Behavioural Nudges for Water Conservation: Experimental Evidence from Cape Town

Abstract

The applied behavioural literature has found green nudges to be effective mechanisms for reducing residential water and energy consumption. This study estimates the impact of several such behavioural interventions on residential water consumption in Cape Town, South Africa-a setting characterised by both water austerity and extreme income inequality. In this setting, in addition to testing the ability of behavioural messages to encourage water conservation in times of water scarcity, we are able to examine whether average treatment effects are heterogeneous across the income spectrum. Ultimately, the behavioural messages are found to have a significant effect on water saving: resulting in an average reduction of water usage of between 0.6%-1.3% across the various treatments. As water restrictions intensified throughout the intervention period, these decreases in consumption represent decreases over and above those attributed to rising tariffs and increasingly stringent physical restrictions. In terms of the most effective message, the results indicate that publicly recognising water conservation or appealing to households to act in the public interest are the most effective motivators for water conservation. Treatment effects were found to vary across income groups. Specifically, wealthier households were more responsive to social incentives such as public recognition of water conservation, appeals to the public good and social norm comparisons.

Full text

Behavioural Nudges for Water Conservation:
Experimental Evidence from Cape Town
Kerri Brick
Samantha De Martino
Martine Visser
University of Cape Town
World Bank
University of Cape Town
December 2017
FINAL DRAFT MANUSCRIPT
Abstract – The applied behavioural literature has found green nudges to be effective mechanisms for
reducing residential water and energy consumption. This study estimates the impact of several such
behavioural interventions on residential water consumption in Cape Town, South Africa – a setting
characterised by both water austerity and extreme income inequality. In this setting, in addition to testing
the ability of behavioural messages to encourage water conservation in times of water scarcity, we are
able to examine whether average treatment effects are heterogeneous across the income spectrum.
Ultimately, the behavioural messages are found to have a significant effect on water saving: resulting in
an average reduction of water usage of between 0.6%-1.3% across the various treatments. As water
restrictions intensified throughout the intervention period, these decreases in consumption represent
decreases over and above those attributed to rising tariffs and increasingly stringent physical restrictions.
In terms of the most effective message, the results indicate that publicly recognising water conservation or
appealing to households to act in the public interest are the most effective motivators for water
conservation. Treatment effects were found to vary across income groups. Specifically, wealthier
households were more responsive to social incentives such as public recognition of water conservation,
appeals to the public good and social norm comparisons.
Index Terms—Behavioural Economics, Nudges, Water Conservation

ACKNOWLEDGEMENTS
The authors would like to thank the Water Research Commission for partially funding this study. We
would further like to thank the Norwegian Research Council, the National Research Foundation in South
Africa and the Environment for Development (EfD) Initiative for additional funding to scale up the study
over a six-month period. We would also like to thank the African Climate & Development Initiative
(ACDI) for funding Martine Visser’s Research Chair. Our deep gratitude to the Reference Group of the
WRC Project (Mr Bhagwan, Prof. Rivett, Dr Ziervogel, Assoc. Prof. Jansen, Dr T Booysen, Ms Hay, Ms
Pereira, Ms Davis, Mr Tyrrell and Mr Millson) for the assistance and the constructive discussions during
the duration of the project. We would also like to thank the City of Cape Town’s Water Demand
Management Departments, the Billing Department and the Utilities Directorate for support throughout the
study.

1. INTRODUCTION
Cape Town is reeling from the onslaught of sustained drought. Dam levels stand at 37% of capacity
(December 2017), with the last 10% being unusable (City of Cape Town 2017). The city is barrelling into
the hot summer season with dams at a fraction of capacity. As part of their response to the water crisis,
the local municipality has implemented increasingly stringent water restrictions, which include physical
restrictions on usage as well as higher tariff rates. Under the current (level 5) restrictions, individuals are
mandated to use less than 87 litres of water per person per day with a maximum ceiling of 10 kl per
household. Water usage for non-essential purposes – like irrigation – is prohibited.
Against this background, this study evaluates the impact of eight behavioural messages in reducing
residential water consumption.1 Behavioural messages were sent to domestic waters over a six-month
period – starting in November 2015. While Cape Town is now besieged by drought in December 2017,
this crisis was only beginning to unfold some two-years prior at the start of the behavioural intervention.
In this setting, this intervention provides a test of the ability of behavioural messages to encourage water
conservation in times of water austerity and, furthermore, reinforce efforts by local municipalities (using
more conventional instruments) to reduce water consumption.
South Africa is characterised by extreme levels of income inequality. In this setting of water austerity and
income inequality, green nudges might be an especially useful adjunct to more traditional demand-side-
management (DSM) measures. Firstly, in contrast to more traditional DSM tools, green nudges do not
feel punitive and regressive to poor households (who are not able to obviate the hardships associated with
higher tariffs and physical restrictions) (Datta et al. 2015). Non-monetary incentives can thus be
introduced across the income spectrum. Secondly, as water consumption is subsidised for low-income and
indigent households, these households (who don’t internalise financial incentives) might be more
responsive to non-monetary interventions. Against this background, using a very rich dataset of over
400 000 residential users, we are able to examine how treatment effects differ across income groups.
In terms of the mechanics of the intervention, around 400 000 residential households participated in eight
behavioural treatments over the period November 2015 to April 2016. The messages were sent as inserts
with the monthly utility bill. The messages are classified into two distinct groups:
1 The applied behavioural literature has demonstrated that nudges are a useful mechanism with which to reduce
residential electricity and water consumption (Allcott & Mullainathan 2010; Allcott 2011; Ferraro & Price 2013;
Ayres et al. 2013; Allcott & Rogers 2012; Ferraro and Price, 2011; Ayres et al., 2012; Allcott and Rogers, 2014).

The first group of messages promotes water conservation by addressing informational failures around
price and usage of water.
Informational failures, which are prevalent in a water context, likely play a role in inefficient resource
usage. In the case of usage, water consumption is often unobservable (toilet flush, washing machine,
irrigation, and leaks) and, even in cases where usage is visible, it is not always easily quantifiable, for
example, it is not easy to estimate total water usage during a shower. Given that quantifying the water
used by appliances, toilets, irrigation systems, showers etc. can be both complex and costly, water usage
becomes a de facto unobservable characteristic, receiving less weighting than other preferences (adapted
from Ramos et al. 2015). With respect to price, consumers are not responsive to price signals when price
information is unclear and the pricing system is complex (Ramos et al. 2015; Chetty et al. 2009; Gaudin
2006). This is particularly true of water where (i) conventionally metered consumers pay for water ex-
post and not at the instance of usage and, (ii), the tariff structure used by municipalities are often complex
and nonlinear – such as the inclining block tariff system used by the local municipality where marginal
prices increase with consumption. While elasticity of demand for water has typically been found to be
inelastic (Olmstead et al. 2007), recent literature suggests that consumers are simply inattentive to price
changes and thus fail to respond appropriately (Chetty et al. 2009; Datta et al. 2015; Gaudin 2006). For
example, it has been found that consumers underreact to non-salient taxes and shipping costs (Chetty et
al. 2009; Hossain & Morgan 2006). Anecdotal evidence from focus groups held in Cape Town prior to
the roll-out of the behavioural intervention indicates that consumers, from across the income spectrum,
are generally not aware of the quantity of water they use, the stepped tariff structure or tariff rates.
With this in mind, when operating in an environment with complex (non-salient) pricing and lack of
transparency around usage, households might consume outside of their private optimum. In such a case,
sending households regular feedback on both price and consumption - and making both more salient -
may provide some private signal whereby utility maximizing individuals get closer to their private
optimal water use. For example, Gaudin (2006) finds that elasticity of demand increases by at least 30%
when price information is provided on the bill.
Given this background, the first group of messages are designed to address these informational failures
around usage and price. The tips treatment demonstrates how to reduce water and quantifies the water
saving associated with each water-saving action. The tariff graph treatment provides a visual break-down
of the nonlinear tariff structure and situates the household’s consumption within six tariff blocks. The

