How Effective are Individual Lifestyle Changes in Reducing Electricity

Consumption? - Measuring the Impact of Earth Hour

David Solomon*

University of Chicago

Graduate School of Business

First Draft: April 2007

This Draft: May 2008

Abstract: This paper examines a unique natural experiment where Sydney residents

turned off lights and electrical appliances for one hour. While polls reported 57% of

Sydney participated, statewide electricity use declined by 2.10%, statistically

indistinguishable from zero. This indicates that discretionary household electricity use

like lighting forms only a small component of total electricity consumption, and policies

targeting such use may be of limited impact. Using poll data on participation and

previous estimates of household electricity consumption, evidence indicates that

respondents overstated their involvement by around 36%. This is consistent with

consumers feeling pressure to overstate their preferences for environmental goods.

*5807 S. Woodlawn Ave., Chicago IL 60637. Contact at dsolomon@[Link] .

I would like to thank Sam Hartzmark, Roni Kisin and John List for helpful comments and suggestions. All

remaining errors are my own.

1. Introduction

On March 31st 2007 at 7:30pm, the residents of Sydney, Australia, held an Earth

Hour, where people were urged to turn off their lights and electrical appliances for one

hour. This event was organized by the WWF Australia, an environmental group. A poll

by AMR Interactive, reported in newspaper The Sydney Morning Herald1, found that

57% of respondents had participated in Earth Hour, corresponding to roughly 2.4 million

people. 53 per cent turned off the lights at home, 25 per cent switched off their computer

and 17 per cent turned off the television.

This event provides a unique natural experiment to test two questions of

importance in economics. First, it allows for a measure of the size and importance of

discretionary household electricity consumption, which has implications for the ability of

energy-saving policies aimed at households to achieve cuts in electricity use and

greenhouse gas emissions.2 Second, it provides a test of whether individuals overstate

their preference for social goods in the presence of interviewer scrutiny, and the

reliability of poll evidence in examining preferences for such goods.

In terms of the size of discretionary household electricity use, I find that state-

wide electricity use declined by a statistically insignificant 2.10% during Earth Hour,

equivalent to a drop of 168-173 MW/h. In economic terms, the voluntary cuts to

electricity consumption that Sydney households are willing to bear (even even for an

hour) in the name of environmental events produces only a very small decline in New

1
[Link]
2
The measured drop in Earth Hour will necessarily not capture household use from items of a less-

discretionary nature which were unlikely to be turned off, such as refrigerators, water heaters, pool pumps

etc. For estimates of the size of household use of these appliances, see Bartels and Fiebig (2000)

2
South Wales electricity use. This further suggests that policies targeting discretionary

household electricity use may not be of large environmental impact.

Moreover, the estimated electricity declines suggest that poll respondents

significantly overstated their level of involvement in the event. Using the results of an

AMR Interactive poll of 926 Sydney residents and electricity end-use estimates of

lighting and television from Bartels and Fiebig (2000), it is possible to compare actual

and predicted declines in electricity use and obtain a measure of whether respondents

appeared to have over- or under-stated their level of involvement in the event. I find that

the predicted electricity decline during Earth Hour based on Sydney participation was

236 MW/h. The ratio of the actual and predicted electricity declines suggests that

respondents overstated their level of participation in the event by around 36%. This

indicates that poll results may significantly overstate consumers’ true willingness to make

sacrifices for environmental issues.

This paper adds to the literature in environmental economics that seeks to

estimate of the size and importance of household electricity use. Most of the previous

literature on the subject has focused on either smaller scale individual metering data

(Bartels and Fiebig (2000)) or structural models of consumption (Narayan and Smith

(2005), Larsen and Nesbakken (2004), Saab, Badrb and Nasra (2001), among others). In

terms of measuring household consumption, the Earth Hour experiment is unique in that

it not only covers a very large number of households, but also allows for reduced form

estimates of household consumption, rather than relying on a particular structural model

of household use. Most similar to the current paper is Kendrick and Wolff (2007), who

estimate the effect of daylight saving on total electricity consumption.

