Effects of Short Term Measures to Curb Air Pollution: Evidence from. Santiago, Chile

Transcription

1 Effects of Short Term Measures to Curb Air Pollution: Evidence from Santiago, Chile Jamie Mullins Prashant Bharadwaj March 21, 2013 Abstract: Do temporary restrictions aimed at reducing air pollution lead to better air quality and health in the short run? This paper focuses on examining the short term effects of a particular set of policies, starting in 1997 in Santiago, Chile, which were aimed at lowering pollution by the issuance of an Environmental Episode whenever air pollution reached dangerously high levels. These Episodes involved issuing alerts to warn the general public about poor air quality, imposing driving restrictions and temporarily shutting down stationary emitters of air pollution. Using propensity score matching and difference in differences, we show that Chile s Environmental Episodes approach to addressing severe air pollution events in the short run is quite effective, with the enactment of an Environmental Episode leading to reductions in P M 10 concentrations on the order of 20%. Episode announcements are also significantly related to fewer deaths, especially deaths among the elderly and due to respiratory causes. Three days after implementation, an Episode results in nearly 15 fewer deaths among those aged 64 and over compared to the three days after a pollution peak when no Episode announcements are made. Keywords: Pollution alerts, particulate matter, mortality Department of Economics, UC San Diego Department of Economics, UC San Diego

2 1 Introduction Over the past 50 years, an increasing number of national, regional, and city governments around the world have taken deliberate actions to address local air pollution. Unlike global pollution issues such as the deterioration of the Ozone layer or global temperature rises, local air pollution problems arise largely due to location specific economic development, topography and weather patterns. 1 As a result, government responses to local air pollution have varied greatly in their methods and levels of success. Broadly, these policies can be thought of as either measures that tackle pollution in the long run (switching to cleaner sources of energy, mandatory car emissions standards, et cetera) or measures that tackle pollution problems in the short run (driving restrictions in response to current air quality, temporarily shutting down or reducing usage of stationary emissions sources, et cetera). While the literature examining the impacts of long run pollution abatement policies is rich (see for example: Davis (2008), Auffhammer and Kellogg (2011), Chay and Greenstone (2003)), few studies empirically examine the effects of temporary policy responses to poor air quality. Quantifying such effects are important in light of recent pollution peaks in cities like Beijing, which prompted political leaders to address the issue. However, even in that instance, most of the debate has surrounded long term changes like relocating power plants and updating cars and buses to pollute less (China Daily, March 2013). While long term solutions are certainly important in reducing overall pollution levels, this paper shows that even short run responses to peak pollution episodes can have positive impacts in the near term and can perhaps mitigate some of the societal costs of elevated pollution levels. 2 Santiago, Chile is particularly susceptible to poor air quality 3 : prior to the start of significant government interventions, P M 10 levels within the city of greater than 300µg/m 3 were not uncommon and occasionally levels of more than 500µg/m 3 were measured. Considering that World Health Organization guidelines for P M 10 are currently set at a 24-hr mean value of 50µg/m 3, it is unsurprising that Santiago was widely known for its poor air quality (WHO, 2011). In response to these concerns, and growing public discontent with the air quality in the city of Santiago, the Government of Chile instituted a string of policies in the early 1990s to address worsening air pollution. These actions culminated in the 1997 publication of the Plan de Prevención y Descontaminación Atmosférica (hereafter PPDA), 4 which laid out a region-wide governmental approach explicitly intended to reduce air pollution 1 Though air pollutants are transported globally, the amounts of air pollution attributable to foreign sources is small compared to ambient levels in many cities (EPA, 2010). 2 Examples of costs imposed by air pollution include: degradation of asset valuations (Chay and Greenstone, 2005), diminished quality of life (Luechinger, 2009), reduced experiential values (Carson et al., 1992), property and ecosystem damage (Likens et al., 1996), and reductions in economic output (Ostro, 1983, Graff Zivin and Neidell, 2013). 3 Throughout the rest of this paper we use the terms poor air quality interchangeably with pollution and also with P M 10. P M 10 is one of the main contributors to poor overall air quality in Santiago and levels of P M 10 solely dictate the short term responses we examine. P M 10 is a measure of particulate matter in the ambient air with diameter of 10 micrometers or less. P M 10 levels are usually reported in units of micro grams of such particulates per cubic meter of air, or µg/m 3. 4 PPDA translates as the Plan to Prevent and Reduce Air Pollution. 1

3 in the Santiago Metropolitan Region and mitigate the negative health effects of air pollution exposure among the population. While the PPDA included a number of provisions, central among them was a program under which the government would publicly identify days on which levels of air pollution exceeded (or were expected to exceed) certain threshold levels, and flag such days according to a tiered labeling system. Announcement of the higher levels of such Environmental Episodes (as they are called) were accompanied by mandatory restrictions on driving as well as the shut down of certain major stationary emitters. The analysis in this paper focuses on the effectiveness of the entire suite of policies implemented under Episode announcements rather than any specific component of the various restrictions. 5 One of the main challenges in an empirical examination of such policies is isolating the impact of the policy from other factors that might also be driving pollution levels in the short run. Since Episodes are announced when pollution levels are well above average levels, it is hard to determine whether a subsequent observed drop in air pollution levels is due to the enacted interventions or simply due to the recession from a natural maximum which would have occurred regardless of policy actions. While it is difficult to pin down the causal impacts of such policies in the long run, we can do better in the short run by using the fact that Episodes are only announced on days when pollution levels reach a certain level. By comparing Episode days to days with similar pollution levels but when no Episodes were announced, we are able to ascertain the effectiveness of Episodes in reducing pollution levels and air quality related mortality. Moreover, we exploit the fact that the Episode announcements were not fully implemented before 1997, giving our design a period of time when such policies were effectively not in place. Our empirical methodology consists of propensity score matching (to identify days in the pre-ppda period that look similar 6 to Episode days in the post-ppda period) followed by a difference in differences analysis (outcome variables on days before and after the Episode announcement are differenced across the two regimes) to get at the causal effect of Environmental Episodes on pollution and mortality levels. 7 Finally, we are aware that when examining health impacts of such policies, avoidance behavior becomes an important empirical confounder (Neidell, 2009, Moretti and Neidell, 2011). Our empirical strategy, while effective at dealing with mean reversion in pollution levels, does not resolve the issue of avoidance behavior. Hence, we interpret our estimates on mortality as the full impact of the policy on mortality, capturing deaths avoided through both lower levels of air pollution and avoidance behaviors undertaken by a more informed public. We find that the metropolitan area of Santiago has been able to effectively address high-p M 10 levels on a short term basis through the use of Environmental Episode announcements. Days after an Episode 5 We will use the term Episode to include both the informational and implementational aspects of the policy. Thus, for our purposes, an Episode announcement encompasses both the identification of a high air pollution day, and the automatic counteractive measures taken under such circumstances. 6 Looking similar in this context means having balance in pre-treatment covariates. 7 By differencing before and after an Episode announcement, our methodology differences out other long term pollution fighting measures. 2