financial gain treatment replicates the graphic in the tariff graph mailer and additionally quantifies the
projected monthly and annual savings from reducing consumption and moving into the preceding tariff
block. Given the growing traction of loss aversion (Kahneman & Tversky 1979; Tversky & Kahneman
1992), we also consider whether a loss or gain framing has a greater impact on water conservation. The
loss treatment is identical to the gain treatment, but quantifies the financial dissaving of not reducing
consumption and moving to a lower tariff block.2
The second group of messages promotes water conservation via social incentives and appeals to the
public good.
Social norm messaging is increasingly seen as a useful mechanism for promoting pro-environmental
behaviour. Allcott (2011) and Ferraro & Price (2013) find that social-norm based appeals reduce US
energy and water consumption by 2% and 4.8%, respectively. The literature indicates that the social
comparisons inherent in social-norm messaging facilitates social (observational) learning about the
households’ privately-optimal level of water usage (Allcott 2011). In the social norm treatment, a
household’s consumption is compared to the average household in their neighbourhood.
Social pressure can be leveraged to drive pro-social behaviour – particularly if social recognition is
provided for positive (do-gooding) behaviour. For example, in a lab setting, experiments have
demonstrated that revealing identity increases public-good contributions (Andreoni & Petrie 2004; Rege
& Telle 2004). Randomized field experiments have used social recognition to increase public good
contributions in varying public good contexts, for example, charity contributions, voter turnout and blood
donations (Gerber et al. 2008; Lacetera et al. 2011). In the context of pro-environment conservation,
Yoeli (2009) examines take-up of energy-saving technology in California. Customers whose decisions are
made public are 1.5% more likely to sign up for the technology as compared to those customers whose
decisions remain anonymous. In the intrinsic motivation treatment, households are asked to help the City
save water by supporting a City-led water saving initiative and reducing consumption by 10% over the
2 Loss aversion, where losses are felt more acutely than gains of the same magnitude, is a core component of Prospect
Theory (Kahneman & Tversky 1979; Tversky & Kahneman 1992), a theory of decision making under risk that is now
the leading alternative to Expected Utility Theory. There is a large experimental literature consistent with loss
aversion. Laboratory and field experiments, consisting of lottery-choice tasks for real monetary stakes, find that
participants are more risk averse when real financial losses are at stake (Wik et al. 2004; Yesuf & Bluffstone 2009;
Harrison & Rutström 2009; Tanaka et al. 2010; Liu 2013). Demonstrating that loss aversion impacts on individuals’
decision-making outside of the lab, Liu (2013) finds that farmers who are more loss averse adopt new farming
technologies later.

study period. In the social recognition treatment, households that succeed in reducing their usage by 10%
are publicly recognized on the City's website.
The final treatment is framed around the public-good dilemma synonymous with the current water crisis.
Water scarcity is a public good dilemma: while a significant aggregate reduction in water consumption is
needed to obviate the impact of the drought, the benefits of water saving are shared equally by all
households irrespective of individual contribution, creating an incentive to free ride (Hasson et al. 2010;
Brekke et al. 2008). However, evidence from lab experiments shows that, while the dominant strategy in
linear public good games is for each player to contribute nothing, subjects make positive but suboptimal
contributions to public goods (Cherry et al. 2005). The public good treatment appeals to all households to
conserve as much water as possible. This public good appeal might alter the moral cost of water usage or,
alternatively, generate conditional cooperation whereby households alter their usage in the belief that
others will do the same (Allcott 2011; Gächter & Herrmann 2009).
In terms of the results, the social recognition and public good framings outperformed the other messages.
We thus find public recognition of water-conservation efforts and appeals to the public good to be the
most effective motivators of pro-environmental behaviour. More broadly, over the intervention period, all
treatments successfully induced a reduction in household consumption. Reductions ranged from 0.6% for
the tips treatment to 1.3% for the social recognition and public good treatments. Furthermore, social
incentives were found to be particularly effective among wealthier households. In particular, in the fifth
quintile (wealthiest households), the social norm treatment induced a reduction of 1.2%, with the
treatment effect increasing to 1.9% for the public good message and 2.3% for the social recognition
treatment. An important implication of these findings is the conclusion that behavioural messages are a
useful mechanism even under conditions of water scarcity and can promote water saving on top of more
conventional DSM tools. Our findings suggest that behavioural nudges are a useful addition to the
toolbox of DSM instruments available to policy makers and could realise significant savings at local and
national level if implemented during strategic periods.
The paper proceeds as follows: the study design and sample are described in Section 2. Section 3 details
the randomization procedure. Section 4 provides an analysis of pre-intervention consumption and
considers whether treatment and control groups have comparable trends in the pre-intervention period.
The data and estimation procedure is described in Section 5. Experimental results are presented in Section
6, while Section 7 concludes.

2. EXPERIMENTAL SETTING AND DESIGN
The study was conducted in Cape Town, South Africa, in collaboration with the local municipality, the
City of Cape Town. The first mailers were delivered in November 2015 as inserts with the monthly
municipal bill. Messages were printed on blue paper to be as distinctive as possible. Households received
inserts on a monthly basis between November 2015 and April 2016. As will be discussed shortly, one of
the treatments asked households to reduce consumption by 10% over the intervention period. Households
in this treatment received their feedback in May 2016.
The intervention period (November 2015 – April 2016) coincided with the introduction of more stringent
water restrictions. While the City permanently imposes level 1 water restrictions (to encourage a 10%
water savings), level 2 water restrictions were implemented from 1 January 2016. These included tighter
restrictions around water usage (for example, irrigation to take place on certain days and between certain
hours) as well as a tariff increase. The water restrictions were pre-empted by a city-wide media campaign
led by the City of Cape Town. The campaign was initially around the water shortages (and the need to
conserve water) but, once level 2 water restrictions were approved in December 2015, focused on the
impending tariff increase and water shortages.
2.1. TREATMENTS
Message 1: Tips
The tips treatment consisted of a one-page tip sheet which provided information on ways to reduce usage.
This information was adapted from the Water By-Law (City of Cape Town 2010) and the Smart Living
Handbook (City of Cape Town 2011) (City of Cape Town sources that are widely available to the public).
In addition to providing information on how to save water, the leaflet also quantified (where possible) the
water saving associated with each water-saving action (savings were contextualized for a family of four).
Following Allcott (2011), the tips were divided into quick fixes and smart purchases. The tip sheet is
replicated in Appendix A.1. The remaining seven mailers all included the tip sheet on the back page.
Message 2: Tariff graph
The second treatment augmented the tips treatment by providing a graphical breakdown of the bill. As
evident from Appendix A.2, the insert provided information on both tariff rates and nonlinear pricing
structure, and, furthermore, situated the household’s consumption within the six tariff blocks.
Message 3: Financial gain

