Using pledges to improve the effectiveness of environmental information campaigns: The case of biowaste recycling

Significance

Effective climate action requires sustained citizen engagement in pro-environmental behaviors. Behavioral interventions show promise in promoting pro-environmental actions, yet their long-term effects warrant further investigation. This longitudinal field study, using real-time data on biowaste sorting, finds that a pledge enhances the effectiveness of an environmental campaign to promote waste sorting. Following an ‘action-and-backsliding’ pattern before stabilizing, the effect persists for four years. Our study is pioneering in showing that a pledge intervention can generate a pro-environmental impact lasting for years, contributing to the debate on the persistence of behavioral interventions. A direct policy implication is that soft commitments can improve the effectiveness of informational environmental campaigns at minimal cost and with enduring effects.

Abstract

Through a field study (N = 1,519) that uses a technology to record real-time data on waste sorting, we find that offering the opportunity to sign a pledge increases the effectiveness of an environmental campaign. With a timespan of over four years, the pledge increased waste sorting participation by 4.55 to 5.10 percentage points (SD = 0.1997). The effect is greater immediately after the campaign (around 9 to 10 pp during the first 15 wk), but it remains sizable and statistically significant 150 to 210 wk after signing (3.11 to 4.45 pp).

Recycling is an important policy for reducing CO₂ emissions and mitigating climate change (1). Citizen engagement in pro-environmental activities is essential for environmental policies like recycling, saving energy, or reducing water consumption. Companies, governments, and NGOs invest in campaigns to promote pro-environmental behavior. We study whether being given the opportunity to sign a soft commitment, i.e., a nonbinding pledge to recycle (2), can increase the effectiveness of an environmental campaign aimed at encouraging waste sorting.

Traditionally, waste sorting in urban areas has been anonymous. However, new technologies make monitoring easier and more feasible. At the end of 2018 and the beginning of 2019, electronic bins for collecting biowaste were installed in the municipality of Palma, Spain.^* These bins were locked, and a personal card had to be scanned to open the lid. This feature provided real-time data on individual participation in waste sorting. We take advantage of this technology to run an intervention evaluating the effectiveness and dynamics of soft commitments with regard to fostering recycling.

For our study we partnered with the company in charge of waste management in the municipality. The company was running an environmental campaign that we used to perform the intervention. The campaign was conducted by a group of environmental educators who informed and encouraged citizens to sort biowaste in various neighborhoods. The educators invited citizens to participate in our study. A quasi-random subsample of the participants was given the opportunity to sign a soft commitment, a form where participants pledged to sort their biowaste. Citizens who were given the option to sign the pledge make up our treatment group, regardless of whether they actually signed it. Our control group is the participating households that were not given the opportunity to sign the pledge. This setting offers us the possibility to neatly estimate the ability of the soft commitment to improve the effectiveness of the pro-environmental campaign.

Our total sample included 1,519 households (46.71% in the control group and 53.29% in the treatment group). The two groups saw a sharp increase in recycling immediately after the campaign. While control and treatment groups behaved identically during the preintervention period, the increased participation in waste sorting that followed the campaign was greater in the group that was given the opportunity to sign the pledge. Averaging over a long period (210 wk after the campaign) we show that the pledge increased participation in waste sorting by 5.10 percentage points (0.1997 SD). In relative terms, the effect represents a 30% increase compared to the control group. The effect is greater immediately after the campaign (around 9 to 10 percentage points during the first 15 wk), but it remains sizable and statistically significant 150 to 210 wk after (4.49 percentage points). Finally, by running a follow-up feedback intervention, we observe that it did not increase recycling in the long-run, and do not find evidence of any interaction with the pledge.

While our paper addresses the ability of pledges to reinforce a waste sorting campaign in achieving its target, a more comprehensive analysis of the environmental consequences of our intervention should also consider its potential spillovers on nontargeted behavior (e.g., waste generation or energy conservation). These spillovers can be positive—reinforcing other proenvironmental actions—or negative—undermining other proenvironmental behaviors. While uncommon in pro-environmental interventions (3–5), negative spillovers require particular attention due to their potential to undermine the effectiveness of environmental campaigns. In contrast, positive spillovers imply that pro-environmental effects of interventions extend beyond the targeted behavior. Although we cannot address this question for the soft commitment intervention, an exploratory analysis using aggregate data to investigate the relationship between biowaste sorting and waste generation in our setting suggests the absence of negative spillovers on this domain.

