The Effect of Metro Expansions on Air Pollution in Delhi

The Delhi Metro (DM) is a mass rapid transit system serving the National Capital Region of India. It is also the world?s first rail project to earn carbon credits under the Clean Development Mechanism of the United Nations for reductions in CO2 emissions. Did the DM also lead to localized reduction in three transportation source pollutants? Looking at the period 2004?2006, one of the larger rail extensions of the DM led to a 34 percent reduction in localized CO at a major traffic intersection in the city. Results for NO2 are also suggestive of a decline, while those for PM25 are inconclusive due to missing data. These impacts of pollutant reductions are for the short run. A complete accounting of all long run costs and benefits should be done before building capital intensive metro rail projects.


2
The Delhi Metro (DM) is an electric-based mass rapid rail transit system mainly serving the Indian National Capital Territory (NCT) of Delhi. The NCT of Delhi covers an area of 1,483 square kilometers and has a population of 16.8 million people according to the Indian Census of 2011, making it one of the world's most densely populated cities. 1 The DM was introduced in 2002 and since then it is being continually extended within the NCT and adjoining areas. As of 2012, its total route length was 190 kilometers, and annual ridership was 0.7 billion (DMRC 2012). 2 In this paper we examine whether this important mode of public transportation has had any impact on air pollution in Delhi. We identify the immediate localized effect of extending the DM rail network on air pollution measured at two different locations within the city: ITO, a major traffic intersection in central Delhi, and Siri Fort, a mainly residential neighborhood in south Delhi.
Air pollution is measured in terms of three criteria pollutants, namely, nitrogen dioxide ( 2 NO ), carbon monoxide (CO), and fine particulate matter ( 2.5 PM ).
An impact study of the DM on air pollution is important for two reasons.
First, there is substantial scientific evidence on the adverse effects of air pollution on human health. Block et al. (2012) provide a review of epidemiological research that shows the link between air pollution and damage to the central nervous system, which may manifest in the form of decreased cognitive function, low test scores in children, and increased risk of autism and of neurodegenerative diseases such as Parkinson's and Alzheimer's. They also cite other studies which show that air pollution causes cardiovascular disease (Brook et al. 2010) and worsens asthma 1. According to a worldwide ranking of cities by City Mayors Statistics, Delhi ranked thirteenth in terms of population density with 11,050 persons per square kilometer. Mumbai ranked first, and Beijing ranked twelfth, with population densities of 29,650 and 11,500, respectively. The data are compiled from various sources and are the most recent available.  (Auerbach and Hernandez 2012). Turning to recent research in economics, Tanaka (2015) finds that regulations to curb pollution from coal-based power plants in China led to 3.29 fewer infant deaths per 1000 live births, amounting to a 20 percent reduction in infant mortality rate. 3 Ghosh and Mukherji (2014)  Board (CPCB), the national authority responsible for monitoring and managing air 3. Greenstone and Hanna (2014) find that environmental regulations in India have been effective in reducing air pollution. However, in contrast to Tanaka (2009), they find an insignificant impact of reduced air pollution on infant mortality. For reasons discussed in their paper, they advise readers to be cautious when interpreting their result on infant mortality. quality in India, finds that pollution in Delhi is positively associated with lung function deficits and with respiratory ailments (CPCB 2008a;CPCB 2008b). Guttikunda and Goel (2013)  5. Appendix table S1.1 in the supplementary appendix S1 presents the CPCB limits along with those prescribed by the WHO. WHO (2000) maintains that its limits are to be interpreted as guidelines, and individual countries may have different standards based on prevailing exposure levels, and social, economic, and cultural considerations. We measure violations vis-à-vis the CPCB limits, as they should be more relevant within the Indian context. 6. During the same period at Siri Fort, 2 NO , CO, 2 SO and 3 O , exceeded prescribed limits, 13, 26, 0, and 5 percent of the time, respectively.

