Introduction

Household energy consumption is a substantial contributor to carbon emissions1,2, with the highest consumption per capita identified in Gulf Cooperation Council (GCC) countries3,4. GCC countries are resource-rich countries outside temperate climate zones, which tend to use large and mounting amounts of energy for cooling and humidity control year-round5. The climatic conditions of GCC countries make it a particularly challenging context for household energy reduction. Due to extreme heat and humidity, the structural reliability of buildings demands 24-h air conditioning (AC) use, reducing the possibility for alternative, less carbon-intensive, cooling options (e.g., ceiling fan)6. The potential range of behavioral change adjustments to reduce energy consumption is mainly restricted to purchasing energy-efficient AC (mostly for property owners), increasing the AC set temperature (to 25 °C or 77 °F), or reducing usage of other more energy-intensive appliances (e.g., washing machine, TV, laptops, lights)7. Furthermore, electricity prices in these oil and gas-producing countries are only a fraction of the cost compared to resource-poor countries and are often subsidized8, limiting the impact of price-based interventions.

Here we present a pre-registered behavioral field experiment conducted in the United Arab Emirates (N = 100 households; 26,400 observations over 9 months, from September 2019 to May 2020) focused on reducing household energy consumption. To the best of our knowledge, no behavioral interventions for household energy reduction targeting GCC countries have been conducted thus far. We test a theoretically-grounded intervention comparing a traditional social norms message (group 1) with a social norms message added to an attachment priming (group 2), versus a control group (group 3). The rationale behind this experimental design is twofold.

First, investigating social norms holds substantial promise for driving down household energy use by leveraging the inherent human tendency to align with the actions and expectations of peers. Social norms essentially serve as the tacit directives that guide societal conduct used to influence human behavior9,10, consistently corroborated in a variety of settings, including energy savings11,12,13,14,15,16,17,18,19 as well as other climate change mitigation behaviors20,21. When individuals are made aware of the energy-conserving practices prevalent among their peers—a concept known as descriptive norms—along with being encouraged that such practices are socially endorsed—known as injunctive norms13—they are often motivated to modify their own habits. Providing households with comparative feedback on their energy consumption relative to others harnesses the motivational effects of social comparison and social influence, which are powerful determinants of behavior in everyday life. Nonetheless, boosting the effectiveness of social norms is crucial on two grounds. On the one hand, behavioral interventions (as all climate action) need to step up their impact to meet the targets of the Paris Climate Agreement and the 2030 Sustainable Development Goals22. On the other hand, there may be climatic conditions that challenge the effect of social norms to reduce energy usage, which have been established mostly in the US and Europe11,12,13,14,15,16,17,18,19,20,21, with few exceptions23,24,25. Behavioral adjustments to save energy as a reaction to social norm messaging may be easier to achieve under moderate weather than under extreme heat, which may require additional motivational triggers to elicit and maintain reduced energy consumption. Past research has focused on applying similar social norm formats to different contexts, but there has been relatively little exploration on augmenting the effectiveness of social norms messaging.

Second, we introduce the concept of secure attachment as an auxiliary to social norms. Secure attachment is a central concept in psychological theory, which delves into the primary emotional bond that forms between humans, often observed between a mother and her child26. Attachment is understood as a natural behavioral system that is hardwired to protect individuals from perceived dangers by ensuring closeness to nurturing and supportive figures27. When people experience a sense of secure attachment, that is, feeling more assured and less vulnerable to the world around them, they are more prone to show trust and compassion towards others28. Thus, the sensation of being securely attached kindles the mental frameworks that encourage human connections29, which is particularly important in amplifying the impact of social norms. This sense of secure attachment has been shown to aid in the efforts to counteract climate change30,31, and we further suggest it can leverage social norms that underline the urgency of changing our behaviors. Given that climate change stems from collective behaviors and necessitates a shift in group actions, fostering secure attachment could play a pivotal role in encouraging communal responses to this global issue, boosting social norms messaging. Thus, we contribute to the limited evidence about motivational complements32—meaning, which different motivational elements should be combined in interventions to amplify behavioral impact. There is an underlying assumption that more elements in a message will increase its effectiveness, but this is not necessarily true33. It depends on the motivational compatibility of the elements included. The stimuli that are combined must produce a motivational synergy to be more effective, meaning they must be psychologically aligned in their motivational power, otherwise, they may cancel each other's effects32. This has been identified primarily in research focused on identifying elements that may act as substitutes34, meaning elements that have competing (not additive) effects. Fewer studies have focused on identifying elements that work as complements, producing a motivational synergy. A rare example of such research, also examining energy usage, has examined which type of social norm may be more effective—descriptive, injunctive, or both in combination13. Our work goes beyond the analysis of the variation of social norms messages to examine the added value of different stimuli. The use of attachment theory in this setting is both pioneering and promising. There is both compelling psychological theory and evidence26,27,28,29 suggesting that social norms could be complemented by a motivator like attachment security that is aligned in promoting human interconnectedness.