The financial gain treatment replicated the visual from the tariff graph treatment and additionally
provided information around the potential financial savings (gain) from moving into a lower tariff block.
Given the complex tariff structure and households’ uncertainty around the volume of water consumed
over the month, the purpose of the financial framing was to make the financial gains from saving water
more transparent. The insert is provided in Appendix A.3.
Message 4: Financial loss
The loss treatment replicated the information from the gain framing, but framed it as a financial loss
(provided information around the financial dissaving from not moving into a lower tariff block). In this
way, the link between inefficient usage and financial cost was made explicit. The insert is replicated in
Appendix A.4.
Message 5: Social Norm message
The social norm message (appendix A.5) graphically compared the household's average daily water
consumption to that of the average for the neighbourhood. This comparison was presented in both a
descriptive text and a bar graph.
Message 6: Intrinsic motivation
This treatment asked households to support a City-led water-saving initiative by reducing their water
consumption by 10% over the study period. Households received an initial announcement message which
noted that “as you used X kl this month, you need to keep your monthly consumption around Y kl.” The
consumption value (X kl) was derived from the current bill. Thereafter, this consumption value remained
unchanged and households were reminded each month that their target level of consumption is X kl. The
insert is provided in appendix A.6.
Message 7: Social recognition
As with Message 6 (intrinsic motivation), this treatment encouraged households to reduce their water
consumption by 10% to support a water-saving initiative that was recently launched by the City.
However, the message further stated that successful households would be publicly recognised on the
City’s website. In comparison to message 6, where the motivation to conserve water is internal, this
framing explores whether the opportunity to be socially recognised as one of the best performers (water
savers) promotes conservation. If people desire to appear to their society that they are doing good deeds,
then it follows that the opportunity to be socially recognised promotes conservation. The insert is
provided in appendix A.7.

Message 8: Public good
The message highlighted the public good context by encouraging households to voluntarily reduce their
water consumption in order to reduce the stress on water resources and prevent future water restrictions.
See Appendix A.8.
2.2. SAMPLE
The sample includes domestic water users living in free-standing houses with access to an uncontrolled
water supply that is metered by a credit meter. By focusing on free-standing houses only, we avoid
households which are served by bulk water meters (as is the case with blocks of flats). Furthermore,
households that receive an electronic bill are excluded from our sample amid logistical difficulties in
sending them the inserts. Finally, so as not to put low-income households in a vulnerable state, only
households consuming more than the six-kilolitre (at the time) free monthly allocation were eligible to
receive a message in any particular month (i.e. households in the first tariff block are not messaged).
Power calculations were conducted using City of Cape Town consumption data for the period November
2014 to April 2015. We chose to use the months which coincide with our study to allow for seasonality
effects (consumption typically increases in the summer months). In addition to excluding individuals
consuming below six kilolitres in a given month, we excluded the 95th percentile to control for outliers.
We include two power calculations: one where we look at the mean consumption over the period with an
unbalanced panel and one where we use the balanced panel. With respect to the unbalanced panel, we can
detect a 1.5% change in means per treatment with a minimum sample size of 18 579 per treatment arm
(with 80% power). With respect to the balanced panel (a sample of households whose consumption we
observe in each month), we are able to detect a 1.5% change in means per treatment with a sample size of
14 104 households per arm (with 80% power).
October 2015 consumption data was used to randomize households into treatment and control groups.
Importantly, when randomizing, households identified as having indigent status were not allocated to the
tariff graph, financial gain and financial loss treatments. As these households receive a subsidy on their
bill, the information on the inserts would be inaccurate (for example, specifying a total bill and
monthly/annual savings which do not account for the subsidy). In terms of the mechanics, before
randomizing, households in tariff block 1 and indigent households were excluded from the sample. The
remaining sample was stratified on both suburb and tariff block and then randomized into control and
treatment groups. Thereafter, the same procedure was followed with the previously excluded indigent

households, who were then randomly allocated into the control group and tips, social norm, intrinsic
motivation, social recognition and public good treatments. In all subsequent analysis, these indigent
households are excluded from the sample.
The sample sizes allocated to the control and treatment groups in October 2015 (for the first wave of
mailers sent in November 2015) are reflected in column I of Table 1. As will be discussed, there are
several reasons a treated household is not eligible to receive an insert in a particular month – for example,
their bill reflects an estimated reading, they have slipped into tariff block 1 or have an extensive billing
period. With this in mind, a number of households that were allocated to the intrinsic motivation and
social recognition treatments during the October 2015 randomisation did not receive an insert in
November 2015 (the first month of the intervention). These households, who were not eligible to receive
an insert in the month of November, were reallocated equally and randomly across the remaining
treatments (tips, tariff graph, financial gain, financial loss, social norm and public good) (and received
their first insert in December 2015, the second month of the intervention (assuming they were eligible in
that month)). The updated treatment allocation, for the December 2015 – April 2016 intervention period,
is reflected in column II.
Table 1. Treatment allocation in October 2015 & reallocation in November 2015
I
II
Households
(October 2015)
Households
(November 2015)
Control
48 206
51 113
Tips
49 928
52 833
Tariff graph
34 000
36 888
Financial gain
33 687
36 584
Financial loss
34 073
36 990
Social norm
40 001
33 043
Intrinsic motivation
40 058
32 724
Social recognition
44 174
38 557
Public good
44 421
47 316
Total
368 548
366 048
As mentioned, while the numbers in Table 1 reflect the total number of households allocated to each
treatment, in any given month, a household is not eligible to receive an insert if: the household has an
estimated meter reading (so as not to give households inaccurate information), has a billing period of
greater than 35 days (which is usually indicative of a problematic bill) and/or moves into the lowest tariff
block.