The use of voluntary commitments to shape individual behavior has recently gained interest in economics and management (for a review, see ref. 2) and is well established in the field of social psychology (for a discussion on the foundations of commitment theory see refs. 6–10). Commitments involving a self-imposed penalty for not complying have proven useful when it comes to helping individuals overcome self-control problems in several contexts, e.g., increasing their savings (11, 12), going to the gym (13), quitting smoking (14), reducing alcohol consumption (15), and treating digital addiction (16). When voluntary commitments do not carry a penalty for noncompliance, they are labeled as soft commitments, which generally involve the signing of a pledge to embrace the desired behavior. Despite their nonbinding nature, soft commitments can modify conduct through psychological mechanisms, such as the desire for consistency (17, 18), cognitive dissonance (19), guilt aversion (20), or individuals’ preferences to keep promises (21). If these behavioral mechanisms work, the lack of a penalty can be seen as an advantage because it leads to higher take-up. Soft commitments have been found to work in laboratory experiments (20–23), yet results in the field seem to depend on context and design. For instance, while some works found positive effects in improving savings (24) and student effort (25), others found null results when it came to promoting mask-wearing during the COVID-19 crisis in Bangladesh (26). When it comes to pro-environmental behavior, pledges are a very popular mechanism. For instance, the European Commission specifically encourages pledges and commitments in its European Climate Pact (https://climate-pact.europa.eu/about/about-pact_en).^† Despite their pervasiveness, evidence on their effectiveness is mixed. They have been found to encourage towel reuse in hotels (27) and promote public transportation (28) but not to save shower water (29). On the other hand, commitments were effective at reducing energy consumption only if made public (30, 31). Due to the difficulty in monitoring waste sorting, evidence on the effects of soft commitments in this domain is restricted to specific contexts, like single-family homes (32–37), student residences (38, 39), on-campus college housing (40), complexes with collective bins (41), and retirement homes (38). This difficulty also limited sample size and time horizon of previous studies.^‡ Overall, the evidence from these previous studies remains inconclusive. While some studies (32–34, 38) found positive results, others (35–37, 40, 41) obtained null results (we consider only the cases where commitments were not combined with other mechanisms). Our study is a pioneer in its evaluation of the effect of soft commitments on a large scale, in a densely populated urban area, and over a long time horizon.

Beyond identifying a positive effect of soft commitment on waste sorting, our main contribution lies in demonstrating its long-term impact, which persists for 4 y after the intervention. Our study provides empirical evidence that a pledge intervention can generate a pro-environmental impact that endures for several years. This finding highlights the potential of light-touch behavioral interventions to drive sustained behavioral change, strengthening the effectiveness of environmental campaigns. More broadly, this paper advances the literature on the long-term persistence of behavioral interventions, an area where prior research has emphasized the need for further investigation (44–51). By leveraging the real-time features of our data, we also provide a characterization of the dynamics of soft commitment effects on waste sorting. Analyzing the effects at a weekly level, we identify an action-and-backsliding pattern similar to those documented in other domains, such as energy and water conservation (49, 51). This pattern is characterized by an initial surge in the promoted behavior, followed by a subsequent decline over time. In our case, backsliding stopped after approximately 10-20 wk and did not fully offset the initial gains. Finally, we exploit the temporal dimension of our data to analyze time-of-day transition matrices, which suggest that habit formation (52, 53) is relevant for recycling, which could explain the persistence of the effect (44).

In addition to its main contribution regarding the long-term effects of soft commitments, our article also offers further insights. First, while previous field studies have already shown that soft commitments can be effective in several domains (24, 25, 27, 28, 30–34, 38, 54–58), our study strengthens the evidence by providing field results. Thus, it contributes to reinforcing scientific consensus, testing the robustness of earlier findings, and assessing their applicability across different contexts. Second, although our field study does not identify the specific factors that determine the effectiveness of soft commitments, we provide insights into their mechanisms by separately estimating their effects on the extensive and intensive margins. We show that the effect is primarily driven by an increase in the number of households that started recycling (extensive margin) rather than by increased adherence among households that were already recycling (intensive margin). Among studies that estimate these two margins for soft commitments, the extensive margin does not always drive the effect. While soft commitments primarily operate through the extensive margin in charitable donations (57) and savings (55), they influence the intensive margin in the domains of donations (58) and gambling avoidance (56). Finally, since waste sorting contributes to reducing CO₂ emissions, our findings add to the evidence on the effectiveness of behavioral mechanisms in combating climate change (59). Given the low cost of incorporating a pledge into face-to-face pro-environmental campaigns, this result has important practical implications for the design of such interventions.

This paper is organized as follows. Sections 1.1 and 1.2 describe the intervention design and the data. Section 2 details the identification strategy and presents the main results, while Section 3 explores the robustness of the findings together with further results. Section 4 concludes.

1. Materials and Methods

All data, codes, and replication materials are available at https://osf.io/jkh3w/.

1.1. Intervention Design.

Our intervention was conducted within an existing environmental information campaign, which was run by a team of environmental educators hired by the municipal waste management company (our partner). Environmental educators informed citizens about the introduction of biowaste separation. On the basis of our partnership agreement, the environmental educators were instructed by the research team on how to recruit participants and run the intervention. The project was approved by the IRB of the Universitat de les Illes Balears (reference 128CER19).

1.1.1. The setting.

With around 420,000 citizens, Palma is the largest city in the Balearic Islands, Spain. Our partner, a municipally owned company called EMAYA, is in charge of managing and collecting urban waste. In 2018, EMAYA started a program to introduce biowaste recycling in the city, a type of waste that had not previously been sorted. Initially, the bins were only brought into some specific areas of the city, and bin installation took place in two stages, the first starting in November 2018, and the second in March 2019.

To avoid improper sorting, which would compromise the processing of biowaste, the installed bins were locked, and the lid could only be opened by scanning a personal card (residents’ city transportation cards). Most of the residents in the city have the card, as it is free and its use reduces fares on public transport. Households without any card-holding members were excluded from the study because they could not be matched with administrative data (Section 1.2). Importantly, every time that a bin is used, it records the number of the scanned card. This provides real-time individualized data on waste sorting.

1.1.2. Implementation.

The team of environmental educators set up information points at different locations around the area in which the program was implemented. Banners were used to gain visibility, and recycling kits were given to encourage citizens to approach the educators.