Results show that, compared to standard social norm message, the social norm plus attachment message (i) helped households save significantly more energy during the intervention period (9.98% vs. 6.11%); (ii) was more effective for already efficient households; (iii) was more effective during days of higher temperatures; and (iv) was linked to energy reduction twice as persistent after the intervention was discontinued. An unexpected event, for which there is no pre-registered hypothesis, was that the COVID-19 local lockdown took place in the final months of the follow-up period. From the start of the lockdown onward, there were no significant differences between the experimental conditions in the average total daily energy consumption.

Results

Over 9 months, we examined the total daily kWh in an international University campus in the United Arab Emirates. We randomly allocated 100 campus housing apartments to a control group (no social norm message), a standard social norm message (with descriptive and injunctive norms), and an enhanced social norm message (with descriptive and injunctive norms + attachment security priming) (full methodological details presented in “Methods” below). During the experimental period, four messages were delivered to the apartments according to their experimental group, once per week over 4 weeks. After the 4-week intervention period, messages were discontinued and total daily kWh was monitored for the following 7 months.

Household energy consumption at baseline (pre-intervention period: 4 weeks)

Baseline average daily electricity usage during the 4-week pre-experiment period was 7.65 kWh among households in our study sample. Note that this is much lower than the average usage of U.S. OPOWER program households (ranging from 19 to 60 kWh)35, which is likely because these are all moderate-sized apartments with at most two bedrooms rather than single-unit houses. Statistical results indicate no significant differences in electricity usage among experimental arms prior to the intervention (pairwise Wilcoxon–Mann–Whitney tests p values 0.390, 0.885, 0.356) (Supplementary Table 1).

Household energy reduction per condition (intervention period: 4 weeks)

Table 1 presents the average treatment effect (ATE) during the 4-week intervention period. The result of the main specification is presented in Col 4, where we controlled for weather-related variables, apartment fixed effects, day of the week fixed effects, and week fixed effects. For robustness, we run the regressions using several different specifications. In Col 1, we only controlled for apartment fixed effects and day of the week fixed effects. In Col 2, we further controlled for week fixed effects. Additionally, we added building-week fixed effects in Col 3.

Table 1 Average treatment effect (ATE) estimates
Full size table

Across all specifications, we observe that both interventions lead to large reductions in electricity usage, and the magnitudes of the coefficients are quite stable. Compared to households in the control group, households receiving the standard social norm message reduced daily electricity usage by 0.502 kWh (95% CI: [−0.630, −0.375], p = 0.000), and households received the enhanced social norm message with an attachment security priming had a bigger reduction of 0.820 kWh (95% CI: [−1.046, −0.595], p = 0.000). In other words, there was an average of 6.11% reduction in daily electricity usage for the group of households that received the standard social norm messages, versus a 9.98% reduction for the group of households that received the enhanced social norm messages with attachment priming. The difference in treatment effects between these two treatment arms is highly statistically significant (p = 0.001).

High-frequency analysis

Given the daily frequency of our data, we take advantage of this feature to conduct high-frequency analysis. Figure 1 plots a visual assessment of the 3-day moving window ATE for each day of the 4-week intervention period. This 3-day moving average is used to smooth over idiosyncratic variation, following Alcott and Rogers35. Figure 1 shows that both trends were negative throughout the intervention period. Compared to the control group, there was a further reduction in electricity usage among households receiving the enhanced social norms message starting from the second round of messages.