Utility bills are delivered in twenty batches (portions) over the course of the month. During November
2015 (the launch month), because of an operational glitch at the printers, a subsample of eligible
households in the social recognition and public good treatments did not receive inserts.3 These
households remained in their allocated treatments and received their first mailers in December 2015. With
respect to the social recognition treatment, December consumption (reflected in the December bill) was
used as the reference value.4 Likewise, households in the public good treatment that missed their
November 2015 insert, received their first insert in December 2016.
2.3. TIMELINE OF ROLL-OUT
As previously mentioned, the first messages were delivered in November 2015, as inserts with the
monthly municipal bill. Households received the inserts monthly between November 2015 and April
2016, during the warmer summer period when water usage typically increases:
Households in the intrinsic motivation and social recognition treatments were asked to reduce
consumption by 10% over the intervention period. These households were provided feedback in their
May 2016 bill (Appendices A.6.1-A.6.4 and A.7.1-A.7.4). Households in the social recognition treatment
that successfully met the target (reduced consumption by 10%) qualified to have their information posted
on the City’s website. These households were given two months to opt out (the feedback mailer provided
a contact number should the household members prefer not to have their information published).
Thereafter, the names were published on the City’s website.
3 English speaking households in portions 1-5 in the social recognition treatment and all households in portion 1-4
in the public good treatment.
4 As evident from appendix A.7, households are asked to reduce their consumption by 10% between November and
April. Eligible households (not subject to the operational glitch) receive an announcement in November 2015
(specifying their target level of consumption), monthly reminders between December – March 2016 and, finally,
feedback in May 2016 (Appendix A.7.1 – A.7.2). For example, a household might have received the following
announcement in November 2015: as you used 30 kl this month, you need to keep your monthly consumption around
27 kl. The reference consumption value (of 30 kl) is derived from the November bill which accompanies the insert.
Thereafter, households receive a reminder in December, January, February and March about their target level of
consumption. For example: as you used 30 kl in the month when the initiative was launched, you need to keep your
monthly consumption around 27 kl. Likewise, eligible households that did not receive an insert in November 2015
(because of the technical glitch) received an identical announcement with their December bill. While the text of the
insert remained the same, the consumption value referenced in the bill corresponded to the accompanying December
bill. These households received monthly reminders in January, February, March and April 2016. Thereafter, both
groups received feedback in May 2016.

3. RANDOMISATION
Using October 2015 consumption data (the data used for the randomization), we test whether the
treatment and control groups are balanced in terms of several demographic characteristics, namely:
monthly consumption, daily average consumption, property value, number of billing days (over the
month) and tariff block (in October 2015). Table 2 provides the descriptive mean and standard deviation
for all treatments and control as well as the p-value from the t-tests of equal means (H0: equality of
means). As evident from Table 2, the treatment and control groups are balanced in terms of monthly
consumption and daily average consumption. As we do find significant effects for property value and
billing period, we control for these variables in all regressions. Property value is controlled for using
quintile property value dummies. For billing period, we include the number of billed days per month. As
mentioned, the sample was stratified on both suburb and tariff block during the randomization. As such,
we use tariff block and suburb fixed effects in all regressions (Bruhn & McKenzie 2009).
Table 2. Demographic characteristics by treatment group
Treatment
mean
Treatment sd
Control
mean
Control sd
Ttest
(p-value)
Obs.
Monthly
Tips
21.96
39.91
22.73
102.56
0.358
296927
Graph
22.31
35.83
22.73
102.56
0.272
296927
Gain
22.73
53.29
22.73
102.56
0.532
296927
Loss
23.04
58.30
22.73
102.56
0.803
296927
SN
20.83
42.75
22.73
102.56
0.090
296927
IM
22.11
108.12
22.73
102.56
0.701
296927
SR
21.33
55.84
22.73
102.56
0.149
296927
PG
22.55
45.84
22.73
102.56
0.962
296927
Daily average
Tips
0.67
0.95
0.67
2.25
0.634
296927
Graph
0.71
1.23
0.67
2.25
0.216
296927
Gain
0.70
0.98
0.67
2.25
0.966
296927
Loss
0.71
2.06
0.67
2.25
0.385
296927
SN
0.64
0.74
0.67
2.25
0.243
296927
IM
0.65
0.96
0.67
2.25
0.869
296927
SR
0.65
0.89
0.67
2.25
0.443
296927
PG
0.68
0.90
0.67
2.25
0.062
296927
Property value
Tips
735986.76
1130441.20
718659.14
1065660.40
0.041
274965
Graph
1003011.50
1297181.70
718659.14
1065660.40
0.041
274965
Gain
992206.73
1276251.00
718659.14
1065660.40
0.267
274965
Loss
996404.17
1278483.90
718659.14
1065660.40
0.170
274965

SN
679493.72
995739.89
718659.14
1065660.40
0.994
274965
IM
679318.72
1000443.00
718659.14
1065660.40
0.906
274965
SR
714320.60
1110126.30
718659.14
1065660.40
0.208
274965
PG
695613.65
1082580.70
718659.14
1065660.40
0.006
274965
Billing days
Tips
33.13
25.16
33.43
27.55
0.128
296927
Graph
32.43
22.22
33.43
27.55
0.021
296927
Gain
32.75
24.04
33.43
27.55
0.568
296927
Loss
32.67
23.87
33.43
27.55
0.074
296927
SN
32.49
22.34
33.43
27.55
0.000
296927
IM
32.73
24.32
33.43
27.55
0.002
296927
SR
32.79
23.45
33.43
27.55
0.000
296927
PG
33.26
27.48
33.43
27.55
0.050
296927
Tariff block
Tips
3.17
1.03
3.15
1.01
0.001
296927
Graph
3.27
1.03
3.15
1.01
0.000
296927
Gain
3.26
1.02
3.15
1.01
0.012
296927
Loss
3.28
1.03
3.15
1.01
0.000
296927
SN
3.14
1.01
3.15
1.01
0.212
296927
IM
3.16
1.01
3.15
1.01
0.001
296927
SR
3.15
1.02
3.15
1.01
0.251
296927
PG
3.19
1.07
3.15
1.01
0.000
296927
Notes: Means, standard deviations and p-values calculated using October 2015 consumption data and November 2015
treatment allocation (Table 1, column II). October 2015 consumption data used for randomizing into treatment and control.
4. PRE-INTERVENTION ANALYSIS
One of the key identifying assumptions of the difference-in-difference model is that water usage trends
would be the same in the control and treatment groups in the absence of the treatments and that the
intervention induces the deviation from this common trend (Angrist & Pischke 2008). In this section, we
consider whether treatment and control groups have comparable pre-intervention trends.
Given seasonality in water consumption (to be discussed shortly), we use the same month of the
preceding year as baseline consumption. Figure 1 graphically depicts water use trends in the pre-
intervention period (December 2014 – April 2015). Means are reported in Table 3. From the figure, it is
evident that there is a seasonal component to water usage: specifically, mean monthly consumption is
higher in the warmer summer months. The increase over the summer months is due to both an absolute
increase in consumption commensurate with both warmer weather and the holiday season (for example,
filling pools and increasing the frequency of irrigation) as well as an increase in the billing period

