Abstract
We conduct a randomized controlled trial to evaluate the impact of training university students in Bogotá to use a smartphone app that displays real-time, location-specific air quality data. The training increased participants’ acquisition of information about air quality, their knowledge about avoidance behavior, and, most important, their reported avoidance behavior. Back-of-the-envelope calculations suggest that if scaled to the entire city of Bogotá, the training could reduce premature cardiovascular, cerebrovascular, and respiratory deaths by 51%–61% a year among the additional 3%–8% of the city’s population incentivized to undertake avoidance behavior, a benefit valued at US$11–$13 million.
1. Introduction
A product of decades of industrialization, urbanization, and motorization, chronic severe air pollution is now a global phenomenon. Today, 99% of the world’s population live in places that do not meet World Health Organization air quality standards (WHO 2021a). The consequences for human health have been grave. Each year, air pollution causes 5–9 million premature deaths and many more cases of bronchitis, asthma, and other cardiopulmonary illnesses (GBD 2021; Vohra et al. 2021; World Bank 2022). The Global South is the epicenter of this problem, accounting for roughly 95% of the mortality attributed to air pollution (World Bank 2022). Although indoor air pollution from the use of biomass fuels is a major contributor, most of these deaths and illnesses are caused by ambient pollution (Landrigan et al. 2018). Moreover, in most developing countries, ambient air pollution is getting worse. One study predicts that without aggressive interventions, the number of deaths from such pollution will increase by 50% by 2050 (Lelieveld et al. 2015).
Unfortunately, a variety of structural and institutional factors limit the effectiveness of regulatory initiatives aimed at controlling ambient air pollution in developing countries (Blackman 2010). Therefore, perhaps the most practical and cost-effective strategy for reducing illness and death caused by ambient air pollution in the short to medium term is to reduce exposure, which entails encouraging people to avoid outdoor activities, wear face masks, and engage in other avoidance behaviors on days when air pollution is severe. In turn, this requires disseminating timely location-specific information about air quality, along with recommendations about how to avoid exposure. Information and communication technologies (ICTs) are a logical means of providing such information.
Recently, smartphone apps that display real-time information on air quality have become available. For example, the U.S. Environmental Protection Agency’s AirNow app provides historical, real-time, and predicted data for most cities in the United States. Sameer, Air Quality China, and AirRater provide similar data for cities in India, China, and Australia. Apps like IQAir’s AirVisual compile this information for cities around the world. However, the extent to which these apps can actually affect avoidance behavior and other outcomes is not clear; to our knowledge, a rigorous evaluation of an intervention that entails using an air quality smartphone app has yet to appear.
Here we report on a randomized controlled trial aimed at evaluating the effect of training university students in Bogotá, Colombia—a city with chronic severe air pollution—to use Aire Bogotá, an air quality smartphone app developed by the city government. The training, which amounted to a bundled intervention, had three components: (1) information on the app; (2) information meant to motivate use of the app, including on air pollution, its health effects, and avoidance behaviors; and (3) a six-week interactive email campaign aimed at reinforcing the first two elements. We randomly assigned a sample of 578 students to a control group or a treatment group that received the training. A baseline survey, administered in person in March 2020 just before the training, and an endline survey, administered remotely three months later, collected information on sociodemographic characteristics and behavioral and attitudinal outcomes. These surveys were administered in proctored sessions capped at 30 participants.
We find that the training boosted participants’ acquisition of information about air quality, their knowledge about avoidance behavior, and most important, their reported avoidance behavior. The potential benefits are significant. Back-of-the-envelope calculations suggest that if scaled to the entire city of Bogotá, the training could reduce premature cardiovascular, cerebrovascular, and respiratory deaths by 51%–61% a year among the additional 3%–8% of the city’s population incentivized to undertake avoidance behavior, a benefit valued at US$11–$13 million.
In interpreting these findings, it is important to keep in mind that bundled smartphone app treatments like ours have advantages and disadvantages. The main advantage is efficacy. Considerable research has demonstrated that even if users can be convinced to download a smartphone app that provides targeted, salient, actionable information related to their health (including on smoking cessation, exercise, mental health, and air pollution), subsequent levels of engagement with the app can be low or decline over time (Guertler et al. 2015; Regmi et al. 2017; Torus et al. 2020; Delmas and Kohli 2021). For that reason, randomized controlled trials evaluating such apps often use bundled interventions intended to motivate participants and habituate them to using the app (Araban et al. 2017; O’Connor et al. 2020) just as experiments evaluating all manner of policies often rely on bundled interventions to boost efficacy (e.g., Stephens and Toohey 2022; Leight et al. 2022; Wheeler et al. 2022).
A disadvantage of bundling is that multiple treatment arms are required to disentangle the effects of the treatment components. Owing to resource constraints, our experiment featured a single treatment arm. As a result, we are not able to disentangle the effects of the three components of the bundled treatment, including the effect of the Aire Bogotá app separate from the information session and email campaign.1 Given that we find that as a bundle, the components of our treatment had significant effects, disentangling the impacts of these components would be a fruitful area for follow-up research.
This article makes three contributions to the literature on the efficacy of air quality ICTs. The first concerns the type of intervention we study. To our knowledge, this is the first rigorous study of an intervention focused on an air quality smartphone app and is among only a handful of studies examining “personal” air quality ICTs such as smartphone apps, text messages, and personal air quality monitors that provide information on air quality at specific times and locations. Because they can provide information tailored to each recipient’s specific needs, such ICTs hold promise for motivating people to engage in avoidance behavior and helping them plan it. The lion’s share of studies of air quality ICTs focus on radio, television, newspapers, webpages, and other impersonal mechanisms that offer more generic content to all users. It is not clear whether findings from studies of impersonal air quality ICTs or of other types of personal air quality ICTs generalize to interventions focused on air quality smartphone apps. Impersonal air quality ICTs not only disseminate more generic information but also affect different (albeit overlapping) populations and likely generate different spillover effects. Other types of personal air quality ICTs like text messages entail different user interactions. For example, recipients of text messages do not need to request or search for the air quality information they receive, whereas users of smartphone apps and air quality monitors do.