The interaction between educators and citizens took place as follows. First, environmental educators informed citizens about the introduction of biowaste separation, responded to citizens’ questions about the process, and encouraged them to recycle. Afterward, citizens received the recycling kit, which included a small bin and recycling bags. Finally, citizens were given the option to participate in our study by signing an informed consent release (SI Appendix, Figs. S1 and S3).

If the citizens agreed to participate in the study, they signed an informed consent. Afterward, they were asked for some personal and household information, including the households’ number of weekly disposals of unsorted waste. Based on that self-reported number, educators informed them about their expected number of disposals if biowaste were regularly sorted (MENUCO: minimum expected number of uses of the container). Thus, the number was relevant for informing participants—both in the control and treatment groups—of the criteria that would qualify them as recycling households. Specifically, the MENUCO was set equal to 1 if the number of weekly disposals of nonrecyclable waste were 1 or 2, set equal to 2 if there were 3 or 4, and was set equal to 3 if there were 5 or more such disposals. The numbers were agreed upon with the waste management company on the basis that roughly half of nonrecyclable waste corresponds to biowaste.

For participants assigned to the control group, the interaction ended here. For the treatment group, instead, the educator gave the subjects the option of committing to recycling biowaste by signing a form. Then, in front of the educator, the participant decided whether to commit (YES/NO) and filled in the corresponding form with his/her full name, fiscal ID, and signature (SI Appendix, Figs. S2 and S4). The simple design of our pledge form may be interpreted as a cheap talk script (60, 61) that helps the subject better focus on the task at hand. However, the interaction with the educator, explaining the meaning of the pledge and acting as a witness, makes the commitment more meaningful than a cheap talk script. Indeed, the use of a voluntary pledge in front of a witness is an important feature shared with solemn oaths (8, 62).

Note that, as usual in the evaluation of soft commitments, our treatment is defined as being offered to sign the soft commitment, rather than actually signing it. Doing so not only prevents selection bias but also provides the most appropriate analysis from a policy-making perspective. Equivalently, one can consider that we are estimating an intended-to-treat effect. In any case, 782 out of 810 households signed the soft commitment, a 96.54% acceptance rate, which makes the effect on those who were given the opportunity to sign the commitment and those who actually signed the commitment identical.

For treatment assignment, we used a quasi-random procedure based on order of arrival. Participants were allocated to control or treatment condition in alternating order (zipper strategy). That is, if one participant was allocated to the control group, the next was allocated to the treatment group, and so on.^§ We chose this mechanism to facilitate the work of the environmental educators.^¶ Although this assignment procedure does not formally qualify as random, it is exogenous and expected to work as good as random (63). Moreover, it was unpredictable to participants and researchers (participants were unaware about the control trial nature of the intervention, while recruitment and assignment was external to the research team). Reassuringly, as will be shown below, by means of balancing tests and placebo tests, the mechanism produced two ex ante identical groups, working as a purely stochastic process.

1.1.3. Follow-up feedback.

In February 2020, between 12 and 54 wk after the commitment campaign, we ran a feedback experiment. It consisted of sending a letter to all participants in the soft commitment intervention, reminding them about biowaste sorting and their MENUCO. Additionally, the letters to the treated group contained absolute feedback, which consisted of classifying the household according to its engagement in recycling (“Satisfactory,” “Can improve,” “Unsatisfactory,” or “Very unsatisfactory”). Assignment to the treatment group was randomized by computer within the two groups of the soft commitment intervention (i.e., a 2 × 2 design). This process divided the initial sample (N = 1,519) into four experimental groups: no soft commitment and no feedback (N₁ = 347), soft commitment and no feedback (N₂ = 416), feedback and no soft commitment (N₃ = 362), and soft commitment and feedback (N₄ = 394). SI Appendix, section 2.E provides further details on this follow-up intervention.

1.2. Data.

To conduct the study, we used three different types of data: field data, administrative data, and biowaste disposal data.

1.2.1. Field data.

During the campaign, the educators collected the following information: the participant’s name and surname, national identification number, number of members in the household, MENUCO, treatment assignment, and address.^# We used the latter variable for retrieving income data from the Spanish census. Such data is publicly available from the Spanish National Statistics Bureau (median income is provided for areas of 1,000–2,500 residents, making them quite accurate).

Field data were manually collected and handwritten on the informed consent sheet (SI Appendix, Fig. S3), on the soft commitment sheet (SI Appendix, Fig. S4), and in the educators’ log files. Environmental educators recruited 1,878 households for the study from January 28 to November 18, 2019.

1.2.2. Administrative data.

Waste is generated at the household level. However, electronic bins are able to provide data on the individual level, via card scanning. We sent the list of participants to the local body responsible for issuing and managing the cards. They returned to us anonymized data containing the card numbers of the participants and their cohabitants. This procedure allowed us to aggregate biowaste disposal at the household level.

We used postal addresses (street and number), national identification numbers, and complete names to match participants’ data to their transportation card numbers. The matching process was successful for 80,88% of the participants recruited initially (unmatched cases might be households with no members holding a card or inaccuracies in handwritten field data). This process yields a sample size of 1,519 households, divided into 709 (46.68%) in the control group and 810 (53.32%) in the treatment group.