(number of days billed in a month) as the City of Cape Town has a lower staff contingent over the holiday
season (pers. comm.). Because of this seasonal component to water usage, we use month fixed effects in
all specifications to control for seasonal effects. However, while there is an element of seasonality in
water use, this seasonal trend is common across treatment and control groups, lending credibility to the
parallel trends assumption needed for difference-in-difference estimation.
We test whether the treatment and control groups are balanced in the pre-intervention period across the
now familiar demographic characteristics: monthly consumption, daily average consumption, number of
billing days (over the month) and property value. The estimates reported in Table 4 are based on a
regression of the outcome variable/characteristic as the dependent variable and dummy variables for the
four treatment groups (omitting the control group) as explanatory variables (Bhanot 2015). Following
Bruhn & McKenzie (2009), we control for stratification by including tariff block and suburb dummy
variables in the regressions.5 The table indicates that the treatment and control groups are balanced on key
characteristics such as consumption, daily average consumption and property value. We do find
significant effects for the social norm treatment for billing period – however, as previously mentioned, we
control for this in all subsequent regressions.
5 As households frequently changed tariff blocks throughout the course of the study, we designated the household
the tariff block the household was in in October 2015 when the randomization was conducted. The same procedure
was used for suburb.

Figure 1. Mean monthly consumption by treatment group

Figure 2. Mean billing days by treatment group
Table 3. Pre-intervention means
Baseline means
Dec 2014
22.76
Dec 2014 – April 2015
25.04
1st quintile
19.07
2nd quintile
19.61
3rd quintile
22.56
4th quintile
26.28
5th quintile
37.05
Short-run baseline mean calculated using December 2014; Long(er)
run and quintile averages calculated for the period December 2014 –
April 2015.

Table 4. Balance test regressions for pre-intervention period (December 2014-April 2015)
I
II
III
IV
Monthly
consumption
Daily
average
Property value
Billing
days
(kl)
(kl)
(ZAR)
Tips
0.058
0.000
-1884.255
0.002
0.084
0.003
4254.426
0.008
Graph
0.076
-0.001
-332.013
-0.003
0.082
0.004
4385.315
0.008
Gain
0.015
0.001
-3733.279
-0.011
0.083
0.004
4019.783
0.008
SN
0.083
0.003
982.063
-0.024***
0.095
0.004
4187.096
0.009
IM
0.162*
0.004
3483.136
-0.002
0.096
0.004
4714.47
0.008
SR
0.058
0.001
1690.624
-0.002
0.089
0.004
4373.004
0.009
PG
0.005
0
-7261.963
-0.009
0.081
0.003
4478.801
0.008
Constant
15.645***
0.525***
930410.261***
30.310***
0.129
0.005
7887.119
0.011
R-squared
0.417
0.196
0.693
0.016
Observations
849555
849555
797635
849555
Treated
744878
744878
699164
744878
Control
104677
104677
98471
104677
Tips
107005
107005
100537
107005
Graph
107600
107600
100937
107600
Gain
106268
106268
99806
106268
Loss
107235
107235
100540
107235
SN
72312
72312
67929
72312
IM
71882
71882
67690
71882
SR
85973
85973
80571
85973
PG
86603
86603
81154
86603
Fpvalue
Notes: Regressions include tariff block and suburb fixed effects. Standard errors are
clustered at the suburb level. Regressions are run for the pre-intervention period of
December 2014 – April 2015.
While figure 2 suggests that treated and control households have comparable pre-intervention trends,
following Abramitzky & Lavy (2014), we use pre-intervention data from December 2014 to April 2015 to
determine whether the treatment and control groups have differential time trends with respect to water

usage. The estimated results are reflected in Table 5. Panel A reflects the results of a constant linear time
trend model which allows for an interaction of the trend with the treatment indicator, while, in Panel B,
the linear time trend variable is replaced by a series of month dummies as well as an interaction of the
treatment indicator with each of these time dummies (Abramitzky & Lavy 2014). While the results from
both models confirm the presence of a time trend with respect to water usage, in general this trend is
identical for treated and non-treated households.
The results in Panel A suggest that, on average, water consumption decreased by 263-266 litres per month
between December 2014 and April 2015. This decline in consumption is expected given that consumption
generally peaks in January/February and declines as South Africa moves into the rainy season. As evident
by the interaction term (Treatment X Pre-trend), this trend does not differ significantly between treatment
and control groups (except for the PG treatment at the 10% level).
The estimates in the dummies model (Panel B) are largely consistent with the linear trend model: most of
the interaction terms (of the treatment indicator with the month dummies) are insignificant except for
February and March 2015 where consumption in treatment households is significantly different relative to
control. Overall the models indicate that treatment and control households followed the same trend with
respect to water usage in the year preceding the intervention (December 2014-April 2015). However, as
water usage trends do differ across some of the treatment and control groups (specifically, in February
and March 2015), we control for trend in the regressions.

Table 5. Difference in the time trend of water usage in treatment and control in Dec 2014-Apr 2015
I
II
III
IV
V
VI
VII
VIII
Monthly
consumption
Monthly
consumption
Monthly
consumption
Monthly
consumption
Monthly
consumption
Monthly
consumption
Monthly
consumption
Monthly
consumption
(kl)
(kl)
(kl)
(kl)
(kl)
(kl)
(kl)
(kl)
Tips
Graph
Gain
Loss
SN
IM
SR
PG
Panel A
Pre-intervention
trend
-0.265***
-0.263***
-0.263***
-0.264***
-0.264***
-0.263***
-0.266***
-0.266***
0.039
0.038
0.039
0.038
0.038
0.039
0.039
0.038
Treatment
0.093
0.048
-0.048
-0.028
0.056
0.139
0.01
-0.18
0.132
0.13
0.144
0.136
0.14
0.156
0.145
0.144
Treat x pre-trend
-0.017
0.007
0.018
0.008
0.006
0.004
0.018
0.059*
0.028
0.025
0.03
0.029
0.028
0.03
0.031
0.031
Constant
16.340***
16.442***
16.388***
16.405***
16.358***
16.317***
16.522***
16.546***
0.191
0.19
0.197
0.189
0.188
0.19
0.209
0.185
R-squared
0.415
0.417
0.412
0.414
0.414
0.415
0.416
0.42
Observations
211682
212277
210945
211912
176989
176559
190650
191280
Treated
107005
107600
106268
107235
72312
71882
85973
86603
Control
104677
104677
104677
104677
104677
104677
104677
104677
Fpvalue
0
0
0
0
0
0
0
0
Panel B
Treatment
-0.084
-0.023
-0.091
-0.055
0.032
0.089
0.002
0.012
0.11
0.1
0.103
0.11
0.117
0.116
0.126
0.111
14-Dec
-1.013***
-1.024***
-1.028***
-1.022***
-1.017***
-1.027***
-1.015***
-1.018***
0.202
0.201
0.202
0.201
0.202
0.203
0.203
0.202