3
 Second, it presents further evidence on the problems of taking consumers stated

preferences for social goods as indicating the actual value they place on them. A long

literature in this area relates to hypothetical bias, the tendency of people to overstate their

preferences for goods when asked hypothetical questions (see for instance List and

Shogren (1998) and List (2001)). This paper documents even more basic problem - that

in questions of environmental goods, consumers cannot even be relied on to accurately

report their own actions ex post (rather than their valuations of goods ex ante). In a

broader sense, it shows importantly that consumers may feel pressure in surveys to claim

to have preferences for environmentalism, even if they privately do not (over and above

whether they are being forced to make hypothetical estimates)

There are a number of reasons why consumers might overstate their involvement

in such an event. Levitt and List (2007) develop a model where an individual’s choice of

behavior depends in part upon the moral cost of the action, which is a function of the

level of scrutiny the actions receive - people are less likely to undertake morally costly

actions when there is more scrutiny. If being perceived to be environmentally unfriendly

has moral costs, then respondents who did not take part in Earth Hour may hesitate to

reveal this fact when faced with a poll situation. This could result in them either declining

to participate in the poll at larger rates than those who were involved3, or lying to

interviewers about their level of involvement (if these two alternatives are seen as less

3
Merkle and Edelman (2000) argue that different response rates do not necessarily affect prediction error,

although they do not examine the possibility that refusal may vary systematically across groups, rather than

randomly.

4
morally costly to them than admitting to being environmentally unfriendly). Either way

would see the poll results overstating true participation in the event.

Indeed, the polls themselves suggest such overstatement. Although 57% of

respondents took part in some level, only 53% turned off lights, notwithstanding that this

was the main promotion point of the event. In other words, at least 7% of claimed

participation in the event came from people who didn’t even turn off any lights, but

claimed participation based on switching off some ‘other appliance’, which may or may

not have otherwise even been on in the first place. While this is only suggestive, it is

consistent with respondents preferring to classify themselves as participating (or pollsters

facing similar pressure), even when they had only limited levels of involvement.

This finding is related to the literature in political science on the accuracy of

opinion polls (see Converse and Traugott (1986)), and in particular exit polls in elections,

where respondents are asked to report on past actions rather than opinions per se. There is

evidence that exit polls may overestimate the support for socially desirable causes

(Traugott and Price (1992)), and that this effect is larger for face-to-face interviews than

secret ballots (Bishop and Fisher (1995)). As a more positive policy result, the results of

this paper would suggest that more accurate measures of preferences can be obtained by

using less intrusive questioning techniques, such as handing out anonymous

questionnaires rather than conducting face-to-face or phone interviews.

Using intraday data on 8 years of New South Wales electricity consumption, I

find that there is limited evidence that Earth Hour caused a statistically significant

decrease in electricity consumption. Following Kellogg and Wolff (2007), I estimate

electricity use with a baseline regression that controls for year, month, day-of-week and

5
time-of-day fixed effects, daylight saving, retail electricity price, and weather-related

variables, as well as interaction terms. Furthermore, I show that simply taking the

difference between actual and predicted consumption (as press accounts purported to do4)

can produce erroneous inferences based on the presence of omitted variables. While the

controls above seem likely to control for much of household use (such as lighting,

heating and cooling etc), it is unclear how well they explain industrial electricity usage.

This is likely to fluctuate with firm inventory requirements, short term demand, and a

variety of other economic variables for which daily and intraday measures are hard to

come by, and which may not be captured by time fixed effects.

Consistent with this, while a Dummy Variable for Earth Hour shows a reduction

in electricity use of 6.33%, or 551-567MW/h, over 67% of this effect is due to omitted

factors common throughout the entire day. When an Earth Day dummy is included as

well, the Earth Hour effect is statistically indistinguishable from zero, with a point

estimate of a drop 2.10%, or 168-173 MW/h.

While it is possible that some of this Earth Day effect may be coming from

consumers switching off appliances earlier in the day in response to the event, I present a

variety of evidence that this is unlikely to be a large driver of the difference between the

two estimates. Firstly, equivalently large percentage declines as during the 7:30-8:30pm

4
The effect of this was reported as being a drop in electricity use in the Central Business District of 10.2%,

according to an analysis by Energy Australia, that purported to control for time-of-day, weather, and month

effects. Results, without details on the specifications used, are available at

[Link]

2+Earth+Hour+wrap+[Link]

6
Earth Hour were observed as early as 5am and persisted fairly uniformly throughout the

day, indicating that the bulk of the declines were due to variables operating before

sunrise. Second, an effect of turning off lights in preparation for the event ought to

become more pronounced as Earth Hour approached. In fact, a dummy variable for the

6:30-7:30pm hour shows a weakly positive sign (after controlling for the day effect),

suggesting that if anything consumers brought forward electricity demand that they might

otherwise have had, thus the drop during the hour itself ought to have been even more

pronounced.