4 announcement experience significantly lower levels of air pollution compared to similar days with no announcement. This is most starkly shown in Panel A of Figure 3 below. For example, we find that day 2 after an Episode announcement has approximately 25% percent lower P M 10 levels than we would expect had the Episode not been implemented. Hence, in the short term, the policies as executed after the PPDA appear to have had a significant impact in improving air quality in Santiago. Our results on mortality are generally supportive of the idea that improvements in air quality in the short run reduce deaths and in particular deaths among the elderly and deaths attributed to respiratory ailments. We find that 3 days after an Episode announcement, there are approximately 15 fewer (cumulative) deaths above the age of 64, 8 and nearly a quarter of this reduction is due to decreases in deaths due to respiratory causes. This is in line with the idea that poor air quality harms health via respiratory illness. Such results also corroborate other findings demonstrating that even short term exposure to air pollution can impose a significant cost on human health (Graff Zivin and Neidell, 2013, Neidell, 2009, Currie et al., 2009, Schlenker and Walker, 2011). This study adds an important dimension to the broader literature examining the intersection of air quality and health. While studies have shown that air quality matters for health and other outcomes of interest to Economists like productivity and school attendance (Graff Zivin and Neidell, 2013, Currie et al., 2009) these papers do not directly tackle the impacts of air quality policies. Most of the studies examining the impact of air quality policies on health have examined long term policies like the role of the Clean Air Act (Chay and Greenstone, 2003, Sanders and Stoecker, 2011) or the NO x Budget Trading Program (Deschenes et al., 2012). Hence, this paper s main contribution is a novel analysis of the environmental and health impacts of short term policy measures to tackle air quality. Our paper also adds to the literature measuring the effects of localized governmental policies aimed at curbing air pollution. Most of these studies and policies focus on driving restrictions and have found mixed results. Using data from São Paulo, Bogotá, Beijing, and Tianjin, Lin et al. (2011) find that driving restrictions may reduce the incidence of extremely high concentrations of air pollutants, but that such restrictions do not, on average, significantly improve overall urban air quality. Davis (2008) concludes that one-weekday-per-week driving restriction in Mexico City did not significantly improve air quality in that city, and finds that the restriction likely drove the purchase of additional (often older and dirtier) vehicles by households to get around the restrictions. In addition to examining the environmental impacts of similar but short term policies in Santiago, we extend our analysis to examine the health impacts of such policies. Quantifying this natural externality of pollution abatement policies, especially in the context of a developing country, is a novel addition to the literature. 9 8 This is equivalent to a 3.3 deaths per 100,000 relative reduction in death rates for those 65 and over. 9 As this paper addresses the short-term effectiveness of Episode announcements in Santiago, it is important to note the paper by Troncoso et al. (2012), which finds that the announcement of Environmental Episodes leads to significant reductions of particulate matter, CO, NO x, and O 3, while having no effect on SO 2. In addition to an examination of the health impacts of the Santiago policies, this paper adds to the Troncoso et al. (2012) analysis in a number of other important ways. First, a broader time period, including the initial introduction of the PPDA is examined, allowing us to take advantage of greater overall variation in pollution. Second, we explicitly account for mean reversion of ambient pollutant levels and the fact 3

5 2 Background Santiago s geographic position in a basin at the foot of the Andes mountain range leads to the frequent occurrence of temperature inversion layers over the city. 10 Such inversion layers reduce vertical atmospheric mixing, thereby trapping pollutant emissions near the ground. Thermal inversions are common in summer as well as in winter; however, in the winter months (April-August), the inversion layers tend to be much closer to the ground, consistently leading to increased levels of air pollution at ground level (Gramsch et al., 2006, Rutllant and Garreaud, 1995). Since the 1960 s Santiago has periodically suffered from periods of air pollution far in excess of levels considered healthy for the region s ever growing population (now estimated at 7 million people). Beginning in the late 1980s, the government of Chile implemented a string of policies intended to address air pollution in the greater Santiago area. This included the establishment of an automated network of pollution monitors in 1988, the mandated inspection and ranking of stationary emissions sources based on the concentration of pollutants emitted in 1990, and the requirement that all new cars have catalytic converters in 1992 (Chilean Ministry of the Environment, 2007). As a result of these, and other programs, pollution in the city has declined over the subsequent decades. 11 The policies that are the focus of this paper were put on the books in 1990, and involve the Chilean Government preemptively identifying days on which air pollution was expected to be particularly severe. 12 Such days, known as Environmental Episodes (or simply: Episodes ), were dubbed Pre- Emergency Episodes if forecast P M 10 concentrations exceeded 240µg/m 3, and Emergency Episodes if P M 10 levels over 330 µg/m 3 were anticipated (Supreme Decree 32,1990). For the purposes of our analysis, it is important to note that although the policy of identifying and announcing Critical Episodes was technically established in the early 1990s, evidence suggests that it was not vigorously implemented until much later. Figure 2 shows the shift from lax execution of Critical Episodes coincided roughly with the implementation of the PPDA. 13 In the pre-ppda period, we see 148 days when P M 10 was in excess of 240µg/m 3 (i.e. an Episode was warranted under the law), but less than 40% of these days were announced as Episodes. Conversely, after the passage of the PPDA, the government announced that Episode announcements are likely to correspond with natural peaks in pollution levels. Third, we address the mutli-day effects of Episode announcement, which are not discussed by Troncoso et al. (2012) but prove to be a significant portion of the overall benefit of such announcements. Fourth, the approach of this paper allows for the examination of the treatment effects of Episode announcement, while Troncoso et al. (2012) are restricted to characterizing pollution levels contemporaneous with Episode announcements. Finally, we examine and are able to identify mortality impacts of the Episode program, something that Troncoso et al. (2012) does not consider, but is clearly of great interest. 10 Santiago is also flanked to its West by the significant Coast Range, to the North by the Chacabuco mountains, and to the South by the Cantillana range (Prendez et al., 2011). 11 See Figure 1 for a plot of annual mean P M 10 levels in the 1990s and 2000s. 12 Critical Episodes must be announced by 7pm on the prior evening. Thus, such announcements are based on predicted levels air pollution. 13 The PPDA also introduced a new level of Environmental Episode implemented when P M 10 concentrations were expected to be > 195µg/m 3. The new level is called an Alert Episode, and is generally omitted from the analysis of this paper because it came with only minimal additional driving restrictions and no implications for stationary emitters. The protocols that come into force upon the announcement of an Alert Episode are summarized in Table A.9. 4