2.5
PM was not recorded at Siri Fort.

5
Another reason for restricting focus to only these pollutants is that while 2 NO , CO, and 2.5 PM are mainly generated from transportation sources, 2 SO and 3 O are not. In one of the first pollution inventory studies for Delhi, Gurjar et al. (2004) infer that during their study period (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000), transport sector contributed about 82 percent of nitrogen oxides ( x NO ), 7 and 86 percent of CO. In another study for Delhi conducted in 2007, NEERI (2010 reports that the contribution of vehicles shares. On the other hand, NEERI, and Guttikunda and Calorie, report that vehicular emissions were responsible for 0.3 percent and 3 percent, respectively, of 2 SO . 9 None of these studies look at 3 O . However, it is known that 3 O is not directly emitted by motor vehicles, but is created through a complicated nonlinear process wherein oxides of nitrogen and volatile organic compounds react together in the presence of sunlight (Sillman, 1999 PM . To the extent that one of the main channels through which the DM is likely to affect air pollution is through its impact on overall levels of vehicular emissions, we focus our attention on these three pollutants. Moreover, Delhi is a heavily motorized city, 10 and the consequent vehicular emissions are a 7. x NO refers to both nitrogen monoxide (NO) and nitrogen dioxide ( 2 NO ).
8. The contribution of vehicles towards particulate matter reported in NEERI (2010) and in Guttikunda and Calorie (2012) includes the contribution of road dust as well. 9. Gurjar et al. (2004) do not report this figure for 2 SO .
10. Among the 44 reported million-plus cities in India, Delhi had the largest number of registered motor vehicles during 2011-12 with 7.4 million vehicles (TRW 2013). Goel et al. (2015) estimate that of all registered cars and 6 matter of serious concern.
Theoretical research from transport economics (Vickery 1969;Mohring 1972 To be able to attribute changes in a pollutant measure to the DM, we use the Regression Discontinuity (RD) approach. As we explain below, due to the presence of sporadic sources of pollution in Delhi (such as spontaneous burning of waste), this approach is not ideal for short periods of analysis. We therefore rely on a three-year study period and argue that these sporadic sources of pollution cancel each other within this timeframe, resulting in reliable estimates.
Our analysis reveals that soon after some of the larger extensions of the DM there were significant reductions in at least some transportation source pollutants. identifying assumption.