15-Jan
5.406***
5.407***
5.405***
5.405***
5.405***
5.405***
5.411***
5.414***
0.274
0.275
0.275
0.274
0.274
0.274
0.274
0.274
15-Feb
1.277***
1.271***
1.266***
1.263***
1.275***
1.273***
1.270***
1.267***
0.296
0.296
0.296
0.297
0.296
0.296
0.296
0.296
15-Mar
0.484***
0.483***
0.472***
0.478***
0.482***
0.474***
0.474***
0.476***
0.132
0.132
0.132
0.132
0.131
0.132
0.132
0.132
15-Apr
0
0
0
0
0
0
0
0
.
.
.
.
.
.
.
.
DEC2014 x treat
0.109
-0.056
-0.057
-0.046
-0.043
-0.029
-0.052
-0.296**
0.105
0.096
0.11
0.103
0.108
0.113
0.118
0.116
JAN2015 x treat
0.124
0.158
0.106
0.084
0.02
0.07
0.096
0.078
0.154
0.141
0.15
0.142
0.148
0.152
0.158
0.156
FEB2015 x treat
0.209*
0.254**
0.291**
0.15
0.234**
0.252**
0.113
0.11
0.115
0.115
0.122
0.107
0.115
0.121
0.121
0.123
MAR2015 x treat
0.190**
0.116
0.152
0.079
0.011
0.042
0.170*
0.06
0.09
0.091
0.093
0.085
0.094
0.107
0.093
0.099
APR2015 x treat
0
0
0
0
0
0
0
0
.
.
.
.
.
.
.
.
Constant
14.346***
14.455***
14.408***
14.425***
14.367***
14.335***
14.530***
14.551***
0.162
0.167
0.175
0.166
0.171
0.168
0.179
0.165
R-squared
0.432
0.435
0.43
0.431
0.432
0.433
0.434
0.438
Observations
211682
212277
210945
211912
176989
176559
190650
191280
Treated
107005
107600
106268
107235
72312
71882
85973
86603
Control
104677
104677
104677
104677
104677
104677
104677
104677
Fpvalue
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
Notes: Regressions include tariff block and suburb fixed effects. Standard errors are clustered at the suburb level. Regressions are run for the
pre-intervention period of December 2014 – April 2015.

5. DATA AND ESTIMATION
This analysis is based on a sample of households that were eligible to receive inserts between
November 2015 and April 2016 (more specifically, in any particular month, these households are not
in tariff block one, they have an actual – as opposed to an estimated – meter reading and their billing
period was at most 35 days). We obtained monthly consumption data (measured in kilolitres) from the
City of Cape Town for the intervention period (November 2015 – April 2016) as well as the pre-
intervention period, which consists of the corresponding months from the previous year (November
2014 – April 2015).
If the treatment, T, is a behavioural nudge to reduce water consumption in household k, we want to
estimate the causal effect of T on the water consumption of household k in the City of Cape Town. To
understand the causal effect, we need to measure what would happen on average to a household, k, in
time t=1, randomly chosen from the population in the City of Cape Town, if we provided a
behavioural nudge via insert in their monthly municipal bill, , as opposed to not providing a
behavioural nudge via insert in their monthly municipal bill, .
! "
#$%&'
(&' ) ! "
#$%&'
(&*
In order to compare the effect of our treatments on water consumption in household k, we use
difference-in-difference methods to compares the changes in outcomes over time between treated
households, T = 1, and control households, where T = 0. Instead of simply taking a before and after
estimate of the impact of the project, DID compares the before-and-after outcomes for the households
that received the insert (the first difference) and the before-and-after outcomes for the households that
did not receive the insert but were exposed to the same set of economic and environmental conditions.
Then the difference between the difference in outcomes for the treated and the comparison is
calculated. In a two-period setting, the average program impact is estimated as follows:
++ , ! "
%&'
() "
%&*
(-
', . ) /! "
%&'
0) "
%&*
0-
', 1 (1)
where t = 0 before the intervention is implemented and t = 1 after implementation, Yt
T is the outcomes
for treated households at time t, Yt
C the outcome for non-treated households at time t, T1 = 1 denotes
treatment and, finally, T1 = 0 indicates allocation to control (Khandker et al. 2010).
As evident from equation 1, the difference-in-difference estimator is based on a comparison of treated
and control households both before and after the intervention – with the average treatment effect
being calculated as the difference between the outcomes for the treatment group after netting out the
difference in outcomes for the control group. A central assumption of any difference-in-difference
strategy is the parallel trend assumption, whereby the outcome in treatment and control groups would
Ti=1
Ti=0

display the same trend in the absence of treatment. In this case, the outcome change in the control
group (! "
'
0) "
*
0-
', 1 ) acts as the appropriate counterfactual (Khandker et al. 2010). While
unobserved heterogeneity (differences in ability, income or motivation across treatment and control
groups which might affect the outcome) could result in selection bias, if unobserved heterogeneity is
time invariant, then, once again, the outcome change in the control represents the counterfactual.
Figure 1 is supportive of the parallel trends assumption.
The DID estimate can also be calculated within a regression framework. Our reduced form expression
for the causal effect of a reduction in water consumption due to receiving an insert is estimated by the
below regression (Abramitzky & Lavy 2014):
"
#% ,2 34'-567896:8/;<=>6?<@A>#3 4BC<>8 3 4DC<>8 E-567896:8/;<=>6?<@A>#
3 4F-56:A 3 4FGH3 IH
where -567896:8/;<=>6?<@A># is a treatment variable indicating if the household is in a
treatment group, either before or after the intervention is received and C<>8 is a time variable
indicating whether treatment has commenced. The coefficient 4D on the interaction between the
treatment and time variables C<>8 E-567896:8/;<=>6?<@A># gives the average DID effect of the
program. The variables -567896:8/;<=>6?<@A># captures the underlying differences between
treatment and control groups while C<>8 captures the underlying differences between the two-time
periods. Given the trend analysis in section 4, we control for the effect of trend. Finally, GH are a
vector of controls.
As outlined previously, our main controls are informed from the balance and pre-intervention trend
regressions as well as the stratified randomization. These include: property values, baseline tariff
block (the tariff block the household was in in October 2015 when the randomization was conducted),
month and suburb fixed effects and an indicator for indigent status.6 In addition, we also include a
control if the household was a “late receiver”, a status given to households in the social recognition
and public good treatments that received their first insert a month later because of a technical glitch.
We also control for the amount of times the household appears in the panel (frequency) and lag
variables for both the bill and, separately, marginal tariff rate (tariff rate of the household’s highest
tariff block). These lag variables control for the effect of the previous bill on consumption in the
current month. For example, if a household was charged at a higher rate in November, either due to an
increase in tariff or due to higher consumption/presence of a leak, any subsequent behaviour change
6 COCTindigents households are excluded from the sample. We control for additional indigent households by
including a dummy variable in the regression analysis.