6 Critical Episodes on nearly all high-pollution days. 14 We conclude that the policy of Environmental Episodes practically came into force beginning in 1997, and proceed with our analysis, treating 1997 as the working implementation date of the policies of interest. 15 Practically speaking, this means that we are treating all pre-1997 days as if they were untreated by the Episode program of interest. If the pre-ppda Episodes were in fact effective, this strategy would result in a conservative estimate of the effectiveness of short term environmental regulations post PPDA. 16 As described earlier, Episode announcements include the automatic implementation of a set of restrictions and government actions intended to reduce air pollution and improve health outcomes. In addition to the severe driving restrictions and mandated shutdowns of large shares of stationary emissions sources, a number of other protocols come into force upon the announcement of an Episode. These include extra street sweeping, cancellation on physical education in schools, and restrictions on residential heating fuels. 17 As public knowledge and air pollution levels are contemporaneously impacted by an Episode, we are unable to attribute effects to specific protocols, or separate the effects of avoidance behavior from other ameliorative impacts of Episode announcement. Empirically we thus examine the gross impacts of the Episode policy package of announcement and prescribed government actions. 3 Empirical Approach and Data In order to examine the short term effects of Episode announcements, we must develop an approach that allows us to control for mean reversion. Because Episodes are typically announced at times when air pollution is above mean concentrations, on average we would expect air pollution levels to fall on subsequent days whether or not any (effective) actions were taken. Failing to take account of this fact will lead to upward bias in our estimated effects. Additionally, other factors, including weather shifts, may lead to differential changes in pollution levels that are unrelated to Episode announcements. In order to account for mean reversion (and perhaps other effects that might change outcomes, but are unrelated to Episode announcements) we will utilize a difference-in-differences approach, comparing changes in outcomes from before to after an Episode to changes in outcomes from before to after 14 To drive this point home, note the similarities between 1996 v and 1994 v in Figure 2. Each pair had comparable numbers of high-p M 10 days, yet we see in Figure 2 that the years after the implementation of the PPDA (namely: 1998 & 1999) have dramatically more Pre-Emergency and Emergency level Episode announcements. Thus, years with similar incidence of air pollution days were met with quite different government actions, suggesting that the vigilance with which Episode announcements were made was different before versus after the implementation of the PPDA. 15 As a check of this approach, we use our empirical framework to asses the potency of the pre-ppda Episodes. We find that such Episodes were relatively ineffectual, even when they were implemented (results not presented). 16 As a robustness check, we omit Pre-1997 Episodes from our matching group. We find little change in our ultimate results. If anything, our results are slightly stronger when excluding these days (which suggests some, minimal, effectiveness of pre-1997 Episodes). This finding is in line with our expectation that including these pre-ppda Episode days results in an underestimate of the true effect. Table A.5 contains our mainline results reevaluated without Pre-1997 Episode days considered in the matching process. 17 See Table A.9 for a complete list of Episode protocols. 5

7 another (similar) day when no Episode was announced. In the remainder of this section, we briefly describe the data used in our analysis and then lay out our empirical methods. 3.1 Data The empirical analysis in this paper relies on a panel data set which covers , and was created by merging day-level data from a number of administrative and observational sources. The data set is upon the foundation of observational measures of Santiago P M 10 concentrations, which were collected by the the MACAM (pre-1997) and MACAM-2 (post-1997) monitor networks and maintained by Chile s Ministry of the Environment. 18 In order to ensure the most consistent comparison of P M 10 levels before and after 1997, we only use data from sites with monitors in place during both periods. In practice, data from monitors at Parque O Higgins 19, La Paz 20, and Los Condes are used for the entire period from , because only these locations were monitored under both networks. 21 P M 10 data is aggregated to the average daily level by station, and the citywide mean across stations is the focus of our analysis. See Figure 1 for a plot of P M 10 levels in Santiago over the period of study. As weather conditions are expected to covary with many of the outcomes of interest in this study, observational weather controls are of critical importance. Daily minimum, maximum, and mean levels of a large number of meteorologic variables (temperature and precipitation for example) are calculated from hourly observational data obtained from the U.S. National Climate Data Center s Summary of the Day data series, and merged with the daily pollution data for the entire period from Administrative data on the dates and levels of Episode announcements from the Santiago Metropolitan Region s Ministry of Health is also added to the panel. Unfortunately, mortality data for the Santiago Metropolitan region was only available starting in This data was obtained from the Chilean Ministry of Health s Department of Statistics and Health Information, and includes data on each death in Chile for the period , including date of death, age of the deceased, and International Classification of Diseases (ICD) codes for primary and secondary causes of death. 22 This data is aggregated to the daily level for the Metropolitan Region and added to the panel. It is important to note however, that the limited temporal scope of the available mortality data means that we will have slightly fewer matched observations for analyses involving mortality data. 18 When the original 6 monitors (there are now 9) in the MACAM-2 network replaced the 5 monitors in the MACAM network, they were deliberately spread more widely throughout the city, with placements intended to capture traditional hotspots and provide observations on representative pollution levels(gramsch et al., 2006). Due to the adjustment of monitor locations between the networks, only 3 sites were monitored over the entire period of the study. Our primary analysis uses just these three monitors, although using all available monitors to compute citywide pollution measures does not substantially alter our results. 19 This site is also referred to as Santiago. 20 Also referred to as Independencia in some sources. 21 See Figure A.1 for a comparison of monitor placement under the MACAM and MACAM-2 networks. See Table A.1 for a robustness check of our monitor selection. Our results are not driven by monitor selection. 22 ICD-9 codes are used in the data through 1996, and ICD-10 coding is used thereafter. Our results account for the changing classification. 6