Estimation Equation
We measure pollution using data from monitoring stations at two different locations within the city, ITO and Siri Fort. 14 Table 2 presents pollution statistics at each location, along with weather conditions at Safdarjung, Delhi. ITO has much higher pollution compared to Siri Fort: Average hourly 2 NO and CO at ITO are 3.2 and 1.5 times their respective levels at Siri Fort. This is not surprising given that ITO is a major traffic intersection, while Siri Fort is a mainly residential area.
Ideally, we would have liked to know weather conditions specific to each location.
However, we only have hourly weather data for Safdarjung, which is fortunately located between ITO and Siri Fort. We use this as the best available proxy for weather conditions at each location. As the dynamics of pollution are likely to be different across the two locations, and also because they are at different distances from the various line expansions, we estimate impacts at each location separately.
At each location we estimate the impact of a particular metro extension using a time series of hourly pollutant data lying within a symmetric window around that extension's opening date. We also ensure that there are no other extensions within this window. Thus, a window is characterized by a location l (ITO or Siri Fort) and an extension m. The RD approach is implemented by estimating the following OLS regression within each window: , l m t y is pollutant level (in logs) in hour t, at location l, when studying the effect of extension m. m t DM is the discontinuity dummy for extension m: Within 14. In talking to experts at the CPCB, we were told that a monitoring station measures the quality of ambient air passing by it, and it is not possible to demarcate a precise catchment area for which the quality measure would apply. Given that the monitoring stations at ITO and Siri Fort are approximately 9 kilometers apart, we believe that they each measure air quality in two distinct geographies within the city. Some evidence for this is provided in table 2, which shows that average pollutant levels are very different across the two locations. each window it takes the value 1 for all time periods after the extension date, and 0 for periods before it. 15 , t x is the vector of covariates and includes controls for weather; 16 for hour of the day, 17 day of the week, and interactions between these two; and for public holidays and festivals such as Diwali. 18 P(t) is a third-order Our identification strategy is similar to that used in Chen and Whalley (2012) (henceforth CW). 19 CW look at the effect of the introduction of the Taipei Metro (TM) on air quality in Taipei City. While they use the discontinuity arising from the opening of the metro system, we exploit future discontinuities arising from various extensions of the network. Unlike CW, we do not use the first opening of the metro for two reasons. First, we do not have pollution data that 15. We exclude the 24-hour data pertaining to the day of the extension because we do not know the exact hour when the new line became operational. 16. Controls for weather include current and up to 4-hour lags of temperature, relative humidity, wind speed, and rainfall and quartics of both current and 1-hour lags of these weather variables. 17. Figures S1.1a through S1.1c in the supplementary appendix S1 show how the pollutants at ITO vary by hour of the day and by season. Once again we see that pollution is higher in winter compared to summer. Also, there is substantial intraday variation with peak levels reached between 8 pm and 2 am. Similar patterns are observed at Siri Fort. One plausible reason for why pollution peaks during these night hours could be a citywide ban on the entry and movement of heavy goods vehicles (mostly diesel powered trucks) between 6:00 am and 9:00 pm. The substantial intraday variation in pollution calls for inclusion of hour of the day fixed effects in order to improve the precision of our estimates. In most specifications season-fixed effects are not included because of short window lengths, typically nine weeks. Choice of window lengths is discussed later in this section. 18. Diwali is a Hindu festival that falls in winter, typically in October or November. It spreads over several days and is celebrated with an ostentatious bursting of firecrackers. It has been documented that air pollution in Delhi shoots up during and immediately following Diwali (CPCB 2012). It is therefore important to control for this source of pollution. 19. Before Chen and Whalley, Davis (2008) used similar identification to estimate the effect of driving restrictions on air pollution in Mexico City. dates back to the time when the metro was introduced. Second, even if we had this data, it would be incorrect to use opening ridership discontinuity for Delhi. This is because there was an unprecedented jump in metro ridership when it was first opened, a large part of which was due to joy rides. 20 These joy rides are expected to die out as the novelty of the metro fades away. By using discontinuities in ridership that occur a couple of years after the metro first started, we believe that to a large extent we avoid capturing effects arising from one time rides, and the impact that we measure is closer to the steady-state short-term effect.
One of the challenges that we faced in estimating equation (1) is the presence of segments of missing observations in each pollutant series. The last column of table 2 shows the share of missing observations. 21 The best pollutant series is CO at ITO, for which 14 percent of observations are missing.

2.5
PM , which is only recorded at ITO, has 42 percent missing observations. For the RD strategy to be effective, there cannot be too many missing observations around the extension dates. Therefore, to begin with, we restrict our analysis to only those extensions for which there is a symmetric window of at least nine weeks around the date of extension, wherein missing observations in each included week do not exceed 20 percent of the potential observations. 22 Then we look at other window lengths, and finally, for those pollutants with relatively good data, we analyze the entire series. In order to ensure correct inference in the presence of serial correlation in pollution, in all our specifications we use standard errors clustered at 20. "On the first day itself, about 1.2 million people turned up to experience this modern transport system. As the initial section was designed to handle only 0.2 million commuters, long queues of the eager commuters wishing a ride formed at all the six stations . . . Delhi Metro was forced to issue a public appeal in the newspapers asking commuters to defer joy rides as Metro would be there on a permanent basis" (DMRC 2008). 21. We were informed by a CPCB official that missing data could be due to many reasons: power cuts, instrument failure, software malfunction when transferring data to storage device, and disruption in telephone. According to the official, none of these reasons are systematically linked to high or low pollution episodes. Later on, in section IV, we examine whether there is evidence for a systematic pattern to missing observations. 22. At the hourly frequency, the number of potential observations in a week is 168 = 24*7. The 20-percent rule implies that each week in our estimation window has at least 134 = 0.8*168 observations. one week. 23

