contents lists available at sciencedirect journal of...

ARTICLE IN PRESS

Contents lists available at ScienceDirect

Journal ofEnvironmental Economics and Management

Journal of Environmental Economics and Management 56 (2008) 207–220

0095-06

doi:10.1

� Cor

E-m1 H

The US

seventy

journal homepage: www.elsevier.com/locate/jeem

Daylight time and energy: Evidence from an Australian experiment

Ryan Kellogg a, Hendrik Wolff b,�

a Department of Economics, University of Michigan, USAb Department of Economics, University of Washington, 524 Condon Hall Box 353330, Seattle, WA 98195-3330, USA

a r t i c l e i n f o

Article history:

Received 30 April 2007Available online 15 August 2008

Keywords:

Energy

Daylight saving time

Difference-in-difference-in-difference

96/$ - see front matter & 2008 Elsevier Inc. A

016/j.jeem.2008.02.003

responding author. Fax: +1 206 685 7477.

ail address: [email protected] (H. W

istorically, DST has been most actively implem

observed year-round DST during World War II

countries worldwide. For more information

a b s t r a c t

Several countries are considering using daylight saving time (DST) as a tool for energy

conservation and reduction of greenhouse gas emissions, and the United States

extended DST in 2007 with the goal of reducing electricity consumption. This paper

assesses DST’s impact on electricity demand by examining a quasi-experiment in which

parts of Australia extended DST in 2000 to facilitate the Sydney Olympics. Using

detailed panel data and a difference-in-difference-in-difference framework, we show

that the extension did not reduce overall electricity consumption, but did cause a

substantial intraday shift in demand consistent with activity patterns that are tied to the

clock rather than sunrise and sunset.

& 2008 Elsevier Inc. All rights reserved.

I say it is impossible that so sensible a peopleyshould have lived so long by the smoky, unwholesome, andenormously expensive light of candles, if they had really known that they might have had as much pure light of thesun for nothing.

Benjamin Franklin, 1784

1. Introduction

One principal socio-economic problem is the optimal allocation of individuals’ activities—sleep, work, and leisure—over24 h of the day. In today’s world of artificial lighting and heating, people set their active hours by the clock rather than bythe natural cycle of dawn and dusk. In one of the earliest statistical treatments in economics, ‘‘An Economical Project,’’Benjamin Franklin [9] observes that this behavior wastes valuable morning daylight as people sleep until long after sunrise,and requires expensive candles to illuminate the night. Franklin calculates that this misallocation causes Paris to consumean additional 64 million pounds of tallow and wax annually.

Governments have since attempted to address this resource allocation problem through the mechanism of daylightsaving time (DST).1 Each year, we move our clocks forward by 1 h in the spring, and adjust them back to Standard Time inthe fall. The intuition behind this DST adjustment relies on the premise that people’s activities shift forward with the clock,so that, during the summer, the sun appears to set 1 h later and the ‘‘extra’’ hour of evening daylight cuts electricitydemand.

Today, heightened concerns regarding energy prices and greenhouse gas (GHG) emissions are driving interest inextending DST in several countries, including Australia, Canada, Japan, New Zealand, and the United Kingdom[17,22,24–26]. The United States has already passed legislation to extend DST by 1 month, beginning in 2007, with the

ll rights reserved.

olff).

ented in times of energy scarcity. The first application of DST was in Germany during World War I.

and implemented extensions during the energy crisis in the 1970s [7]. Today, DST is observed in over

on the history of DST, see [6,20].

www.sciencedirect.com/science/journal/yjeem

www.elsevier.com/locate/jeem

dx.doi.org/10.1016/j.jeem.2008.02.003

mailto:[email protected]

ARTICLE IN PRESS

R. Kellogg, H. Wolff / Journal of Environmental Economics and Management 56 (2008) 207–220208

specific goal of reducing electricity consumption by 1% during the extension [8]. Therefore, the United States now switchesto DST in March rather than in April. The energy legislation also explicitly calls for research into the energy impacts ofextending DST, and suggests reverting to the prior DST system if it is demonstrated that the policy goal will not be achieved.Beyond this federal initiative, California is considering even more drastic changes—year-round DST and double DST—thatare predicted to save up to 1.3 billion US dollars annually [5].

Our study challenges the energy conservation predictions that have been used to justify these calls for the expansion ofDST. Across the studies and reports we surveyed, estimates of an extension’s effect on total electricity demand range fromsavings of 0.6% to 3.5%. The most widely cited savings estimate of 1% is based on an examination conducted over 30 yearsago [27]. Arguably, these findings are not applicable today. For example, the widespread adoption of air conditioning hasaltered intraday patterns of electricity consumption. Further, the 1% savings estimate may be confounded by other energyconservation measures enacted during the oil crisis.

More recent efforts to predict the effect of extending DST on electricity demand employ simulation models, which usedata from the status quo DST system to forecast electricity use under an extension. One prominent recent study is beingused to argue in favor of year-round DST in California [4]. It predicts three benefits of an extension: (1) a 0.6% reduction inelectricity consumption, (2) lower electricity prices, driven by a reduction in peak demand, and (3) a lower likelihood ofrolling blackouts. However, this study is not based on firm empirical evidence; it instead uses electricity consumption dataunder the current DST scheme to simulate demand under extended DST. It may therefore fail to capture the full behavioralresponse to a change in DST timing.2

Our study obviates the need to rely on simulations by examining actual data from a quasi-experiment that occurred inAustralia in 2000. Typically, three of Australia’s six states observe DST beginning in October (which is seasonally equivalentto April in the northern hemisphere). However, to facilitate the 2000 Olympics in Sydney, two of these three states beganDST 2 months earlier than usual. Because the Olympics can directly affect electricity demand, we focus on the state ofVictoria—which extended DST but did not host Olympic events—as the treated state, and use its neighboring state, SouthAustralia, which did not extend DST, as a control. We also drop the 2-week Olympic period from the 2-month treatmentperiod to further remove confounding effects. Using a detailed panel of half-hourly electricity consumption and prices over7 years, as well as the most detailed weather information available, we examine how the DST extension affected electricitydemand in Victoria.

Our treatment effect estimation strategy is based on the difference-in-difference (DD) framework that exploits, in boththe treatment state and the control state, the difference in demand between the treatment year and the control years. Weaugment the standard DD model to take advantage of the fact that DST does not affect electricity demand in the mid-day.This allows us to use changes in mid-day consumption to control for unobserved state-specific shocks via a difference-in-difference-in-difference (DDD) specification. We show that this technique allows us to employ a mild identifyingassumption that is more appropriate for the data than that of a standard DD model.

Our results confirm policy-makers’ expectations that the extension of DST causes electricity demand to decreasesignificantly in the evening. However, we also find an opposing effect in the morning: the Australian extension significantlyincreased electricity consumption between 07:00 and 08:00. Overall, the evening decrease in demand did not outweigh themorning increase, so that total electricity consumption in Australia was not reduced as a result of the DST extension. Theseeffects are consistent with waking and sleeping behaviors that are tied to the clock, rather than to sunrise and sunset. Inparticular, the residents of Australia do not appear to have substantially altered the clock time at which they awokefollowing the extension of DST; they therefore rose before sunrise and needed electric power for lighting.

These results contradict the claims made by prior studies that extending DST will conserve energy, and indicate thatproposals in Australia to extend DST permanently are unlikely to reduce energy use and GHG emissions. Furthermore, themorning peak demand caused by Australia’s 2000 extension is associated with significantly higher wholesale electricityprices, indicating that the steep morning ramp-up in demand likely caused an increase in generation costs. This outcomeundercuts claims that extending DST leads to generation efficiencies by smoothing the hourly demand profile.

While we cannot directly apply our results to other countries without adjustments for behavioral and climaticdifferences, this study raises concern that the recent DST extension in the United States is unlikely to result in energyconservation. To investigate the degree to which our results extend to the US, we reconstruct the simulation model thatwas used to forecast energy savings for California [4], and apply it to the Australian data. Noting that Victoria’s latitude andclimate are similar to those of central California, we find that the simulation systematically overstates energy savings inboth the morning and evening, casting further suspicion on claims that extending DST in California and the rest of theUnited States will reduce electricity consumption.

2. Background on daylight saving time in Australia

The geographical area of interest is the southeastern part of Australia. Three states in the southeast of the mainlandobserve DST: South Australia (SA), New South Wales (NSW), and Victoria (VIC). DST typically starts on the last Sunday in

2 Rock [21] also uses a simulation model, and finds that year-round DST decreases electricity consumption by 0.3% and expenditures by 0.2%.

However, his study does not include non-residential electricity use, which accounts for 64% of US total electricity consumption [28].

ARTICLE IN PRESS

Table 1Characteristics of capital cities in southeast Australia

Capital State Population in

millions

Population in

millions

Per capita in

1000 AUD

Latitude

south

Longitude

east

Average

sunrise

Average

sunset

Sydney NSW 4.3 6.5 41.4 33150 151110 5:50 17:45

Melbourne VIC 3.7 4.8 39.3 371470 1451580 6:20 18:10

Adelaide SA 1.1 1.5 33.4 341550 1381360 6:50 18:35

All data are for 2000. Sunrise and sunset times are in East Australian Standard Time, and averaged during September.

R. Kellogg, H. Wolff / Journal of Environmental Economics and Management 56 (2008) 207–220 209

October and ends on the last Sunday in March. Queensland, the Northern Territory, and Western Australia do not observeDST. Table 1 provides summary statistics and geographical information for the capitals of these states, where thepopulations and electricity demand are concentrated.3

In 2000, NSW and VIC started DST 2 months earlier than usual—on 27 August instead of 29 October—while SAmaintained the usual DST schedule. The extension was designed to facilitate the Olympic Games that took place in Sydney,in the state of NSW, from 15 September to 1 October.4 Specific rationales for the extension included easing visitormovements from afternoon to evening events, and reducing shadows on playing fields during the late afternoon [16]. Noneof the justifications for the extension were related to curbing energy use.

In the analysis that follows, we define the treatment period to be 27 August to 27 October 2000, exclusive of theOlympic period from 15 September to 1 October. While we discuss our rationale for excluding the Olympics in Section 4.1,we note here that we exclude 28 October because, in the control year of 2001, this date marks the beginning of theregularly scheduled DST period in both VIC and SA. For ease of exposition, we will also use the term treatment dates

to refer to 27 August to 27 October, exclusive of 15 September to 1 October, in any year, including the controlyears.

3. The Australian data and graphical results

3.1. Data

Our study uses detailed electricity consumption and wholesale price panel data, obtained from Australia’sNational Electricity Market Management Company Limited (NEMMCO).5 These consist of half-hourly electricitydemand and wholesale prices by state from 13 December 1998 to 31 December 2005. Wholesale prices aremarket prices paid by utilities to generators, while end-use customers instead pay a regulated price for electricityand are not exposed to fluctuations in wholesale prices. Therefore, these prices do not affect electricityconsumption.

Because electricity demand is heavily influenced by local weather conditions, we use two data sets from the Bureau ofMeteorology at the Australian National Climate Centre. The first consists of hourly weather station observations in Sydney,Melbourne, and Adelaide—the three cities that primarily drive electricity demand in each state of interest. The data cover 1January, 1999 to 31 December, 2005 and include temperature, wind speed, air pressure, humidity, and precipitation. Thesecond data set consists of daily weather observations, including the total number of hours during which the sun shines,unobstructed by clouds, each day.

Table 2 provides summary statistics for each of these variables during the treatment dates for 1999–2001, and alsoreports the frequency of school vacations and holidays. Additional details regarding the data set as well as our procedurefor dealing with missing observations are provided in Appendix A at the online archive.

3.2. The impact of the DST extension on electricity consumption and prices

The goal of the empirical analysis is to examine the effect of the extension of DST on electricity use and prices. Prior to adiscussion of the econometric model, much can be learned from the graphical analysis presented in Fig. 1. Panel (a) displaysthe average half-hourly electricity demand in SA during the treatment dates in 1999, 2000, and 2001. The load shape in SA,

3 A figure displaying the relevant geographic area is contained in an appendix that is available through JEEM’s online archive of supplementary

material, which can be accessed at http://www.aere.org/journal/index.html.4 The decision to start DST 3 weeks prior to the beginning of the Olympic Games was intended to avoid confusion for athletes, officials, media, and

other visitors who would likely arrive prior to the opening of the Games. VIC adopted the NSW timing proposal to avoid inconveniences for those living

near the NSW–VIC border. However, SA did not extend DST in 2000 due to the opposition of the rural population [16,17].5 NEMMCO data can be obtained at http://www.nemmco.com.au/data/market_data.htm.

http://www.aere.org/journal/index.html

http://www.nemmco.com.au/data/market_data.htm

ARTICLE IN PRESS

Table 2Summary statistics: 1999–2001, treatment dates only

State Variable Unit 2160 observations per state, per year

1999 2000 2001

Mean Std. dev. Mean Std. dev. Mean Std. dev.