8 Table A.8 presents mean values for a number of variables used in the analysis. 3.2 Identifying Treatment (Episode) Days In order to examine magnitudes of an Episode s impact on air quality and mortality outcomes, it is necessary to observe an Episode treatment in isolation from confounding factors and additional treatments. A factor that could prove problematic for our analysis is the clustering of Episode announcements. Often, Episodes would be announced repeatedly over a sequence of days when air pollution levels were particularly high. In order to avoid confounding the impacts of one Episode with those of another, our analysis focuses on Episode days which are neither preceded nor followed by another Episode for a five day period. Thus, under our separation criteria, if an Episode is announced on Day 1, no other Episodes may occur between day -5 and day While this approach limits the number of Episodes used in the analysis to 35 (a total number of 91 post-1997 Episodes were announced during the period examined), it helps ensure that the estimated impacts are appropriately attributed. Additionally, our results are generally robust to the use of other separation criteria and to the inclusion of all Episodes in the analysis Identifying Appropriate Counterfactuals Given that the impacts of mean reversion on pollution levels are likely to be non-trivial, a simple event study approach, comparing outcome variables before and after an Episode, will not be sufficient. Instead, a comparison group is needed. Since the focus of this investigation is on the short-term impacts of Episodes as they were structured under the PPDA protocols (i.e and after), it may seem that the obvious comparison group is the days on which Episodes were announced before the PPDA. We are, however, interested in assessing the impacts of Episode announcements, and not the differential impacts of Episodes following the PPDA. Thus, limiting our comparison group to pre-ppda Episodes does not serve the needs of this analysis. Also, we showed earlier that implementation of Episodes was quite lax and inconsistent prior to 1997, which suggests that pre-ppda Episodes, as a group, may not be the best comparisons for evaluations of post-ppda Episodes. Conversely, Figure 2 shows that the Chilean government s zealous implementation of Episodes in 1997 and after likely robs us of closely comparable non-episode days in the post-ppda period. All this leads us to look in the pre-1997 period (when many days likely should have had Episode announcements) for days that are comparable to Episode days in and after As most days in the period from are not similar to days after 1996 on which Episodes were announced, a direct comparison of outcomes would lead to high levels of bias (Abadie, 2005). Instead, we must use 23 It is worth noting here again that, beginning in 1997, there was another Episode level titled Alert. These Episodes were quite frequent and came with much less severe restrictions (see Table A.9 for a comparison), thus we ignore Alert Episodes in our analysis. We use the term Episode to refer only to Pre-Emergency and Emergency Environmental Episodes. 24 See Table A.2. 7

9 matching techniques to identify days from the pre-1997 period that are similar to Episode days in the period after the PPDA. Similarity in this case will be based on pre-treatment characteristics of the days leading up to an Episode, but even such covariates are much to rich to find exact matches amongst days in the pre-ppda period. Following Rosenbaum and Rubin (1983), we use propensity scores to reduce the dimensionality of the matching problem. Matching on propensity scores provides a quantitative approach to linking post-ppda-episode days to similar days on which an Episode of interest was not announced. Each Episode in the post-ppda period is matched, based on a Logit generated propensity score, to a number of days in the pre-ppda period. 25 The propensity score can be thought of as an estimated probability that a given day would have had an Episode announcement. The value of the score is generated by running a Logit model on predetermined characteristics of days in the post-ppda period on which Episodes were announced, and using the estimated coefficients to predict the probability that a given day in the pre-ppda would have had an Episode announcement. The following specification of the Logit model is estimated in order to generate propensity scores for days in the sample: where: y t = α + 5 (β j P M10 t j + ˆγ j X t j ) + ˆδ DOW t j=1 +ˆθ month t + ε t (1) y t : indicator variable =1 if an Episode was announced on day t P M10 t : Mean, P M 10 concentration on day t X t : vector of observed weather variables on day t including mean temperature, average windspeed, and precipitation DOW t : vector of dummy variables for each day of the week month t : vector of dummy variables for each month of the year ε t : error term Weather variables for day t are excluded because their levels are not determined at the time of the Episode announcements. 26 Day of the week was included as it likely captures some emissions infor- 25 As noted, we allow days in the pre-ppda period on which Episodes were announced to be matched to ensure the best possible matches. If these pre-ppda Episodes did have positive effects on the outcomes of interest, the inclusion of these days will bias our estimates downward. 26 While the inclusion of day-of weather variables might be justified because they are unaffected by treatment, or because they serve as proxies for forecast values that officials may have consulted, such inclusion does not markedly change our results. We therefore stick with propensity scores based strictly on lagged values. 8

10 mation that may improve match quality. 27 Month dummies were included to capture both seasonal variation in weather patterns, which likely contribute to fluctuations in air pollution, and seasonal variation in Episode announcements connected to attitudes within government. Once the above Logit model is estimated, each post-ppda Episode and pre-ppda day is plugged into the model, and the resulting ŷ t is the propensity score for day t. In our headline result each post-ppda Episode is then matched to the the five pre-ppda days with the most similar propensity scores. This method is known generally on the five Nearest Neighbor approach. 28 Matching each Episode to multiple pre-ppda days reduces the variance of our estimates, while limiting the number of matches reduces the possibility of using poor matches in our analysis. We enforce common support on propensity scores between the Episode and matched groups, which leads us to drop one Episode day from our analysis because its propensity score is above that of any comparison day. Enforcing common support is important theoretically and empirically to ensure that matches are in fact similar (Heckman et al., 1999). 3.4 Difference-in-Differences Now that we have identified a set of matched days in the pre-ppda period, we have an appropriate control group against which to compare the outcomes of our treatment group of days with a post- PPDA Episode announcement. Since the goal of this exercise is to identify the effects of an Episode on several different variables, the outcome variables we will compare between the treatment and control groups will be differences over time, across the date of the Episode (or Episode matches). Using a difference-in-differences (hereafter DID ) strategy helps ensure comparability of outcomes on similar days from different periods given the different air quality and mortality characteristics between the periods. 29 The DID approach controls for long-term mean changes in our outcome variables 30 better than a time trend in the propensity score. Additionally, Smith and Todd (2005) find DID estimators better address population and measurement method mismatch than do estimates based on comparisons of levels. Conceptually, our first difference (of the DID) will be changes in pollution levels or number of deaths (our outcome variables of interest) from before to after an Episode (or matched day). Our first results are the traditional comparison of the means (of these differences) between the treated (Episode) days 27 Additionally, although it is not technically a factor that should be considered by authorities, the day of the week may impact the government decision of whether to announce and Episode or not. 28 See Table A.4 for a demonstration of the robustness of our results to alternative matching procedures including Caliper and Kernel based methods. Additionally, note that we match with replacement in order to maximize the quality of our matches to reduce the bias of our estimates, though possibly at the cost of higher variance in our estimations (Smith and Todd, 2005, Abadie and Imbens, 2005). This means that some pre-ppda days are matched to multiple Episodes. When this occurs, such days are appropriated higher weightings in the empirical analysis. 29 Mean P M 10 levels are different in different years, thus spikes in P M 10 are different between years, both in absolute levels and in relation to the thresholds warranting Episode announcements. 30 Year on year, air quality was improving and the number of deaths in Santiago was growing as populations increased. 9