Plausibility of Identifying Assumption
Identification of the metro effect breaks down if we have not accounted for an event that has a discontinuous effect on air quality. 24 One example is a citywide strike by private bus operators called on the same day as the extension of the metro. If this happens it would be impossible to disentangle the effect of the metro from that of the strike. We have studied the chronology of events in the city and do not find occurrences of such events on any of the extension dates. Here we discuss some of the other likely threats to identification. 25 Government policies aimed at reducing pollution may have an abrupt impact. One such policy, implemented only in Delhi, was the mass conversion of diesel fueled buses to compressed natural gas (CNG). However, this happened in 2001, well before our study period began, and is therefore not problematic. In 2005, Delhi moved from Bharat Stage-II to the stricter Bharat Stage-III emission standards. Although this regulatory change was implemented in the middle of our study period, it is unlikely to have led to a sudden change in pollution. This is because the improved norms are only applicable for vehicles manufactured after the new standards were adopted. Given that new vehicle registrations happen uniformly over time, adoption of stricter emission standards should not lead to a sudden drop in vehicular emissions. 26 We do not know of any other regulatory 23. Although, both Chen and Whalley (2012) and Davis (2008) use standard errors clustered at five weeks, we cluster at one week. This is because our analysis is based on shorter windows of five and nine weeks (due to missing data), while they use two-year horizons. Also, for all pollutants in our data, the auto correlation in daily average pollutant level is less than 0.5 beyond seven days. Clustering at one week should therefore be sufficient. Nonetheless, we re-estimated tables 4 and 5 by clustering at two weeks and found similar results.
24. An event that has a gradual effect on pollution will be captured by the time polynomial P(t) and therefore does not impede our analysis. 25. For them to be problematic, the discontinuous effects do not have to necessarily happen on the extension date. Discontinuous effects arising anywhere within our short windows would be problematic for estimating the correct causal effect of the metro extension. 26. Emission standards in India are adopted in a phased manner with stricter norms first being implemented in major cities, including in Delhi, and then extended to the rest of the country after a few years. Given that inter-state freight that plies through Delhi continues to follow the more relaxed emission standards, the impact of Bharat Stage-III 13 change implemented between 2004 and 2006 that may have had a discontinuous effect on pollution.
Another concern could be that construction activity undertaken to build the new rail lines may have added to localized pollution in the period preceding the metro extension and this would then over-estimate the DM effect. On speaking to officials from the DMRC we were told that such construction activity is typically completed fifteen to thirty days prior to the opening of a new line so as to conduct trial runs to ensure safety of passengers. Therefore, at least for the shorter window lengths, we do not expect this issue to be a problem. Another worry could be that metro officials choose the extension dates in a systematic manner to coincide with either high-or low-pollution days. We think that this is highly unlikely. Given the public enthusiasm for the metro and the recognition of economies of scale in its operation, the DMRC has always been eager to open a new line once it had met all safety requirements.
Finally, Delhi is characterized by a multitude of pollution sources.
According to Guttikunda and Calorie (2012), domestic sources, such as burning of biofuel for cooking and heating, use of diesel generator sets, waste burning, and construction, together account for 20, 19, and 26 percent of collects it as part of the National Air Quality Monitoring Program (NAMP). We use hourly pollution data recorded at two monitoring stations in the city, namely at ITO and at Siri Fort. 27 Both are immobile stations that operate on electricity. They provide comparable data as they were bought from the same manufacturer and followed the same monitoring protocol throughout our study period. Hourly data on weather conditions at Safdarjung, Delhi, were obtained from The National Data Center of the India Meteorological Department. Our choice of study period (2004)(2005)(2006) was dictated by the overlapping period for which we had both pollution and weather data. The Delhi Metro Rail Corporation (DMRC) provided us with data on metro ridership.

IV RESULTS
Before presenting the impact estimates, we investigate whether the data validate a sudden increase in metro ridership at the time of each extension.