Victoria Demand MW 5131.86 528.87 5347.71 554.17 5405.90 553.66

Price AUD/MWh 19.22 6.34 43.30 179.29 29.11 84.04

Temperature 1C 13.51 4.46 12.12 3.90 12.11 3.73

Precipitation mm/h 0.07 0.50 0.14 0.73 0.05 0.24

Wind m/s 4.72 2.94 5.57 2.92 4.69 2.61

Pressure hPa 1018.18 6.30 1011.97 7.17 1012.09 6.17

Sunshine h/day 6.78 3.89 5.90 3.71 5.78 3.43

Humidity % 71.02 17.18 72.51 15.83 72.58 17.11

Employment in 1000 2192.72 14.14 2272.06 12.05 2289.02 11.46

Non-working day % of days 0.31 0.46 0.24 0.43 0.33 0.47

School-vacation % of days 0.00 0.00 0.00 0.00 0.09 0.28

South Australia Holiday % of days 0.00 0.00 0.00 0.00 0.00 0.00

Demand MW 1324.23 185.70 1398.49 201.43 1428.66 197.66

Price AUD/MWh 54.12 166.53 56.27 178.97 27.50 17.85

Temperature Celsius 15.95 4.81 14.41 3.69 13.48 3.20

Precipitation mm/h 0.00 0.00 0.12 0.54 0.12 0.48

Wind m/s 4.22 2.53 5.05 2.87 4.73 2.88

Pressure hPa 1017.93 6.53 1014.51 6.89 1013.79 6.32

Sunshine h/day 8.53 3.12 7.20 3.54 6.38 3.31

Humidity % 62.99 19.20 68.52 17.76 70.46 16.93

Employment in 1000 668.76 2.69 684.22 2.43 682.85 2.33

Non-working day % of days 0.42 0.49 0.27 0.44 0.44 0.50

School-vacation % of days 0.11 0.31 0.00 0.00 0.20 0.40

Holiday % of days 0.02 0.15 0.02 0.15 0.00 0.00

Abbreviations: MW ¼ megawatts; AUD/MWh ¼ Australian dollars per megawatt-hour; mm ¼ millimeters; hPa ¼ hectopascal. Note that the maximum

wholesale electricity price is capped at 5000 AUD/MWh from 1999–2000, and at 10,000 AUD/MWh in 2001. The cap is designed to mitigate generator

market power [15].

1000

1200

1400

1600

Con

sum

ptio

n in

Meg

awat

ts

0 2 4 6 8 10 12 14 16 18 20 22 24

19992001

4200

4600

5000

5400

5800

Con

sum

ptio

n in

Meg

awat

ts

0 2 4 6 8 10 12 14 16 18 20 22 24

19992001

Hour (Standard Time) Hour (Standard Time)

2000 2000

Fig. 1. Average half-hourly electricity demand in South Australia and Victoria during the treatment dates (a) South Australia (control) and (b) Victoria

(treated in 2000).


the control state, is very stable over these 3 years, featuring an increase in consumption between 05:00 and 10:00, a peakload between 18:00 and 21:00, and then a decrease in load until about 04:00 on the following morning.6 Notably, SA’sdemand in 2000 appears unaffected by the DST extension in its neighbors VIC and NSW.7

6 The ‘‘zigzag’’ pattern that occurs between 23:00 and 02:00 in both states is due to centralized off-peak water heating that is activated by automatic

timers, set to standard time [19].7 Hamermesh et al. [11] examine spatial coordination externalities triggered by time cues. Their results imply that SA in 2000 may have adjusted its

behavior in response to the treatment in VIC. In particular, their model predicts that people in SA would awaken earlier in the morning to benefit from

aligning their activities with their neighbors in VIC. However, the effects are small, and panel (a) of Fig. 1 does not show evidence of such a time shift.

ARTICLE IN PRESS

4500

5000

5500

6000

0

100

200

300

400

500

Pric

e in

Dol

lars

per

MW

h

0 2 6 8 10 12 14 16 18 20 22 24Hour (Standard Time)

Con

sum

ptio

n in

Meg

awat

ts

Prices1999

Demand2000

4

20002001

Fig. 2. Average half-hourly electricity prices and demand in Victoria during the treatment dates.


The 2000 load shape in VIC is quite different from the loads in 1999 and 2001, as shown in panel (b). Consistent with theprior literature, the treatment of extended DST dampens evening consumption. However, the 2000 treatment also raisesmorning demand to a peak that is even higher than the 2001 evening peak load. The intraday shift is consistent with theexpected effects of DST’s 1-h time shift: less lighting and heating are required in the evening, and more in the morning. Inparticular, the large increase in demand from 07:00 to 08:00 closely matches environmental variables at this time of theday. During the treatment period, the latest sunrise in Melbourne (on 27 August) occurs at 07:51, and the average sunriseoccurs at 06:55. Further, the 07:00–08:00 interval is the coldest hour of the day; the average temperature for this hour isonly 9 1C. Therefore, when people shift their activities forward with the clock in response to the imposition of DST, theyawaken in cold, low-light conditions, driving an increase in electricity demand that persists even 1 h after sunrise.Extending DST only conserves energy if this morning increase in consumption is outweighed by the evening decrease;however, in Fig. 1 it is not clear that this is the case.

Panel (b) of Fig. 1 also casts doubt on claims that extended DST brings additional benefits, in the form of higher systemreliability and lower prices, due to a more balanced load shape. While the extension does reduce the evening peak load inVIC in 2000, it creates a new, sharp peak in the morning that is even higher than the evening peak in 2001. This morningpeak is also coincident with a large spike in wholesale electricity prices, as shown in Fig. 2. Morning price spikes occurredon every working day during the first 2 weeks of the extension, suggesting that the generation system was initially stressedto cope with the steep ramp in demand.8

Furthermore, this analysis suggests that, contrary to common claims, the extension of DST is not likely to reduce GHGemissions from electricity generation. The quantity of these emissions is dictated primarily by the quantity of energygenerated: without a reduction in electricity consumption, significant reductions in GHG emissions are unlikely. Hollandand Mansur [12] note, however, that changes in the variance of the load shape can affect production of GHGs, even holdingtotal electricity production constant. If the generation units that serve peak load are dirtier (cleaner) than the baseloadunits, then an increase in variance can slightly increase (decrease) emissions. However, because the extension of DSTappears to merely shift the peak load from the evening to the morning, rather than smooth the daily load, such effects seemunlikely to be significant here.

Our graphical analysis does not, of course, account for important determinants of electricity demand, such as weatherand holidays. To obtain an unconfounded estimate of the effect of extended DST on electricity use, we employ a formaleconometric analysis, which we now describe in detail.

4. Empirical strategy for measuring the effect of DST on electricity use

4.1. Identification

While we have noted that the DST extension was implemented solely to facilitate the Olympic Games, and that we arenot aware of any energy-based justifications for it, identification of the extension’s effect on energy use is made difficult by

8 Because the Australian electricity market is integrated across state boundaries, demand shocks in VIC caused by extended DST affect not only

wholesale prices in VIC, but also prices in SA. We therefore do not undertake a formal analysis of extended DST’s effect on prices, because a control state

does not exist.

ARTICLE IN PRESS

1

1.1

1.2

1.3

1.4

1.5

1.6

1.7

0Hour (Standard Time)

1999200020012002

200320042005

Log(

VIC

dem

and

/ SA

dem

and)

22202 4 6 8 10 12 14 16 18

Fig. 3. Ratio between average VIC demand and average SA demand during the treatment dates, 1999–2005.


the presence of potentially confounding factors. In particular, there are reasons to suspect that the Olympics may havechanged electricity consumption in Australia significantly, even absent a DST extension. The 2000 Games were the mostheavily visited Olympics event in history, school vacations were rescheduled to facilitate participation in carnival events,and the Games were watched on public mega-screens and private televisions by millions of Australians in Sydney andelsewhere.

Our identification strategy incorporates several features designed to account for these potential confounds, and benefitsfrom observations during the treatment year and the control years in both the treated and the non-treated state, as well asfrom the detailed half-hourly frequency of our data. First, we exclude the 17 days of the Olympic Games from the definitionof the treatment period; this allows us to avoid many of the biases noted above. Second, even with the Olympics excludedfrom the treatment, electricity demand may have been affected before and after the games by, for example, pre-Olympicconstruction activities and extended tourism. To control for these, we ignore NSW (where the Olympics took place), andfocus on the change in electricity demand in VIC relative to that in SA.9 This strategy eliminates the impact of anyconfounders that operate on a national level, and accounts for all differences between the two states that are constant overtime.

Further, to control for unobservables that may have affected VIC and SA differentially over time, we use relative demandin the mid-day as an additional control. That is, because DST does not affect demand in the middle of the day, variations instate-specific mid-day demand levels that are not explained by observables such as weather can be attributed to non-DST-related confounders. Thus, our model is robust against transient state-specific shifts in demand that affect the overall levelof consumption in any state on any day, but do not affect the shape of the half-hourly load pattern. We verify theassumption that DST does not affect mid-day demand by examining changes from standard time to DST in non-treatmentyears. We discuss this verification, as well our choice of 12:00–14:30 as mid-day, in Appendix B at the online archive.

These features of our model imply that a mild identifying assumption is sufficient for our regressions to produce anunbiased estimate of the extension’s effect. We assume that, conditional on the observables and in the absence of thetreatment, the ratio of VIC demand to SA demand in 2000 would have exhibited the same half-hourly pattern (but notnecessarily the same level) as observed in other years. Support for this is found by plotting the ratio of consumption in VICto that in SA for 1999–2005, as shown in Fig. 3. The demand ratio exhibits a regular intraday pattern in all non-treatedyears, even without controlling for observables. Moreover, the levels of these curves change non-systematically, fromsmallest to largest, over 2002, 2000, 2001, 1999, 2004, 2003, and 2005. These level shifts are consistent with the existenceof transient state-specific shifts in consumption that must be controlled for using demand in the mid-day.

As an alternative strategy to control for unobservables that affect each state differently in different years, we alsoconsidered taking advantage of demand data for the months adjacent to the treatment dates: August and November. Thatis, we considered using August and November each year to control for non-DST-related state-specific shocks to demand

9 To further analyze whether visitors before and after the Olympic Games spent extended vacations in VIC or SA, we collected tourism information.

These data show that, while NSW was affected by tourism during the Olympics, VIC and SA were unaffected. Additional details are provided in Appendix D

at the online archive.

ARTICLE IN PRESS

1.2

1.25

1.3

1.35

1.4

1.45

1999

log(

VIC

Dem

and)

- lo

g (S

A D

eman

d) 01 Aug - 26 AugTreatment dates28 Oct - 30 Nov

2000 2001 2002 2003 2004 2005

Fig. 4. Ratio between average VIC demand and average SA demand, August–November, 1999–2005.


during the treatment dates. However, this strategy is valid only if the state-specific demand shocks are persistent overseveral months—if a shock causes VIC’s demand to be relatively large in 2001 during the treatment dates, then the shockmust also cause VIC’s demand to be relatively large in August and November.

Fig. 4 instead demonstrates that state-specific demand shocks vary unpredictably across months and years. For example,in 2001, the ratio of VIC demand to SA demand does not vary over August–November. However, in 1999 the ratio is largerduring the treatment dates than it is in August or November, and in 2002 the ratio decreases monotonically from August toNovember. This lack of stability implies that the data cannot support an identification strategy that relies on observationsfrom months adjacent to the treatment period. Indeed, when we estimate a model based on this strategy we findstatistically significant treatment effects that are implausibly large—1–2% increases in demand during the mid-day (andoverall).10 Given that both intuition and evidence instead indicate that DST does not affect mid-day demand, we eschew the‘‘adjacent months’’ strategy in favor of the ‘‘within-day’’ strategy that uses mid-day demand to control for state-specificshocks.

4.2. Difference-in-difference-in-difference (DDD) estimation

We implement our identification strategy using a DDD framework [10]. This technique is illustrated in Table 3, whichdisplays the raw DDD estimate of the DST extension’s impact on electricity consumption. Each cell contains the meanlogarithm of electricity consumption per half-hour for the indicated state, year, and hours, as well as the standard error andnumber of observations. The top panel A concerns consumption during the treated hours; that is, all 24 h of the dayexcluding the mid-day hours of 12:00–14:30. This panel shows that there was approximately a 2.0% increase in electricityconsumption in VIC, the treated state, during the extension, compared to a 2.4% increase in the control state SA.These values imply that, in a DD estimate, the extension of DST decreased electricity consumption by a statisticallyinsignificant 0.4%.

This DD estimate, however, does not identify the impact of the extension if there were state-specific demand shifts inVIC and SA during these time periods. To examine this possibility, we repeat the above exercise in panel B of Table 3, usingthe within-day control period of 12:00–14:30. Our DD estimate during this control period also indicates a slight decrease inconsumption, with a magnitude of 0.2%. Though statistically insignificant, this estimate is of the same order of magnitudeas that obtained during the treated hours, suggesting that it is important to control for state-specific demand shifts in ouranalysis.