The second contribution to the literature concerns methods. Ours is one of only a few studies of the efficacy of an air quality ICT to use experimental methods. Studies of impersonal air quality ICTs rely on quasi-experimental methods. Among analyses of personal air quality ICTs, to our knowledge, only Araban et al. (2017) and Hanna et al. (2021) use randomized controlled trials. Lyons et al. (2016) rely on quasi-experimental methods, and Oltra et al. (2017) and Haddad and de Nazelle (2018) use small-sample focus groups. The best-known advantage of experimental studies is that they control for unobserved confounding factors. In addition, compared with quasi-experimental studies, they allow the evaluation of a wider array of outcomes. Most quasi-experimental studies of ICTs use secondary data to measure easily observed but infrequent behaviors, such as attendance at a public zoo or sporting event. By contrast, we use original survey data to consider a range of outcomes, including those related to acquiring air quality information and multiple avoidance behaviors that concern everyday activities (e.g., wearing a face mask and closing windows).
Our third contribution concerns geography: we add to the relatively thin literature on information-based air quality interventions in the Global South, where the problem we study is most urgent and where the rapid diffusion of cell phones over the past decade (Silver et al. 2019) has created new, seemingly low-cost opportunities to address that problem using air quality apps. Among the 14 studies summarized in the review of the literature in the next section, only five focus on developing countries (Araban et al. 2017; Liu, He, and Lau 2017; Hanna et al. 2021; Greenstone et al. 2022; Barwick et al. 2024).
2. Literature
A fast-growing literature evaluates the efficacy of air quality ICTs, including impersonal and personal ones. Studies of impersonal ICTs, which tend to focus on air quality alerts disseminated through conventional electronic and print media, mostly find that they boost avoidance behavior. Researchers have found that air quality alerts reduce attendance at zoos and botanical gardens in Southern California (Graff Ziven and Neidell 2009), lower attendance at baseball games in South Korea (Yoo 2021), cut the use of outdoor recreation facilities by the elderly and other sensitive groups in Atlanta (Noonan 2014), reduce the use of bicycles by 14%–35% in Australia (Saberian, Heyes, and Rivers 2017), decrease hospital admissions for asthma in England (Janke 2014), and double online queries for face masks with filters in China (Liu, He, and Lau 2017). Two recent studies conclude that over the past two decades, the rollout of automated real-time air quality monitoring and disclosure systems across Chinese cities has boosted indicators of avoidance behaviors, including purchases of air purifiers and online searches for face masks (Greenstone et al. 2022; Barwick et al. 2024). The evidence on the effects of impersonal informational mechanisms is not uniformly positive, however. Semenza et al. (2008) and Stieb, Paola, and Neuman (1996) find that air quality alerts in Canada, Texas, and Oregon have little effect on self-reported avoidance behavior.
The literature on personal ICTs is far more limited, and the results are mixed. On one hand, Araban et al. (2017) find that a bundled intervention consisting of daily text messages on air quality, motivational interviewing, and printed educational materials boosted avoidance behavior in a sample of pregnant women in Tehran. Hanna et al. (2021) report that in Mexico City, SMS air quality alerts tailored to recipients’ locations increased the probability that recipients stayed indoors with windows closed on perceived high-pollution days. Oltra et al. (2017) find that in Barcelona, individual air quality monitors increased awareness of and motivation for avoidance behaviors more than impersonal information dissemination mechanisms. On the other hand, Lyons et al. (2016) find that AirAware, a targeted personal air pollution information system in the United Kingdom that delivers texts, emails, and voicemail messages to high-risk persons, increased emergency room admissions for respiratory conditions (which they attribute to the system exacerbating participants’ anxiety about air pollution). Haddad and de Nazelle (2018) report that in the United Kingdom, individual air pollution monitors and smartphone apps did not affect travel-related behaviors or attitudes in a group of pilot testers.
3. Background
Air Quality in Bogotá
The air quality monitoring network in Bogotá (Red de Monitoreo de Calidad del Aire de Bogotá) consists of 19 stations that provide hourly data on six air pollutants and seven weather variables. Of the six pollutants for which the World Health Organization has established guidelines (WHO 2021b)—coarse particulate matter (PM10), fine particulate matter (PM2.5), nitrogen dioxide (NO2), ozone (O3), sulfur dioxide (SO2), and carbon monoxide (CO)—Bogotá regularly fails to meet both annual and 24-hour guidelines for three (PM10, PM2.5, and NO2), all by a considerable margin (Appendix Figure A1). Two of these pollutants, PM2.5 and NO2, cause more than 2,750 premature mortalities each year in the city (Blackman, Bonilla, and Villalobos 2023). Vehicles are the source of 81% and 98% of combustion emissions of PM2.5 and nitrogen oxides, and trucks and buses are the main sources (SDA 2020).
Importantly, air quality exhibits considerable variation over time and across Bogotá’s neighborhoods, implying that the temporally and spatially specific “personal” information from the Aire Bogotá app has value in planning avoidance behavior (Appendix Figures A1–A4). Notwithstanding this variability, on average, air quality is worse during early morning and evening hours, in the first and last quarters of the year (when thermal inversions trap air pollution at ground level), and in the southwestern part of the city.
The Aire Bogotá App
Created by the municipal environmental agency (Secretaría Distrital del Ambiente) and launched in January 2020, Aire Bogotá is a free interactive smartphone app that provides a range of information on air quality in the city. Perhaps most important, it displays either real-time concentrations or a color-coded air quality index called the IBOCA (Índici Bogotano de Calidad del Aire y Riesgo en Salud) for three pollutants—PM2.5, PM10, and O3—at the city’s 19 air quality monitoring stations, with interpolated information for points in between, such as health clinics, public transit stations, and museums. The app also provides historical data on air quality for the past seven days, predictions for the next 48 hours, and health recommendations based on the index. Appendix Figure A5 compiles four screenshots illustrating the app’s capabilities: real-time data on air quality at the monitoring stations (panel A), real-time data at points of interest selected by the user (here, health clinics) (panel B), historical air quality data (panel C), and air quality predictions (panel D).
4. Experimental Design, Theory of Change, and Data
We designed an experiment to assess the effects of training university students to use the Aire Bogotá app on their acquisition of air quality information, knowledge about avoidance behavior, actual avoidance behavior, dissemination of air quality and environmental information, and attitudes about the environment.
Sample
Our sample comprised students 18 years of age or older studying at universities in Bogotá. We focused on university students for two reasons. First, we expected virtually all students to have smartphones, to be comfortable with and habituated to using digital technologies, and to have easy access to Wi-Fi networks that would enable them to use the app at no cost. Second, we expected them to have relatively flexible schedules that would lower the costs of avoidance behaviors, such as limiting outdoor activities and adjusting travel on severe air pollution days. In Section 8, we consider the implications of our focus on university students for external validity.