Ridership Discontinuities
For each month, figure 3 shows the percentage change in average daily ridership on the DM over the previous month. 28 The exact magnitudes of change are given in the last column of table 1. Except for the introduction of the Yellow Line and the first extension of the Blue Line, the figure shows a significant rise in average daily ridership for the month (or for the following month) 29 of each extension.
27. Under the NAMP there is one other monitoring station located at the Delhi College of Engineering (DCE) in north Delhi. We do not use data from this station because our identifying assumption is unlikely to hold at this location. Compared to ITO and Siri Fort there are many more erratic sources of pollution at DCE. This is because, (a) it is surrounded by Badli, a major industrial township; (b) all along its periphery there are other small scale industrial production units; (c) during our study period the college building itself was undergoing repair and renovation; and (d) DCE is in a mainly rural part of Delhi where sporadic burning of biomass and wood is widespread. 28. Actual daily ridership, instead of average daily ridership in a month, would have been ideal in order to check the sudden increase in ridership at each extension date. However, this data was not available for our study period. 29. For the second extension of the Red Line and the first extension of the Blue Line, we see the surge in ridership in the month following the one in which the extension took place. This is because these extensions were introduced on the last day of the month, and one would therefore expect average ridership to increase only in the following month.
The absence of a significant rise in ridership for the introduction of the Yellow Line (3 percent increase) may be attributed to the fact that it was the first segment of the north-south corridor and also a short segment (3 additional stations) that connected the university to the existing Red Line at a time when the university was closed for the holiday season. Further, besides the university station, the two other stations on this segment are relatively rich neighborhoods where many people may continue to prefer private over public transportation. For the first extension of the Blue Line, the insignificant rise in ridership (5 percent increase) may be attributed to the relatively lower population density in south-west Delhi where the extension took place. 30 Given the necessity of observing a large surge in ridership in order to identify the DM effect, we exclude these two extensions from our analysis.
The largest jump in ridership is seen for the first extension of the Yellow Line (76 percent increase), which connects areas having a high population density (North-East and Central districts) to the hub of government offices in Central Secretariat. A large surge in ridership is also seen for the introduction of the Blue Line (56 percent increase) which is the longest extension among all the extensions considered here.
Given this ridership pattern we expect to see larger effects for the first extension of the Yellow Line and the introduction of the Blue Line. We also expect larger effects at ITO than at Siri Fort because of its relative proximity to the line expansions and also because it is a major traffic intersection whereas Siri Fort is mainly residential.