We obtain the DDD estimate of the effect of the extension by taking the difference between the DD estimatesin the two panels of Table 3. We find that electricity consumption in the treated state during the treated year andtreated hours decreased by 0.2%. While this result suggests that the impact of extending DST is of small magnitude,the estimate is imprecise, with a standard error of 1.5%. To improve our precision, we employ a regressionframework, below, in which we may control for other factors that affect electricity consumption, such as day-of-weekand weather.

10 The specification is the same as Eq. (1), as described in Section 4.3, except that treatment dummies are included for all 24 h of the day, and the id are

replaced with fixed effects for the interaction of state with year.

ARTICLE IN PRESS

Table 3DDD estimate of extended DST’s impact on the log of electricity consumption

Location/year Control period in 1999 and 2001 Treatment period in 2000 Time difference for location

(A) Treated hours 00:00–11:59 and 14:30–24:00

Victoria 8.554 8.574 0.020

(Treated state) (0.001) (0.002) (0.003)

[5332] [1935]

South Australia 7.202 7.226 0.024

(Control state) (0.002) (0.003) (0.004)

[5332] [1935]

Location difference at point in time: 1.352 1.348

(0.003) (0.004)

Difference-in-difference �0.004

(0.005)

(B) Control hours: 12:00–14:30

Victoria 8.603 8.624 0.021

(Treated state) (0.004) (0.006) (0.007)

[620] [225]

South Australia 7.263 7.286 0.023

(Control state) (0.005) (0.008) (0.009)

[620] [225]

Location difference at a point in time: 1.340 1.338

(0.006) (0.010)

Difference-in-difference �0.002

(0.012)

DDD �0.002

(0.015)

Note: Cells display the mean logarithm of electricity consumption per half-hour for the indicated state, year, and hours. Standard errors are provided in

parentheses and the sample sizes in squared brackets. Data apply only to the dates of 27 August to 27 October in each year. The difference-in-difference-

in-difference (DDD) estimate is the subtraction of the DD estimate of panel B from that of panel A.


4.3. Treatment effect model

Our specification of the treatment effect model is drawn primarily from the DD literature [2,14]. We augment thestandard DD model by estimating a DDD specification [10] because our control structure is three-fold:

(a)
cross-sectional overstates (with VIC as the treated state and SA as the control), (b) temporal over years (with the untreated years in SA and VIC as controls), and (c) temporal within days (with the mid-day hours as ‘‘within-day’’ controls)
The reference case model uses data from VIC and SA during 27 August to 27 October in 1999, 2000, and 2001; these datescorrespond to the period when DST was observed in VIC in 2000 and Standard Time was observed in 1999 and 2001. Ourspecification is given by Eq. (1), in which states are subscripted with i, and time is jointly subscripted with d (for date) and h

(for hour):

lnðqidhÞ ¼ Tidhbh þ did þ Xidhah þWidhjh þ �idh. (1)

The dependent variable qidh for each observation is the logarithm of electricity demand in state i, day d, and half-hour h

(in clock time). We use the log of demand rather than its level to account for the large difference in size between SA andVIC. As indicated in Table 2, the average electricity demand in VIC is approximately four times that in SA. Given thisdifference, a linear model will not be robust to proportional shifts in demand that are common to both states.

The covariates of primary interest are the indicator variables Tidh for the treatment period. These are equal to one in VICduring the treatment period in 2000 for all half-hours except those in the mid-day, and are zero otherwise. The variables did

are fixed effects for the interactions between each date in the sample period and indicator variables for each state. Thesevariables effectively force the ‘‘omitted’’ treatment effect in the mid-day to be zero for every day of the treatment period.The dummy variables Xidh include 48 half-hour dummies, and interactions of these dummies with indicator variables forthe following: state, year, day of week, holidays, school vacations, the interaction of state with week of year, and theinteraction of state with a flag for the Olympic period. The weather variables Widh are also interacted with half-hour

ARTICLE IN PRESS


dummies and include a quadratic in hourly heating degrees, daily hours of sunlight, the interaction of sunlight withtemperature, hourly precipitation, the interaction of precipitation with temperature, and the average of the mid-dayheating degrees.11,12 All weather variables enter the model lagged by 1 h.

In Eq. (1) the treatment effect parameters to be estimated are given by bh and the percentage change in electricitydemand in each half-hour h is given by exp(bh)�1 (parameters for the mid-day half-hours are omitted). To obtain theoverall effect of the DST extension, we use two separate methods. First, we estimate a version of Eq. (1) in which we poolthe treatment effect overall half-hours of the day. That is, instead of including separate treatment effect dummies for eachhalf-hour, we only specify a single treatment dummy that is equal to one for all non-mid-day hours in VIC in 2000.

In the second method, we first estimate the half-hourly coefficients bh (again excluding the mid-day hours) and thenaggregate them to obtain the overall percentage change in demand caused by the DST extension. Denoting the vector ofhalf-hourly treatment coefficients as b, the overall effect y is given by (2)

y ¼ f ðbÞ ¼

P48h¼1 expðbhÞohP48

h¼1oh

� 1. (2)

That is, y is the weighted sum of the half-hourly percentage effects, where the weights oh are the average of the baseline1999 and 2001 half-hourly demands during the treatment dates.

Given a set of half-hourly treatment effect estimates b, the point estimate of y is given by f ðbÞ and its estimated varianceis given by the delta method, per equation:

VðyÞ ¼ rbf ðbÞTCovðbÞrbf ðbÞ. (3)

In (3), rbf ðbÞ denotes the gradient of f ðbÞ with respect to each element bh, and CovðbÞ denotes the estimated covariancematrix of b. Our estimation of CovðbÞ allows the disturbance eidh to be both heteroskedastic and correlated within each day,

Eð�idh�idhjZÞ ¼ s2idh; Eð�dj�dkjZÞ ¼ rdj8jak,

Eð�d�Td0 jZÞ ¼ 0 8dad0,

where Z ¼ [T,d,X,W]. This block-diagonal covariance structure accounts for both autocorrelation and common shocks thataffect both states contemporaneously. We therefore use the clustered sample estimator to obtain the covariance matrix of b

[1,2,29]. As an alternative, we also estimate the model using the Newey and West [18] estimator with 50 lags.13

5. Results

5.1. Reference case results

We use two separate methods to obtain the overall effect of the DST extension. First, we estimate a version of (1) inwhich the treatment variables are pooled overall half-hours of the day. These results indicate that the extension of DST didnot significantly affect overall electricity consumption in VIC in 2000. Our point estimate indicates a 0.02% increase inconsumption due to the extension, with a clustered standard error of 0.43.

Second, we estimate a separate treatment effect for each half-hour. These results are displayed graphically in Fig. 5; atabular version is presented in Appendix C at the online archive. Extending DST affects electricity consumption in a mannerconsistent with the preliminary graphical analysis: there is a substantial, statistically significant transfer of consumptionfrom the evening to the morning. This behavior agrees with the expected effects of DST’s 1-h time shift. Less lighting andheating are required in the evening, from 17:00 to 19:30; however, demand increases in the morning—particularly from07:00 to 08:00—driven by reduced sunlight and lower temperatures.

We aggregate the half-hourly estimates using (2) to yield an estimate of y, the overall effect of the extension. Consistentwith the result of the pooled model, we find that the extension of DST did not conserve electricity, as shown in the firstcolumn of Table 4. The point estimate of the percentage change in demand overall hours of the treatment period is +0.09%with a clustered standard error of 0.40. This result is very similar to that of the pooled model, above. Because the unpooledspecification (1) yields a better fit to the data than the pooled specification (pooling of the half-hourly treatmentcoefficients is strongly rejected), the remainder of the paper will focus on estimates using the unpooled model, withaggregation per Eq. (2).

11 Our final specification pools some hours to improve efficiency of the weather models. This does not impact the reported estimates of the treatment

effects.12 Heating degrees are calculated as the difference between the observed temperature and 18.33 1C (65oF). The motivation behind squaring the

heating degree is that, as the temperature deviates from 18.33 1C, cooling or heating efforts increase nonlinearly. This functional form is consistent with

other electricity demand models [3].13 Fifty lags allow the errors to be correlated over slightly more than 1 full day. Tests of AR(p) models on e suggest that the disturbances are correlated

over the first 6 h of lags, but not beyond that. However, the coefficient on the 48th lag is significant. Also, note that the DDD specification considerably

decreases the autocorrelation of the dependent variable, relative to a standard DD.

ARTICLE IN PRESS

Table 4Summary of estimated treatment effects, aggregated over all half-hours

All days September October Working days Non-working days

Percent change in demand 0.09 0.39 �0.06 0.43 �0.82

Standard error (0.40) (0.43) (0.47) (0.44) (0.43)

[0.38] [0.43] [0.43] [0.41] [0.50]

Standard errors clustered on date are in parentheses and Newey-West standard errors are in brackets.

Table 5p-Values for rejection of electricity saving hypotheses

Null hypothesis September estimate (+0.39%) October estimate (�0.06%) Pooled estimate (+0.09%)

Clustered std. error Newey-West Clustered std. error Newey-West Clustered std. error Newey-West

Electricity savings

y ¼ �1.0% 0.001 0.001 0.048 0.047 0.007 0.004

y ¼ �0.6% 0.023 0.019 0.257 0.290 0.085 0.069

Electricity neutrality

y ¼ 0.0% 0.371 0.344 0.895 0.738 0.814 0.804

−10.0

−8.0

−6.0

−4.0

−2.0

0.0

2.0

4.0

6.0

8.0

10.0

6Hour (Clock Time)

Per

cent

Cha

nge

in C

onsu

mpt

ion

7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

Fig. 5. Half-hourly effects of extending DST on electricity use. 95% confidence intervals are indicated, with standard errors clustered by day. Effects in

mid-day half-hours are zero by assumption.


5.2. September vs. October

We also examine the impact of the DST extension separately for September and October. Because September in thesouthern hemisphere is seasonally equivalent to March in the northern hemisphere, this examination has policyimplications beyond Australia—the recent change to DST in the United States concerns an extension into March, as DST isalready observed in April in the US Prior studies have found that such an extension reduces electricity consumption by 1%in the US and by 0.6% in California. In contrast, we estimate that the extension of DST into September in Australia increased

electricity demand by 0.39%, as shown in Table 4.14

14 More precisely, we estimate the effects of DST separately for the pre-Olympic and post-Olympic treatment periods, which we refer to loosely as the

months of September and October. That is, for each half-hour h, we interact the treatment dummy Tidh with indicators for each month and estimate

separate coefficients bSeph and bOct

h . We then aggregate these half-hourly estimates to yield estimates of the overall treatment effects Sep and Oct. The point

estimate of Oct indicates that the extension reduces electricity demand by 0.06%. While the difference between the September and October estimates is

ARTICLE IN PRESS

Table 6Robustness tests: estimated overall effect of the DST extension in VIC

Reference

case

CEC

weather

Include data to

2005

Exclude week before/

after Olympics

Queensland as

control state

Run in Standard

Time

Percent change in demand 0.09 0.08 0.12 0.42 �0.08 0.51

Standard error (0.40) (0.39) (0.35) (0.50) (0.28) (0.41)

Standard errors are clustered on date.


To formally compare our estimates to the previous literature, we define three null hypotheses: (1) y ¼ �1.0%, (2)y ¼ �0.6%, and (3) y ¼ 0.0%, and test whether they are rejected by our estimates. Table 5 displays p-values for rejection ofeach null hypothesis in a two-sided test, given both our pooled and unpooled results. Even with clustered standard errors,our estimate of the effect of the DST extension in September rejects the most modest energy savings estimate in theliterature of 0.6% [4] at a 5% level. Over the entire treatment period (September and October), we reject a 1% reduction indemand at a 1% level, and reject a 0.6% reduction at a 10% level. These rejections are strengthened with the use of Newey-West standard errors.

In summary, the results indicate that extending DST did not significantly reduce electricity demand in VIC. In Septemberin particular, the extension was more likely to have increased than decreased electricity consumption.

5.3. Robustness

Our results are robust to many alternative specifications, as shown in Table 6. Our results are invariant to the choicebetween the two alternative weather models in [3,4]. Further, our results do not change appreciably if we include morerecent data, use Queensland as a control state rather than SA, exclude the weeks immediately preceding and following theOlympics from the treatment period definition, or estimate (1) in Standard Time rather than clock time. This robustness isunderlined by the precise fit of our model: the adjusted R2 across all models is greater than 0.99.

Regression equation (1) contains over 1800 parameters. While the point estimates and the standard errors for thetreatment effects are our primary interest, most of the other coefficients are significant and carry signs that agree withintuition. For example, weekends, holidays, and vacations lower electricity consumption, and deviations from the basetemperature of 18 1C increase electricity consumption, consistent with the effects of air conditioning (when above 18 1C)and heating (when below 18 1C).

The weights oh used to calculate y are based on the average of the 1999 and 2001 half-hourly demands. As analternative set of weights, we also use the estimated half-hourly counterfactual demand in 2000, given byexp{dVICd+XVICdhaidh+WVICdhfih}. Doing so does not affect our estimate of y.