Third, a spokesperson for Integral Energy, the company that maintains the

electrical network, claimed that “We noticed a steep decline in the first five minutes of

Earth Hour, between 7.30 and 7.35pm”5, confirming the intuition behind the event that

the focus was on consumers turning off lights for that particular hour. Finally, I show that

by simply comparing actual and predicted consumption, over 27.5% of days in the

sample have two consecutive periods with consumption ‘Earth Hour’ sized gaps more

than 6.36% below predicted values (the larger of the two Earth Hour period drops, using

the base specification without any dummy variables). This confirms both that it is

problematic to simply attribute any gap in consumption to Earth Hour, and that

notwithstanding the long list of controls, omitted variables relating to industrial

production still need to be accounted for, such as by a day fixed effect. This conclusion is

consistent with Kellogg and Wolff (2007), who find a similar need to control for omitted

variables when working with data from the same source and a similar list of controls.

5
[Link]

7
 The rest of the paper is as follows: section 2 describes the data, section 3 presents

estimates of likely changes in electricity use from poll results and Bartels and Fiebig

(2000) estimates, section 4 presents regression tests of the electricity drop during Earth

Hour, section 5 compares Earth Hour drops with declines during the rest of the day, and

section 6 concludes.

2. Data

Data on New South Wales electricity consumption are obtained from

[Link] at a half hourly frequency between January 1, 1999, and

March 31, 2007. These include quantity consumed (in Megawatt hours) and retail price

of electricity (in Australian Dollars). Weather data are obtained from the National

Climatic Data Center at [Link] They give measures from the Sydney

Airport station, including temperature (in degrees Fahrenheit), wind speed, classification

of level of cloud cover, at an hourly (or shorter) frequency. Summary Statistics and

correlations of the non-categorical variables are presented in Table I.

[Insert Table I here]

3. Estimated Effects from Bartels and Fiebig (2000)

To estimate the likely drop in electricity use, it is possible to combine the survey

evidence on Earth Hour participation with the Bartels and Fiebig (2000) estimates of

residential end-use. The end-use estimates apply to an average working day in August,

and so to the extent that they differ from a Saturday in March the estimates will be noisy.

8
On March 31st 2007, sunset in Sydney was at 5:52pm, compared with sunset times from

5:15pm to 5:36pm during August in 1997 (in Bartels and Fiebig (2000)), suggesting that

lighting use estimates from 7:30pm to 8:30pm should not differ greatly due to the

different seasons. The effect of the different time of year and day of week on computer

use and television use is less clear, although these make up a much smaller component of

the total estimated effect.

The AMR Interactive Poll found that 53% of respondents turned off lights, 25%

turned off a computer, 17% turned off the television, and 25% turned off ‘other

appliances’. In estimating electricity use from computers, I assign the likely use the same

as for televisions (as Bartels and Fiebig (2000) do not provide a specific computer

estimate). I ignore the ‘other appliances’ category, because it is not clear what this

corresponds to – it is unlikely that households switched off some more energy-intensive

appliances such as refrigerators, freezers, pool pumps, or hot water systems. The level of

participation for cooktops, ovens, dryers and dishwashers is less obvious. To the extent

that these are not included, the estimates will underreport the likely drop in electricity

consumption. Further, while the event was targeted largely at households, to the extent

that any businesses took part the true likely drop will again be greater than the numbers

here.

Table II presents the results of these estimates. Estimated electricity use is

calculated assuming the survey participation data covers only Sydney (population 4.28m

according to June 2006 Australian Bureau of Statistics figures, in column 3), or all of

New South Wales (population 6.82m according to June 2006 Australian Bureau of

Statistics figures, in column 4). The estimated drops are 236 MW/h and 374 MW/h. I take

9
the lower of the two, which assumes only Sydney residents took part, as a conservative

estimate of the likely electricity reduction based on the poll numbers.

[Insert Table II here]

4. Regression Tests of Earth Hour Effects

There are a variety of approaches that can be used to model electricity demand.