The imbalance in the size of the two groups comes from the start of the recruitment process, when national ID numbers (the most effective matching variable) were only requested from the participants in the treatment group (SI Appendix, Fig. S5 shows that the imbalance came from the first 10 wk of recruitment). This difference in sample sizes during the first weeks of recruitment implies that the average length of the period for which we can observe preintervention outcomes for the control group (17.24 wk) and the treatment group (15.26 wk) differs significantly (P-value < 0.01).^‖ As the reason for the imbalance was exogenous and only related to the effectiveness of the matching process, it should not affect the estimation of the treatment effect, only its precision. Nevertheless, in SI Appendix, section 2.C.2, we exhaustively analyze the implications of this imbalance and provide clear evidence that it does not affect the results. Among other arguments, we show that our results would not change if we corrected the imbalance by repeating the matching process after omitting the national IDs of the treated group in the same period for which this information was not available for the control group.

1.2.3. Biowaste data.

The following information was recorded during each use of electronic bins: a user identifier (anonymized), a bin identifier, the date, the time when the card was scanned, the time when the lid was opened, and the time when the lid was closed.

We consider a period of 210 wk after the campaign. Given the staggered recruitment process, this period meant different calendar dates based on when participants were recruited. Bins installation was completed in two phases. The first one finished on November 17, 2018, and the second on March 24, 2019. Most participants (1,179) were recruited after the installation of the electronic bins in their neighborhood (77.62% of the sample), which means we can observe their bin usage before being recruited. Participating households were included in the analysis from the moment that bins were made available in their neighborhoods.

1.2.4. Outcome variables and descriptive statistics.

An advantage of our study is that we can track waste sorting at an individual level in densely populated areas. However, our data do not provide information on the actual amount of waste disposed; they provide only information on bin usage. To construct the outcome variables, we consider bin usage to be a proxy for recycling behavior. We find this step to be legitimate after cross-checking data on lid openings and the aggregate amount of biowaste that was collected. The two measures follow a very similar pattern over time and have a correlation above 0.95 (SI Appendix, Fig. S6).^**

Three outcome variables are constructed at the week-household level using bin registries: #Uses, which aggregates the weekly number of lid openings; DoF which divides #Uses over the MENUCO to measure the degree to which households fulfilled their commitment (with truncation at 1 denoting full compliance); and %Weeks, which considers weekly participation in waste sorting (i.e., a dummy variable taking value one every week that there is a disposal).^†† The latter captures regular participation in waste sorting, and we argue it is the most relevant and accurate outcome we can obtain with the available data. First, participation measures are more convenient from an environmental perspective because preventing waste generation (precycling) is preferable to recycling. Cardinal measures like #Uses and DoF do not depend solely on recycling but also on waste generation (e.g., households with more leftovers make more disposals, which is not environmentally better). Second, one of the limitations of our data (which measures lid openings but not the amount of waste disposed) makes the measure of participation a more reliable outcome than measures based on the number of uses. Weekly participation is hardly affected by household heterogeneity in waste practices. By contrast, since the amount of waste is not observable, cardinal measures are likely to be affected by such heterogeneity. For instance, some households might make frequent small disposals while others might make less frequent larger disposals but recycle the same. Participation is also less affected by seasonality (more weekly disposals during warmer periods) and household size. If that was the case, the participation outcome would be unaffected by such an effect. Finally, DoF has an additional source of inaccuracy as it is based on a self-report. All in all, we use %Weeks to display results in the main text. The results are similar when we consider #Uses and DoF as outcomes (SI Appendix, Table S4).^‡‡

Panel A in Table 1 shows the descriptive statistics and balancing tests. Sample sizes change slightly for the self-reported variables due to the existence of missing values. Overall, there are 1,483 households (685 corresponding to the control group) with all controls available. The average household in our study includes 2.9 people (SD = 1.268), has a MENUCO of 1.8 (SD = 0.861), and a median monthly income of €1,717.68. The control and treatment groups are balanced in terms of their observable characteristics. Panel B in Table 1 compares the outcomes of the control and treatment groups during the preintervention period. The table shows the three outcomes described above (%Weeks, #Uses, and DoF) as well as the fraction of households that never used the bins (%Inact. Users). Reassuringly, the assignment mechanism based on subjects’ arrival worked as good as random, as the two groups exhibited identical intervention outcomes before recruitment.

Table 1.

Descriptive statistics, balancing tests, and preintervention outcomes

	N	Control	Treatment	Diff	P-value
	(1)	(2)	(3)	(4)	(5)
Panel A: Descriptive statistics and balancing tests
Size	1,505	2.874	2.891	0.018	0.7858
		(0.0485)	(0.0442)	(0.0656)
MENUCO	1,493	1.854	1.797	−0.057	0.1992
		(0.0332)	(0.0301)	(0.0448)
%Phase = 2	1,519	0.423	0.464	0.041	0.1083
		(0.0186)	(0.0175)	(0.0255)
Income	1,519	1,708.25	1,725.95	17.70	0.2808
		(12.15)	(11.07)	(16.41)
Panel B: Preintervention outcomes (placebo tests)
%Weeks	1,179	0.081	0.082	0.001	0.948
		(0.0095)	(0.0092)	(0.0133)
#Uses	1,179	0.174	0.167	−0.006	0.850
		(0.0243)	(0.0228)	(0.0333)
%Inact. Users	1,179	0.816	0.819	0.003	0.894
		(0.0164)	(0.0155)	(0.0225)
DoF	1,160	0.068	0.066	−0.002	0.852
		(0.00862)	(0.00783)	(0.0116)

Column (1) displays the number of households for which we observe each variable. Columns (2) and (3), show the averages for control and treatment households, respectively. Column (4) shows the difference in means and column (5) its corresponding P-value for the t-test of equal means. Size refers to the self-reported number of people living in the household. MENUCO refers to the minimum expected number of uses of the container per week. %Phase = 2 shows the proportion of households living in the areas where bins were installed later (second phase). Income is a proxy for household income imputed from using the median income at the census tract level. For Panel B, %Weeks refers to the proportion of weeks the container is used at least once. #Uses refers to the weekly average number of uses of the container. %Inact. Users is the percentage of households that never used the container. DoF is the degree of fulfillment computed as #Use/MENUCO (truncated at 1). SE in parentheses.