As a final check of our estimates, we evaluate whether extending DST causes a relatively greater reduction in electricityconsumption on weekends and holidays than on working days. This would be consistent with the intuition that, on non-working days, less early activity mitigates the morning increase in demand. We estimate that electricity consumption onworking days increased by 0.43% during the extension, while consumption on weekends and holidays decreased by 0.82%.This difference is significant at the 1% level.

6. Evaluation of the simulation technique

It is natural to ask whether the simulation technique used in [4] to predict energy savings in California would haveaccurately predicted the outcome of the Australian DST extension. A successful validation would lend credence to themodel’s results in California, and suggest that California may experience reduced energy use due to an extension, even ifAustralia did not.

The simulation approach uses data on hourly electricity consumption under the status quo DST policy to investigate theimpact of a DST extension. This procedure first employs a regression analysis using status quo data to assess how electricitydemand in each hour is affected by weather and light, and then uses the regression coefficients to predict demand in theevent of a 1-h time shift, lagging the weather and light variables appropriately. The consistency of the simulation resultsrelies on the assumption that extending DST will not cause patterns of activity that are not observed in the status quo,which may not hold in practice. For example, to simulate demand under extended DST at 07:00 in March in the US, themodel must rely on observed status quo behavior at 07:00 under similarly cold and low-light conditions. Without a DST

(footnote continued)

significant at only the 30% level, the sign of the difference is intuitive: in October there is more morning sunlight and temperatures are warmer, so the

morning increase in demand is mitigated.

ARTICLE IN PRESS

Fig. 6. Actual and simulated electricity consumption in VIC, September 1999–2001.


extension, these conditions are observed only in mid-winter. The simulation will be inaccurate if people behave differentlyin the morning in mid-winter than they do in spring under extended DST.15

In contrast, in the Australian quasi-experiment, we have already estimated the effect of the DST extension directly, bycomparing observations under both the status quo and the extension. We can therefore evaluate the simulation’sperformance by re-estimating its first stage using status quo observations, forecasting electricity demand under anextension, and then comparing these results to those estimated from actual data.

The first stage of the simulation model is a regression of hourly electricity demand, qdh, on employment, weather, andastronomical sunlight and twilight variables, for a full year of observations:

qdh ¼ ah þ bh Employmentd þ ch Weatherdh þ dh Lightdh þ udh.

The disturbance ud is correlated across the h ¼ 1,y,24 hourly equations per the Seemingly Unrelated Regression method[30]. The regression allows the weather and light coefficients to vary across the 24 h of the day, and the weatherspecifications are very detailed, involving several lags and moving averages of half-hourly temperatures, with differentcoefficients for hot, warm, and cold conditions.16 Once the vectors of regression coefficients are estimated, they are used inthe second stage of the model to forecast electricity consumption under a DST extension. This is accomplished by laggingthe weather and light variables by 1 h and by adding the first stage realized error term to construct the followingprojection:

qsimdh ¼ ah þ bh Employmentd þ ch Weatherdh�1 þ dh Lightdh�1 þ udh.

We apply the first stage of the CEC model to the Australian data for all of 1999 and 2001, and then simulate electricityconsumption under extended DST in VIC in September 1999 and 2001 (we are unable to simulate demand under anextension in 2000 using the CEC’s method because we do not observe demand under Standard Time in that year). Fig. 6illustrates the simulated demand, as well as actual demand (under standard time), in both years. The simulations predict asubstantial decrease in demand in the evening and only a minor increase in demand in the morning, with overall energysavings of 0.43% in 1999 and 0.41% in 2001. Both the hour-by-hour and overall results closely align with the 0.6% savingspredicted for California in the original study (see Fig. 7). The results disagree, however, with the actual outcome of theAustralian DST extension in 2000. Fig. 6 also includes, in bold, the realized demand in VIC under the 2000 treatment. Inboth 1999 and 2001, the simulation fails to predict a morning increase in electricity consumption similar to that observedin 2000, and also overestimates evening savings. The simulated decrease in overall consumption is inconsistent with whatactually happened in VIC. Based upon our DDD estimate of a 0.39% increase in consumption in September presented earlier,

15 It may be the case that, as winter turns into spring, people move their waking time earlier with the sun, and do not adjust their behavior after the

sudden time shift imposed by extended DST. Thus, there could be more early morning activity under extended DST in the spring than under Standard Time

in the winter, resulting in higher electricity use.16 Details of the definition on these variables, the estimation of the model, and the simulation are explained in CEC [4]. We make minor changes to the

CEC specification to account for our half-hourly, rather than hourly data, and for the fact that we observe humidity, precipitation, and daily unobstructed

sunshine, but not hourly cloud cover. Computer code is available in Appendix E at the online archive.

ARTICLE IN PRESS

Status CuoDST

Clock Time

MW

45000

40000

35000

30000

25000

20000

15000

2:00

3:00

4:00

5:00

6:00

7:00

9:00

8:00

10:00

11:00 1:0

02:0

03:0

04:0

05:0

06:0

07:0

08:0

09:0

010

:0011

:00no

on1:00

Fig. 7. Simulation of DST in California, March 1998–2000 (CEC [4]). Status quo demand is observed under Standard Time.


we reject the simulated 0.41% savings at a 10% significance level. The simulation is unable to predict the substantialintraday shifts that occur due to the early adoption of DST, a result that holds even after we attempt to improve the model’sfit by introducing higher order terms for the continuous variables or by selecting a smaller first-stage sample in which lightand weather conditions most closely resemble the extension period in September.

7. Conclusions

Given the economic and environmental imperatives driving efforts to reduce energy consumption, policy-makers inseveral countries are considering extending daylight saving time (DST), as doing so is widely believed to reduce electricityuse. Our research challenges this belief, as well as the studies underlying it. We offer a new test of whether extending DSTdecreases energy consumption by evaluating an extension that occurred in the state of Victoria, Australia, in 2000. Usinghalf-hourly panel data on electricity consumption and a difference-in-difference-in-difference treatment effect model, weshow that while extending DST did reduce electricity consumption in the evening, these savings were negated by increaseddemand in the morning. These effects are consistent with the alignment of daily activities to the clock rather than the sun,so that the 1-h shift associated with DST caused people to awaken in darkness.

Further, the evidence does not support the existence of two additional DST extension benefits that have been discussedin the prior literature: a reduction in electricity prices and a reduction in the likelihood of blackouts, driven by a morebalanced hourly load shape. The data instead show that the Australian DST extension substantially increased wholesaleprices and caused a sharp peak load in the morning.

From an applied policy perspective, this study is of immediate interest for Australia, which is actively considering usingDST as a tool for energy conservation and greenhouse gas (GHG) emission abatement. Moreover, the lessons from Australiamay carry over to the United States and to California in particular, as Victoria’s latitude and climate are similar to those ofcentral California. The 2007 DST extension in the United States causes DST to be observed in March—a month that isanalogous to September in Australia, when our results suggest that DST increases rather than decreases overall electricityconsumption. Moreover, there is little reason to believe that the lack of electricity demand reduction associated withAustralia’s 2000 extension was driven by the extension’s temporary nature. Because the demand response appears to becaused by behavioral changes rather than long-term investments, significant differences between short- and long-termeffects are unlikely. Finally, we find that the simulation model that supported a DST extension in California overestimatesenergy savings when we apply it to Australia. This casts suspicion on its previous policy applications, and provides furtherevidence that the US extension is unlikely to achieve its energy conservation goals.

Given the evidence presented in this paper, why do policy-makers believe that extending DST will conserve energy?While reliance on prior literature, particularly the 1975 US DOT study [27], may play a role in this belief, a deeper reasonmay come from attempts to extrapolate from DST’s effects in the summer to its effects in the spring and autumn. Thereexists a reasonable intuition for why DST may reduce electricity use in the summer: the days are sufficiently long that DSTshould not cause people to awaken in the dark, and the ‘‘extra’’ sunlight in the evening will reduce energy use. In the springand fall, however, the days are shorter and this intuition is no longer valid: DST causes the sun to rise after 07:00 and peopleawaken in the dark. A failure to recognize this distinction between the seasons may therefore result in a belief thatextending DST will conserve energy.

ARTICLE IN PRESS


As for the future status of DST, the 2005 US energy bill prescribes that the recent extension be repealed should studiesdemonstrate that it does not reduce energy consumption. However, extending DST may have non-energy impacts that areeconomically important and, given the lack of a significant effect on electricity use, be worthy of greater consideration inthe setting of DST policy. For example, recent papers have found that DST significantly affects stockmarket trading andtraffic accidents [13,23]. In addition, extending DST may directly affect welfare by altering the opportunities for outdoorrecreation. A deeper understanding of these impacts would be both economically interesting and of value to policy-makersin assessing the overall merit of extending DST.

Acknowledgments

We are indebted to Michael Anderson, Maximilian Auffhammer, Severin Borenstein, Jennifer Brown, Kenneth Chay,Michael Hanemann, Ann Harrison, Guido Imbens, Enrico Moretti, Jeffrey Perloff, Susan Stratton, Muzhe Yang, DavidZilberman, and two anonymous referees for valuable discussions and suggestions. We further thank Alison McDonald fromNEMMCO and Lesley Rowland from the Australian Bureau of Meteorology for helping us understand the electricity andweather data, and Adrienne Kandel from the California Energy Commission (CEC) for conversations regarding the details ofthe CEC simulation model.

References

[1] M. Arellano, Computing robust standard errors for within-groups estimators, Oxford Bull. Econ. Stat. 49 (1987) 431–434.[2] M. Bertrand, E. Duflo, S. Mullainathan, How much should we trust differences-in-differences estimates?, Quart. J. Econ. 119 (2004) 249–275.[3] J. Bushnell, E. Mansur, Consumption under noisy price signals: a study of electricity retail rate deregulation in San Diego, J. Ind. Econ. 53 (2005)

493–513.[4] California Energy Commission (CEC), Effects of daylight saving time on California electricity use, Report P400-01-013, 2001.[5] California Joint Senate Resolution, Bill No. SJRX2-1, 25 June 2001.[6] M. Downing, Spring Forward: The Annual Madness of Daylight Saving Time, Shoemaker & Hoard, Washington, DC, 2005.[7] Emergency Daylight Savings Time Energy Conservation Act, US Public Law 93-182, December 1973.[8] Energy Policy Act of 2005, US Public Law 109-58, July 2005.[9] B. Franklin, An economical project, essay on daylight saving, J. Paris (April 1784).

[10] J. Gruber, The incidence of mandated maternity benefits, Am. Econ. Rev. 84 (1994) 622–641.[11] D.S. Hamermesh, C.K. Myers, M.L. Pocock, Time zones as cues for coordination: latitude, longitude and letterman, NBER Working Paper No. 12350,

2006.[12] S.P. Holland, E. Mansur, Is real-time pricing green?: The environmental impacts of electricity demand variance, Rev. Econ. Stat. 90 (2008) 550–561.[13] M.J. Kamstra, L.A. Kramer, M.D. Levi, Losing sleep at the market: the daylight saving anomaly, Am. Econ. Rev. 90 (2000) 1005–1011.[14] B.D. Meyer, Natural and quasi-experiments in economics, J. Bus. Econ. Stat. 13 (1995) 151–161.[15] National Electricity Market Management Ltd. (NEMMCO), An Introduction to Australia’s National Electricity Market, 2005.[16] New South Wales Legislative Assembly Hansard, Standard Time Amendment Bill, Second Reading (26 May 1999, article 40, and 2 June 1999,

article 9).[17] New South Wales Legislative Assembly Hansard, Standard Time Amendment (Daylight Saving) Bill, Second Reading (13 September 2005, article 44).[18] W.K. Newey, K.D. West, A simple, positive definite, heteroskedasticity and autocorrelation consistent covariance matrix, Econometrica 55 (1987)

703–708.[19] H. Outhred, personal correspondence, 20 July 2006.[20] D.S. Prerau, Seize the Daylight: The Curious and Contentious Story of Daylight Saving Time, Thunder’s Mouth Press, New York, 2005.[21] B.A. Rock, Impact of daylight saving time on residential energy consumption and cost, Energy and Buildings (1997) 63–68.[22] Scoop Independent News, Minister Urged to Consider Early Daylight Saving, 14 August 2001.[23] N. Sood, A. Ghosh, The short and long run effects of daylight saving time on fatal automobile crashes, B.E. J. Econ. Anal. Pol. 7 (February 2007).[24] The Energy Conservation Center, Japan, Report on the National Conference on the Global Environment and Summer Time, 2006.[25] The Toronto Star, Get set for darker November mornings, 21 July 2005.[26] United Kingdom House of Commons Hansard, Energy Saving (Daylight) Bill, 26 January 2007.[27] US Department of Transportation (US DOT), The Daylight Saving Time Study: A Report to Congress by the US Department of Transportation, GPO,

Washington, DC, 1975.[28] US Department of Energy, Energy Information Administration (US EIA), Electric Power Annual 2005, DOE/EIA-0348 (2005), revised November 2006.[29] J.M. Wooldridge, Cluster-sample methods in applied econometrics, Am. Econ. Rev. Pap. Proc. 93 (2003) 133–138.[30] A. Zellner, An efficient method of estimating seemingly unrelated regression equations and tests for aggregation bias, J. Am. Stat Assoc. 57 (1962)

348–368.

