)

.576*

.368*

.120*

Income Assistance After Age 18 ( 1978-1982) Sibling (Unadjusted) Neighbour (Unadjusted)Neighbour (Adjusted)

.612*

.174*

.000*

Held back by Grade 12 Sibling (Unadjusted) Neighbour (Unadjusted)Neighbour (Adjusted)

.675*

.369*

.111*

A Guaranteed Annual Income Experiment

• A population-based guaranteed annual income experiment in Dauphin, MB from 1974 to 1978 substantially reduced poverty.

To study long-term effects: • Propensity matching created a 3-to-1 control

for all Dauphin and rural municipality residents.

• A second set of controls age-and sex-matched residents living in the test site before the experiment

Research Questions

1) Did the elimination of income insecurity when children were particularly vulnerable affect their lives after the experiment ended?

2) Will young children in families receiving assistance have experienced better health and social outcomes as adolescents?

Provincial Centres:

1) Anchored in a ‘place’2) Feedback to ministries/data suppliers

essential3) Deliverables negotiated4) “Smiling persistence” to obtain data

sets

From Health Research to Social Research

1) Have reviewed issues of measurement, data organization, and analytical strategy

2) Expanded ability to examine outcomes through different stages in the life course