Fig. 1: High-frequency treatment effects during the intervention period.
figure 1

Dashed vertical lines represent each of the four messages delivered to the apartments. The figure plots the 3-day running mean treatment effects for each day during the 4-week intervention period for each group that received intervention, as estimated by Eq. (2). Driscoll and Kraay standard errors were used. Errors were not clustered, given that the randomization was performed at the apartment level.

Full size image

Action and backsliding

Moreover, previous field experiment studies have shown the presence of “action and backsliding”: an immediate decline in usage following each message, coupled with an upward trend a few days later. To examine the “action and backsliding” pattern by experimental arms more closely, we borrowed the strategy from Allcott and Rogers35 (detailed in the “Methods” section). Figure 2 reports the dynamic change in treatment effect since the day of message arrival. In other words, we plotted the difference between the usage on the day of message arrival and a subsequent day, which implies the usage on the day of message arrival has been normalized to zero in this figure. Since the control group also received placebo messages, we also included the control group when plotting the figure. The backsliding patterns are similar for the control arm and the intervention arm receiving the standard social norm message (Fig. 2). Compared to these two arms, the backsliding of the intervention arm receiving the enhanced social norm message is not only smaller in magnitude but also appears to be delayed. Compared to the day when households received the weekly message, households in social norm intervention arm further reduced energy consumption by 0.172 kWh/day (95% CI: [−0.363, 0.019], p = 0.075) 1 day after and 0.161 kWh/day (95% CI: [−0.472, 0.150], p = 0.297) 2 days after, and then backslide on day 3 with an increase of 0.103 kWh/day (95% CI: [−0.179, 0.385], p = 0.461). For households receiving social norm + attachment intervention, after receiving message, they continued to reduce energy consumption until day 3 by 0.253 kWh/day (95% CI: [−0.427, −0.079], p = 0.006), and then they started to backslide slightly. It is also worth noting that no actual backslide took place among households in the social norm + attachment group, as their energy consumption on subsequent days never exceeded the energy consumption on the day of message reception. This visual pattern provides insights into the dynamics of the different forms of social norms messaging.

Fig. 2: Backsliding after receiving messages.
figure 2

The figure plots the treatment effects on each day after households received message during the 4-week intervention period, as estimated by Eq. (3). Driscoll and Kraay standard errors were used. Errors were not clustered, given that the randomization was performed at the apartment level.

Full size image

Heterogeneity by energy use

In both experimental arms, each household knew from the message they had received whether they had consumed more (or less) energy than their average neighbors in the previous week (or baseline month in the case of the first message). For the already energy-efficient households, there was the risk of increasing energy consumption after they became aware of their energy-efficient status, and thus creating a “boomerang effect”36,37. It is therefore important to investigate whether the enhanced social norm message can help households sustain energy efficiency.

Figure 3 shows the quantile treatment effects with their 95th confidence intervals for both intervention arms. Throughout the percentile spectrum, the treatment effects for both intervention arms are negative, and thus, we do not observe a boomerang effect with our intervention. However, there still exist heterogeneous treatment effects by household energy-use level. Among households that received the standard social norm message, we observe a downward sloping curve, indicating that there was more electricity usage reduction among the heavy users. In contrast, the enhanced social norm message led to even higher electricity usage reduction among the already efficient households. Specifically, the effect of standard social norm treatment range from −0.443 kWh/day (95% CI: [−0.850, −0.037], p = 0.033) for the bottom decile to −0.565 kWh/day (95% CI: [−1.056, −0.073], p = 0.024) in the top decile, and the effect of the enhanced treatment range from −0.915 kWh/day (95% CI: [−1.330, −0.499], p = 0.000) for the bottom decile to −0.720 kWh/day (95% CI: [−1.222, −0.218], p = 0.005) in the top decile.

Fig. 3: Quantile treatment effects.
figure 3

The figure plots the conditional ATE for households in different percentiles of their distribution of the baseline electricity usage, where a 5% interval was specified in quantile regression with fixed effects. The covariates are the same as those delineated in Eq. (1). Shading indicates 95% confidence intervals. Errors were not clustered, given that the randomization was performed at the apartment level.