We used print and digital social media to recruit a sample of students. A total of 665 students at 24 universities participated in our baseline sessions, and 578 participated in our endline sessions. Of the 578 students that took both the baseline and endline surveys, 536 responded to all questions. Attrition is balanced across the treatment and control groups (Appendix Table A1). In the final sample of 536 students, the treatment group comprised 226 participants (42%), and the control group, 310 participants (58%). Although randomization was designed to assign roughly half of the sample to each group, actual assignment percentages differ because randomization was at the baseline session level.
Timeline
Our experiment proceeded as follows (Appendix Figure A6). The Aire Bogotá app was launched in January 2020. In February, we recruited our sample. Between March 2 and March 14, we conducted 30 in-person baseline sessions with a total of 665 participants. At each session, we administered the baseline survey and then conducted either a treatment or a control (placebo) information session. Participants were randomly assigned to treatment and control groups at the baseline session level. Of the 30 baseline sessions, 14 sessions with a total of 272 participants featured the treatment materials, and 16 sessions with 393 participants featured the control materials. In the six weeks after their baseline session, participants engaged in an interactive email campaign. Finally, between May 11 and June 19, we conducted 46 remote endline sessions with 578 participants.
Our experiment coincided with the beginning of the COVID-19 pandemic in Bogotá. The first case in the city was reported on March 6, 2020; most universities closed March 16, and a national lockdown began on March 20. Our planned baseline sessions were nearly completed when the Rosario Experimental and Behavioral Economics Lab, which hosted them, was shuttered on March 16. By that time, we had completed baseline surveys for 665 participants, representing 89% of our planned baseline sample of 750 participants. We discuss potential effects of the pandemic on our study below.
Treatments
Participants assigned to the treatment group attended an in-person information session lasting approximately 20 minutes (Appendix Section 5) that covered the following topics:
■ Air quality in Bogotá: this regularly fails to meet international standards and is worse than that in most Latin American cities.
■ Effects of air pollution on human health: this includes a variety of short- and long-term illnesses and, in Bogotá, approximately 2,000 deaths per year, 14% of all deaths in the city.
■ Basic information on air pollution: this includes the most important types, temporal variation over the course of the year and the day, and spatial variation in Bogotá.
■ Avoidance behavior to reduce health risks from air pollution: this includes wearing an N95 mask, limiting outdoor physical activity, closing windows when and where air quality is particularly poor, seeing a doctor promptly when experiencing cardiorespiratory symptoms, and avoiding tobacco products.
■ Aire Bogotá app download and installation: this includes instructions, the main types of information it provides, and its functionality.
■ Aire Bogotá app use to reduce exposure: this includes determining when and where to engage in avoidance behavior.
■ Email campaign: this includes an offer to participate in a six-week interactive email campaign for which the Aire Bogotá app would be needed.
The purpose of the email campaign was to train participants to use the Aire Bogotá app, habituate them to using it, and reinforce the informational treatment. Participants who agreed to receive emails were sent one per week during the six weeks after the baseline session. Each email contained a brief bullet-point summary of selected key messages from the baseline information session about the health effects of air pollution and avoidance behavior (see Appendix Section 5). Each email also included a question about air quality at a specific time and location in Bogotá: for example, “What was the IBOCA for PM2.5 at the Barrios Unidos air quality monitoring station on March 21 at 9:00 p.m?” To answer these questions, participants needed to query the Aire Bogotá app and submit and answer using a SurveyCTO link within 24 hours of receiving the email.
Participants assigned to the control group received a placebo information session on art history, an offer to participate in a six-week placebo email campaign for which they would need a free app called DailyArt, and instructions on how to download, install, and use the app. The purpose of the placebo treatment was to minimize differential attrition by ensuring that participants in the control group had an opportunity to earn compensation comparable to those in the treatment group. Compensation is discussed in Section 4.
Outcomes
In the baseline and endline surveys, we collected information about three sets of primary outcomes (Table 1). The first set concerned the acquisition of air quality information. Respondents indicated whether they had installed the Aire Bogotá app on an electronic device (installed app), whether they had used the app to search for air quality information (info searched app), and whether they had used other means to do that (info searched other). At baseline—before the treatment was administered—only 3% of our participants had installed the app on a device, and only 1% had used it to search for air quality information in the previous two weeks (Table 1). Nevertheless, during this time, 37% had sought information on air quality from another source.
Variables and Means at Baseline
The second set of primary outcomes concerned knowledge about avoidance behaviors. Respondents indicated whether they knew that their own behavior could reduce the health risks from air pollution (know) and whether they knew that specific behaviors could reduce those risks, including restricting outdoor activities (know outdoors), changing their travel mode or route (know travel), wearing a mask with a filter (know mask), closing windows (know windows), and using an air purifier (know air purifier). At baseline, 91% of participants knew that changing their own behavior could have health benefits. Participants’ knowledge about specific avoidance behaviors ranged from a low of 22% for restricting outdoor activity to a high of 57% for wearing a mask with a filter.
The third set of primary outcomes concerned avoidance behaviors. Respondents reported whether they had changed any behavior specifically because of poor air quality in the previous two weeks (behavior), and if so, what specific type of behavior they had changed (behavior outdoors, behavior travel, behavior mask, behavior windows, and behavior air purifier). At baseline, few participants—only 14%—reported engaging in any type of avoidance behavior because of poor air quality in the two weeks before the baseline session. The most common avoidance behaviors were restricting outdoor activity (7%) and changing travel modes or routes (5%).
In addition to these three sets of primary outcomes, we collected information on three sets of secondary outcomes (Appendix Table A5). The first set had to do with providing warnings about air quality to friends, family, and others, the second with discussing environmental issues with others, and the third with attitudes about various environmental issues. We discuss the effects of our intervention on these secondary outcomes in Section 1 of the Appendix.
Theory of Change
We present a theory of change that describes hypothesized causal pathways between our intervention and the three sets of primary outcomes. The theory of change consists of four components: the intervention, intermediate outcomes, final outcomes, and moderators (Figure 1). For the sake of simplicity, we hypothesize that the relationships among these components depends only on participants’ perceived benefits and costs and on their baseline levels of knowledge. More specifically, we hypothesize that the intervention boosts the acquisition of air quality information and spurs avoidance behavior by increasing the perceived benefits of undertaking these actions, by reducing the costs, or both. The intervention boosts a participant’s knowledge when her baseline level of knowledge is low. We emphasize that this theory of change amounts to a set of interrelated hypotheses, not empirically verified relationships. The purpose is simply to motivate and provide a heuristic conceptual framework for the empirical analysis.