Narayand and Smyth (2005) use a cointegration approach to estimate annual electricity

consumption in Australia, while Saab, Badrb and Nasra (2001) use autoregressive models

to estimate monthly electricity consumption in Lebanon. It is not clear that autoregressive

models are necessarily appropriate for modeling much higher frequency consumption,

where autoregression may be drowned out by within-day cyclical factors. Closest to the

current question is the work of Kellogg and Wolff (2007), who use a panel regression

framework to estimate the effects of daylight saving on half-hourly data from Australian

electricity consumption. The structure and choice of variables used here is based on

Kellogg and Wolff (2007). To test whether Earth Hour caused a significant drop in

electricity use, I use a regression framework. I estimate the equation:

ln(Consumtiont ) = α + β1 EarthDayt + β 2 EarthHourt + β 3 PreEarthHourt

+ β 4 PostEarthHourt + β 5Controls + ε t

where ln(Consumption) is the natural log of New South Wales electricity consumption in

MW/h (for that half-hour period), EarthDay is a Dummy Variable that equals 1 on March

31st 2007 and zero otherwise, EarthHour, PreEarthHour and PostEarthHour are

10
Dummy Variables that equal 1 for demand from 7:30pm-8:30 pm, 6:30pm-7:30pm, and

8:30pm-9:30pm respectively on March 31 2007, and zero otherwise. Controls includes a

large number of variables that affect electricity consumption. Dummy Variables are

included for year, month, day of the week, time of day (per half hour, for a total of 47

variables), and whether Daylight Savings is occurring. Additional controls are included

for the retail price of electricity in Australian Dollars, and the most recently available

Sydney Airport weather observations: the quadratic temperature (that is, (T-65)2, where T

is the temperature in degrees Fahrenheit), wind speed, and Dummy variables for the level

of cloud cover. Additional controls are included for the interaction of the half-hour time

dummy variables with daylight saving, day of the week, month, and the weather

variables. The interaction terms thus allow these variables to have different effects at

different times of the day, rather than forcing them to be constant.

[Insert Table III here]

The results of these regressions are presented in Table III. Columns 1 examines

the effect of EarthHour variable without day fixed effects. EarthHour is statistically

significant. The coefficient is –0.065, with a t-statistic of –1.90, significant at a 10%

level. The point estimate for EarthHour corresponds to a decrease in electricity

consumption of 6.33% relative to what it would otherwise have been, equal to 551MW/h

from 7:30-8pm and 567 MW/h from 8-8:30pm 6.

6
Percentage changes in consumption are obtained by converting predicted log consumption to predicted

consumption by taking an exponent, and then taking the percentage changes. This is equivalent to exp(β)-1,

11
 Column 2 examines the average drop during the whole day of March 31st.

EarthDay shows a statistically significant coefficient in both specifications, of –0.046

(t-statistic of –6.43, percentage drop of 4.66%).

Column 3 examines the effect of EarthHour once a day fixed effect is controlled

for. Once the EarthDay fixed effect is controlled for, the EarthHour effect is greatly

reduced, and loses any remaining significant explanatory power. The EarthHour

coefficient is -0.021 (t-statistic of –0.60, percentage decrease of 2.10%, decreases of 173

MW/h and 168 MW/h).

These results show quite strongly that the Earth Hour effect is of only weak

statistical significance, and even this is sensitive to the inclusion of a fixed effect for the

whole day, whereupon the drop in electricity use during Earth Hour is statistically

indistinguishable from zero. Controlling for the day fixed effect reduces the estimated

EarthHour effect by 67.6%..7

Column 4 examines the question of whether electricity consumption was larger

before and after Earth Hour, and indicates that there was not a significant increase in

consumption before or after due to substitution effects. The coefficients on PreEarthHour

and PostEarthHour are insignificant in all cases. The magnitude of these changes is of

where β is the EarthHour coefficient in Table III. To convert percentage declines to MW/h declines, the

percentages above are multiplied by the predicted value of log consumption from the same regression if the

EarthHour variable were equal to zero during the hour in question.

7
The EarthHour coefficients and significance are even weaker if the EarthDay coefficient is substituted

with a Dummy that equals 1 on March 31st after 6:00am (if the effects were due to people’s waking

activity), or for a Dummy that equals 1 between 5:30pm and 10:00pm (to compare Earth Hour with the

immediate surrounding times).