Full size image

Heterogeneity by daily maximum temperature

A central motivation for this study was to examine social norms effectiveness in regions with extreme climate. The average daily maximum temperature during the pre-intervention and intervention periods is 38.6 °C [101.5 °F] (Table 2). We further divide our daily data into two subgroups by this average. The first two columns of Table 2 show separate regression results by these subgroups. There is an even larger treatment effect during days with higher temperatures for both intervention arms: for the standard social norm arm, the ATE is −0.521 kWh (95% CI: [−0.606, −0.436], p = 0.000) compared to −0.504 kWh (95% CI [−0.803, −0.204], p = 0.002) in cooler days, whereas for enhanced social norm arm, the ATE is −0.915 kWh (95% CI: [−1.101, −0.730], p = 0.000) compared to −0.767 kWh (95% CI: [−1.162, −0.372], p = 0.001) in cooler days, suggesting this effect difference is more pronounced in the enhanced social norm message intervention arm as temperature rises. The results are similar if we divide the subgroups by median daily maximum temperature during the same period (Table 2, Col 3 and 4). However, these differences are not statistically significant.

Table 2 Heterogeneity of impacts by daily maximum temperature
Full size table

Long-term household energy reduction per condition (follow-up period: 7 months)

During the 4-week intervention period, messages were sent out each week, after which messages were discontinued. How persistent are treatment effects after the intervention ends?38 To answer this question, we plot the treatment effects and their 95% CIs by 4-week intervals in Fig. 4 (specification in the “Methods” section). We observe that both treatment groups continued to have a reduced energy consumption level compared to the control group after the end of the intervention. For the standard social norm intervention arm, the treatment effect was negative and significant up to the 12th week after the start of the intervention, or 8 weeks after the intervention ended. The ATEs during these 12 weeks range from −0.484 kWh/day (95% CI: [−0.852, −0.116], p = 0.011) to −0.904 kWh/day (95% CI: [−1.082, −0.726], p = 0.000). For the enhanced social norm treatments, the significance of the treatment effect persisted much longer, lasting 24 weeks after the start of the intervention, or 20 weeks after the intervention ended. The ATEs during these 24 weeks range from −0.522 kWh/day (95% CI: [−0.852, −0.191], p = 0.002) to −1.165 kWh/day (95% CI: [−1.418, −0.911], p = 0.000). It should be noted that the COVID-19 lockdown has coincided with the 24th week after the start of our intervention.

Fig. 4: Persistence of treatment effects.
figure 4

This figure plots the ATES for each month (measured in 4-week intervals) from the end of the intervention for each experimental arm, as estimated by Eq. (4). X-axis denotes the number of weeks since the intervention started (i.e., first message sent). Weather conditions, apartment fixed effects, day of the week fixed effects, and week fixed effects are taken into consideration. Shading indicates 95% confidence intervals. Driscoll and Kraay's standard errors were used. Errors were not clustered, given that the randomization was performed at the apartment level.

Full size image

To fully explore this panel data, Fig. 5 plots ATEs for each day using a 7-day moving window over the 4-week intervention period and the subsequent 7-month follow-up period in our study. The treatment effects for both experimental arms were largely negative till the start of the COVID-19 lockdown. Also, prior to the start of the COVID-19 local lockdown, the gap between the trends indicates that there exists a persistent difference in treatment effects between the two experimental arms. The size of the gap ranges from −0.907 kWh/day (95% CI: [−1.032, −0.782], p = 0.000) to −0.042 kWh/day (95% CI: [−0.152, 0.068], p = 0.389) between the start of intervention on October 9th, 2019, and the start of the first COVID lockdown on March 26, 2020.

Fig. 5: High frequency treatment effects over a 7-month follow-up.
figure 5

The figure plots the 7-day running mean treatment effects for each day during the 4-week intervention period and the subsequent 7-month follow-up period for each experimental arm, as estimated by Eq. (2). Dashed vertical lines represent each of the four messages delivered to the apartments. The yellow shaded area indicates local lockdowns due to the COVID-19 pandemic. Driscoll and Kraay's standard errors were used. Errors were not clustered, given that the randomization was performed at the apartment level.