Theory of Change
Intervention
As noted already, our intervention is a bundle of three components: information on the app, information meant to motivate its use, and a six-week email campaign aimed at reinforcing the first two components.
Intermediate Outcomes
We hypothesize that the intervention boosts the two intermediate outcomes—acquisition of air quality information and knowledge about avoidance behavior—in two ways. First, it reduces participants’ postintervention cost of acquiring air quality information (by giving them information about how to install and use the app). Second, it improves their knowledge about that topic (if their baseline level of knowledge is not already high).
Final Outcomes
We hypothesize that the intervention boosts avoidance behavior via the two channels shown in Figure 1: direct and indirect. The intervention has a direct positive effect on avoidance behavior by raising the perceived benefits of such behavior. The intervention has an indirect positive effect on avoidance behavior by boosting intermediate outcomes—namely, participants’ acquisition of air quality information and their knowledge about this avoidance behavior—which, in turn, lower the cost and/or raise the perceived benefit of such behavior.
Moderators
We focus on two participant characteristics that moderate the direct and indirect effects of the intervention: health self, an indicator of whether the participant has cardiopulmonary conditions, and exercise outdoors, an indicator of whether she exercises outdoors at least some days of the week.
We hypothesize that the effect of our intervention is stronger for participants for whom health self = 1. The broad reason is that such individuals face higher health risks from exposure to air pollution (Manisalidis et al. 2020). As a result, they are more likely to acquire air quality information (because the benefits are higher), and that information boosts avoidance behavior (because the costs are lower)—the indirect effect in Figure 1. Such participants are also more likely to undertake avoidance behaviors for any given level of air quality information (because the perceived benefits are higher)—the direct effect in Figure 1. Finally, health self has countervailing moderating effects on avoidance knowledge. On one hand, participants with cardiopulmonary conditions may be more likely to absorb and retain such knowledge. On the other hand, they may be more likely to have high baseline levels of such knowledge.
Finally, we hypothesize that the effect of our intervention is stronger for participants for whom exercise outdoor = 1. The reasons are similar to those for participants for whom health self = 1. Participants who exercise outdoors and have been made aware of the heightened risk of doing so by our intervention are more likely to acquire air quality information (because the benefits are higher), and this information boosts avoidance behavior (because the costs are lower)—the indirect effect. These participants are also more likely to undertake avoidance behaviors for any given level of air quality information (because the perceived benefits are higher)—the direct effect. A caveat is that for participants who regularly exercise outdoors, and for whom not doing so is costly, the benefits and costs of avoidance behavior may be higher than average. Finally, exercise outdoors likely has countervailing moderating effects on avoidance knowledge for the same reasons that health self does: on one hand, participants who exercise outdoors may be more likely to absorb and retain such knowledge, but on the other hand they may be more likely to have high baseline levels of such knowledge.
Logistics
Study participants were compensated: they received COL$30,000 (US$9.25) for attending the baseline survey and information session, COL$40,000 (US$12.30) for attending the endline survey session, and COL$6,000 (US$1.85) for each email question answered correctly.2 By attending the baseline and endline sessions and answering every email question correctly, participants could earn a maximum of COL$142,000 (US$43.78). Payments for baseline sessions were made in cash immediately after the session. Payments for correct responses to questions in the email campaign and for the endline session were made using money transfer smartphone applications.
To reduce inattention and ensure adherence to study protocols, the baseline and the endline sessions were conducted in proctored group meetings with a maximum of 30 participants. Baseline sessions were conducted in person at the Rosario Experimental and Behavioral Economics Lab in Bogotá’s city center. Because of COVID-19 social distancing requirements, endline sessions were conducted online using a web conferencing platform (Zoom). In-person baseline and remote endline sessions were proctored by at least two members of the study team, who checked identification to verify that participants were the university students who had been invited; obtained consent; introduced, explained, and monitored engagement with the surveys; answered procedural questions; and, after completion of the baseline survey, presented the informational treatments. Administered using SurveyCTO online software, the baseline and endline surveys elicited information on the outcomes described here and on sociodemographic characteristics (Table 1). An average of 22 students participated in each baseline session, and an average of 13 students participated in each endline session.
Sociodemographic Characteristics
At baseline, just over a third of the students in our sample came from homes in the lowest or second-lowest estratos—socioeconomic level categories used by Colombian municipal governments (Table 1).3 Slightly more than half were male, 83% lived with their immediate family, 36% had mothers who had attended college, and just under a quarter held full- or part-time jobs in addition to attending university. Twenty-three percent smoked tobacco products, just under a fifth had a cardiopulmonary condition that could be exacerbated by air pollution, and 57% had an immediate family member with such a condition. A significant share lived with household members vulnerable to the effects of air pollution: 13% lived with children younger than five, and 30% lived with adults older than 60. Almost one-third exercised outdoors at least some days of the week. Forty-two percent lived in the southwestern part of Bogotá, which has the city’s most severe air pollution.
Although participants were randomly assigned to either the treatment or the control group at the baseline information session level, it is useful to check for balance on their observable characteristics. Only one covariate—estrato 1 & 2—is (weakly) correlated with treatment assignment controlling for remaining covariates (Appendix Table A2a). Although no normalized differences exceed the standard threshold for a small effect of 0.25 (Imbens and Rubin 2015), for two covariates, estrato 1 & 2 and employed, simple t-tests indicate that means at baseline are significantly different for treatment and control groups (Appendix Table A2b). To control for residual correlations, we include participant characteristics as covariates in the regressions used to generate treatment effect estimates and analyze treatment effect heterogeneity (see equations [1] and [2]).
Noncompliance
Study participants exhibited two types of noncompliance. The first concerned installation of the Aire Bogotá app and was two-sided. Of the 226 participants in the treatment group, 2% never installed the app; another 16% installed it at some point during the experiment but had uninstalled it by the time of the endline survey (Table 2). Of the 310 participants in the control group, 3% had the app installed at the time of the endline survey, and 11% had installed it at some point before the endline but had since uninstalled it.
Treatment Assignment Noncompliance (%)
The second type of noncompliance concerned treated participants’ engagement in the six-week interactive email campaign and was, by definition, one-sided. On average, treated participants did not respond to 2.5 (41%) of the six emails they were sent. On average, 7% of treated participants’ responses were incorrect. In the next section, we discuss the implications of these types of noncompliance for the consistency of our treatment effect estimates.