11 and matched control days. These results are presented in Table 3. No additional controls are included at this stage of the analysis because the propensity score matching effectively controls for all the inputs which are matched over. 3.5 Regression Analysis The strict comparison of means relies on perfect matching (i.e.- exact balance in the distributions of all observable and unobservable covariates between the treatment and matched groups) of our treatment and control groups, and correct specification of the propensity score estimation. Table 1 shows the balance on our observable matching covariates, and demonstrates that our matching procedure has given us good balance, but definitely not perfect balance. In order to control for remaining pre-treatment differences between our treatment and control groups we implement a mixed method like those laid out in Imbens (2004), which involves running regressions using the samples and weightings generated in the matching procedure. 31 In addition to controlling for remaining differences between our comparison groups, regression analysis, with proper controls, will address correlations between the matching covariates and our outcomes of interest, which has the potential to increase the precision and/or reduce the bias of our estimates (Imbens, 2004). Approached from another perspective, we are using propensity score matching prior to a difference-in-difference analysis to better meet the necessary assumptions regarding parallel evolution of paths between the treatment and control groups (Abadie, 2005) Although we run a number of regression specifications (results from some are presented as robustness checks below), our main specification is simply a traditional DID OLS regression assessing the interaction of indicator variables for treatment and being a member of the treated group. The key to this analysis is that it is run, not on the whole sample, but on the matched treatment and control groups obtained via propensity score matching. Such an analysis is analogous to regressions run following the implementation of a randomized control trial, as additional controls are still useful even when it is believed that treatment and control groups are closely comparable. 32 Below is the specification of our mainline regression analysis where the coefficient of interest is the interaction term: Y t = β 0 + β 1 Ep t + β 2 P ostp P DA t + β 3 (Ep t P ostp P DA t ) 5 + ˆγ j X t j + ˆδ DOW t + ˆθ month t +ε t (2) j=1 31 See also: Hirano and Imbens (2001). 32 In the context of an RCT, the close comparability of treatment and control group is achieved through random assignment. While we don t have random assignment here, the goal of our matching procedure is to mimic random assignment of Episode days amongst days warranting an Episode. 10

12 where: Y t : outcome variable day t Ep t : indicator variable =1 if day t received an Episode announcement P ostp P DA t : indicator variable=1 if day t is in a year >1996 (i.e.- after the implementation of the PPDA) X t : vector of observed weather variables on day t including mean temperature, average windspeed, and precipitation. DOW t : vector of dummy variables for each day of the week month t : vector of dummy variables for each month of the year ε t : error term 4 Results The essence of our results is presented in Panels A and B of Figure 3 which shows the average movement of air pollution and mortality through time across the threshold of an Episode announcement in treatment (post-ppda) and control (pre-ppda) periods. In each of the graphs, the vertical line represents the announcement of an Episode, while the horizontal axis groups days by the amount of time before or after an Episode. In the case of the Matched line, these data represent pre-ppda days drawn using the matching methodology described in the previous section. Visually it is clear that relative to similar days in the pre-ppda period, implementing an Episode at the peak of pollution levels in the post-ppda period drastically reduces P M 10 levels. Panel A also shows that our matching methodology for the five days before an Episode announcement works quite well. Since the matching was done purely based on pollution and weather patterns, the match in Panel B for deaths is not as close. However, this stays true to the way the Episode announcements worked in that only pollution levels and meteorologic conditions were taken into account while making that decision. The simplest regression analog of Panel A - Figure 3 is Table 2, where the difference in means by lagged days are taken across treatment and (matched) control groups. As Panel A - Figure 3 suggests, the effect of an Episode announcement on pollution levels is large and statistically significant. Particularly, we see that the day of an Episode experienced average P M 10 levels that were approximately 22.5 µg/m 3 lower than would have been expected without the Episode announcement. This is a very striking effect given that the mean city-wide level of P M 10 on post-ppda days when Episodes were announced was 112 µg/m 3. This suggests that an Episode announcement, along with all the government actions and restrictions such an announcement brings into force, leads to immediate P M 10 concentration 11

13 reductions of approximately 17% from anticipated levels, with additional air quality benefits continuing for several subsequent days. The magnitude of this day-of estimated effect compares to the high-end of the estimated effects presented by Troncoso et al. (2012) for Pre-Emergencies, but given that our analysis includes only stand alone Episodes, this is to be expected. In order to assess the short-term effectiveness of Episode announcements on health outcomes, we estimate regressions using cumulative daily death counts among the elderly as the outcome variable (in Table A.6, we consider other age groups). Columns 2 and 4 of Table 2 show that deaths and death rates among the elderly appear to decrease after an Episode announcement. Three days after an Episode, the cumulative decrease in deaths appears to be around Hence, it appears that Episode announcements are associated with significant decreases in elderly mortality. As Table A.6 suggests, we do not see any changes in deaths or death rates for causes that are likely not linked with short term exposure to high levels of air pollution. For instance, Table A.6 shows that deaths due to cancer are not affected as a result of Episode announcements. 34 As discussed in the previous section, an alternative way to assess the effectiveness of the Episode days is to simply compare the means across treated and control days. The results from this approach are shown in Table 3. This table shows largely similar effects although the mortality results are less precisely estimated. This is anticipated as the direct differences-in-means approach of Table 3 assumes perfect balance and attempts to mimic a randomized control trial. In Table 2 we show that the use of regression controls can account for some of the remaining differences between our comparison groups. Such an approach leads to increased precision of our estimates as expected. Table 2 in general verifies the results presented in Table 3, but the estimates are more precise. Given the advantages of such mixed methods as outlined in Imbens (2004), we use the results in Table 2 as our preferred estimates. 4.1 Robustness Checks Our results are robust to a broad range of specification checks. As mentioned earlier, our results are not driven by changing monitors across the pre and post-ppda regimes. As Table A.1 shows, including all monitors that are available regardless of when the monitors came online does not drastically change our results. This is not altogether surprising given the lack of variation in pollution across communas within the city of Santiago. 35 Since the results hinge on the variables we use to match and identify control groups, we explore whether our results are robust to the addition of more covariates in the matching process. In Table A.3, we use 5 days of lags for each of: P M 10, daily max P M 10, wind speed, precipitation, temperature, maximum temperature, minimum temperature, atmospheric pressure, dew-point, and the square and 33 Corresponding to a lower death rate of approximately 4 per 100,000 individuals over the age of We do see a small reduction in infant accidental deaths coincident with Episode announcements. This could be due to driving restrictions keeping mothers (and children) home on Episode days. 35 Los Condes is a bit of an exception with lower pollution than most of the other communas in Santiago, but robustness checks confirm that our results are not changed substantially by the exclusion of Los Condes monitor data (results not shown). 12