Full size image

Discussion

To the best of our knowledge, this is a pioneering field experiment conducted in the United Arab Emirates, which broadens our understanding of how the application of behavioral interventions, specifically social norms messaging, can effectively reduce household energy consumption even in regions with extreme climatic conditions like those of GCC countries. It extends the literature on social norms by showing that such interventions can work outside the temperate zones where they have been predominantly tested11,12,13,14,15,16,17,18,19,20,21, suggesting that such strategies are robust across diverse environmental contexts. The closest previous paper to our research in terms of climate conditions and geographical location is the one by Sudarshan25 during the summer months in India. This study found that providing households with weekly reports comparing their energy consumption to that of their peers led to a 7% reduction in usage. This outcome closely aligns with our findings for the standard social norm group, which showed a 6.11% reduction, further supporting the credibility of our results.

Moreover, by incorporating the psychological concept of attachment security into traditional social norms messaging, this research suggests that combining motivational factors can significantly enhance social norms effectiveness. This integrated approach has led to a reduction in energy consumption markedly greater than what is achieved through standard social norm messages alone (9.98% versus 6.11%). Even under severe heat, where energy use might be deemed essential for comfort, the combined social norms and attachment security messaging prompted higher energy reductions than traditional social norms messages. The results suggest that the impact of the enhanced messaging is comparable to what would be expected from a substantial electricity price increase39,40, indicating that our intervention could be a potential strategy for energy policy where price interventions may not be feasible or desirable.

Furthermore, this research addresses concerns about the boomerang effect36,37 by showing that the enhanced social norm message led to even higher electricity usage reduction among the already efficient households. Additionally, one of the most noteworthy contributions is the evidence showing that the motivational enhancement can lead to sustained changes in behavior for a period extending well beyond the conclusion of the intervention, a notable achievement in behavioral intervention strategies, which often struggle with long-term persistence38.

In conclusion, the broad impacts of this paper lie in its empirical support for the efficacy of behavioral interventions in promoting energy conservation across different climate regimes. The work underscores the potential of psychological insights to inform practical applications that not only lead to immediate benefits in terms of reduced energy consumption but also contribute to sustained behavior change, offering a model for large-scale implementation in similar environments.

This study has some limitations. Our findings emphasize the value of theory-driven interventions41, promoting the idea that the application of robust psychological theories in designing interventions can offer valuable insights, enhance the efficacy and cost-effectiveness of such programs, and ultimately support more impactful climate action. Nonetheless, although attachment priming has been extensively tested and its underlying psychological mechanisms established27,28,29, we did not evaluate the mental process underlying the observed behavioral changes. We inferred an increase in empathy due to the attachment security stimulus, based on previous research30,31, but an elicitation of pro-environmental or pro-social preferences could have also occurred.

Another potential limitation is the small number of participating households in our experiment. To mitigate this concern, we performed two tests: (a) Fisher’s randomization inference test and (b) sensitivity analysis based on Cohen’s f. We conducted Fisher’s randomization inference test with 2000 iterations. In terms of the number of iterations, Young42 found little difference when the number of replications goes beyond 200th. The p value of randomization inference is the percentage of hypothetical treatment effects larger than the observed treatment effect. If we observed the main results by pure chance, then we would expect to see large p values from the test results. As Supplementary Table 2 shows, all p values from our randomization inferences are smaller than 0.01, that is, less than 1% of the hypothetical treatment effects are larger than our observed treatment effects, as shown in the main results. Given the intraclass correlation coefficient of the household energy consumption in our sample is 0.586, we used a more conservative correlation coefficient 0.6 in our power analysis based on Cohen’s f. By using software G*Power43,44, we found that the sample (N = 99) provided 80% power to detect a medium effect size of 0.246 at a significant level of p = 0.05.

Moreover, we cannot say with certainty that there was no neighborly discussion about the messages being delivered to the apartments. What we can say with certainty is that the content of the messages was not visible to other neighbors. Messages were taped facing down in each apartment. Regardless, even if neighborly discussions did occur, this cannot be used to explain or predict any behavioral change in a particular direction. Even if neighbors compared the messages received and noticed different content was being delivered, that does not explain why groups at the aggregate level behaved differently. At best, it could have made people suspect they were part of a study and likely either cancel any experimental effect or make people all behave the same way to “please” the experimenters, i.e., all reduce energy consumption similarly. It is most likely that households were aware they were being observed, particularly during the 4 weeks that they received a message at their door once per week. An experimenter effect in this scenario would predict that all (or most) households would reduce energy consumption as the socially desirable thing to do. But an experimenter effect cannot explain why experimental groups behaved differently. This can be attributed to the different messages received, because all other aspects of the study were kept constant between groups.