1

Appendix A: Data processing

Electricity data are missing for occasional half-hours, so we estimate the missing

observations via interpolation using adjacent half hours. Weather data are also missing for some

occasional hours as well for four entire days (none of which fall in within the treatment dates in

any year, except for the air pressure variable). While we estimate weather for isolated missing

hours via interpolation, we estimate weather for unobserved days via a regression analysis using

information from the daily-level weather dataset. In the two primary states of interest, Victoria

and South Australia, the missing values we impute do not comprise more than 0.6% of the data

for any variable. Details and code for this procedure are available from the authors upon request.

Schedules for most school vacations, state holidays, and federal holidays were obtained

from the Australian Federal Department of Employment and Workplace Relations, The

Department of Education and Children's Services (SA), and The Department of Education and

Training (VIC). For years in which information was not available from the above institutions, the

dates were obtained by internet search. Employment data were obtained from the Australian

Bureau of Statistics’ Labor Force Spreadsheets, Table 12. Sunrise, sunset, and twilight data were

sourced from the U.S. Naval Observatory, and the days and times of switches to and from DST

were obtained from the Time and Date AS Company.

While our data are provided in Standard Time, we conduct our analysis in clock time. We

therefore convert our data to clock time, which, for most affected observations, requires a simple

one-hour shift. However, at the start of a DST period, the 02:00-03:00 interval (in clock time) is

missing. To avoid a gap in our data, we duplicate the 01:30-02:00 information into the missing

02:00-02:30 half hour, and likewise equate the missing 02:30-03:00 period to our 03:00-03:30

observation. Further, when a DST period terminates, the 02:00-03:00 period (in clock time) is

observed twice. Because our model is designed for only one observation in each half-hour, we

average these observations.

2

Appendix B: Justification of using 12:00 to 14:30 as the control period

Our identification strategy uses the assumption that electricity demand in the mid-day is

not affected by DST. The purpose of this appendix is to offer regression results to justify this

assumption and to explain our choice of 12:00 to 14:30 as the mid-day control period in our

analysis.

In VIC and SA, we observe “typical” switches from standard time to DST in late October

of 1999 and 2001-2005. These observations allow us to examine DST’s effect on mid-day

electricity consumption by performing a regression discontinuity analysis of demand near the

date of each switch. Specifically, we form a regression sample consisting of half-hourly demand

observations during the week before and week after the switch to DST in each year. We then

regress, separately for each half-hour, the logarithm of demand on an indicator variable for when

DST is in effect, state-specific within-year time trends, fixed effects for the interaction of state

and year, fixed effects for day of week, fixed effects for holidays and vacations, and weather

variables.

Before discussing the estimated effect of DST in the mid-day, we first note that this

specification produces estimates which show that DST increases demand in the morning and

decreases demand in the evening. For example, during 07:00-07:30, we estimate that demand

increases by 5.9% following the switch to DST, with a standard error of 1.0% (we report

standard errors clustered on year, though these are not appreciably different from OLS standard

errors). During 18:30-19:00 we estimate a decrease in demand of 4.9%, with a standard error of

1.7%. The signs and statistical significance of these results are consistent with intuition, and

indicate that this specification has sufficient statistical power to resolve non-zero effects where

they exist.

In the mid-day, however, the effect of switching to DST is statistically insignificant.

Table B displays estimates of the percentage effect of switching to DST, along with standard

errors, for several half-hour intervals. These results are robust to the addition of another week of

3

data before and after each switch to DST, and to the addition of quadratic state-specific within-

year time trends. The point estimates are smallest in magnitude from 12:30-14:00, and increase

both before and particularly after this time period.

Selection of this 12:30-14:00 interval as the base period for estimation of the main

specification, equation (1), ultimately yields an estimate of θ of +0.3% (which is statistically

indistinguishable from zero). Expanding the base period symmetrically around 12:30-14:00

includes within it hours of the day in which our regression discontinuity estimates in table B

indicate that DST is likely to increase demand. Therefore, as we expand the base period, the

estimates of θ decrease. This is demonstrated by the fact that the reference case estimate of θ,

which uses the longer 12:00-14:30 interval as the base period, is +0.1%. We choose this interval

to be our base period, rather than 12:30-14:00, to be conservative in our final estimate.

Appendix C: Half-hourly estimation results

Table C displays the estimated percentage impact of the DST extension on electricity

demand in each half hour: these are the point estimates given by 1)ˆexp( −hβ and correspond to

figure 5. Note that the large effects in the late-night hours are caused by centralized off-peak

water heaters in Melbourne [19]. These are triggered by timers set on Standard Time—groups of

heaters are activated at 23:30 and 01:30. Each turns off on its own once its heating is complete.

During the DST extension, each heater turns on one hour “late” (according to clock time). This

drives the negative, then positive, overnight treatment effects. Regressing equation (1) in

Standard Time, rather than clock time, eliminates these overnight effects, and produces a point

estimate that the extension increased overall electricity consumption by 0.51%.

4

Appendix D: On Tourism in Australia

Figure D1 displays tourism data for VIC and SA, demonstrating that the 2000 Olympics

did not significantly impact tourism in the third and fourth quarters of 2000. Tourism data for

Sydney in NSW (Figure D2), however, shows that tourism increased in September 2000, and

that there was no such increase in 1998 or 1999 (Australian Bureau of Statistics, 2001a, 2001b).

Moreover, anecdotal evidence from Melbourne newspapers shows that Melbourne (the most

frequently touristed location in VIC) did not experience any change in tourism before, during, or

after the Olympic Games in 2000. Further details on tourism may be found in the Australian

Bureau of Statistics’ special report on Tourism related to the Olympics (2001b).

Appendix E: Stata Code for the estimation and the simulation of the CEC model * CEC Model clear cd "C:\My Documents\Berkeley\ARE\Diss\Ideas\DST\AustraliaData\Analysis\CEC-Refurbished" log using "C:\My Documents\Berkeley\ARE\Diss\Ideas\DST\AustraliaData\Analysis\CEC-Refurbished\CEClogCEC.smcl", replace set mem 800m set more off set matsize 2000 set maxvar 7000 use "C:\My Documents\Berkeley\ARE\Diss\Ideas\DST\AustraliaData\Analysis\DSTFullDataset.dta" *************************** * Fill up data holes, such that VIC1 is (and partially SA1) complete data sets ************************************ sort Province Year Month Date Day Hour drop if (Province =="SA1"| Province =="QLD1" | Province =="NSW1") *************************** * Generate Variables that are needed for the CEC - SUR ************************************ * Re-define treatment variables to exclude Olympic period gen TreatPerI=0 gen TreatPerII=0 replace TreatPerI=1 if (Province=="VIC1" | Province=="NSW1") & Year==2000 & Month ==8 & Day ==27 & Hour >=4 replace TreatPerI=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(28Aug2000) & Date<=d(8Sep2000) replace TreatPerI=1 if Province=="VIC1" & Date>=d(28Aug2000) & Date<=d(14Sep2000) replace TreatPerII=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(7Oct2000) & Date<=d(28Oct2000)

5

replace TreatPerII=1 if (Province=="VIC1" | Province=="NSW1") & Year==2000 & Month ==10 & Day <=29 & Hour <=3 gen Treat = TreatPerI + TreatPerII gen TreatPer2001I=0 gen TreatPer2001II=0 replace TreatPer2001I=1 if (Province=="VIC1" | Province=="NSW1") & Year==2001 & Month ==8 & Day ==26 & Hour >=4 replace TreatPer2001I=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(27Aug2001) & Date<=d(8Sep2001) replace TreatPer2001I=1 if Province=="VIC1" & Date>=d(27Aug2001) & Date<=d(14Sep2001) replace TreatPer2001II=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(3Oct2001) & Date<=d(27Oct2001) replace TreatPer2001II=1 if (Province=="VIC1" | Province=="NSW1") & Year==2001 & Month ==10 & Day <=28 & Hour <=3 gen TreatPer1999I=0 gen TreatPer1999II=0 replace TreatPer1999I=1 if (Province=="VIC1" | Province=="NSW1") & Year==1999 & Month ==8 & Day ==30 & Hour >=4 replace TreatPer1999I=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(31Aug1999) & Date<=d(11Sep1999) replace TreatPer1999I=1 if Province=="VIC1" & Date>=d(31Aug1999) & Date<=d(18Sep1999) replace TreatPer1999II=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(3Oct1999) & Date<=d(27Oct1999) replace TreatPer1999II=1 if (Province=="VIC1" | Province=="NSW1") & Year==1999 & Month ==10 & Day <=28 & Hour <=3 gen CECRegression =0 replace CECRegression =1 if (Province=="VIC1" | Province=="NSW1") & (Date<=d(26Aug2000) | (Date>=d(30Oct2000) & Date<=d(31Dec2001))) gen Weekday = dow(Date) tab Weekday, gen(WeekdayD) /* WeekdayD1=Sunday, ..., WeekdayD7=Saturday*/ sort Province Year Month Date Day Hour gen TempShort = 0.45*Temp + 0.45*Temp[_n-1] + 0.10*Temp[_n-2] gen TempShortQuadr = TempShort*TempShort gen TempShortCubic = TempShort*TempShort*TempShort egen AvTemp49HouresT = ma(Temp), t(49) gen TempMADay = AvTemp49HouresT[_n-24] gen TempLong = 0.60*(TempMADay) + 0.30*(TempMADay[_n-48]) + 0.10*(TempMADay[_n-96]) gen TempLongH = 0 gen TempLongW = 0 gen TempLongC = 0 replace TempLongH = TempLong if Temp > 21.1111 replace TempLongW = TempLong if (Temp >= 10 & Temp <= 21.1111) replace TempLongC = TempLong if Temp < 10 destring SunshineHrs, replace gen SunTemp = SunshineHrs * Temp /* sunshine reduces elec on cold days, increases it on hot days */ drop AvTemp49HouresT drop TempMADay TempLong * Keep only VIC, and reshape wide by Hour (to enable SUR) keep if inlist(Province, "VIC1") sort Province Year Month Date Day Hour /* browse if (Demand ==. |Price==. | Date ==. |Hour ==. |Month==. | Year ==. |Day ==. |DST ==.|Temp ==. |Wind ==. |Pressure==. | /* */ Precip ==. |SunshineHrs ==. |PctTwilight ==. |PctDaylight==. | Employment ==. |Olympic ==. |OffDay ==. |TransDayOff ==. | /* */ TreatPerI ==. |TreatPerII ==. |Treat ==. |Weekday==. | WeekdayD1 ==. |WeekdayD2 ==. |WeekdayD3 ==. |WeekdayD4==. | WeekdayD5==. | WeekdayD6==. | WeekdayD7 ==. |TempShort ==. |TempShortQuadr ==. |TempShortCubic ==. |TempLongH==. | TempLongW ==. |TempLongC ==. |SunTemp==.) & (Province =="VIC1")

6

*/ *Check to see missing data * egen a=rmiss(Province-SunTemp) * tab a * browse if a>=1 & Province=="VIC1" * drop a ********************************************* /* This is for the SUR - Regression, such that observations are ligned up by clock-time and not by Standard Time !! */ ********************************************* *Generate a Clock Time Variable from 0-47 gen DST_Hour=Hour[_n+2] if DST ==1 replace DST_Hour=Hour if DST ==0 *At the end of DST period, clock time 5 and 6 is observed twice. * Solution: Take the average of the half hours foreach var of varlist Demand Price Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII /* */ TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II Treat TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp { replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(28Mar1999) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(28Mar1999) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(26Mar2000) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(26Mar2000) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(25Mar2001) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(25Mar2001) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(31Mar2002) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(31Mar2002) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(30Mar2003) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(30Mar2003) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(28Mar2004) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(28Mar2004) & DST_Hour ==6 & Province =="VIC1" replace `var'= (`var' + `var'[_n+1])/2 if DST == 1 & Date ==d(27Mar2005) & DST_Hour ==5 & Province =="VIC1" replace `var'= (`var' + `var'[_n-1])/2 if DST == 0 & Date ==d(27Mar2005) & DST_Hour ==6 & Province =="VIC1" } * Finally we need to drop the Half Hours 5 and 6 for the end of DST that appear twice. drop if DST == 1 & Date ==d(28Mar1999) & DST_Hour ==6 drop if DST == 0 & Date ==d(28Mar1999) & DST_Hour ==5 drop if DST == 1 & Date ==d(26Mar2000) & DST_Hour ==6 drop if DST == 0 & Date ==d(26Mar2000) & DST_Hour ==5 drop if DST == 1 & Date ==d(25Mar2001) & DST_Hour ==6 drop if DST == 0 & Date ==d(25Mar2001) & DST_Hour ==5 drop if DST == 1 & Date ==d(31Mar2002) & DST_Hour ==6 drop if DST == 0 & Date ==d(31Mar2002) & DST_Hour ==5