12
the similar size as the EarthHour effect in this specification : EarthHour, PreEarthHour

and PostEarthHour show coefficients of –0.021, 0.010 and -0.010 (column 4,

corresponding to percentage changes of –2.10%, 0.99% and –0.98%)..

The interpretation of the results in Table III is that the lower consumption during

Earth Hour was driven primarily by factors common to the entire day, rather than people

actually turning off their lights during the main period of the event. Since the estimates

are for statewide electricity consumption, if it is assumed that only Sydney residents took

part, then the estimated drop measured in MW/h will be correct, but it will represent a

larger proportion of Sydney electricity use (relative to New South Wales). On the

assumption that only Sydney residents took part (62.9% of the New South Wales

population) and that total electricity use is spread evenly per capita, this would amount to

a Sydney-wide reduction of 3.34% of total electricity use under the column 3

specification.8 If it is assumed that that all of New South Wales took part, then 2.10%

represents the entire decline as a percentage of electricity use (over the relevant area).

Moreover, the omitted variable controlled EarthHour / EarthDay specification

produces estimates of the Earth Hour effect that are smaller than the estimated effects

from Table II. The poll-use estimate of 234 MW/h estimate based on Sydney

participation represents a 36-40% overstatement of true participation levels based on the

168-173 MW/h decline. This is consistent with Earth Hour non-participants either lying

to pollsters about their involvement, or refusing to participate in the poll at a higher rate

than Earth Hour participants.

8
Based on Australian Bureau of Statistics 2006 Population estimates,

13
5. Predicted and Actual Electricity Use Around Earth Hour

To show further that the apparent electricity decline during Earth Hour was in fact

mainly an effect over the whole day, Figures 1 and 2 present visual evidence that the

declines during Earth Hour were not unusual even within the day of March 31st, 2007.

Figure 1 plots actual electricity consumption versus the consumption predicted from

regressions of consumption on just the Control variables (including the interaction

variables).

[Insert Figure 1 here]

Consistent with Table III, the gap between predicted and actual electricity

consumption is large throughout the whole day, but is not especially large during the two

data points corresponding to Earth Hour. In other words, the claimed size of the drop due

to Earth Hour is not visible in the data. In order to better see the size of the decline during

the day, Figure 2 plots the difference between predicted and actual consumption as a

percentage of the predicted value.

[Insert Figure 2 here]

Based on the raw residuals from the base regression (rather than the estimated

effect of the Dummy variable, as was calculated earlier), the drop in electricity

14
consumption during Earth Hour was 6.36% (547 MW/h) lower than expected from 7:30-

8pm, and 6.18% (518MW/h) lower than expected from 8-8:30pm. On the other hand,

equivalently large percentage declines were observed virtually throughout the day,

starting as early as 5am, as seen in Figure 2. For drops this early, it suggests that the

variables causing the day-long decline were present even before sunrise. It is stretching

credibility to claim that changes in electricity use this early in the morning were due to

consumers reacting to Earth Hour at this point in the day, and consequently the entire day

of March 31st looks unusual, not just Earth Hour. In fact, directionally consistent with

intertemporal substitution, the only points in Figure 2 that look unusual are 6-6:30pm and

6:30-7pm, where actual use was closer to predicted values. This pattern makes it difficult

to attribute the early decline during the day to consumers switching off appliances early

in the day as part of Earth Hour, as it would require them to have turned off appliances

early, turned them back on from 6-7pm, and turned them off again at 7:30pm.

The second question is whether the declines from predicted values during Earth

Hour are unusual in terms of the whole sample. Ranking all periods in terms of

percentage declines from predicted values, the 7:30-8pm period was ranked only in the

86.17th percentile of largest unexpected declines (that is, just inside the top 14% of

biggest unexpected declines), while the 8-8:30pm period was ranked 86.86th percentile of

unexpected declines. Another way to test the significance of the results is to find out how

many days in the sample observed two consecutive periods with unexpected declines

greater than 6.36% (the higher of the percentage declines between the two Earth Hour

periods). There were a total of 831 days with two consecutive periods having unexpected

declines of more than 6.36%, out of a total of 3011 days. In other words, slightly more

15
than 27.5% of days in the sample have Earth Hour sized differences between predicted

and actual electricity consumption, making it highly problematic to simply attribute the

difference between predicted and actual consumption during Earth Hour as being due to

the event itself, and indicating that some method (such as a day fixed effect) is needed to

control for omitted variables.