Another possible limitation is that we were not able to establish the potential duration of the lasting effects under normal circumstances; during the follow-up period, an unexpected event affected energy usage behavioral patterns. The last months of the follow-up period unpredictably included the COVID-19 lockdown, after which there were no significant differences in energy usage between the experimental conditions. This suggests that (behavioral) interventions to influence energy demand may have limited resilience to stress and risky hazards45, which may involve negative emotions and thoughts that interfere with regular behavioral habits. It is also possible that households saved energy by turning down the AC when they left for work/went out during the day. When lockdown happened, there may have been less room for behavioral adjustment. We cannot claim for sure that the lockdown is the reason why no between-group differences were identified thereafter. But we can say that from the lockdowns onwards, no differences were found. Whether the intervention effects ran their course, or the lockdown produced homogeneous patterns of energy consumption behavior (or both) cannot be conclusively interpreted.

Furthermore, while it is reasonable to deduct that the reduction in energy use observed in this study came from everyday behavior change, given the nature of on-campus housing, we cannot gauge the exact strategies households employed to reduce consumption since no survey or sociodemographic information could be collected from the residents. We cannot provide a definitive answer about how energy reduction was achieved. But the situation here is similar to many interventions or policies, in which the mechanisms underlying the behavioral change are not fully understood. For instance, interventions using social norms to promote timely tax returns, how is this achieved? Do people hire an accountant? Do they take a day off work? It is often unclear the mediating psychological or behavioral processes that promote the visible change, but nonetheless this is an important point to be addressed in future research. Future research should also investigate whether these results may not be limited to extreme heat and humidity, but perhaps generalizable to extreme weather conditions, including harsh winters.

Lastly, both the lack of sociodemographic controls and the specific accommodation setting where experiment was conducted further recommend caution in the interpretation of the results and limit the generalizability of the results.

Methods

Ethics approval

This study was approved by the Institutional Review Board (IRB) of the New York University Abu Dhabi on 9th February 2019 (HRPP-2019-37) with the title “Reducing energy consumption on campus housing.” This study did not employ an a priori informed consent procedure to avoid selection bias, which has been proven to increase sevenfold the effects of behavioral interventions46. We employed an opt-out procedure: at the entrance of both housing buildings, a public message was displayed informing that a study would be conducted in the following months, registering the energy consumption from each apartment. The households unwilling to participate could opt out by informing the research team (email provided for this purpose—three households chose to opt out).

Study sample

One hundred households (apartments) were included in this study, distributed over two buildings. These apartments are located on the campus of an international university in the United Arab Emirates, where research staff lived (alone or with their partners/families) and paid a flat fee for utilities (i.e., same cost regardless of energy consumption). The 100 households were randomly allocated to different experimental groups using the Excel random function. An Excel sheet with all apartments was constructed, and an Excel random number function was generated. Individual apartments were assigned to experimental groups according to the random number assigned. Randomization was done at the apartment level and not at the floor level or building level.

The only information provided to the research team was the number of bedrooms per apartment and the daily total consumption of energy in kWh. No sociodemographic information could be collected about the residents, e.g., no information about the total number of residents per apartment, gender distribution, average age, or presence of children. However, randomization was conducted at the (individual) apartment level, which is expected to assure that any differences between households are equally (or very similarly) distributed between experimental groups. An additional suggestion that randomization was effective in guaranteeing this between-group homogeneity in aspects such as household size, quality of appliances, or energy class was the test for baseline (pre-intervention) energy consumption, which was not statistically different between groups.

The study ran for 264 days, from 9/11/2019 to 5/31/2020. An announcement was placed at the entrance of the building stating that energy consumption would be monitored in the following months, and households which did not allow for this monitoring could email the research team to opt out. Access to the buildings was granted by the Facilities Management to the research team to place the weekly messages during the experimental period.

One household was excluded from the estimation because we observed some negative daily electricity usage. Descriptive statistics at baseline are presented in Supplementary Table 1. For a balance check, we run pairwise Wilcoxon–Mann–Whitney tests (two-sided).