5. Estimations
We estimate intention-to-treat (ITT) effects using ordinary least squares (OLS) to fit regressions of the form
1where Y is the outcome at endline, treated is a binary indicator of whether a participant received the treatment information session, y is the outcome at baseline, x is a vector of covariates, β is a parameter or vector of parameters, and ϵ is an error term. Our estimated treatment effect is given by ϵ1. We include y and x as independent variables to increase the precision of our treatment effect estimate, β1 (Twisk et al. 2018; Glewwe and Todd 2022). The elements of x are estrato 1 & 2, male, education mother, live with immediate family, employed, health self, health family, smoke, household members<5, household members>60, exercise outdoors, and three region fixed effects (Table 1). For robustness, in Appendix Tables A3 and A4, we report results from simplified regressions that omit the vector of participant characteristics variables. We cluster standard errors at the baseline session level.
To evaluate treatment effect heterogeneity, we use OLS to fit regressions of the form
2where z is a vector of covariates that is a subset of x. Here, too, we cluster standard errors at the baseline survey session level.
Three factors could in principle bias our results (we discuss several additional factors in Section 8). First, as noted, our study participants exhibited two types of noncompliance, one of which was two-sided. Not all the participants in the treatment group were fully treated (not all installed the app and kept it installed for the duration of the experiment, and not all fully participated in the email campaign) and some of the participants in the control group were partly treated (some had installed the app). Although the scope of this noncompliance was not extreme—only 2% of the treatment group never installed the app and only 3% of the control group had it installed at baseline—the implication of noncompliance in each group is that our treatment effect estimates are likely to be biased downward (Gertler et al. 2016). Hence our ITT effect estimates can be considered lower bounds on true effects.
Second, we estimate treatment effects for multiple closely related outcomes: three outcomes related to acquisition of air quality information, six related to avoidance knowledge, and six related to avoidance behavior. To control for multiple hypothesis testing, we report both standard p-values and sharpened false discovery rate (FDR)–adjusted q-values (Anderson 2008) that limit the expected proportion of false rejections of the null hypothesis across related outcomes (Tables 3–5).
Intention-to-Treat Effect Estimates: Acquisition of Air Quality Information
Intention-to-Treat Effect Estimates: Knowledge and Behaviors
Intention-to-Treat Effect Heterogeneity
Finally, we have 30 clusters corresponding to baseline sessions that serve as our unit of randomization. To control for potential too-few clusters problems, we report p-values generated by the wild cluster bootstrap procedure (Cameron, Gelbach, and Miller 2008; Roodman et al. 2019) (Appendix Tables A3b and A4b).
6. Results
Main Effects
Simple t-tests show that for several variables in each of our outcome categories, endline means are significantly different for the treatment and control groups, providing suggestive evidence that our treatment had an impact (Appendix Table A2b). Here we report results of estimations (equation [1]) that provide more precise treatment effect estimates.
These estimates suggest that the treatment led to substantial changes in all three of our categories of primary outcomes: acquisition of air quality information, knowledge about avoidance behavior, and avoidance behavior. Regarding the acquisition of air quality information, the treatment led to very large increases in use of the Aire Bogotá app to gather information on air quality but a smaller and less statistically significant effect on the use of other means to do that. Specifically, it led to an 84 percentage point increase in the probability of having the Aire Bogotá app installed and a 29 percentage point increase in the probability of using it to search for air quality information, but only a (weakly significant) 5 percentage point increase in the probability of using other means to search for air quality information (Table 3).
The treatment had substantial effects on participants’ knowledge about avoidance behaviors. It led to a 3 percentage point increase in the probability of knowing that changing one’s behavior can reduce the adverse health effects of air pollution (3% increase over a baseline level of 91%) (Table 4). As for knowing that specific behaviors can reduce such effects, the treatment spurred a 20 percentage point increase in the probability of knowing about restricting outdoor activities (93% increase over a baseline level of 22%), a 15 percentage point increase in the probability of knowing about wearing a mask with a filter (27% increase over a baseline level of 57%), and a 31 percentage point increase in the probability of knowing about closing windows (103% increase over a baseline level of 30%).
What might explain variation in the economic and statistical significance of estimated treatment effects related to knowledge about specific avoidance behaviors? In general, we observe large significant effects for avoidance behaviors that were discussed in the treatment information session and for which baseline levels of awareness were relatively low (restricting outdoor activity and closing windows), and we observe smaller and/or statistically insignificant effects for behaviors that either were not mentioned in the treatment information session (changing travel, using an air purifier) or were mentioned in the baseline treatment session but for which baseline levels of awareness were relatively high (wearing a mask with a filter).
The treatment had substantial effects on participants’ reported avoidance behavior. It led to a 9 percentage point increase in the probability of changing any behavior as a result of poor air quality (68% increase over a baseline level of 14%). With regard to specific behaviors, the treatment spurred a 5 percentage point increase in the probability of wearing a mask with a filter (218% increase over a baseline level of 2%) and a 7 percentage point increase in the probability of closing windows (221% increase over a baseline level of 3%) (Table 4).4 The treatment had weakly significant effects on the probability of exercising outdoors, changing travel behavior, and using an air purifier. We discuss potential bias in these results due to self-reporting and the effects of the COVID-19 pandemic in Section 8.
What might explain variation in the magnitudes of estimated treatment effects related to avoidance behavior? Ceiling effects are unlikely to be the explanation: baseline levels of all our specific avoidance behaviors (wearing masks, closing windows, etc.) were less than 10%. Instead, we hypothesize that the explanation may have to do with the unobserved financial or psychological costs associated with the behavior. Treatment effects for behaviors for which one would expect such costs to be substantial (limiting outdoor activities, changing travel, using an air purifier) are only weakly significant.
In terms of magnitude, the effect of our treatment on the probability of changing any behavior as a result of poor air quality—a 68% increase—is comparable to effects reported in the literature on other personal ICTs for similar outcomes (Appendix Table A11). Araban et al. (2017) find that a bundled intervention, including text messages, boosted a self-reported Likert-scale measure of a group of avoidance behaviors by 81%, and Hanna et al. (2021) find that text messages increased the probability of having done something different in the past week by 62%. However, the effects of our treatment on the probability of closing windows—a 221% increase—is considerably larger than the only similar treatment effect reported in the literature of which we are aware: Hanna et al. (2021) find that text messages increased the probability of staying home with closed windows by 88%. The discrepancy may be at least partly attributable to the fact that Hanna et al. (2021) measure the effect of an ICT on closing windows and staying home, whereas we measure the effect only on closing windows. As for the effect of our treatment on restricting outdoor activities, other studies have found significant (albeit somewhat modest) effects of impersonal air quality alerts ranging from 6% to 35%.