(such as a reduction in consumption) will be reflected in the December bill.
All standard errors are clustered at the suburb level. We use a fixed effects model to control for the
unobserved heterogeneity across treatment and control that does not vary over time.
6. RESULTS
The DID estimates are provided in Table 6. The results in Table 6 (for the full intervention period of
December 2015 – April 2016) point to two main findings. Firstly, despite households’ exposure to
water restrictions (implemented from 1 January 2016), extensive media campaign and tariff increase
(in Jan 2016), all mailers successfully induced a reduction in household consumption. Secondly, the
social recognition and public good framings consistently outperformed the other messages over the
long(er) run. In other words, public recognition of water-conservation efforts and appeals to the
public good are the most effective motivators of pro-environmental behaviour.
As previously discussed, the behavioural messages were divided into two categories. The first
category of messages promoted water conservation by addressing informational failures around price
and usage of water (tips, graph, gain and loss treatments). The second group of messages promoted
water conservation via social incentives and appeals to the public good (social norm, intrinsic
motivation, social recognition and public good treatments). Focusing on Regression III in Table 6, the
tips, graph, gain and loss messages (group I: informational messages) reduced consumption by
between 160 – 209 litres per household per month on average (tips: 160 litres, graph: 168 litres, gain:
209 litres, loss: 176 litres). The social preference messages reduced consumption between 262 litres –
319 litres per household per month on average (social-norm: 208 litres, intrinsic-motivation: 262
litres, social-recognition: 315 litres and public-good: 319 litres).
Using the pre-intervention mean consumption values in Table 3, we estimate that reductions ranged
from 0,6% for the tips treatment to 1.3% for the social recognition treatment. The implication being
that the messages are a useful mechanism in the context of water scarcity and promote addition water
saving (water saving on top of more conventional DSM tools).
Figure 3 graphically illustrates the DID regression coefficients for regression III. Households
receiving the social-recognition and public-good messages reduced consumption by an average 315
and 319 litres per month, respectively (Table 3, regression III). This reduction equates to a reduction

in consumption of around 1.3% and this is a significantly greater reduction compared to what was
achieved via the tips, graph, gain, loss, social norm and intrinsic-motivation messages.7
While it may be surprising that most of the framed treatments did not perform considerably better
than the tips treatment, it is possible that households were more responsive to the tips treatment than
would typically be the case given the City’s extensive multi-media campaign around the drought,
water shortages and need to conserve water. Previous behavioural nudge studies did for instance not
find tips to have a significant impact on incentivizing water savings.
Figure 3. Graphical representation of the coefficients (regression III, Table 6)
Finally, we consider heterogeneous effects of the treatments across different subgroups of water users.
Households were divided into five quintiles (low to high) depending on their property value and the
regressions are replicated for these subgroups. The full set of regression results is provided in
Appendix B.1 – B.5. For ease of reference, we focus on the results for regression III, which are
collated in Table 7. In addition, Figure 4 illustrates the coefficients for regression III, per quintile.
7 As discussed in section 2.2, there is a late receiver group from the social-recognition and public-good
treatments. These households (although eligible), did not receive an insert in November 2015 because of an
operational glitch at the printers. These households received their first inserts in December 2015. In the social-
recognition treatments, these households received their last insert in April 2015 (compared to March 2015 for
the non-late receivers). Both groups received their feedback in May 2015. As there is a seasonal aspect to
consumption, we run a series of regressions excluding these households as a robustness check. When running
the same regression as regression III in Table 6, but excluding these late receives, the coefficients of both the
social-recognition and public-good treatments remain significant at the 1% level. However, the magnitude of
the effect is slightly dampened (SR: -0.282; PG: -0.252).

As before, we find households to be most responsive to the social preference treatments (social norm,
intrinsic motivation, social recognition and public good). More so, we find that, in general, the
wealthiest households are most responsive to the treatments.
Examining the results of the social preference treatments, we find no significant impact in the first
quintile (poorest households). Households in the second quintile receiving the social-norm treatment
reduced consumption by an average 317 litres (1.6%) per month, increasing to reductions of 341 litres
(1.5%) and 444 litres (1.2%) in the third and fifth quintiles, respectively. The social-recognition and
public-good treatments are particularly effective amongst the fifth quintile (wealthiest homes), where
treated households reduce consumption by an average 852 litres (2.3%) and 702 litres (1.9%),
respectively.
Figure 4. Graphical representation of the coefficients (regression III in Tables B.1 – B.5)
Table 6. Long(er) run effects: December 2015 – April 2016

I
II
III
IV
V
Consumption
Consumption
Consumption
Consumption
Consumption
(KL/month)
(KL/month)
(KL/month)
(KL/month)
(KL/month)
Tips x Post
-0.152*
-0.152*
-0.160**
-0.156**
-0.163**
0.081
0.081
0.081
0.072
0.078
Graph x Post
-0.171**
-0.171**
-0.168**
-0.183**
-0.183**
0.08
0.08
0.079
0.071
0.076
Gain x Post
-0.203***
-0.203***
-0.209***
-0.194***
-0.206***
0.076
0.076
0.078
0.068
0.074
Loss x Post
-0.177**
-0.177**
-0.176**
-0.179**
-0.181**
0.084
0.085
0.084
0.074
0.08
SN x Post
-0.298***
-0.298***
-0.284***
-0.293***
-0.328***
0.092
0.092
0.092
0.082
0.088
IM x Post
-0.297***
-0.297***
-0.262***
-0.288***
-0.326***
0.085
0.085
0.087
0.078
0.084
SR x Post
-0.332***
-0.332***
-0.315***
-0.353***
-0.386***
0.09
0.09
0.092
0.083
0.088
PG x Post
-0.293***
-0.292***
-0.319***
-0.313***
-0.326***
0.082
0.082
0.083
0.072
0.078
Trend
-0.332***
-0.332***
-0.126***
-0.352***
-0.381***
0.016
0.016
0.013
0.016
0.014
Billing period
0.665***
0.665***
0.678***
0.694***
0.663***
0.011
0.011
0.011
0.011
0.011
Indigent
-0.500***
-0.272
-0.463**
-0.448**
0.184
0.176
0.193
0.204
Frequency (-1)
-0.603***
0.039
Billed amount (-1)
0.005***
0
Tariff rate (-1)
0.131***
0.012
Constant
8.047***
8.086***
6.541***
5.957***
6.394***
0.301
0.301
0.338
0.308
0.352
R-squared
0.22
0.22
0.226
0.242
0.23
Observations
1776253
1776253
1752140
1752140
1752052
Treated
1558919
1558919
1537977
1537977
1537903
Treat1
223913
223913
220752
220752
220739
Treat2
225428
225428
222364
222364
222347
Treat3
223012
223012
219954
219954
219946
Treat4
224992
224992
221845
221845
221842
Treat5
152025
152025
150015
150015
150010

Treat6
151013
151013
149056
149056
149049
Treat7
180432
180432
178219
178219
178208
Treat8
178104
178104
175772
175772
175762
Control
217334
217334
214163
214163
214149
Clusters
674
674
674
674
674
Fpvalue