Outcome: daily total kWh per household

The primary outcome under analysis was the daily total kWh used per apartment, given to the research team in an Excel datasheet by the Campus Facilities Management at the end of each week (on Fridays). The Facilities Management was blinded to the randomization of the apartments to different experimental groups.

Experimental procedure

The image included in the message was selected among dozens of images that were piloted (via TurkPrime n = 289 https://www.cloudresearch.com/) to identify the image that promoted the greatest sense of attachment security. Visual imagery is particularly effective in priming secure attachment28,29. We tested 30 images, including images related to (1) mother–child bonding, (2) nature scenarios, and (3) mother–child bonding in nature. Each image was evaluated using five-point Likert scales based on the following questions: “How much does this picture makes you feel…loved and protected? Safe and secure?” The image selected (shown below in Message Stimuli) represented the prototypical primal bonding (a mother) and portrays Earth as our mother, of whom we are all children. This was the image identified in the pilot testing as promoting more feeling secured and loved. Given that the image selected produced the psychological effect of attachment security, this established construct validity for the message to be used in field studies to assess predictive validity to produce real-world behavior changes as predicted by the theory. This picture has been tested successfully in the field in past research30. Our contribution here is the application of attachment security to energy reduction and its synergetic link with social norms messaging. The images selected (i.e. pregnant woman with an Earth-shaped bell—shown below) were evaluated the highest on these three items.

The messages were taped to each apartment door, faced down. This means the content was not visible to other households. When participants arrived at their apartment the day the message was delivered, they would see an A4-size sheet of paper taped to their entry door.

The messages were designed as shown below. The control group (A) only received tips on how to save energy. The standard social norm group (B) received the same tips plus comparative information about their energy consumption versus neighbors in apartments with the same number of bedrooms. When the household was using more energy than their neighbors, the message would depict a red bar with a frowny red face. The enhanced social norm included the elements of a standard social norm message plus an image priming attachment security (C)–which had been determined in the pilot study. The image is the only difference between the enhanced social norm + attachment treatment and the standard social norm treatment. A single change in experimental stimuli guarantees the internal validity of the results. More than one additional change between experimental groups weakens the quality of the experimental protocol (Fig. 6).

Fig. 6: Experimental messages per group.
figure 6

A Control Group; B Standard Social Norm Group; C Enhanced Social Norm + Attachment Group.

Full size image

Secure attachment priming indeed does not change or increase individuals’ attachment to people or change participants’ relationships or what they’ve learned from those relationships28,29. However, secure attachment priming works by activating an association or representation in memory; it makes some mental structures become more salient and active. Although attachment security is typically developed within the context of a supportive, reliable relationship, that is generally assumed to take considerable time, there is ample evidence that several methods can be used to create short-term changes in people’s sense of security. Priming is one of them. Security priming entails exposing individuals to stimuli designed to activate a sense of love, comfort, and safety. Then, by way of “spreading activation,” this creates a sense of security similar to what is induced by the presence of supportive others who provide love, comfort, and security (termed attachment figures).

Estimation strategy for main analysis

For the main analysis, we estimated the ATE of each intervention with a fixed effect model using the following specification:

$${y}_{{it}}={\beta }_{1}+{\beta }_{2}{{{{\rm{Post}}}}}_{t}+{\beta }_{3}{{{{\rm{Arm}}}}}_{i}* {{{{\rm{Post}}}}}_{t}+{M}_{t}\cdot {{\Gamma }}+{v}_{i}+{\sigma }_{d}+{\mu }_{w}+{e}_{{it}}$$
(1)

where \({y}_{{it}}\) is apartment i’s electricity use on day t. \({{{{\rm{Arm}}}}}_{i}\) is a categorical variable indicating which treatment arm apartment i is in. The regression sample period consists of a 4-week pre-intervention (baseline) period and a 4-week intervention period. \({{{{\rm{Post}}}}}_{t}\) is an indicator variable that equals one if day t falls under the 4-week intervention period and equals zero before the first message sent out. \({M}_{t}\) is a vector of weather conditions on day t, including maximum temperature, humidity and their square terms. And \({v}_{i}\), \({\sigma }_{d}\), and \({\mu }_{w}\) are apartment fixed effects, day-of-the-week fixed effects, and week fixed effects, respectively.