2. Results

Fig. 1A plots the evolution of the proportion of households that recycled each week during the study period separately for the control (dashed line) and treatment (solid line) groups. Vertical bars display the sample size, which depends on recruitment and the availability of electronic bins in participants’ neighborhoods (Section 1.2.3). The horizontal axis shows the number of weeks before and after the information campaign (i.e., recruitment). Thus, t = 0 corresponds to the intervention week, t < 0 to preintervention weeks (before recruitment), and t > 0 to postintervention weeks (after recruitment).

Fig. 1.

Weekly evolution of recycling behavior in the control and treatment groups (A) and of average treatment effect (B) up to 210 wk after the intervention. The x-axis shows the number of weeks from the intervention, with negative values being the preintervention weeks and positive values being the postintervention weeks. In A, lines represent the percentage of households recycling in the control and treatment groups, while the bars (right axis) indicate the sample size. In B, green circles represent the treatment effect for week $j\in \{1,...,210\}$ after passing by the table, which is estimated as ${\beta }_1+{\beta }_3^j$ from the regression $y_{\mathit{it}}={\beta }_0+{\beta }_{1}S{C}_i+\sum _{j=1}^{210}{\beta }_2^{j}WeeksAfte{r}_{\mathit{it}}^j+\sum _{j=1}^{210}{\beta }_3^{j}S{C}_i\times WeeksAfte{r}_{\mathit{it}}^j+\theta X_i+{ϵ}_{\mathit{it}}$, where $WeeksAfte{r}_{\mathit{it}}^j$ is a dummy variable taking value one when the weeks elapsed since the intervention for household i at period t is equal to j, and $S{C}_i$ is an indicator taking value 1 if household i was given the opportunity to sign a soft commitment. The solid line represents a smoothed approximation of the treatment effect over time, estimated using a random-effects model with a third-degree polynomial time trend: $y_{\mathit{it}}={\beta }_0+{\beta }_{1}S{C}_i+\sum _{p=1}^3{\alpha }_p{\left(W{A}_{\mathit{it}}\right)}^p+\sum _{p=1}^3{\gamma }_{p}S{C}_i\times {\left(W{A}_{\mathit{it}}\right)}^p+\theta X_i+{ϵ}_{\mathit{it}}$, where $W{A}_{\mathit{it}}$ is the number of weeks elapsed since the intervention for household i at time t. Dashed lines account for the 95% confidence band. SE are clustered at the household level. X_i includes household characteristics (MENUCO, number of inhabitants, bin installation phase, and household income, as well as fixed effects for household postal code and recruitment week).

The figure shows a sharp increase in waste sorting at t = 0, which indicates the important effect of the environmental campaign. Before being informed, less than 10% of households were recycling, but immediately after the campaign, close to 40% of households in the control group did so, and nearly 50% in the treatment group. This increase was sharp, and it clearly emerged at the intervention week, showing that the information campaign was responsible for the surge in recycling. Second, from t = 0 onward, we see a positive gap between the treatment and control groups. This gap represents the ability of the soft commitment to improve the effectiveness of the environmental campaign. Third, the effect of the campaign steadily declined over time for both groups. By contrast, the gap remains quite stable, suggesting that the dynamic effect of the pledge persisted 210 wk after having been offered the soft commitment. The differing dynamics between the effects of the pledge and the environmental campaign resulted in a notable contrast in the long-term outcomes for households in the control and treated groups. While the environmental campaign initially had a much stronger impact—about three times greater than the pledge—recycling levels in the control group eventually returned to their original levels after 170 wk, with around 10% of households recycling. In contrast, recycling rates in the treated group remained significantly higher than their initial levels. This implies that the soft commitment not only enhanced the campaign’s impact but also extended its duration.

We confirm the stability of the effect after almost four years by estimating the week-by-week treatment effect and the smoothed curve of the treatment effect in Fig. 1B. The confidence band indicates that the effect remains statistically significant throughout the entire study period, confirming that a light-touch mechanism (being given the opportunity to sign a pledge) is highly persistent, with an effect size of around 4.5 percentage points.^§§ We see another remarkable pattern in Fig. 1: despite the effect being persistent over time, it seems to backslide until around weeks 15 to 20, after which it stabilizes. The econometric analysis in the two subsections below analytically confirms the insights from Fig. 1.

2.1. Average Treatment Effect.

We average the outcome variable over the 210 wk ($y_i=\frac{1}{210}\sum _{t=1}^{210}{y}_{\mathit{it}}$) and estimate the average treatment effect (ATE), following ref. 64. Consequently, the main outcome captures the proportion of weeks participating in waste sorting. The benchmark estimates for the effect of the pledge are obtained through an OLS estimation of the following equation:

$$\begin{array}{c}\hfill y_i={\beta }_0+{\beta }_{1}S{C}_i+\theta X_i+{ϵ}_i,\end{array}$$ [1]

where $S{C}_i$ is an indicator taking value 1 if household i was given the opportunity to sign a soft commitment and zero otherwise; X_i is a set of household-specific controls (MENUCO, number of household members, and household income) and fixed effects (postal code and recruitment week); and ϵ_i is an error term.^¶¶ For households recruited after bin installation, we also add a specification in which we control for their waste sorting before being recruited (i.e., the prerecruitment value of the dependent variable). The main coefficient of interest is β₁, which captures the average treatment effect across the 210 wk period of offering the soft commitment on y_i. Considering the fractional or count nature of our outcomes, the residuals do not follow a normal distribution. Thus, the usual Huber-Eicker-White sandwich correction for SE was applied.