14 cube of temperature as covariates in the matching process. As Table A.3 shows, the addition of these covariates, if anything, improves the precision of our estimates, though this analysis covers a smaller group of Episodes due to the enforcement of the common support requirement. In Table A.4, we continue to explore the sensitivity of our results to different matching methods. Table A.4 shows that the choice of matching method is not consequential to our overall set of results. A potential concern is that including pre-ppda Episodes as part of the control group could bias our results. As mentioned earlier, the inclusion of pre-ppda Episodes, if anything, biases our results towards finding no effect of post-ppda Episodes. In Table A.5, we exclude all pre-ppda Episodes from our control group and find that our results are indeed stronger. However, as Figure A.2 indicates, it is not clear that the pre-ppda announcements had any impact on pollution levels in the near term and hence, our preferred specification includes pre-ppda Episodes as part of the control group to ensure that matches are of the highest possible quality. 5 Conclusion In this paper we analyzed the short term environmental and health consequences of pollution Episode announcements in Santiago, Chile. As Episode announcements included temporary restrictions on mobile and stationary emitters, we show that such restrictions are effective in reducing pollution levels, and that the combined informational and pollution-reducing aspects of the Episode measures reduce mortality in the short run. Going forward, we hope to understand more deeply the precise mechanisms behind the overall effects we see. For example, one aspect we cannot address in this paper is that the Episode restrictions are a bundle of interventions rather than one specific intervention. From a public policy effectiveness standpoint, it would be important to separately identify the effectiveness of each restriction. The bundled approach also makes it harder to analyze the economic costs of such restrictions. While pollution levels and mortality are reduced by these restrictions, they must certainly come at a cost. In future work, we hope to gain a better understanding of the costs of such restrictions. While we find some significant effects on mortality, the analysis would be better suited to examining a more sensitive measure of human health, such as hospitalizations. Hence obtaining data on hospitalizations in cities where such short term restrictions are active would be very insightful. 13

15 References A. Abadie. Semiparametric Difference-in-Differences Estimators. The Review of Economic Studies, 72 (1):1 19, A. Abadie and G. W. Imbens. Large Sample Properties of Matching Estimators for Average Treatment Effects. Econometrica, 74(1): , M. Auffhammer and R. Kellogg. Clearing the Air? The Effects of Gasoline Content Regulation On Air Quality. The American Economic Review, 101(6): , M. Caliendo and S. Kopeinig. Some Practical Guidance for the Implementation of Propensity Score Matching. Journal of Economic Surveys, 22(1):31 72, R. Carson, M. Conaway, W. Hanemann, R. Kopp, K. Martin, R. Mitchell, S. Presser, and P. Ruud. Valuing Winter Visibility Improvement in the Grand Canyon. In Association of Environmental and Resource Economists Meeting. New Orleans, K. Chay and M. Greenstone. Does Air Quality Matter? Evidence From the Housing Market. Journal of Political Economy, 113(2), K. Y. Chay and M. Greenstone. Air Quality, Infant Mortality, and the Clean Air Act of Technical report, National Bureau of Economic Research, J. Currie, E. A. Hanushek, E. M. Kahn, M. Neidell, and S. G. Rivkin. Does Pollution Increase School Absences? The Review of Economics and Statistics, 91(4): , L. Davis. The Effect of Driving Restrictions On Air Quality in Mexico City. Journal of Political Economy, 116(1):38 81, O. Deschenes, M. Greenstone, and J. S. Shapiro. Defensive Investments and the Demand for Air Quality: Evidence From the NOx Budget Program and Ozone Reductions EPA. International Transport of Air Pollution. Our Nation s Air - Staus and Trends through 2008, pages 43 44, J. Graff Zivin and M. Neidell. The Impact of Pollution On Worker Productivity. American Economic Review, E. Gramsch, F. Cereceda-Balic, P. Oyola, and D. Von Baer. Examination of Pollution Trends in Santiago de Chile with Cluster Analysis of PM10 and Ozone Data. Atmospheric Environment, 40 (28): ,

16 J. J. Heckman, R. J. LaLonde, and J. A. Smith. The Economics and Econometrics of Active Labor Market Programs. Handbook of Labor Economics, 3: , K. Hirano and G. W. Imbens. Estimation of Causal Effects Using Propensity Score Weighting: An Application to Data On Right Heart Catheterization. Health Services and Outcomes Research Methodology, 2(3): , Z. Huanxin. Beijing Pollution Fight Greater Than for Olympics. China Daily, March 13, G. W. Imbens. Nonparametric Estimation of Average Treatment Effects Under Exogeneity: A Review. Review of Economics and Statistics, 86(1):4 29, G. Likens, C. Driscoll, and D. Buso. Long-Term Effects of Acid Rain: Response and Recovery of a Forest Ecosystem. Science, 272(5259): , C. Lin, W. Zhang, and V. Umanskaya. The Effects of Driving Restrictions On Air Quality: São Paulo, Bogotá, Beijing, and Tianjin. In Annual Meeting of the Agricultural and Applied Economics Association, S. Luechinger. Valuing Air Quality Using the Life Satisfaction Approach*. The Economic Journal, 119 (536): , E. Moretti and M. Neidell. Pollution, Health, and Avoidance Behavior. Journal of Human Resources, 46(1): , M. Neidell. Information, Avoidance Behavior, and Health. Journal of Human Resources, 44(2): , B. Ostro. The Effects of Air Pollution On Work Loss and Morbidity. Journal of Environmental Economics and Management, 10(4): , M. Prendez, G. Alvarado, and I. Serey. Air Quality Monitoring, Assessment and Management, chapter Some Guidelines to Improve the Air Quality Management of Santiago, Chile. InTech, J. M. Robins, Y. Ritov, et al. Toward a Curse of Dimensionality Appropriate (Coda) Asymptotic Theory for Semiparametric Models. Statistics in Medicine, 16(3): , P. R. Rosenbaum and D. B. Rubin. The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika, 70(1):41 55, J. Rutllant and R. Garreaud. Meteorological Air Pollution Potential for Santiago, Chile: Towards an Objective Episode Forecasting. Environmental Monitoring and Assessment, 34(3): ,

17 N. J. Sanders and C. F. Stoecker. Where Have All the Young Men Gone? Using Gender Ratios to Measure Fetal Death Rates. Technical report, National Bureau of Economic Research, W. Schlenker and W. R. Walker. Airports, Air Pollution, and Contemporaneous Health. Technical report, National Bureau of Economic Research, J. A. Smith and P. E. Todd. Does Matching Overcome LaLonde s Critique of Nonexperimental Estimators? Journal of Econometrics, 125(1): , R. Troncoso, L. de Grange, and L. Cifuentes. Effects of Environmental Alerts and Pre-Emergencies On Pollutant Concentrations in Santiago, Chile. Atmospheric Environment, World Health Organization: Air Quality and Health, September Chilean Ministry of the Environment: Gestion de La Calidad Del Aire: Varias Decadas de Esfuerzo, Chilean Ministry of Health: Supreme Decree 32,