Our estimator of interest, \({\beta }_{3}\), measures the ATE in the households of the intervention arms. In consideration of serial correlation in error terms within an apartment over time, and “small N, large T” feature of our dataset, we conducted the estimation using a fixed effects model with Driscoll and Kraay standard errors. Errors are not clustered, given that the randomization is performed at the apartment level47.

Estimation strategy for high-frequency analysis

To take advantage of the daily frequency of our dataset, we conducted high-frequency analysis by running the following set of regressions:

$${y}_{{it}}={\tau {{{\rm{Arm}}}}}_{i}+{\theta y}_{{ib}}+{\pi }_{t}+{\varepsilon }_{{it}},\,\forall t\in [d-n,d+n]$$
(2)

where \({y}_{{ib}}\) is the average daily usage for apartment i during the pre-intervention baseline period, and \({\pi }_{t}\) is date fixed effects. n equals 1 for the 3-day running mean treatment effects for each day. And n equals 3 for the 7-day running mean treatment effects for each day.

Examining of “action and backsliding”

To examine the “action and backsliding” pattern by treatment arms, we borrowed the strategy from Allcott and Rogers35 and run the following regression:

$${y}_{{it}}= {\Phi {{{\rm{Poster}}}}}_{t}* {{{{\rm{Arm}}}}}_{i}+\tau {{{{\rm{Arm}}}}}_{i} +{\beta }_{1}{{{{\rm{Arm}}}}}_{i}* {M}_{t}+{M}_{t}\cdot {{\Gamma }}+\theta {y}_{{ib}}+{\mu }_{w}+{\varepsilon }_{{it}}$$
(3)

where \({{{{\rm{Poster}}}}}_{t}\) is the number of days after receiving the latest message, and our coefficient of interest \(\Phi\) measures the change in treatment effect from the day of message arrival and subsequent days. Figure 2 plots the value of \(\Phi\). We include the control arm in the graph since these households also received a placebo message on the day of message arrival.

Assessing the persistence of treatment effects

To investigate how persistent the treatment effects are, we obtained the measurements from the following regression:

$${y}_{{it}}={\beta }_{1}+{\beta }_{2}{{{{\rm{Month}}}}}_{t}+{\beta }_{3}{{{{\rm{Arm}}}}}_{i}* {{{{\rm{Month}}}}}_{t}+{M}_{t}\cdot {{\Gamma }}+{v}_{i}+{\sigma }_{d}+{\mu }_{w}+{e}_{{it}}$$
(4)

where “Month” is measured in 4-week intervals.

Deviations from the original pre-registered protocol

There are some deviations from the original pre-registered protocol, which we present next. Some deviations offered less than initially proposed, while some other deviations were substantial improvements from the original protocol.

The initial plan was to conduct a factorial randomized controlled trial, testing between-subjects (units) the main effect of human attachment (1), the main effect of social norms (2), the interactive, combined effect of human attachment + social norms (3) and a control group (4) (i.e., basic energy saving tips). We did implement this design in the field and collected data from these 4 groups. However, when the time came to write the paper, the focus became strictly to showcase how social norms can be improved, as this was considered the most compelling contribution to the literature. Thus, the focus of the paper was centered around how attachment could act as a boost to social norms, already been previously established to be effective. The standalone effect of Attachment was dropped.

Initially, it was specified that this trial would involve a 4-week baseline period, a 4-week intervention, and a 4-week follow-up. The change was that the follow-up has substantially extended to 9 months.

The blinding procedures also changed. Originally, triple blinding was proposed. However, the co-author responsible for the data analysis was aware of the experimental conditions. This procedure was revised because the person responsible for data analysis was included in the team several years after the experiment was conducted, and an explanation of the procedural details was explicitly guided.

Regarding sample size, initially, it was reported that 201 apartments would be included. This was an initial miscalculation. In fact, only 133 apartments were tested in this research. The 33 excluded were related to the Attachment group. Therefore, the final apartments sample under analysis is 100.

In terms of the analysis plan, initially, a very simple plan was detailed (“Controlling for baseline (time 1), we will use between-subjects ANOVA (time 2 and time 3), P value 0.05 two-tailed”). However, with the integration of an expert applied econometrician, the data analysis was heavily improved.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.