Columns (1) to (3) in Table 2 show that, on average, during the 210 wk period pledges increased the proportion of households that sorted their waste by around 5 percentage points (0.1997 SD). Considering the proportion of households that recycled in the control group [17.3%, the intercept in column (1)], the effect represents a 30% increase. This number can be read as the increase in the effectiveness of the pro-environmental campaign that is obtained by giving citizens the chance to sign a pledge. Column (2) adds controls for household characteristics and fixed effects (recruitment week, postal code, and bin installation phase), while column (3) includes the preintervention value of the dependent variable as a control (for the subsample with these data available). As it would be expected with random assignment, the change in the coefficient of interest between columns (1) to (3) is negligible. Overall, these results confirm that the soft-commitment notably improved the effectiveness of the campaign.

Table 2.

Average treatment effect and long-lasting effects of the soft commitment

	ATE			Dynamic effects
	(1)	(2)	(3)	(4)	(5)
SC	0.0534***	0.0510***	0.0455***	0.0449***	0.0311**
	(0.0136)	(0.0139)	(0.0151)	(0.0138)	(0.0153)
Pre-int.			0.373***		0.414***
			(0.0405)		(0.0368)
SC × 15 wk				0.0563***	0.0621***
				(0.0196)	(0.0219)
SC × 16 to 50 wk				0.0152	0.0284
				(0.0161)	(0.0176)
SC × 51 to 100 wk				0.00503	0.0178
				(0.0132)	(0.0145)
SC × 101 to 150 wk				−0.00663	0.00577
				(0.00925)	(0.0105)
15 wk				0.250***	0.235***
				(0.0143)	(0.0157)
15 to 50 wk				0.141***	0.127***
				(0.0117)	(0.0123)
51 to 100 wk				0.0767***	0.0666***
				(0.00904)	(0.00967)
101 to 150 wk				0.0431***	0.0411***
				(0.00662)	(0.00731)
Constant	0.173***	0.0413	0.140	−0.0340	0.0786
	(0.00948)	(0.0908)	(0.122)	(0.0867)	(0.115)
N	1,519	1,483	1,153	7,415	5,765
Adj. R2	0.00928	0.0366	0.130	0.128	0.196
Controls	No	Yes	Yes	Yes	Yes
FE	No	Yes	Yes	Yes	Yes

Estimation of Eq. 1 in columns (1) to (3) and of Eq. 2 in columns (4) and (5) with the proportion of weeks recycling after the intervention as the dependent variable. Columns (2) to (5) include ZIP-code and Intervention week fixed effects (FE) and household controls. Household controls include the number of people living in the household (Size), the Minimum Expected Number of Uses of the Container per week (MENUCO), the bin installation phase (Phase) and household income (imputed from the census tract, Income). Pre-Int refers to the average value of the outcome variable up to 40 wk before the intervention (only available for a subsample). Robust SE in parentheses [clustered at the household level for columns (4) and (5)].*P < 0.10, **P < 0.05, ***P < 0.01

Our study cannot determine which parameters explain the difference from previous works that found no recycling effects of soft commitments. However, we provide insights on how they operate in this context by separately estimating the intensive and extensive margins. Using a two-part model (65), we show that the effect is mainly driven by an increase in the number of households that started to recycle (extensive margin) rather than by achieving more adherence among households that already recycled (intensive margin). See SI Appendix, section 2.A for further details. The extensive margin also drives effects of monetary incentives for recycling (66) but not necessarily when considering soft commitments on other domains (55, 56, 57, 58).

We also analyzed the possibility that soft commitments work differently depending on the value of the MENUCO by interacting the treatment dummy and the MENUCO. This analysis reveals that the effect of the pledge does not differ across MENUCO groups (SI Appendix, section 2.F).

2.2. Dynamic Effects.

Columns (4) and (5) examine the dynamics of the treatment effect. To do this, we break the data down over five time periods and average the outcome variable: the first 15 wk after the campaign, weeks 16 to 50, weeks 51 to 100, weeks 101 to 151, and weeks 151 to 210. This division of time is based on the dynamics observed in Fig. 1B, where the treatment effect is greater during the first 10 to 20 wk after the intervention and remains quite constant afterward. The results are similar when we divide into different periods of time. This leads to the following regression:

$$\begin{array}{c}\hfill \begin{array}{cc}\hfill y_{\mathit{ij}}& \hfill ={\beta }_0+{\beta }_{1}S{C}_i+{\beta }_2{W}_i^{1-15}+...+{\beta }_5{W}_i^{101-150}\hfill \\ \hfill & \hfill +{\beta }_{6}S{C}_i\ast W_i^{1-15}+...+{\beta }_{9}S{C}_i\ast W_i^{101-150}\hfill \\ \hfill & \hfill +\theta X_i+{ϵ}_{\mathit{ij}},\hfill \end{array}\end{array}$$ [2]

where $W_i^j$ takes value 1 when the observation comes from period j and zero otherwise.