5. Conclusion

The Earth Hour natural experiment presents a number of lessons for economics,

and in particular environmental economics. While it is unclear whether the aim of the

event was to actually reduce electricity significantly or simply raise awareness, the

reductions in electricity use nonetheless provide a valuable measure of the size of

discretionary household electricity use. The estimated drop in consumption is sensitive

to whether a day effect is included to control for possible omitted variables, but all

measures suggest that discretionary household electricity use is only an economically

small component of total electricity use. While I argue that the most reliable measure is

of a decrease in statewide electricity use of 2.10%, even excluding the day control

produces an estimated effect of only 6.33%.

At a basic level, this indicates that the vast majority of electricity use comes from

either industrial use, or household use that is more difficult to change in the short term.

This is consistent with Bartels and Fiebig (2000), who find that the largest components of

annual household electricity use are water heaters (2666-3885 kW/h depending on type),

refrigerators (993kW/h for households with only a single refrigerator), freezers (648

16
kW/h) and pool pumps (1311 kW/h), as opposed to light use (560 kW/h). Given that the

large focus of Earth Hour was on switching off the lights, the modest decrease in

electricity use reflects the fact that environmental policies targeting household light use

are unlikely to have a large effect. Indeed, the estimates in Bartels and Fiebig (2000)

would have suggested that Earth Hour would almost certainly have had a larger effect on

electricity use had it been billed as a ‘turn off your pool pump for an hour’ event, or a

‘take a cold shower for the day’ festival.

This large-scale natural experiment thus suggests that policies aimed at reducing

electricity use may be better targeted at areas other than residential light use.

Notwithstanding that changes here are easy to make, the aggregate effect of any change is

likely to be small. Turning off the lights altogether provides an upper bound for the likely

effect of energy efficient light bulbs, for instance.

Furthermore, the 36% discrepancy between estimated electricity reductions and

likely values implied by poll measures mean that caution must be taken in inferring

consumers preferences for environmental goods from poll results. The current findings

suggest that consumers feel pressure to overstate their preferences for environmentalism

when surveyed by pollsters. This is particularly troubling in this context, because the

question involved was not one of complex valuations of goods, but merely to truthfully

report ex post levels of involvement. Data limitations on poll responses mean that it is not

possible to distinguish between explanations involving non-participants either selectively

refusing to take part in the poll or simply overstating their level of involvement.

In either case, theoretical arguments in Levitt and List (2007) and previous

empirical results from the political science literature (Bishop and Fisher (1995)) indicate

17
that this problem could be reduced by designing surveys that subject respondents to less

scrutiny. This could involve (as in Bishop and Fisher (1995)) anonymous questionnaires

that are placed in sealed boxes upon completion, rather than phone interviews with an

interviewer. Nonetheless, the results in this paper present further evidence in favor of the

desirability of using market-based, revealed preference estimates of consumers’

environmental preferences, rather than the stated preference measures implied by polls

and surveys.

Indeed, the Earth Hour experiment suggests that consumers feel pressure to

overstate their preferences for environmental causes relative to which costly actions they

are actually willing to undertake, and that making changes to light and appliance use will

likely result in only small changes to total electricity consumption. Neither of twould

likely be heartening to the event’s or

18
References

Bartels, Robert and Denzel G. Fiebig (2000), “Residential End-Use Electricity Demand:

Results from a Designed Experiment”, Energy Journal, 21 (2), pp51-81

Bishop, George F. and Bonnie S. Fisher (1995), “Secret Ballots and Self-Reports in an Exit

Poll Experiment, The Public Opinion Quarterly, 59 (4), pp. 568-58

Converse, Philip E. and Michael W. Traugott (1986), “Assessing the Accuracy of Polls

and Surveys”, Science, 28 November 1986, 234 (4780), pp. 1094-1098

Griffin, James M., (1974), “The Effects of Higher Prices on Electricity Consumption”,

The Bell Journal of Economics and Management Science, 5 (2), pp515-539

Kellogg, Ryan and Hendrik Wolff (2007) “Does Extending Daylight Saving Time Save

Energy? Evidence From an Australian Experiment”, Working Paper, Center for the Study

of Energy Markets

Levitt, Steven D. and John A. List (2007), “What Do Laboratory Experiments Measuring