For all three categories of primary outcome variables, simplified main effects regressions that omit participants’ characteristic covariates (Appendix Tables A3a and A4a) and those that use wild cluster bootstrap standard errors (Appendix Tables A3b and A4b) generate results that are qualitatively similar to those summarized above. These results indicate that among the acquisition of air quality outcomes, the most robust effects were for the probability of installing the app and the probability of using it to search for information on air quality. Among the knowledge about avoidance behavior outcomes, the most robust effects were for knowing that restricting outdoor activities, wearing a mask with a filter, and closing windows could reduce adverse health effects; and among the avoidance behavior outcomes, the most robust effects are for wearing a mask with a filter and closing windows.
Treatment Effect Heterogeneity
Here we report on tests of our hypotheses about the moderating effects of health self and exercise outdoors on both intermediate and final outcomes. To make the analysis tractable, we focus on one representative outcome in each of the categories of primary outcomes. For “acquisition of air quality information,” we focus on info searched app, an indicator of whether the participant used the Aire Bogotá app to search for air quality information in the previous two weeks. For “avoidance knowledge,” we focus on know, an indicator of whether participants know that changing any type of behavior can reduce adverse effects of air pollution. (Recall that at baseline, 90% of participants knew this. Hence, ceiling effects likely come into play.) For “avoidance behavior,” we focus on behavior, an indicator of whether the participant changed any behavior because of poor air quality in the previous two weeks.
In general, we find only limited support for the hypotheses that health self and exercise outdoors moderate the effect of intervention (Table 5). Of the six coefficients on interaction terms reported in Table 5, only two are significant, both at the 10% level (after correcting for multiple hypothesis testing). The coefficient on treated × health self is significant in the regression explaining the intermediate outcome know, suggesting that the effect of the intervention on knowledge about avoidance behavior was stronger among participants with cardiopulmonary conditions. The coefficient on treated × exercise outdoors is significant in the regression explaining info searched app, suggesting that participants who exercise outdoors were more likely to use the app to search for air quality information.5
7. Health Benefits
From a policy perspective, our most important findings are that training people to use the Aire Bogotá app spurred avoidance behavior, specifically a 5 percentage point increase in the probability of wearing a mask with a filter on poor air quality days and a 7 percentage point increase in the probability of closing windows on such days.6 What are the policy implications of these findings? Do they suggest that scaled-up trainings in the Aire Bogotá app could have significant health benefits? Rigorously addressing this question would require an evaluation of a scaled intervention, which is beyond the scope of the present study. To get a preliminary indication, we develop back-of-the-envelope estimates of the number of premature deaths that could be avoided by making Aire Bogotá trainings more widely available in Bogotá. We use a standard air pollution epidemiological simulation model to calculate avoided premature deaths (Ostro 2004; Enriquez, Larsen, and Sánchez-Triana 2018; Faustini and Davoli 2020) along with a simple benefit transfer model to value avoided deaths (World Bank 2022). An important caveat is that both models require strong parametric assumptions. Again, our goal is simply to develop a rough preliminary estimate of the potential health benefits from scaling our intervention.
Scope
As noted above, in an experimental setting, our intervention had robust positive effects on two specific types of avoidance behavior: wearing face masks and closing windows. For the epidemiological simulation, we focus on the effects of one of these behaviors—wearing a mask with a filter—on one type of health outcome: premature death due to exposure to PM2.5. We focus on wearing a mask and not closing windows because quantitative evidence on the effect of closing windows on pollution exposure is limited (Rajagopalan et al. 2020). Of the pollutants for which the Aire Bogotá app provides information, we focus on particulate matter and not O3 because common face masks do not reduce exposure to O3 (only face masks with an activated charcoal filter, which are quite rare, do that). Finally, we focus on PM2.5 and not PM10 because ambient levels of PM2.5 and PM10 are highly correlated, and disentangling their effects on human health is challenging (Janssen et al. 2013). To avoid double counting, most epidemiological analyses of the effect of particulate matter focus only on PM2.5 (e.g., Liu, Wang, and Zheng 2021; Vohra et al. 2021; World Bank 2022).
Epidemiological Model
The epidemiological model is described in detail in Section 2 of the Appendix. The foundation of the model is a set of hazard ratios from econometric studies of the effect of short-term exposure to PM2.5 on premature death caused by health conditions exacerbated by air pollution, namely, cardiovascular, cerebrovascular, and respiratory illness. We use hazard ratios from a recent systematic literature review and metanalysis (Orellano et al. 2020). Our epidemiological models combine these hazard ratios with (1) estimates of the effect of our informational intervention on exposure to PM2.5 in Bogotá, and (2) observed historical daily data from Bogotá on PM2.5 levels and on average (pre-intervention) deaths attributable to cardiovascular, cerebrovascular, and respiratory illness.
Scenarios
Our estimates of the effect of our informational intervention on exposure to PM2.5 in Bogotá rest on three key assumptions. First, we assume that if our intervention were scaled to the entire city of Bogotá, our estimated 5 percentage point treatment effect for mask wearing (Table 4) would apply to the city’s population; that is, an additional 5% of the city’s population would wear masks on days with relatively high pollution levels and this effect would not attenuate over time. As discussed in Section 8, scaled trainings could be provided by, for example, replacing the live information sessions used in our experiment with an online video and/or printed materials along with automated email campaigns. We believe this scaling assumption is not unreasonable given that our study sample was composed of young adults, a relatively healthy population that is arguably less inclined to engage in avoidance behaviors than the average Bogotá citizen. Second, we assume that this 5% of the city’s population would wear a mask on all days when PM2.5 exceeds the World Health Organization’s air quality guidelines for PM2.5: 15 mg/m3 (WHO 2021b). Finally, following a recent study of the efficacy of inexpensive cloth face masks, we assume that mask wearing reduces PM2.5 exposure by 39% (Neupane, Bajracharya, and Khatry 2023).
To assess the sensitivity of our simulation results to these key parametric assumptions, we estimate avoided deaths for three scenarios (Appendix Table A9). The middle scenario assumes the parameterization described above: the scaled treatment effect is 5 percentage points, the masking threshold is 15 mg/m3, and mask efficiency is 39%. The low scenario assumes the scaled treatment effect is 2.5 percentage points, the masking threshold is 20 mg/m3, and mask efficiency is 20%. Finally, the high scenario assumes the scaled treatment effect is 7.5 percentage points, the masking threshold is 10 mg/m3, and the mask efficiency is 60%.