7

drop if DST == 1 & Date ==d(30Mar2003) & DST_Hour ==6 drop if DST == 0 & Date ==d(30Mar2003) & DST_Hour ==5 drop if DST == 1 & Date ==d(28Mar2004) & DST_Hour ==6 drop if DST == 0 & Date ==d(28Mar2004) & DST_Hour ==5 drop if DST == 1 & Date ==d(27Mar2005) & DST_Hour ==6 drop if DST == 0 & Date ==d(27Mar2005) & DST_Hour ==5 *At the beginning of DST period, clock time 4 and 5 are missing * Solution: Use data of clocktime 3 and dublicate to clocktime 4 & use data of clocktime 6 to dublicate to clocktime 5. expand 2 if Date==d(31Oct1999) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(27Aug2000) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(28Oct2001) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(27Oct2002) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(26Oct2003) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(31Oct2004) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" expand 2 if Date==d(30Oct2005) & (DST_Hour==6 |DST_Hour==3) & Province =="VIC1" sort Province Year Month Day Hour DST_Hour /*Attention, Don't Sort by DST_Hour first, but by Hour !!! (otherwise Last hour of the day becomes first hour */ replace DST_Hour = DST_Hour+1 if (Date ==d(31Oct1999) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(31Oct1999) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(27Aug2000) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(27Aug2000) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(28Oct2001) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(28Oct2001) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(27Oct2002) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(27Oct2002) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(26Oct2003) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(26Oct2003) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(31Oct2004) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(31Oct2004) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = DST_Hour+1 if (Date ==d(30Oct2005) & DST_Hour ==3) & DST_Hour[_n-2] == 2 & Province =="VIC1" replace DST_Hour = DST_Hour-1 if (Date ==d(30Oct2005) & DST_Hour ==6) & DST_Hour[_n+2] == 7 & Province =="VIC1" replace DST_Hour = 0 if Hour == 46 & Date ==d(31Dec2005) & Province=="VIC1" replace DST_Hour = 1 if Hour ==47 & Date ==d(31Dec2005) & Province=="VIC1" /* In order that the reshape works, generate now a DST consistent Date Year Month and Day etc. */ replace Date = Date+1 if DST == 1 & (DST_Hour==0 | DST_Hour ==1) format Date %dD_m_Y drop Year Month Day gen Year = year(Date) gen Month = month(Date) gen Day = day(Date)

8

gen Workday = 1-OffDay foreach var of varlist SunshineHrs Employment Olympic OffDay Workday TransDayOff CECRegression Weekday WeekdayD1 WeekdayD2 /* */ WeekdayD3 WeekdayD4 WeekdayD5 WeekdayD6 WeekdayD7 SchoolVac Holiday{ replace `var'= `var' if DST ==0 replace `var'= `var'[_n+2] if DST ==1 } drop if Date>=d(1Jan2003) drop Hour reshape wide Demand Price DST Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II Treat /* */ TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp, /* */ i(Province Year Month Date Day) j(DST_Hour) * Check to see missing data * egen b=rmiss(Province-WeekdayD7) * tab b * browse if b>1 & Province=="VIC1" drop if Month <=3 drop if Month >=10 drop if Month == 7 drop if Month ==8 & Day <=14 * Instead of "Visibility" and "Cloud Cover" used in the 2001 CEC study, we us "SunshineHrs" and "SunTemp0..SunTemp47" as a proxy. /* Define the Regressors into local `X0' to `X47' as locals */ forvalues i=0(1)47 { local Xì' TempShortì' TempShortQuadrì' TempShortCubicì' TempLongHì' TempLongWì' TempLongCì' /* */ SunTempì' Humidityì' Precipì' Pressureì' Windì' PctTwilightì' PctDaylightì' /* */ SunshineHrs WeekdayD* Employment Workday } /*SUR Estimation */ sureg ( Demand0 X0' ) /* */ ( Demand1 X1' ) /* */ ( Demand2 X2' ) /* */ ( Demand3 X3' ) /* */ ( Demand4 X4' ) /* */ ( Demand5 X5' ) /* */ ( Demand6 X6' ) /* */ ( Demand7 X7' ) /* */ ( Demand8 X8' ) /* */ ( Demand9 X9' ) /* */ ( Demand10 `X10' ) /* */ ( Demand11 `X11' ) /* */ ( Demand12 `X12' ) /* */ ( Demand13 `X13' ) /* */ ( Demand14 `X14' ) /* */ ( Demand15 `X15' ) /* */ ( Demand16 `X16' ) /* */ ( Demand17 `X17' ) /* */ ( Demand18 `X18' ) /* */ ( Demand19 `X19' ) /*

9

*/ ( Demand20 `X20' ) /* */ ( Demand21 `X21' ) /* */ ( Demand22 `X22' ) /* */ ( Demand23 `X23' ) /* */ ( Demand24 `X24' ) /* */ ( Demand25 `X25' ) /* */ ( Demand26 `X26' ) /* */ ( Demand27 `X27' ) /* */ ( Demand28 `X28' ) /* */ ( Demand29 `X29' ) /* */ ( Demand30 `X30' ) /* */ ( Demand31 `X31' ) /* */ ( Demand32 `X32' ) /* */ ( Demand33 `X33' ) /* */ ( Demand34 `X34' ) /* */ ( Demand35 `X35' ) /* */ ( Demand36 `X36' ) /* */ ( Demand37 `X37' ) /* */ ( Demand38 `X38' ) /* */ ( Demand39 `X39' ) /* */ ( Demand40 `X40' ) /* */ ( Demand41 `X41' ) /* */ ( Demand42 `X42' ) /* */ ( Demand43 `X43' ) /* */ ( Demand44 `X44' ) /* */ ( Demand45 `X45' ) /* */ ( Demand46 `X46' ) /* */ ( Demand47 `X47' ) /* */ if CECRegression==1 & Province=="VIC1" estimates store CEC drop if Date >=d(1Jan2003) /*Precict, what would have happened with DST, two month earlier in year 2000 (simulated DST)*/ * Here DSimì'p : predicted Demand_hat by SUR model * DSimì'r : predicted residuals from SUR errors: Attention, in "out of sample" DSimì'r is difference between y_hat and y_observed! * DSimì'wW : Demand without Weather: predicted Demand whereby all Weather and Lighting coefficients from SUR are set to 0 * DSimì' : Simulated Demand whereby all Weather and Lighting variables are laged by one hour * DSimwRì' : Simulated Demand + residuals from SUR, whereby all Weather and Lighting variables are laged by one hour /* For HalfHour 0 , we have to wrap around the day (for DSim), rest remains the same as below forvalues loop */ predict DSim0p, eq(Demand0) predict DSim0r, resid eq(Demand0) gen DSim0wW = DSim0p - ([Demand0]_b[TempShort0]*TempShort0 +[Demand0]_b[TempShortQuadr0]*TempShortQuadr0 +[Demand0]_b[TempShortCubic0]*TempShortCubic0 +[Demand0]_b[TempLongH0]*TempLongH0 +[Demand0]_b[TempLongW0]*TempLongW0 +[Demand0]_b[TempLongC0]*TempLongC0 +[Demand0]_b[SunTemp0]*SunTemp0 +[Demand0]_b[Humidity0]*Humidity0 +[Demand0]_b[Precip0]*Precip0 +[Demand0]_b[Pressure0]*Pressure0 +[Demand0]_b[Wind0]*Wind0 +[Demand0]_b[PctTwilight0]*PctTwilight0+ [Demand0]_b[PctDaylight0]*PctDaylight0) gen DSim0 = DSim0wW +([Demand0]_b[TempShort0]*TempShort46[_n-1]+[Demand0]_b[TempShortQuadr0]*TempShortQuadr46[_n-1]+[Demand0]_b[TempShortCubic0]*TempShortCubic46[_n-1]+[Demand0]_b[TempLongH0]*TempLongH46[_n-1]+[Demand0]_b[TempLongW0]*TempLongW46[_n-1]+[Demand0]_b[TempLongC0]*TempLongC46[_n-1]+[Demand0]_b[SunTemp0]*SunTemp46[_n-1]+[Demand0]_b[Humidity0]*Humidity46[_n-1]+[Demand0]_b[Precip0]*Precip46[_n-1]+[Demand0]_b[Pressure0]*Pressure46[_n-1]+[Demand0]_b[Wind0]*Wind46[_n-1]+[Demand0]_b[PctTwilight0]*PctTwilight46[_n-1]+ [Demand0]_b[PctDaylight0]*PctDaylight46[_n-1]) gen DSimwR0 = DSim0 + DSim0r drop DSim0wW DSim0r DSim0p

10

/* For HalfHour 0 , we have to wrap around the day (for DSim), rest remains the same as below forvalues loop */ predict DSim1p, eq(Demand1) predict DSim1r, resid eq(Demand1) gen DSim1wW = DSim1p - ([Demand1]_b[TempShort1]*TempShort1 +[Demand1]_b[TempShortQuadr1]*TempShortQuadr1 +[Demand1]_b[TempShortCubic1]*TempShortCubic1 +[Demand1]_b[TempLongH1]*TempLongH1 +[Demand1]_b[TempLongW1]*TempLongW1 +[Demand1]_b[TempLongC1]*TempLongC1 +[Demand1]_b[SunTemp1]*SunTemp1 +[Demand1]_b[Humidity1]*Humidity1 +[Demand1]_b[Precip1]*Precip1 +[Demand1]_b[Pressure1]*Pressure1 +[Demand1]_b[Wind1]*Wind1 +[Demand1]_b[PctTwilight1]*PctTwilight1+ [Demand1]_b[PctDaylight1]*PctDaylight1) gen DSim1 = DSim1wW +([Demand1]_b[TempShort1]*TempShort47[_n-1]+[Demand1]_b[TempShortQuadr1]*TempShortQuadr47[_n-1]+[Demand1]_b[TempShortCubic1]*TempShortCubic47[_n-1]+[Demand1]_b[TempLongH1]*TempLongH47[_n-1]+[Demand1]_b[TempLongW1]*TempLongW47[_n-1]+[Demand1]_b[TempLongC1]*TempLongC47[_n-1]+[Demand1]_b[SunTemp1]*SunTemp47[_n-1]+[Demand1]_b[Humidity1]*Humidity47[_n-1]+[Demand1]_b[Precip1]*Precip47[_n-1]+[Demand1]_b[Pressure1]*Pressure47[_n-1]+[Demand1]_b[Wind1]*Wind47[_n-1]+[Demand1]_b[PctTwilight1]*PctTwilight47[_n-1]+ [Demand1]_b[PctDaylight1]*PctDaylight47[_n-1]) gen DSimwR1 = DSim1 + DSim1r drop DSim1wW DSim1r DSim1p /* Finally do predictions for the half-hour 2-47, more efficient in loop */ forvalues i = 2(1)47 { local j = ì'-2 predict DSimì'p, eq(Demandì') predict DSimì'r, resid eq(Demandì') gen DSimì'wW = DSimì'p - ([Demandì']_b[TempShortì']*TempShortì' +[Demandì']_b[TempShortQuadrì']*TempShortQuadrì' +[Demandì']_b[TempShortCubicì']*TempShortCubicì' +[Demandì']_b[TempLongHì']*TempLongHì' +[Demandì']_b[TempLongWì']*TempLongWì' +[Demandì']_b[TempLongCì']*TempLongCì' +[Demandì']_b[SunTempì']*SunTempì' +[Demandì']_b[Humidityì']*Humidityì' +[Demandì']_b[Precipì']*Precipì' +[Demandì']_b[Pressureì']*Pressureì' +[Demandì']_b[Windì']*Windì' +[Demandì']_b[PctTwilightì']*PctTwilightì'+ [Demandì']_b[PctDaylightì']*PctDaylightì') gen DSimì' = DSimì'wW +([Demandì']_b[TempShortì']*TempShort`j'+[Demandì']_b[TempShortQuadrì']*TempShortQuadr`j'+[Demandì']_b[TempShortCubicì']*TempShortCubic`j'+[Demandì']_b[TempLongHì']*TempLongH`j'+[Demandì']_b[TempLongWì']*TempLongW`j'+[Demandì']_b[TempLongCì']*TempLongC`j'+[Demandì']_b[SunTempì']*SunTemp`j'+[Demandì']_b[Humidityì']*Humidity`j'+[Demandì']_b[Precipì']*Precip`j'+[Demandì']_b[Pressureì']*Pressure`j'+[Demandì']_b[Windì']*Wind`j'+[Demandì']_b[PctTwilightì']*PctTwilight`j'+ [Demandì']_b[PctDaylightì']*PctDaylight`j') gen DSimwRì' = DSimì' + DSimì'r drop DSimì'wW DSimì'p DSimì'r } gen index = _n sort index save startAnalysis.dta, replace matrix covb = e(V) matrix coef = e(b)' drop * svmat covb svmat coef gen index = _n sort index save matricesSUR.dta, replace /* *********** start the looop ! forvalues looop=1(2)7 { clear use matricesSUR.dta sort index mkmat covb*, matrix(covb)