Social Preferences Reveal About the Real World?”, Journal of Economic Perspectives,

21 (2), pp153-174

19
List, John A. (2001) Do Explicit Warnings Eliminate the Hypothetical Bias in Elicitation

Procedures? Evidence from Field Auctions for Sportscards” American Economic Review, 91 (5),

pp. 1498-1507

List, John A. and Jason F. Shogrun, (1998), “Calibration of the difference between actual

and hypothetical valuations in a field experiment”, Journal of Economic Behavior &

Organization, 37, pp 193-205

Merkle, Daniel M, and Murray Edelman. (2002) “Nonresponse in exit polls. A

comprehensive analysis”. Survey nonresponse, Groves RM, Dillman DA, Eltinge JL,

Little RJA, eds. (2002) New York: Wiley. 243–257.

Narayan, Paresh K., and Russell Smyth, (2005), “The residential demand for electricity in

Australia: an application of the bounds testing approach to cointegration”, Energy Policy,

33 (4), pp 467-474

Saab, Samer, Elie Badrb and George Nasra (2001), “Univariate modeling and forecasting

of energy consumption: the case of electricity in Lebanon”, Energy, 26(1), pp1-14

Traugott, Michael W. and Vincent Price, (1992) “A Review: Exit Polls in the 1989 Virginia

Gubernatorial Race: Where Did They Go Wrong?”, The Public Opinion Quarterly, 56 (2),

pp.245-253

20
 Table I - Summary Statistics and Correlations

Panel A - Summary Statistics

Mean Std Dev Median Min Max N
ln(Consumption) 9.00 0.17 9.02 8.44 9.49 140604
Price 34.26 179.29 23.04 1.47 9909.03 140604
Temperature 65.26 8.99 66 39 113 140298
Wind Speed 12.00 6.27 10 0 53 140492

Panel B - Correlations
ln(Consumption) Temperature Price Wind Speed
ln(Consumption) 1 0.091 0.126 0.178
Temperature 0.091 1 0.071 0.195
Price 0.126 0.071 1 0.038
Wind Speed 0.178 0.195 0.038 1

This Table presents summary statistics and correlations for the log of New South Wales
electricity consumption (in MW/h), electricity retail price (in Australian $/MW/h) Sydney
Airport Temperature (in degrees Fahrenheit), and wind speed (in knots). The data are from
January 1, 1999 to April 1, 2007, with consumption and price available at a half-hourly
frequency, and temperature at an hourly (or less) frequency, with each consumption matched
to the most recent temperature.

21
 Table II - Estimated Drop in Electricity Consumption from End-Use Estimates

Estimated Bartels &

Fiebig (2000) Use Estimated Earth Estimated Earth
Participation per Household 7:30- Hour Electricity Drop Hour Electricity
Source (Sydney) 8:30pm (Watts) (Sydney, MW/h) Drop (NSW, MW/h)

Lights 53% 300 208.1 331.6

Television 17% 50 11.1 17.7
Computer 25% 50 16.4 26.1

Total 57% 400 235.6 375.4

This Table presents estimates of the likely effect of Earth Hour on New South Wales electricity use.
Participation data come from a poll by AMR Interactive. Estimates of household end-use come from
Bartels and Fiebig (2000) for the period 7:30pm-8:30pm, with the ‘Computer’ category being
approximated by television electricity use. The column ‘Sydney’ is calculated based on participation by
Sydney residents (4.28 million), while the ‘NSW’ column uses participation by all of New South Wales
(6.82 million). The Bartels and Fiebig (2000) mean estimate of 3.27 people per household is used to
covert population figures to household figures.