Table 7. Quintile analysis: December 2015 – April 2016
I
II
III
IV
V
1st Quintile
2nd Quintile
3rd Quintile
4th Quintile
5th Quintile
Consumption
Consumption
Consumption
Consumption
Consumption
(KL/month)
(KL/month)
(KL/month)
(KL/month)
(KL/month)
Tips x Post
0.095
-0.207
-0.057
-0.217
-0.196
0.199
0.158
0.136
0.204
0.223
Graph x Post
0.046
-0.146
-0.253*
-0.042
-0.23
0.207
0.153
0.143
0.176
0.2
Gain x Post
-0.12
-0.339**
-0.145
-0.14
-0.186
0.186
0.154
0.137
0.199
0.236
Loss x Post
0.127
-0.251
-0.179
-0.203
-0.354*
0.228
0.164
0.135
0.182
0.209
SN x Post
0.059
-0.317*
-0.341**
-0.183
-0.444*
0.209
0.162
0.158
0.192
0.247
IM x Post
-0.153
-0.02
-0.307*
-0.218
-0.498**
0.21
0.154
0.172
0.185
0.239
SR x Post
0.215
-0.195
-0.226
-0.302
-0.852***
0.215
0.17
0.164
0.188
0.249
PG x Post
0.051
-0.304*
-0.356**
0.107
-0.702***
0.198
0.176
0.163
0.204
0.253
Trend
-0.089***
-0.045**
-0.114***
-0.165***
-0.198***
0.023
0.019
0.024
0.026
0.032
Billing period
0.559***
0.588***
0.628***
0.719***
0.960***
0.009
0.009
0.018
0.016
0.02
Indigent
-0.072
-0.43
-0.761
-1.226
-8.890**
0.15
0.485
0.659
2.002
3.763
Frequency (-1)
-0.107***
-0.246***
-0.386***
-0.775***
-1.536***
0.029
0.033
0.047
0.046
0.074
Billed amount (-1)
Tariff rate (-1)
Constant
3.170***
3.088***
5.514***
7.093***
10.897***
0.355
0.35
0.672
0.607
0.665
R-squared
0.185
0.253
0.247
0.261
0.273
Observations
323431
320653
330695
340709
326581
Treated
283290
281568
289984
299211
286863
Treat1
40898
39842
42266
42875
41072
Treat2
41105
40816
41736
42991
41632
Treat3
40688
40327
41510
42871
40760
Treat4
40709
41583
41468
42232
41600

Treat5
27847
27307
29025
28731
27734
Treat6
27540
27586
27972
29195
27803
Treat7
32940
32188
33094
35561
33036
Treat8
31563
31919
32913
34755
33226
Control
40141
39085
40711
41498
39718
Clusters
330
347
420
433
440
Fpvalue
.
.
.
.
.

7. DISCUSSION
This study was conducted during the onset of what later evolved into a severe drought in Cape Town.
A large-scale randomised control trial was conducted to test the effectiveness of various behavioural
nudges in reducing residential water consumption. The nudges were delivered to households as inserts
in their monthly utility bills.
At the time of these interventions, the local municipality, the City of Cape Town, introduced water
restrictions and increased tariff rates in order to cut residential water consumption by 10%. As such,
this randomised experiment tests the potential for behavioural interventions to reduce water
consumption over and above structural and pecuniary disincentives.
The behavioural messages were divided into two categories. The first category of messages promoted
water conservation by addressing informational failures around price and usage of water (tips, graph,
gain and loss treatments). The second group of messages promoted water conservation via social
incentives and appeals to the public good (social norm, intrinsic motivation, social recognition and
public good treatments).
The results indicate that all treatments successfully induced a reduction in household consumption.
The overall reductions achieved through the interventions ranged from 0,6% for the tips treatment to
1.3% for the social-recognition treatment. More specifically, the results indicate that publicly
recognizing water conservation (social recognition) or appealing to households to act in the public
interest (public good) are the most effective motivators for water conservation (both resulting in
reductions of 1.3%) and these treatments performs significantly better compared to both the control
group as well as the tips, graph, gain and loss treatment groups.
Households were further divided into five quintiles depending on their property value, allowing for
more detailed heterogeneity analysis. The social recognition and public good treatments are
particularly effective amongst the fifth quintile (wealthiest homes), where treated households reduce
consumption by an average 852 litres (2.3%) and 702 litres (1.9%), respectively. The finding that
social incentives are particularly effective amongst higher income groups signals that treatment
effects are heterogeneous across the income spectrum. Moreover, an understanding of these
heterogeneous effects enables government to target feedback at more responsive groups, an approach
which is both more cost-effective and expedient in terms of meeting policy objectives. In this context,
our study indicates that, while these messages can be used as an additional DSM tool, they need to be
directed at the appropriate income groups.

Further comparing the impact of these behavioural nudges as an instrument compared to pricing
changes, it is important to realise that most studies internationally (Espey et al., 1997; Arbues et al.,
2003) and locally (see Veck and Bill 2000) have indicated price elasticities of demand for water to
range from -0.1 to -0.18 which means that a 10% increase in the price of water would result in a 1-
1.8% reduction in residential demand for water. In line with this, our findings that high income
(quintile 5) households in our social recognition treatment reduced consumption by 2.3% over a six-
month period is equivalent to a 12% price increase in that category (presuming price elasticity of
demand of -0.18).
The study’s findings suggest that savings from behavioural nudges should not be discounted from the
toolbox of DSM instruments available to policy makers and could realise significant savings at local
and national level if implemented during strategic periods. Further social recognition promises to be a
highly effective lever for behavioural change amongst higher income households, which should be
explored in a much wider policy spectrum than only water DSM.

REFERENCES
Abramitzky, R. & Lavy, V., 2014. How Responsive Is Investment in Schooling to Changes in
Redistributive Policies and in Returns? Econometrica, 82(4), pp.1241–1272. Available at:
http://doi.wiley.com/10.3982/ECTA10763.
Allcott, H., 2011. Social norms and energy conservation. Journal of Public Economics, 95(9–10),
pp.1082–1095. Available at: http://linkinghub.elsevier.com/retrieve/pii/S0047272711000478.
Allcott, H. & Mullainathan, S., 2010. Behavior and Energy Policy. Science, 327(5970), pp.1204–
1205.
Allcott, H. & Rogers, T., 2012. The Short-Run and Long-Run Effects of Behavioral Interventions:
Experiment Evidence From Energy Conservation. American Economic Review, 104(10),
pp.3003–3037.
Andreoni, J. & Petrie, R., 2004. Public goods experiments without confidentiality: a glimpse into
fund-raising. Journal of Public Economics, 88(7), pp.1605–1623.
Angrist, J. & Pischke, J.-S., 2008. Mostly Harmless Econometrics : An Empiricist ’ s Companion.
Massachusettts I nstitute of Technology and The London school of Economics, (March), p.290.
Arbués, F., García-Valiñas, M.Á. & Martínez-Espiñeira, R., 2003. Estimation of residential water
demand: A state-of-the-art review. Journal of Socio-Economics, 32(1), pp.81–102.
Ayres, I., Raseman, S. & Shih, A.,