11

mkmat coef, matrix(coef) drop covb* coef merge index using startAnalysis.dta drop _merge index local loop = 1 * here loop just with 2 o (above with 3!) gen test = `looop' if test[1]==1 { keep if Year==1999 & Month ==9 & Province =="VIC1" } if test[1]==2 { keep if Year==1999 & Month ==10 & Date<=d(31Oct1999) & Province =="VIC1" } if test[1]==3 { keep if (TreatPer1999I1 == 1 | TreatPer1999I47 == 1) } if test[1]==4 { keep if (TreatPer2001II1 == 1 | TreatPer2001II47 == 1) } if test[1]==5 { keep if Year==2001 & Month ==9 & Province =="VIC1" } if test[1]==6 { keep if Year==2001 & Month ==10 & Date <=d(28Oct2001) & Province =="VIC1" } if test[1]==7 { keep if (TreatPer2001I1 == 1 | TreatPer2001I47 == 1) } matrix rbhh = rowsof(coef)/48 svmat rbhh reshape long Demand Price DST Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII TreatPer1999I TreatPer1999II TreatPer2001I TreatPer2001II Treat /* */ TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp DSim DSimwR, /* */ i(Province Year Month Date Day) j(DST_Hour) local WandL TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp Humidity Precip Pressure Wind PctTwilight PctDaylight foreach var of varlist `WandL' { gen Sim`var'=`var'[_n-2] } global SimWandL SimTempShort SimTempShortQuadr SimTempShortCubic SimTempLongH SimTempLongW SimTempLongC SimSunTemp SimHumidity SimPrecip SimPressure SimWind SimPctTwilight SimPctDaylight reshape wide Demand Price DST Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II Treat /* */ TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp DSim DSimwR $SimWandL , /* */ i(Province Year Month Date Day) j(DST_Hour) ****************************************** /* Calculate the Standard Errors associated with the predictions! */ gen ONE = 1 global rbb=rbhh[1] /* generate rbhh vectors of zeroes (for the block-diagonal)*/ forvalues i = 1/$rbb { gen ZEROS`i'=0 }

12

mkmat ZEROS*, matrix(Z) /* Define Weather and Light Simulation Values into local `SWL' */ forvalues i=0(1)47 { local SWLì' SimTempShortì' SimTempShortQuadrì' SimTempShortCubicì' SimTempLongHì' SimTempLongWì' SimTempLongCì' SimSunTempì' SimHumidityì' SimPrecipì' SimPressureì' SimWindì' SimPctTwilightì' SimPctDaylightì' } /* Make 48 Matrices XSim (Simulated Weather and Light & daily regressors) from WandL by adding the Daily Regressors plus constant*/ forvalues i=0(1)47 { mkmat `SWLì'' SunshineHrs WeekdayD* Employment Workday ONE, matrix(XSimì') } /* partition coefficent vector into 48 subvectors */ forvalues i=0(1)47 { matrix coefì' = coef[(1+ì'*rbhh[1])..(ì'+1)*rbhh[1],1] } /* Test if Matrix OK: Predict DSim and compare to above variables result D */ forvalues i=0(1)47 { matrix dsì' = XSimì'*coefì' svmat dsì' rename dsì'1 dsì' gen controlì' =dsì'-DSimì' } matrix Xtild0 = XSim0, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild1 = Z, XSim1, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild2 = Z, Z, XSim2, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild3 = Z, Z, Z, XSim3, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild4 = Z, Z, Z, Z, XSim4, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild5 = Z, Z, Z, Z, Z, XSim5, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild6 = Z, Z, Z, Z, Z, Z, XSim6, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild7 = Z, Z, Z, Z, Z, Z, Z, XSim7, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild8 = Z, Z, Z, Z, Z, Z, Z, Z, XSim8, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild9 = Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim9, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild10 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim10, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild11 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim11, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild12 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim12, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild13 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim13, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild14 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim14, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild15 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim15, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild16 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim16, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild17 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim17, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z

13

matrix Xtild18 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim18, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild19 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim19, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild20 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim20, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild21 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim21, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild22 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim22, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild23 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim23, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild24 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim24, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild25 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim25, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild26 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim26, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild27 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim27, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild28 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim28, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild29 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim29, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild30 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim30, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild31 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim31, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild32 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim32, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild33 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim33, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild34 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim34, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild35 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim35, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild36 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim36, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild37 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim37, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild38 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim38, Z, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild39 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim39, Z, Z, Z, Z, Z, Z, Z, Z matrix Xtild40 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim40, Z, Z, Z, Z, Z, Z, Z matrix Xtild41 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim41, Z, Z, Z, Z, Z, Z matrix Xtild42 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim42, Z, Z, Z, Z, Z matrix Xtild43 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim43, Z, Z, Z, Z matrix Xtild44 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim44, Z, Z, Z matrix Xtild45 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim45, Z, Z matrix Xtild46 = Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim46, Z matrix Xtild47 = Z , Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, Z, XSim47 matrix Xtild = Xtild0\Xtild1\Xtild2\Xtild3\Xtild4\Xtild5\Xtild6\Xtild7\Xtild8\Xtild9\Xtild10\Xtild11\Xtild12\Xtild13\ /* */ Xtild14 \Xtild15 \Xtild16 \Xtild17 \Xtild18 \Xtild19 \Xtild20 \Xtild21 \Xtild22 \Xtild23 /* */ \Xtild24 \Xtild25 \Xtild26 \Xtild27 \Xtild28 \Xtild29 \Xtild30 \Xtild31 \Xtild32 \Xtild33 /* */ \Xtild34 \Xtild35 \Xtild36 \Xtild37 \Xtild38 \Xtild39 \Xtild40 \Xtild41 \Xtild42 \Xtild43 /* */ \Xtild44 \Xtild45 \Xtild46 \Xtild47

14

matrix COV=Xtild*covb*Xtild' ******************************************************* reshape long Demand Price DST Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II Treat /* */ TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp DSim DSimwR, /* */ i(Province Year Month Date Day) j(DST_Hour) svmat COV egen rCOVtot= rowtotal(COV*) egen DSimtot_Var=total(rCOVtot) egen DSimTreat = sum(DSim) egen DSimTreatwR = sum(DSimwR) egen DObsTreat = sum(Demand) gen DSimwRtot_t_neutral = (DSimTreatwR-DObsTreat)/sqrt(DSimtot_Var) gen DSimtot_t_neutral = (DSimTreat-DObsTreat)/sqrt(DSimtot_Var) gen DSimwRtot_t_m6pc = (DSimTreatwR-(DObsTreat*0.994))/sqrt(DSimtot_Var) gen DSimtot_t_m6pc = (DSimTreat-(DObsTreat*0.994))/sqrt(DSimtot_Var) gen DSimwRtot_t_mhpc = (DSimTreatwR-(DObsTreat*0.995))/sqrt(DSimtot_Var) gen DSimtot_t_mhpc = (DSimTreat-(DObsTreat*0.995))/sqrt(DSimtot_Var) gen DSimwRtot_t_m1pc = (DSimTreatwR-(DObsTreat*0.99))/sqrt(DSimtot_Var) gen DSimtot_t_m1pc = (DSimTreat-(DObsTreat*0.99))/sqrt(DSimtot_Var) gen PerCh =(DSimTreat-DObsTreat)/DObsTreat*100 gen PerChwR=(DSimTreatwR-DObsTreat)/DObsTreat*100 ** The following Table displays the percentage changes between the CEC-simulation outcomes and the observed Demand ** for the Treatment Period , first for DSim and then for wR (With Residual) respectively sum PerCh PerChwR ** The following Table displays the t-values for electricity neutrality, -1% and -0.5%: ** difference between the CEC-simulation outcomes and the observed Demand ** for the Treatment , first for DSim and then for wR (With Residual) respectively sum DSimtot_t_neutral DSimtot_t_mhpc DSimtot_t_m1pc sum DSimwRtot_t_neutral DSimwRtot_t_mhpc DSimwRtot_t_m1pc save data`looop'output.dta , replace } /* clear set mem 800m set more off set matsize 2000 set maxvar 7000 use startAnalysis.dta, clear reshape long Demand Price DST Temp Humidity Wind Pressure Precip PctTwilight PctDaylight TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II Treat /* */ TempShort TempShortQuadr TempShortCubic TempLongH TempLongW TempLongC SunTemp DSim DSimwR, /* */ i(Province Year Month Date Day) j(DST_Hour) sort Province Year Month Day DST_Hour gen a = _n ***********************

15

* Daily Load Shape Graphs for 2000 *********************** twoway line Demand a if Province=="VIC1" & TreatPerI==1 || /* */ line DSim a if Province=="VIC1" & TreatPerI==1, yaxis(2) clcolor(orange) /* */ title("DST plot--Vicoria TreatPerI") /* */ legend(on order(1 "Demand" 2 "DSim")) graph export VICDemandsTreatPerI.tif, replace twoway line Demand a if Province=="VIC1" & TreatPerII==1 || /* */ line DSim a if Province=="VIC1" & TreatPerII==1, yaxis(2) clcolor(orange) /* */ title("DST plot--Vicoria TreatPerII") /* */ legend(on order(1 "Demand" 2 "DSim")) graph export VICDemandsTreatPerII.tif, replace *********************** * Daily Load Shape Graphs for 2001 *********************** * September Graph *********************** twoway line Demand a if Year==2001 & Month ==9 & Province =="VIC1"|| /* */ line DSimwR a if Year==2001 & Month ==9 & Province =="VIC1", clcolor(orange) /* */ title("DST plot--Vicoria Sept2001 WITH RESIDUAL") /* */ legend(on order(1 "Demand" 2 "DSimwR")) graph export VICDemandswRSep01.tif, replace twoway line Demand a if Year==2001 & Month ==9 & Province =="VIC1"|| /* */ line DSim a if Year==2001 & Month ==9 & Province =="VIC1", clcolor(orange) /* */ title("DST plot--Vicoria Sept2001 NO Residual") /* */ legend(on order(1 "Demand" 2 "DSim")) graph export VICDemandsSep01.tif, replace * October Graph *********************** twoway line Demand a if Year==2001 & Month ==10 & Day<=28 &Province =="VIC1"|| /* */ line DSimwR a if Year==2001 & Month ==10 & Day<=28 &Province =="VIC1", clcolor(orange) /* */ title("DST plot--Vicoria October 2001 WITH RESIDUAL") /* */ legend(on order(1 "Demand" 2 "DSimwR")) graph export VICDemandswROct01.tif, replace twoway line Demand a if Year==2001 & Month ==10 & Day<=28 &Province =="VIC1"|| /* */ line DSim a if Year==2001 & Month ==10 & Day<=28 &Province =="VIC1", clcolor(orange) /* */ title("DST plot--Vicoria October 2001 NO Residual") /* */ legend(on order(1 "Demand" 2 "DSim")) graph export VICDemandsOct01.tif, replace save TempLong.dta, replace *********************** * Average Load Shape Graphs for 2000 *********************** */ use TempLong.dta, clear sort Province Year Month Day DST_Hour gen DSimStandarTime = DSim[_n-2] sort Province DST_Hour Year Month collapse(mean) Demand DSimStandarTime DSim DSimwR, by(Province DST_Hour Year Month) sort Province Year Month DST_Hour twoway line Demand DST_Hour if Month==9 & Year==2000& Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimStandarTime DST_Hour if Month==9 & Year==2000 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September 2000") /* */ legend(on order(1 "2000 Demand" 2 "2000 DSim in Standard Time")) graph export VICAverDemandsStandardTime00.tif, replace

16

twoway line Demand DST_Hour if Month==9 & Year==2000& Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSim DST_Hour if Month==9 & Year==2000 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September 2000") /* */ legend(on order(1 "2000 Demand" 2 "2000 DSim")) graph export VICAverDemands00.tif, replace /* twoway line Demand DST_Hour if Month==9 & Year==1999 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimwR DST_Hour if Month==9 & Year==1999 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September 1999") /* */ legend(on order(1 "1999 Demand" 2 "1999 DSimWR")) graph export VICAverDemandswR99.tif, replace twoway line Demand DST_Hour if Month==9 & Year==2001 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimwR DST_Hour if Month==9 & Year==2001 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September") /* */ legend(on order(1 "Sep2001 Demand" 2 "Sep2001 DSimWR")) graph export VICAverDemandswRSep01.tif, replace twoway line Demand DST_Hour if Month==10 & Year==2001 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimwR DST_Hour if Month==10 & Year==2001 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria October") /* */ legend(on order(1 "Oct2001 Demand" 2 "Oct2001 DSimWR")) graph export VICAverDemandswROct01.tif, replace twoway line Demand DST_Hour if Month==9 & Year==2001 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSim DST_Hour if Month==9 & Year==2001 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September") /* */ legend(on order(1 "Sep2001 Demand" 2 "Sep2001 DSim (NO Residual)")) graph export VICAverDemandsSep01.tif, replace *********** *this following graph does not make sense due to adding of Residial at DSim in the period were Standard Time is not observed in 2000. *rather consistency check, *********** /* twoway line Demand DST_Hour if Month==9 & Year==2000& Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimwR DST_Hour if Month==9 & Year==2000 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria September 2000") /* */ legend(on order(1 "2000 Demand" 2 "2000 DSimWR")) graph export VICAverDemandswR00.tif, replace */ cd "C:\My Documents\Berkeley\ARE\Diss\Ideas\DST\AustraliaData\Analysis" *log close *log using "C:\My Documents\Berkeley\ARE\Diss\Ideas\DST\AustraliaData\Analysis\logCEC.smcl", replace set mem 800m set more off set matsize 2000 set maxvar 7000 clear use TempLong.dta sort Province Year Month Day DST_Hour