Valuation
To value the simulated avoided deaths due to our intervention, we rely on a widely used value of a statistical life (VSL) of 2019 US$1,569,770 per avoided mortality derived from a 2011 systematic meta-analysis of more than 1,000 stated-preference studies of willingness to pay for marginal reductions in mortality risk in more than 30 industrialized and developing countries (Lindheim et al. 2011; OECD 2012). Following standard practice, we adjust this VSL to account for the difference in per capita income in Colombia and for inflation and income growth in Colombia after 2011.7
Results
The simulation of our middle (low, high) scenario suggests that a scaled version of our intervention would reduce cardiovascular, cerebrovascular, and respiratory deaths in the affected population (the 5% of the city’s population incentivized to wear masks on high-pollution days) by 64.3% (61.1%, 51.3%), from a total of 10.5 (5.2, 15.7) deaths per year at baseline to 6.7 (3.2, 8.1) deaths postintervention (Appendix Table A10). The value of these avoided deaths is US$10.57 ($5.0, $12.7) million. Note that baseline deaths vary across scenarios because each scenario assumes a different number of people who receive the intervention (high > middle > low).
8. Discussion
We conducted a randomized controlled trial with university students in Bogotá to evaluate the effect on a range of outcomes of training on a smartphone app that provides information on air quality. We found that the training boosted the acquisition of information about air quality, knowledge about avoidance behavior, and adoption of some avoidance behaviors; specifically, it had robust positive effects on wearing a mask with a filter and closing windows during severe air pollution episodes. We found some (weak) evidence that these effects were moderated by participant characteristics, namely, whether they exercised outdoors and had cardiopulmonary conditions. Finally, back-of-the-envelope calculations suggest that our findings are policy relevant: they indicate that if scaled to the entire city, the training could reduce deaths among those incentivized to undertake avoidance behavior by 51%–61% a year, a benefit valued at US$11–$13 million.
Our study has four main limitations. First, as noted above, our experimental design—a single treatment arm featuring a bundled intervention—does not enable us to disentangle the effects of the three components of our intervention: the Aire Bogotá app, the initial information session, and the email campaign. Having established that as a bundle, these components were effective in boosting intended outcomes, future research could disentangle the separate effects of the components of our bundled treatment. Doing so would clarify the causal mechanisms that explain our findings and would improve the effectiveness and efficiency of similar trainings. As noted above, a stepwise approach of demonstrating the effectiveness of a bundled treatment before attempting to disentangle the effects of its components has been recommended for evaluating other apps related to health (Regmi et al. 2017).
Second, even though our study considers a significantly broader range of outcomes than the quasi-experimental studies of ICTs cited in the first section, unlike these studies, we rely on self-reported outcomes that could be biased upward if respondents tended to provide answers that conform to perceived social norms (Zerbe and Paulhus 1987; Fisher 1993). This bias, in turn, could affect our treatment effect estimates if it was correlated with our treatment; that is, if our treatment created additional incentives for participants to overreport compliance. Unfortunately, we are not able to test for such bias because we do not observe actual outcomes. However, two factors provide reassurance. First, we find robust statistically significant effects for some outcomes but not others. If self-reporting bias were driving our results, we would expect to see more consistently robust significant treatment effects. For example, among our knowledge outcomes, we can discern robust significant effects for wearing a mask and closing windows but not for other avoidance behaviors, including changing travel routes and using an air purifier. In addition, although we are not aware of any evidence on bias of self-reports about pollution avoidance behavior, studies of self-report bias for other types of avoidance behavior have concluded that it is not large.8 Replicating a version of our experiment with observable outcomes, such as travel data from smartphone geolocators, would help clarify the issue.
Third, in principle, the overlap between the COVID-19 pandemic and our experiment could bias our results. Our experimental empirical design controls for cross-cutting factors that affect both treatment and control groups in the same way.9 The pandemic definitely affected both groups. Nevertheless, our results could be biased upward if it had differential effects on those groups—either making positive changes in outcomes more likely in the treatment group or less likely in the control group. That might happen if, for example, our treatment led participants to perceive a positive link between exposure to air pollution and susceptibility to COVID-19, thereby creating additional incentives for them to undertake avoidance behavior, warn others about severe air pollution, and so on.
Unfortunately, we have no way of testing for such perceptions or bias using our survey data. Here, too, several factors provide some reassurance. Our treatment materials were designed well before the pandemic began and made no reference to COVID-19, a potential link between air pollution and COVID-19, or even a link between air pollution and infectious disease. In addition, one of the two robust statistically significant effects of our treatment on avoidance behavior does not comport with the hypothesis of bias induced by the pandemic. We find that our treatment increased the probability that participants had closed windows as a result of air pollution in the previous two weeks. COVID-19 mitigation guidance would recommend the opposite: keeping windows open. Finally, we find it improbable that the positive spillover effects of our treatment for concern about environmental issues other than air pollution arose from concerns about COVID-19. Notwithstanding these factors, we acknowledge that it is possible for our treatment effect estimates to have been biased because of the pandemic. Replicating our experiment after the pandemic has subsided would help clarify the issue.
Finally, we relied on a sample of participants—university students—not representative of the overall population. As a result, the external validity of our findings (the extent to which they generalize to other subgroups and geographies) is not clear. As for subgroups: On the one hand, one might expect university students to be more responsive to interventions like ours because they are relatively comfortable with smartphone technologies and have relatively flexible schedules that enable them to rearrange outdoor exercise and travel during severe air pollution episodes. On the other hand, they may be less responsive because they are relatively healthy and therefore less vulnerable to the effects of air pollution. Further research with different samples is needed to test the external validity of our results.
Despite these limitations, we believe that our study makes a significant contribution to the evidence base on the use of ICTs to reduce exposure of air pollution. It adds to the growing evidence that ICTs providing real-time information on air quality can help reduce exposure to pollution. More specifically, it provides a proof of concept that smartphone app trainings may be an effective means of reducing exposure in developing countries, where air pollution has the most severe effects on human health and where prospects for reducing emissions in the short to medium term are arguably most limited. Moreover, our epidemiological simulations suggest these reductions may have significant health benefits.