18 Tables and Figures Table 1. Balance Table Means t-test Lag Variable Treated Control Percent Bias t p>t 1 P M P M P M P M P M T emperature T emperature T emperature T emperature T emperature T emperature W ind Speed W ind Speed W ind Speed W ind Speed W ind Speed W ind Speed P recipitation P recipitation P recipitation P recipitation P recipitation P recipitation Observations Notes: Results are based on the 34 of 35 post-ppda Episodes meeting our separation criteria, which also satisfy the common support restrictions of the Propensity Score Matching approach. P M 10 values are in terms of concentrations measured in µg/m 3. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for all statistics above. 17

19 Table 2. Difference-in-Differences Matching & Regression Results Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day *** before to Day of Episode (5.31) (2.10) (0.73) (0.50) Difference from *** * * Day -1 to Day 1 (7.16) (3.67) (1.27) (0.89) Difference from *** * * Day -1 to Day 2 (6.74) (4.51) (1.83) (1.13) Difference from *** * * Day -1 to Day 3 (6.33) (5.92) (2.36) (1.48) Difference from *** * ** Day -1 to Day 4 (6.31) (7.42) (2.85) (1.83) Difference from *** * * Day -1 to Day 5 (7.16) (8.70) (3.39) (2.17) 5-Lags Weather Controls Yes Yes Yes Yes DOW Dummies Yes Yes Yes Yes Month Dummies Yes Yes Yes Yes Observations Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: Results are based on the 34 of 35 post-ppda Episodes meeting our separation criteria, which also satisfy the common support restrictions of the Propensity Score Matching approach. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of P M 10 observations from the selected, in-service monitoring stations on a given day. The average city-wide P M 10 level on the 34 Episode days was ~105.1 µg/m 3. All regressions include controls for temperature, wind, and precipitation on the each of the 5 days prior to treatment, and for the month and day-of-week of the treated day. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in

20 Table 3. Comparison of Mean Differences Results Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day *** before to Day of Episode (7.40) (2.63) (0.92) (0.61) [8.15] [2.75] [0.98] [0.62] Difference from *** Day -1 to Day 1 (10.06) (4.52) (1.61) (1.05) [12.47] [4.88] [1.67] [1.14] Difference from *** Day -1 to Day 2 (10.75) (5.75) (2.35) (1.39) [12.35] [6.50] [2.56] [1.55] Difference from *** Day -1 to Day 3 (10.57) (7.39) (3.02) (1.78) [12.09] [8.40] [3.23] [2.00] Difference from ** Day -1 to Day 4 (11.41) (9.42) (3.60) (2.25) [12.24] [10.87] [3.86] [2.56] Difference from *** Day -1 to Day 5 (11.47) (11.06) (4.31) (2.67) [12.60] [12.86] [4.55] [3.04] Observations Treatment Control *** - significant at 0.01 ** - significant at 0.05 * - significant at 0.10 Stars based on unadjusted Standard Errors in Parenthesis ( ) Bootstrapped Standard Errors in Square Brackets [ ] Notes: Results are based on the 34 of 35 post-ppda Episodes meeting our separation criteria, which also satisfy the common support restrictions of the Propensity Score Matching approach. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of P M 10 observations from the selected, in-service monitoring stations on a given day. The average city-wide P M 10 level on the 34 Episode days was ~105.1 µg/m 3. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in Standard errors computed from propensity score matching methods subject to issues as discussed in Caliendo and Kopeinig (2008). Bootstrapped standard errors calculated using 500 replications. 19

21 Figure 1. Annual Mean P M 10 Santiago Notes: Data for P M 10 levels are available from the National Ministry of the Environment. Plotted data are annual averages aggregated from daily means for monitors at the Parque O Higgins, Los Condes, and La Paz monitoring sites. 20

22 Figure 2. Episodes and High-P M 10 Events by Year Notes: Data on historical Episode announcements are available from the Metropolitan Region Ministry of Health. Data for P M 10 levels are available from the National Ministry of the Environment. 21

23 Figure 3. Evolution of P M 10 and Over 64 Deaths: Treated vs. Matched (a) Panel A (b) Panel B Notes: The vertical line at day zero represents the Episode treatment. Data on historical Episode announcements are available from the Metropolitan Region Ministry of Health. Data for P M 10 levels are available from the National Ministry of the Environment. Mortality data are available from the Chilean Ministry of Health s Department of Statistics and Health Information. 22

24 Appendix Table A.1. Robustness: Using All Monitors Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day *** before to Day of Episode (4.94) (2.15) (0.77) (0.52) Difference from *** Day -1 to Day 1 (7.11) (3.78) (1.31) (0.92) Difference from *** Day -1 to Day 2 (7.42) (4.76) (1.87) (1.21) Difference from ** Day -1 to Day 3 (6.54) (6.05) (2.38) (1.53) Difference from *** Day -1 to Day 4 (7.55) (7.62) (2.84) (1.90) Difference from *** Day -1 to Day 5 (7.87) (8.77) (3.37) (2.21) 5-Lags Weather Controls Yes Yes Yes Yes DOW Dummies Yes Yes Yes Yes Month Dummies Yes Yes Yes Yes Observations Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: Results are based on all 35 post-ppda Episodes meeting our separation criteria, which also satisfy the common support restrictions of the Propensity Score Matching approach. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of observations from the selected, in-service monitoring stations on a given day. All monitors for which P M 10 data are available are included in this analysis. These include monitors at: Parque O Higgins, Los Condes, La Paz, Providencia, Cerrillos, El Bosque, La Florida, Cerro Navia, Quilicura, and Pudahuel. All regressions include controls for temperature, wind, and precipitation on each of the 5 days prior to treatment, and on the month and day-of-week of the day of treatment. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in 2009.