The omitted group is the period of time between weeks 151 and 210 and, thus, β₁ captures the average treatment effect in our longest time horizon. The positive and significant coefficient for β₁ confirms that the soft commitment is still effective between 150 and 210 wk after being given the opportunity to sign the pledge, increasing the proportion of households that recycle by 4.49 percentage points in the full sample (3.11 if controlling for the preintervention value of the outcome). However, the effect is two or three times greater in the weeks immediately following the campaign (weeks 1 to 15). The effect declines at the beginning but remains stable and statistically significant afterward, as shown by β₆ being the only significant interaction coefficient.

This confirms the action-and-backsliding pattern observed in Fig. 1B.

The lasting effect of pledges might be surprising, especially considering the relatively short impact often seen in behavioral interventions (50) and the one-shot nature of soft commitments. Thus, the next question is, why does the effect persist as much as it does? First, it should be noted that our intervention took place within an existing environmental campaign. Commitment might be especially effective for reinforcing that campaign. For instance, all participants received a recycling kit which could make them more likely to fulfill the pledge.^## From the perspective of social psychologists, such a lasting effect can be explained by humans’ desire for consistency as a central motivator of behavior (18). Additionally, our pledge incorporates the fundamental elements that make voluntary commitments effective, i.e., it is freely taken, signed, and witnessed (10). In our case, the figure of the environmental educator could be especially significant as a witness, since it was the educator who informed the participant about the relevance of recycling and provided them with a recycling kit designed for the specific purpose of taking the action being committed to. Moreover, behavioral interventions have been found to be especially effective when applied to environmental policies and when administered face-to-face (45, 67), as in the present case. Guilt-aversion (20) and preferences for keeping promises (21)—two factors that have been found to be relevant in explaining the effects of soft commitments in the lab—may be amplified in the context of a face-to-face pro-environmental campaign that provided a free recycling kit.

The type of activity under consideration is probably also relevant to its persistence. Waste sorting is a regular activity performed on a daily basis, and it is prone to habit formation (44, 50). Models of habit formation predict that habit stock accumulates by repeating some actions or choices over time (68). According to these models, induced changes in behavior can be sustained over time if the role of habits is important enough in repeating the task at hand. To analyze the relationship between recycling and habit formation, we draw on the definition of “habit” as a regular pattern of behavior (69, 70). This definition relies on the idea that contextual cues, such as time and location, may trigger habits. Hence, if habits are relevant in the context of recycling, we expect to observe some regularity in household disposals, i.e., disposals being concentrated at a specific time of the day. We look for such time regularities in waste disposals via the transition matrices approach (71). SI Appendix, Table S3 shows the probability of making a disposal during each time window conditional on the time window during which the previous disposal was made. According to these transition probabilities, the most likely event is the repetition of the time a disposal is made. This pattern is stronger when looking at modal time windows for making disposals, i.e., from 6 pm to 8 pm and from 8 pm to 10 pm, which account for approximately 50% of all disposals: more than 40% of the households that make a disposal in one of these time windows make their next disposal in the same time window. These regularities show favorable evidence on the relevance of habits in waste sorting, which might be crucial to the lasting effect of the pledge. Finally, the dynamic effects of the soft commitment—showing an initial decline in the short run and a positive stable effect in the long run—are also consistent with habit formation. The initial decline in the effect of the soft commitment can be explained by some households initially reacting to the treatment but stopping recycling before forming the habit. In contrast, for the households that maintained the behavior for more than 10 wk, the habit seems to have formed, and the effect remains stable. A period of 66 d was found to be necessary for forming habits related to healthy behaviors (72).

3. Robustness Checks, External Validity, and Feedback

In this section, we summarize the additional analyses and discussions available in SI Appendix.

3.1. Robustness Checks.

Our results are robust to alternative specifications and analyses. First, we replicate Fig. 1 and our main estimates in Table 2 by considering the other outcome variables, #Uses and DoF. All previous results are confirmed when considering these alternative outcomes despite the estimations being less precise (SI Appendix, section 2.C.1). For instance, although the size of the effect for the first 15 wk doubles the size of the effect seen after 151 to 210 wk when considering #Uses as the outcome variable, the estimation does not identify a statistically significant decline in the treatment effect.

In SI Appendix, section 2.C.2, we exhaustively analyze the consequences of the imbalance in the size of the control and treatment groups originating from the matching of field records and administrative data (Section 1.2.2). As explained above (see footnote §), the inclusion of recruitment week fixed effects eliminates concerns over the differences in the length of the observed preintervention period. Additionally, for further assurance, we also show that the results are robust to repeating the matching protocol but omitting national ID numbers in the treatment group for the same period that it was unavailable for the control group. By doing this, the imbalance disappears, and the results remain unchanged.

Furthermore, given the specific features of the outcome variables (i.e., their fractional or count nature, as well as zero inflation), we consider other estimation methods in SI Appendix, section 2.C.3. Specifically, we used the proposal by Papke and Wooldridge (73) and a beta distribution to address the fractional nature of the dependent variable and a zero-inflated Poisson to correct for the high prevalence of zeros (SI Appendix, Fig. S7). All results remain unchanged.

Finally, some bin malfunctions were identified (e.g., recording failures). SI Appendix, section 2.C.4 analyzes the impact of these incidents on our results, finding that they have no effect on our conclusions.