Follow-up studies can provide additional information about whether and how best to scale up air quality app trainings. Cost would be a barrier to scaling the type of small-group, in-person, paid trainings used for the present study. However, scaled trainings could likely be provided at lower per beneficiary costs by, for example, replacing the live information session with a prerecorded online video and/or printed materials and automating email campaigns. It would be important to test whether and how the efficacy of such trainings differs from that of the in-person trainings in this study. One means of compensating for any loss in efficacy would be to target the remote trainings to those who are likely to benefit the most (e.g., people with cardiopulmonary conditions that make them susceptible to the effects of severe air pollution).
Our study highlights several directions for future research. First, because our bundled treatment was effective in boosting intended outcomes, future research could aim to disentangle the separate effects of the components of our bundled treatment. This would clarify the causal mechanisms that explain our findings and help improve the effectiveness and efficiency of similar trainings. Second, it would be useful to determine whether and how our results generalize to other geographic settings and subpopulations—particularly those who are most vulnerable to the effects of air pollution and for whom personal ICTs may have the greatest benefits. Third, it would be useful to compare the effectiveness of personal ICTs like the Aire Bogotá app with impersonal ones like air quality alerts disseminated through conventional media. Finally, as noted above, it would be useful to replicate some version of our experiment to test for potential self-reporting and pandemic-related biases.
Acknowledgments
The Inter-American Development Bank provided funding via Technical Cooperation CO-T1560 and Economic Sector Works RG-E1543 and RG-E1499. We are grateful to José Eguiguren-Cosmelli, Estefania Laborde, Emilio Leguizamo, and Maria Paula Medina for research assistance; to Laura Polanco, Dayana Tellez, and Sebastian Balcucho at IPA–Colombia for coordinating field research; to Diego Aycinena, Diego Bermudez, and Andres Zambrano at Rosario Experimental and Behavioral Economics Laboratory for recruiting our sample and helping to administer the surveys; to Gildardo Bermero, Ivette Gómez, Lina Guerrero, Marcelo Korc, Karin Troncoso, Leonardo Quiñones, and participants at CIDE, EfD and LAERE seminars for helpful comments and suggestions; and to Sally Atwater for editorial assistance. The information and opinions presented herein are entirely those of the authors, and no endorsement by the Inter-American Development Bank, its Board of Executive Directors, or the countries they represent is expressed or implied. Innovations for Poverty Action Institutional Review Board provided ethical approval (protocol 15344). The experiment was registered in the American Economic Association Randomized Controlled Trial Registry (AEARCTR-0005348).
Footnotes
↵1 Had we focused only on encouraging use of the Aire Bogotá app, without providing any information about its potential benefits, we could have used a randomized promotion (encouragement) design to identify its effects. That is, we could have used assignment to the treatment as an instrument for use of the app because the exclusion restriction would plausibly have been satisfied: the treatment likely would only have affected outcomes (avoidance behavior, environmental attitudes, etc.) through the app. But our bundled treatment, which combines encouragement to use the app with motivational information, likely had direct effects on our outcomes.
↵2 U.S. dollar amounts assume COL$3,243 = US$1, the exchange rate in January 2020.
↵3 Estratos are used to charge differential fees and taxes for public services and to allocate various benefits (DANE 2020). The six estratos are 1 (low-low), 2 (low), 3 (medium-low), 4 (medium), 5 (medium-high), and 6 (high).
↵4 Masks with a filter were in short supply in Bogotá during our experiment because of increased demand associated with the COVID-19 pandemic (Semana 2021). Nevertheless, we believe our estimated treatment effect for wearing a mask with a filter is plausible. Even though it is large in percentage terms (218%), baseline rates of wearing a mask with a filter were quite low (2%). Therefore, our estimated effect only implies that the treatment caused 10 members of our 226-person treatment group to begin wearing such masks because of poor air quality. That said, given the shortage of masks with filters, we cannot rule out the possibility that some treated participants who reported starting to use such masks during the course of the experiment had in mind masks without filters.
↵5 In principle, the moderating effects of health self and exercise outdoors on info searched app, know and behavior could be dampened if participants for whom health self = 1 (those with cardiopulmonary conditions) or those for whom exercise outdoors = 1 (those who exercise outdoors) have baseline levels of these outcome variables that are so high as to leave little room for improvement. However, our data do not support that hypothesis. Of the three outcome variables included in the heterogeneity analysis, mean baseline levels of two—info searched app and behavior—for participants for whom health self = 1 or exercise outdoors = 1 are all below 22% (Appendix Table A8). For the third outcome variable, know, mean baseline levels for these subgroups exceed 90%, raising the specter of ceiling effects. Nevertheless, we find that health self has a (weak) moderating effect on know (Table 5). We are not able to discern a moderating effect of exercise outdoors on know. But that finding is unlikely to reflect differential ceiling effects for the exercise outdoors = 1 subgroup: we are unable to reject the null hypothesis that baseline levels are equal for participants for whom exercise outdoors = 1 versus zero (Appendix Table A8).
↵6 This section draws on the modeling framework developed for Blackman, Bonilla, and Villalobos (2023).
↵7 We adjust the OECD VSL using the following formula (Narain and Sall 2016):
, where c is the country (Colombia), Y is the per capita GDP, ε is the income elasticity of the VSL, %ΔP is the percentage change in Colombia’s consumer price index from 2011 to 2019, and %ΔY is percentage change in Colombia’s GDP during the same period. VSL is expressed in constant 2019 U.S. dollars adjusted for purchasing power parity. The specific parametric assumptions used to derive our VSL are detailed in Blackman, Bonilla, and Villalobos (2023).↵8 Three studies of bias in self-reports about avoidance behaviors intended to slow the spread of COVID-19, including mask wearing and social distancing—each using a different method to detect deviations between actual and self-reported behaviors (list experiments, crosswise models, and analysis of smartphone location data)—concluded that these deviations are quite small or negligible (Jensen 2020; Gollwitzer et al. 2022; Larsen, Nyrup, and Petersen 2020).
↵9 For example, by restricting economic activity—and in particular motor vehicle use—the Colombian lockdown improved air quality in Bogotá. On average, concentrations of fine particulates fell by more than a third in the first several months of the lockdown, when our experiment was implemented (Blackman, Bonilla, and Villalobos 2023). In addition, the lockdown caused people to remain indoors. The improvement in air quality and the decline in outdoor activities should have reduced our participants’ incentives to learn about avoidance behavior, undertake such behavior, warn others about air pollution, and so on. But unless these factors affected our treatment and control participants differently—and we see no reason to expect they did—they would not bias our results.
This open access article is distributed under the terms of the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: https://le.uwpress.org.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.