25 Table A.2. Robustness: No Separation Criteria Enforced Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day *** before to Day of Episode (3.68) (1.35) (0.51) (0.34) Difference from *** Day -1 to Day 1 (4.83) (2.56) (0.94) (0.64) Difference from *** Day -1 to Day 2 (5.30) (3.55) (1.41) (0.90) Difference from *** Day -1 to Day 3 (5.11) (4.63) (1.86) (1.18) Difference from *** * Day -1 to Day 4 (4.97) (5.72) (2.24) (1.45) Difference from *** Day -1 to Day 5 (7.87) (8.77) (3.37) (2.21) 5-Lags Weather Controls Yes Yes Yes Yes DOW Dummies Yes Yes Yes Yes Month Dummies Yes Yes Yes Yes Observations Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: Results are based on 85 of the 91 post-ppda Episodes which satisfy the common support restrictions of the Propensity Score Matching approach. No separation criteria is imposed. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of observations from the selected, in-service monitoring stations on a given day. All regressions include controls for temperature, wind, and precipitation on each of the 5 days prior to treatment, and on the month and day-of-week of the day of treatment. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in 2009.

26 Table A.3. Robustness: Rich Set of Covariates in Propensity Score Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day * before to Day of Episode (7.51) (2.96) (1.25) (0.74) Difference from *** * * Day -1 to Day 1 (11.29) (5.39) (2.23) (1.35) Difference from *** * Day -1 to Day 2 (8.28) (6.63) (3.36) (1.71) Difference from ** * * Day -1 to Day 3 (8.49) (7.98) (4.25) (2.06) Difference from *** * * Day -1 to Day 4 (11.87) (10.23) (5.22) (2.62) Difference from ** * * Day -1 to Day 5 (9.90) (11.99) (6.39) (3.12) 5-Lags Weather Controls Yes Yes Yes Yes DOW Dummies Yes Yes Yes Yes Month Dummies Yes Yes Yes Yes Observations Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: Propensity Score is generated using a Logit regression of a post-ppda Episode on 5 days of lags of: P M 10, daily max P M 10, wind speed, precipitation, temperature, max temperature, min temperature, atmospheric pressure, dew point, and the square and cube of temperature. Of the 35 post-ppda Episode which met our separation criteria, only the 18 which also satisfy the common support restrictions of the Propensity Score Matching approach are included in this analysis. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of P M 10 observations from the selected, in-service monitoring stations on a given day. All regressions include controls for temperature, wind, and precipitation on each of the 5 days prior to treatment, and on the month and day-of-week of the day of treatment. Pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in 2009.

27 Table A.4. Robustness: Alternative Matching Methods Mean P M10 Cumulative Respiratory Deaths (1) (2) (3) (4) (5) (6) (7) (8) NN5 NN10.05 Caliper Kernel NN5 NN10.05 Caliper Kernel Difference from Day *** *** *** *** ** ** to Day of Episode (5.31) (4.18) (1.97) (1.95) (2.10) (1.79) (0.87) (0.86) Difference from *** *** *** *** * *** *** Day -1 to Day 1 (7.16) (5.78) (2.50) (2.48) (3.67) (3.15) (1.52) (1.50) Difference from *** *** *** *** * *** *** Day -1 to Day 2 (6.74) (5.38) (2.56) (2.57) (4.51) (3.96) (1.90) (1.88) Difference from *** *** *** *** * * *** *** Day -1 to Day 3 (6.33) (5.17) (2.47) (2.44) (5.92) (5.10) (2.42) (2.39) Difference from *** *** *** *** * * *** *** Day -1 to Day 4 (6.31) (5.32) (2.58) (2.53) (7.42) (6.44) (3.05) (3.02) Difference from *** *** *** *** * * *** *** Day -1 to Day 5 (7.16) (5.93) (2.76) (2.77) (8.70) (7.56) (3.57) (3.52) N Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: For ease of comparison, columns 1 & 5 contain the original results from Table 2 using Nearest 5 Neighbors (NN5) matching method. Columns 2&6 contain the results from matching using Nearest 10 Neigbors (NN10), columns 3 & 7 contain matching results using a Caliper matching approach with a 0.05 radius, and columns 4 & 8 contain the results of an Epanechnikov Kernel matching procedure with bandwidth=0.06. Treatments are defined as post-ppda Episodes meeting our separation criteria, and common support conditions are enforced on all procedures. P M10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of P M10 observations from the selected, in-service monitoring stations on a given day. All regressions include controls for temperature, wind, and precipitation on each of the 5 days prior to treatment, and on the month and day-of-week of the day of treatment.pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in 2009.

28 Table A.5. Robustness: Excluding pre-ppda Episodes from Control Mean P M 10 Cumulative Cumulative >64 Cumulative >64 >64 Deaths Resp Deaths Death Rate Difference from Day *** before to Day of Episode (4.92) (2.27) (0.77) (0.55) Difference from *** * Day -1 to Day 1 (6.57) (3.85) (1.36) (0.93) Difference from *** * * Day -1 to Day 2 (6.75) (4.56) (1.93) (1.14) Difference from *** * * Day -1 to Day 3 (6.61) (5.96) (2.46) (1.49) Difference from *** * * Day -1 to Day 4 (7.01) (7.44) (3.01) (1.85) Difference from *** * * Day -1 to Day 5 (7.16) (8.70) (3.39) (2.17) 5-Lags Weather Controls Yes Yes Yes Yes DOW Dummies Yes Yes Yes Yes Month Dummies Yes Yes Yes Yes Observations Treatment Control *** - significant at 0.01; ** - significant at 0.05; * - significant at 0.10 Notes: In our headline results, 4 of the days which are matched from the pre-ppda period were days on which Episodes were announced. Although including such days in our analysis provides conservative estimates of the effects of Episode announcement, we rerun the analysis excluding all Episode days from the matching procedure. Results are based on the 30 of 35 post-ppda Episodes meeting our separation criteria, which also satisfy the common support restrictions of the Propensity Score Matching approach. P M 10 values are in terms of concentrations measured in µg/m 3. Death statistics are reported in total number of deaths per day, and are cumulative beginning at the time of treatment. Death rate statistics are measured in number of deaths per day per 100,000 residents of the sub-population of interest, and are cumulative from the time of treatment. Calculations are based on city-wide averages of the daily means of P M 10 observations from the selected, in-service monitoring stations on a given day. All regressions include controls for temperature, wind, and precipitation on each of the 5 days prior to treatment, and on the month and day-of-week of the day of treatment.pollutant concentration data were obtained from Chile s Ministry of the Environment, and weather data were taken from the NCDC s Summary of the Day data set. Data from was used for the P M 10 estimates above, while data from was used for all mortality related estimates. These date restrictions are imposed due to mortality data availability and changes to the PPDA that were implemented in 2009.

29 Figure A.1. Maps of Santiago with Pollution Monitoring Networks Maps are an adaptions of those produced in PPD (2007). Figure A.2. Evolution of P M 10 : pre-ppda Episodes vs. Matched Notes: Data on historical Episode announcements are available from the Metropolitan Region Ministry of Health. Data for P M 10 levels are available from the National Ministry of the Environment.