22
 Table III - Regressions of Electricity Demand on Earth Hour Variables
Independent Variable is ln(NSW Electricity Consumption in MW/h)

Intercept 8.936 *** 8.937 *** 8.937 *** 8.937 ***

(0.006) (0.006) (0.006) (0.006)
Earth Hour Dummy -0.065 * -0.021 -0.021
(0.035) (0.035) (0.035)
Earth Day (24 hours) Dummy -0.046 *** -0.045 *** -0.045 ***
(0.007) (0.007) (0.008)
(Earth Hour-1) Dummy 0.010
(0.035)
(Earth Hour + 1) Dummy -0.010
(0.035)

Controls
(Time, Day, Month,Year, DL
Saving, Price, Weather) Yes Yes Yes Yes

Interactions
(Day, Month, DL
Saving,Weather)*Time Yes Yes Yes Yes

Total # Control Variables 1256 1256 1256 1256

R2 0.9170 0.9171 0.9171 0.9171

N 140604 140604 140604 140604

This Table presents the results of regressions of New South Wales electricity consumption on a large
number of controls and variables to measure the effect of Earth Hour, March 31st 2007 from 7:30pm-
8:30pm. The dependent variable is the log of New South Wales electricity consumption (in MW/h).
Independent variables are dummy variables that equal 1 during the specified period and zero otherwise.
‘Earth Hour Dummy’ covers observations between 7:30pm and 8:30pm on March 31st 2007, ‘Earth Day
(24 Hours) Dummy’ covers all observations on March 31st 2007, ‘(Earth Hour-1) Dummy’ covers all
observations between 6:30pm and 7:30pm on March 31st 2007, and ‘(Earth Hour+1) Dummy’ covers all
observations between 8:30pm and 9:30pm on March 31st 2007. A large number of additional controls are
included in the regression. Fixed effects are included for time of day (in half hour periods), day of the
week, month, year, whether daylight saving was currently occurring, as well as controls for electricity
retail price (in Australian dollars), Quadratic Temperature (that is, (T-65)^2, where T is the Sydney
Airport Temperature in degrees Fahrenheit) Cloud cover, and wind speed. Interactions are included for
the half-hour time dummy variables with Day of the Week, Month, the Daylight Saving Dummy, and all
the weather variables. The data are taken at a half-hourly frequency from January 1, 1999 to April 1,
2007, top value is the coefficient, and the bottom value in parentheses is the standard error associated
with that coefficient. *, ** and *** indicate significance at the 10%, 5% and 1% level respectively from
a t-statistic.

23
 Figure 1

Predicted vs. Actual Electricity Consumption Around Earth

Hour
9500

Actual
9000
Predicted
Earth Hour
Consumption (MW/h)

8500

8000

7500

7000

6500

6000
0 2 4 6 8 10 12 14 16 18 20 22 24
Hour of Day

This figure presents predicted and actual values of New South Wales electricity consumption (in MW/h)
around Earth Hour, on March 31st 2007 from 7:30pm-8:30pm. Predicted values are estimated from the
equation: ln(Consumptiont ) = α + β1Controls + ε t . Fixed effects are included for time of day (in half hour
periods), day of the week, month, year, whether daylight saving was currently occurring, as well as controls
for electricity retail price (in Australian dollars), Quadratic Temperature (that is, (T-65)^2, where T is the
Sydney Airport Temperature in degrees Fahrenheit) Cloud cover, and wind speed. Additional controls are
included for interactions of the half-hour time dummy variables with Day of the Week, Month, the Daylight
Saving Dummy, and all the weather variables. The y-axis is Electricity consumption in MW/h. The x-axis is
the hour of the day, measured in 24-hour time. The red line is predicted consumption, the blue line is actual
consumption, and points in black correspond to those during Earth Hour.

24
 Figure 2

Percentage Difference Between Actual and Predicted

Electricity Consumption Around Earth Hour

6
% Change
4 Earth Hour

2
Percent Difference

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

-4

-6

-8

-10
Time of Day

This figure presents the difference between predicted and actual values (as a percentage of predicted
values) of New South Wales electricity consumption around Earth Hour, on March 31st 2007 from
7:30pm-8:30pm. Predicted values are estimated from the equation:
ln(Consumptiont ) = α + β1Controls + ε t . Fixed effects are included for time of day (in half hour
periods), day of the week, month, year, whether daylight saving was currently occurring, as well as
controls for electricity retail price (in Australian dollars), Quadratic Temperature (that is, (T-65)^2,
where T is the Sydney Airport Temperature in degrees Fahrenheit) Cloud cover, and wind speed.
Depending on specification, additional controls are included for interactions of the half-hour time
dummy variables with Day of the Week, Month, the Daylight Saving Dummy, and all the weather
variables.. The y-axis is percentage change in electricity consumption, in percent. The x-axis is the
hour of the day, measured in 24-hour time.

25