17

drop TreatPer1999I TreatPer1999II gen TreatPer1999I=0 gen TreatPer1999II=0 replace TreatPer1999I=1 if (Province=="VIC1" | Province=="NSW1") & Year==1999 & Month ==8 & Day ==30 & DST_Hour >=4 replace TreatPer1999I=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(31Aug1999) & Date<=d(11Sep1999) replace TreatPer1999I=1 if Province=="VIC1" & Date>=d(31Aug1999) & Date<=d(18Sep1999) replace TreatPer1999II=1 if (Province=="VIC1" | Province=="NSW1") & Date>=d(3Oct1999) & Date<=d(30Oct1999) replace TreatPer1999II=1 if (Province=="VIC1" | Province=="NSW1") & Year==1999 & Month ==10 & Day <=31 & DST_Hour <=3 gen DSimStandarTime = DSim[_n-2] gen DSimwRStandarTime = DSimwR[_n-2] gen DemandLeft = Demand[_n+2] sort Province DST_Hour Year TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II collapse(mean) Demand DSimStandarTime DSimwRStandarTime DemandLeft DSim DSimwR, by(Province DST_Hour Year TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II ) gen Time = DST_Hour/2 twoway line Demand Time if TreatPer1999I==1 & Province=="VIC1" & Time >=6 & Time <= 23, clcolor(black) ||/* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" & Time >=6 & Time <= 23, clcolor(black) clpat(dash) || /* */ line Demand Time if TreatPer2001I==1 & Province=="VIC1" & Time >=6 & Time <= 23, clcolor(blue) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" & Time >=6 & Time <= 23, clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPerI==1 & Province=="VIC1" & Time >=6 & Time <= 23, clcolor(red) lwidth(thick) /* */ legend(on order(1 "1999 Actual" 2 "1999 Simulated" 3 "2001 Actual" 4 "2001 Simulated" 5 "2000 Actual" )) /* */ , ytitle(Consumption in Megawatts) xtitle(Hour (Clock Time)) xlabel(6 7 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5clockTreatIRefurb.tif, replace twoway line Demand Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) ||/* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) clpat(dash)/* */ legend(on order(1 "1999 Actual" 2 "1999 Simulated" )) /* */ , ytitle(Consumption in Megawatts) xtitle(Hour (Clock Time)) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export clockTreatI99Refurb3_23.tif, replace twoway line Demand Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) clpat(dash)/* */ legend(on order(1 "2001 Actual" 2 "2001 Simulated" )) /* */ , ytitle(Consumption in Megawatts) xtitle(Hour (Clock Time)) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export clockTreatI01Refurb3_23.tif, replace

18

twoway line DemandLeft Time if TreatPerI==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(green) clpat(dash) || /* */ line Demand Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(green) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) /* */ title("Vicoria Treatment Period I, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /* */ , ytitle(Demand in MW) xtitle(Hour) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5.tif, replace twoway line Demand Time if TreatPerI==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) clpat(dash) || /* */ line Demand Time if TreatPer1999I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPer2001I==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) /* */ title("Vicoria Treatment Period I, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /* */ , ytitle(Demand in MW) xtitle(Hour) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5clockTreatIb.tif, replace twoway line Demand Time if TreatPerII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if TreatPer1999II==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) clpat(dash) || /* */ line Demand Time if TreatPer1999II==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) ||/* */ line DSimwR Time if TreatPer2001II==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPer2001II==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) /* */ title("Vicoria Treatment Period II, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /* */ , ytitle(Demand in MW) xtitle(Hour) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5clockTreatII.tif, replace twoway line DemandLeft Time if TreatPerI==1 & Province=="VIC1" , clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" , clcolor(green) clpat(dash) || /* */ line Demand Time if TreatPer1999I==1 & Province=="VIC1" , clcolor(green) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" , clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPer2001I==1 & Province=="VIC1" , clcolor(blue) /* */ title("Vicoria Treatment Period I, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /* */ , ytitle(Demand in MW) xtitle(Hour) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5_AllHours.tif, replace

19

twoway line Demand Time if TreatPerI==1 & Province=="VIC1" , clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if TreatPer1999I==1 & Province=="VIC1" , clcolor(green) clpat(dash) || /* */ line Demand Time if TreatPer1999I==1 & Province=="VIC1" , clcolor(green) ||/* */ line DSimwR Time if TreatPer2001I==1 & Province=="VIC1" , clcolor(blue) clpat(dash)|| /* */ line Demand Time if TreatPer2001I==1 & Province=="VIC1" , clcolor(blue) /* */ title("Vicoria Treatment Period I, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /* */ , ytitle(Demand in MW) xtitle(Hour) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5_AllHoursClock.tif, replace twoway line Demand DST_Hour if TreatPerI==1 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimwR DST_Hour if TreatPer2001I==1 & Province=="VIC1", clcolor(red) || /* */ line DSimwR DST_Hour if TreatPer1999I==1 & Province=="VIC1", clcolor(green) /* */ title("DST plot--Vicoria Treatment Period I in 1999, 2000 and 2001") /* */ legend(on order(1 "2000 Demand" 2 "2001 DSimWR" 3 "1999 DSimWR")) graph export VICAverDemandswR00_01_99.tif, replace twoway line Demand DST_Hour if TreatPerI==1 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSim DST_Hour if TreatPerI==1 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria Treatment Period I") /* */ legend(on order(1 "2000 Demand" 2 "2000 DSim" )) graph export VICAverDemand00_Treat1.tif, replace gen Hour = DST_Hour - 2 twoway line Demand Hour if TreatPerI==1 & Province=="VIC1", clcolor(blue) clpat(dash)|| /* */ line DSimStandarTime Hour if TreatPerI==1 & Province=="VIC1", clcolor(red) /* */ title("DST plot--Vicoria Treatment Period I, in Standard Time") /* */ legend(on order(1 "2000 Demand" 2 "2000 DSim" )) graph export VICAverDemand00_Treat1_StandardTime.tif, replace clear use TempLong.dta sort Province Year Month Day DST_Hour gen DSimStandarTime = DSim[_n-2] gen DSimwRStandarTime = DSimwR[_n-2] gen DemandLeft = Demand[_n+2] sort Province DST_Hour Year Date TreatPerI TreatPerII TreatPer2001I TreatPer2001II TreatPer1999I TreatPer1999II gen TreatIandII = 0 replace TreatIandII = 1 if TreatPerI ==1 | TreatPerII ==1 gen Treat1999IandII = 0 replace Treat1999IandII = 1 if TreatPer1999I ==1 | TreatPer1999II ==1 gen Treat2001IandII = 0 replace Treat2001IandII = 1 if TreatPer2001I ==1 | TreatPer2001II ==1 collapse(mean) Demand DSimStandarTime DSimwRStandarTime DemandLeft DSim DSimwR, by(Province DST_Hour Year TreatIandII Treat1999IandII Treat2001IandII ) gen Time = DST_Hour/2 twoway line Demand Time if TreatIandII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(red) lwidth(thick)|| /* */ line DSimwR Time if Treat1999IandII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) clpat(dash) || /* */ line Demand Time if Treat1999IandII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(black) ||/* */ line DSimwR Time if Treat2001IandII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) clpat(dash)|| /* */ line Demand Time if Treat2001IandII==1 & Province=="VIC1" & Time >=3 & Time <= 23, clcolor(blue) /* */ title("Vicoria Treatment Period I&II, 1999 to 2001") /* */ legend(on order(1 "2000 Demand" 2 "1999 Simulated Demand" 3 "1999 Demand" 4 "2001 Simulated Demand" 5 "2001 Demand" )) /*

20

*/ , ytitle(Demand in MW) xtitle(Hour) xlabel(3 4 to 23) scheme(s1manual) graphregion(fcolor(none) ifcolor(none)) plotregion(fcolor(none) lcolor(none)) graph export FIG5clock_treatIandII.tif, replace

References

Australian Bureau of Statistics (2001a): Tourism Indicators, Report 8634.0, December Quarter

2000, Canberra.

Australian Bureau of Statistics (2001b): Tourist accommodation: an analysis over the Olympic

period. Tourism Indicators, December Quarter 2000.

21

11:00- 11:30- 12:00- 12:30- 13:00- 13:30- 14:00- 14:30- 15:00-11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30

Percent change in demand 1.41 0.44 0.66 0.19 -0.24 0.40 1.35 1.33 1.31Standard error (1.74) (1.43) (1.56) (1.54) (1.41) (1.64) (1.90) (1.37) (1.56)

Standard errors are clustered on year

Half-hour

Table B: Regression discontinuity estimates of the effect of switching to DST, by half-hour:late-October switches in VIC and SA in 1999 and 2001-2005

Half-hour beginning at βh

Standard error

t-statistic exp(βh)-1

Half-hour beginning at βh

Standard error

t-statistic exp(βh)-1

00:00 -0.015 0.007 -2.30 -0.015 12:00 - - - -00:30 0.016 0.007 2.45 0.016 12:30 - - - -01:00 -0.052 0.006 -8.28 -0.051 13:00 - - - -01:30 -0.046 0.006 -7.17 -0.045 13:30 - - - -02:00 0.054 0.006 8.96 0.055 14:00 - - - -02:30 0.074 0.006 12.68 0.077 14:30 0.014 0.003 4.89 0.01403:00 0.071 0.006 12.23 0.074 15:00 0.010 0.004 2.92 0.01003:30 0.066 0.006 10.82 0.069 15:30 0.008 0.004 2.19 0.00804:00 0.056 0.006 9.29 0.058 16:00 0.008 0.004 1.85 0.00804:30 0.045 0.006 7.63 0.046 16:30 0.002 0.005 0.38 0.00205:00 0.032 0.006 5.38 0.032 17:00 -0.015 0.006 -2.51 -0.01505:30 0.024 0.005 4.42 0.024 17:30 -0.028 0.007 -3.99 -0.02706:00 0.019 0.006 3.30 0.019 18:00 -0.050 0.007 -7.18 -0.04906:30 0.016 0.005 3.01 0.016 18:30 -0.068 0.007 -9.50 -0.06607:00 0.080 0.006 13.94 0.084 19:00 -0.058 0.008 -7.51 -0.05607:30 0.083 0.006 14.20 0.087 19:30 -0.028 0.007 -3.85 -0.02808:00 0.028 0.006 5.04 0.029 20:00 -0.011 0.007 -1.54 -0.01108:30 0.013 0.005 2.63 0.013 20:30 -0.007 0.007 -1.06 -0.00709:00 0.009 0.004 2.14 0.009 21:00 -0.002 0.007 -0.30 -0.00209:30 0.006 0.004 1.65 0.006 21:30 0.002 0.007 0.25 0.00210:00 0.003 0.003 0.87 0.003 22:00 -0.009 0.006 -1.45 -0.00910:30 0.006 0.003 1.66 0.006 22:30 -0.030 0.006 -5.30 -0.02911:00 0.004 0.003 1.25 0.004 23:00 -0.128 0.006 -20.60 -0.12011:30 0.002 0.002 1.09 0.002 23:30 -0.131 0.007 -19.07 -0.123

Standard errors are clustered on dateNo effects are estimated for the control period of 12:00-14:30

Table C: Estimated treatment effects by half-hour

The large effects shown during the overnight hours are driven by centralized off-peak water heating that is activated by automatic timers, set to standard time

22

Figure D1: Quarterly Room Nights Occupied in VIC (left panel) and SA (right panel)

Figure: D2: Supply and Demand for Tourist Accommodations in Sydney

Figure 1: Southeastern Australia states and major cities

NSW, VIC, and SA in mainland Australia regularly begin DST on the last Sunday in October each year. In 2000, however, NSW and VIC began DST on 27 August, whereas SA did not begin DST until 29 October.

NSW, VIC, and SA in mainland Australia regularly begin DST on the last Sunday in October each year. In 2000, however, NSW and VIC began DST on 27 August, whereas SA did not begin DST until 29 October.

Source: Australian Bureau of Statistics, 2001. The vertical line indicates the 4th quarter in 2000 (December quarter). The treatment period “September” falls within the 3rd quarter 2000 and the treatment period “October” in the 4th quarter.