3.2. External Validity.

Our findings show that giving people the opportunity to sign a commitment increases the effectiveness of pro-environmental campaigns. Still, two questions about the external validity of these results could be raised. First, it might be that households that approach the table are especially pro-environmental, limiting external validity. SI Appendix, Table S10 shows that participating households exhibited low levels of recycling, indicating they were not particularly environmentally motivated. For example, the control group only used the biowaste bins 17.3% of the weeks during the study period. Thus, external validity is not challenged by having a sample that was especially motivated by pro-environmental concerns.

Second, participants were aware that their recycling practices could be monitored. Although subjects were informed that their data were intended for aggregate rather than individual analysis, one might wonder about their implications for the external validity of the intervention. To address this question, we argue that those concerned about monitoring will focus on fulfilling the MENUCO (the minimum number of times they are told they have to recycle to be classified as recyclers). Consequently, we restrict our analysis to households not complying with the MENUCO to see whether the effect of the soft commitment holds among those less likely to be motivated to recycle by the presence of an external observer. The results of this analysis are consistent with those seen in the main specification (SI Appendix, Table S11). This suggests that a soft commitment would also work if waste sorting were not observable, consistent with the mechanisms of individual self-image (74–76), cognitive dissonance (19), and warm glow (77).

3.3. Feedback Intervention.

The aim of the feedback intervention was twofold. First, we wanted to use our dataset to study the effect of the well-known feedback mechanism for promoting pro-environmental behavior (59), and second, we wanted to analyze whether it positively interacts with a preexisting pledge. As in the soft commitment intervention, we evaluate the effect of the feedback using weekly participation as the main outcome (%Weeks).

In SI Appendix, section 2.E, we examine this follow-up feedback intervention conducted in February 2020. Our analysis focuses on the impact of the feedback from the week immediately after receiving the letter, up to week 210 after recruitment (i.e., the period after receiving feedback that overlaps with the one during which we evaluated the soft commitment). The effects of the feedback intervention were smaller than those of soft commitment and mostly null (SI Appendix, Table S12). Remarkably, the interaction between soft commitment and feedback was never statistically significant. While splitting the sample into four groups for the analysis in SI Appendix, Table S12 may have limited statistical power, the results do not provide evidence that feedback can reinforce the effect of soft commitment.

3.4. Spillovers.

We lack data on secondary outcomes to directly measure potential spillovers arising from the soft commitment’s intervention. Nevertheless, we conducted an exploratory analysis using aggregate data to investigate the relationship between biowaste sorting (targeted behavior) and waste generation (secondary outcome). Among the works addressing that question (78–81) found pro-environmental spillovers resulting in less waste generation, while (82) found a negative spillover resulting in more waste. Specifically, we used panel data to examine the relationship between the proportion of biowaste card users and total waste generation at the neighborhood level. The analysis finds no evidence of positive correlations between the two (after controlling for neighborhood and time fixed effects), suggesting no spillovers between biowaste sorting and waste generation. While these findings do not entirely preclude such possibility, they suggest that soft commitments are unlikely to have generated adverse second-order effects on this domain as they were ultimately used for reinforcing biowaste sorting.^*** Further discussion on spillovers and additional details of the analysis can be found in SI Appendix, section 2.G.

4. Conclusion

Using a technology that allows waste sorting in dense urban areas to be tracked, we evaluate the effect of a soft commitment on promoting pro-environmental behavior. We show that a light-touch mechanism, consisting of being given the opportunity to sign a pledge, has a positive and lasting impact on recycling.

One shortcoming of our study, as explained in Section 1.2.3, is that we do not observe the amount of waste being thrown away but bin usage. More sophisticated technologies would allow us to obtain information on the size of each disposal or even on the specific content of each disposal. Nevertheless, as we show in SI Appendix, Fig. S6, our data provide a good proxy of the amount of biowaste collected, and they represent a notable improvement over previous limitations in waste observability.

Despite our study providing evidence over a long time horizon that pledges improve the effectiveness of environmental campaigns, it cannot discern which parameters explain the difference with respect to some previous works that obtained null results. Although such a question might be better addressed in laboratory experiments and meta-analyses, we consider our study to have at least two relevant insights regarding this matter. First, sample sizes and study areas have traditionally been constrained by the difficulty of observing waste sorting. Our study circumvents this challenge by using a technology that allows to address this question for bigger samples and in the setting of a densely populated urban area. Second, by distinguishing between the intensive and the extensive margin, we advance our understanding of one of the parameters that determines the effectiveness of soft commitments to promote recycling.

The results of our study circumscribe the context of an environmental face-to-face campaign. We cannot assess the effect that offering the soft commitment will have when scaled up to the general population. For instance, a mass mailing campaign to all citizens may result in less adoption of the commitment. This is especially likely considering the increased effectiveness of face-to-face interventions (67). Environmental information campaigns are quite popular when it comes to promoting proper waste separation at the local level. Given that the additional cost of offering to sign the pledge is zero, our intervention can be scaled up to information campaigns promoting recycling to increase their effectiveness at a negligible cost. Considering previous works on voluntary commitments, two changes in the design of the pledge that could improve its effectiveness include, making it public (27, 30, 31) and/or adapting its form to that of a solemn oath (e.g., ref. 62).

An open question for future research is potential spillover of the pledge to other pro-environmental actions. Technologies like the one used in our intervention should make it increasingly feasible to obtain and connect individual data on environmental actions (e.g., waste management or the use of public transport), opening the possibility for future studies to analyze spillover effects of environmental interventions.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Anonymized primary data have been deposited in OSF (https://osf.io/jkh3w) (83). All other data are included in the manuscript and/or SI Appendix.