Impact Estimates: Short Windows
In order to estimate equation (1)  missing values, table 3 shows the maximum window length (in weeks) around each extension on applying the at-most-20-percent-missing-data criteria for each included week and also subjecting the selection to a minimum window length of five weeks. As an example, if we restrict ourselves to good quality data, we are able to examine the effect of the second extension of the Blue Line on 2 NO at ITO using a maximum window length of only thirteen weeks. Of the four extensions characterized by a significant increase in ridership, we are unable to examine the effects of the second extension of the Red Line because of lots of missing observations around its opening date.
For all extensions with at least nine weeks of good quality data, 31 figures S1.2a through S1.2h in the supplementary appendix S1, try to visually examine whether there is a break in a pollutant series at the extension date using a nineweek window around the date. 32 These plots suggest that there is a drop in pollutant level at the time of each extension, and in some cases this drop is large.
Next, we estimate equation (1) to arrive at quantitative estimates. Table 4 shows the results from an estimation of equation (1)  PM at ITO we only have a seven-week window of good quality data, and therefore we look at this shorter window of seven weeks for it. 32. The scatter points are daily averages of residuals obtained from a regression of (log) hourly pollutant on all the right-hand side variables in equation (1)  The overlaid curve depicts the fitted residuals from a regression of the scatter points on the extension discontinuity and a third order time polynomial. 33. When calculating the percentage change we apply the correction suggested by Kennedy (1981) in the context of interpreting the coefficient on a dummy variable in a semi logarithmic equation.
at ITO. The introduction of the Blue Line resulted in a 31 percent decrease in the level of 2 NO at ITO. Its effect on CO at ITO could not be analyzed because of missing data. We had expected the second extension of the Blue Line to lead to smaller declines, and we find that it did not lead to statistically significant reductions in any of the three pollutants. Turning to the effects at Siri Fort, we were only able to examine the second extension of the Blue Line. Our analysis shows that just as for ITO, this extension did not lead to a statistically significant decrease in either 2 NO or CO at Siri Fort. It is to be noted that even where an effect is not statistically significant, its sign is always negative and in some cases the magnitude is not insignificant. Table 5 shows the impacts using a shorter window of five weeks. Compared to the nine-week window, although the magnitude of impact of the first extension of the Yellow Line on 2 NO at ITO is larger, it is still not statistically significant.
The effect on CO at ITO for this extension has increased to 78 percent. Also for the introduction of the Blue Line, the effect on 2 NO at ITO has increased to 55 percent.
Restricting to a shorter window enables us to study the effects of this extension on CO and on 2.5 PM : at ITO it led to a decrease of 56 and 53 percent, respectively.
The results for the second extension of the Blue Line at ITO and Siri Fort are similar to those seen in table 4 and continue to remain statistically insignificant.
As seen in table 3 there are segments of good quality data that span longer than nine weeks. In table 6 we extend the window beyond nine weeks whenever the data permit us to do so. In most cases there is a decrease in magnitude of impact, and none of the effects are significant now. We provide two plausible explanations for the transitory nature of our impact estimates.
One explanation is that some of the sporadic and mobile sources of pollution that characterize Delhi's pollution inventory get captured when we extend the window, and this masks the impacts for longer time periods. Admittedly, this may also happen for shorter windows, and it could even explain the very large magnitudes for some of the estimates seen in tables 4 and 5. 34 As discussed in section II, these sporadic and mobile sources of pollution pose a threat to our identification strategy. However, the fact that when we look at shorter time periods we consistently get negative estimates (in table 4 all estimates are negative, and in table 5, all except one (which is close to zero), are negative), makes us believe that some of the larger extensions did reduce specific transportation source pollutants even if the exact magnitudes of reduction may not be those reported in tables 4 and 5.
Another explanation for the disappearance of effects in table 6 could be that the traffic diversion effects are indeed transitory and over longer time horizons the DM has no impact on pollution. Duranton and Turner (2011) provide evidence in support of this argument. They find that in cities in the United States, increase in road-building and provision of public transport have no impact on vehiclekilometers-traveled. They reason that reduced congestion on roads, experienced soon after new roads are built, has a feedback effect that induces existing residents to drive more. If this is true for Delhi, then it is possible that soon after the larger extensions were initiated, the DM diverted private traffic which lowered pollution (as seen in tables 4 and 5 using shorter windows) and also reduced road congestion. These reductions in turn incentivized the remaining drivers to drive more, and may have also added some new drivers, thus wiping out the initial 34. Additionally, it could also explain some of the inconsistencies in our impact estimates. As pointed out earlier in this section, we had expected the first extension of the Yellow Line to lead to a reduction in both 2 NO and CO.
However, while we see a substantial reduction in CO, the effect on 2 NO is not statistically significant (see tables 4 and 5). There is also a discrepancy across extensions in the sense that while the magnitude of effects on 2 NO and CO at ITO are comparable for the introduction of the Blue Line, for the first extension of the Yellow Line these are quite disproportionate (see table 5). effects on pollution (as seen in table 6) and on road congestion. This explanation is along the lines of the traffic creation effect discussed earlier. Unfortunately, our data and empirical strategy do not allow us to discern with surety which of these explanations is true. However, our subsequent analysis using data for the entire study period suggests that the effects may not be transitory.

Robustness Checks
Here we present some robustness checks for the results seen in table 4. We also present results from a new specification that uses all the data for our study period, ignoring the fact that there are missing observations. For reasons stated below, this is our preferred specification.
Varying the Order of the Time Polynomial: Following CW, we have used a third order time polynomial. However, in specifications similar to ours, Davis (2008) uses a seventh order polynomial. In order to check that our main results are robust to the choice of polynomial order, we ran the regressions in table 4 using a fourth through seventh order polynomial and found similar estimates. The results for the fourth and the seventh order specifications are shown in tables S1.2a and S1.2b in the supplementary appendix S1. 35 Accounting for Persistence in Pollution: Next we examine whether controlling for lagged pollution alters the estimated impacts. We estimate the equation shown below (l, m superscripts have been suppressed): The results are shown in tables S1.3a and S1.3b in the supplementary appendix S1.
As expected, when we account for persistence, the instantaneous impacts are lower. 36 The two instances in table 4 where we had seen significant drops in 35. We also tried interacting the time polynomial with the discontinuity dummy, m t DM . We do not report these results as, in all cases, the discontinuity dummy drops out due to multicollinearity in our dataset. 36. The instantaneous impact is 1  , and the cumulative effect is calculated by iteratively substituting for y in the pollution continue to remain significant, although the magnitude of the cumulative effects is lower.
Artificial Discontinuities: For 2 NO and CO, we implemented the standard placebo test of using data only from the pretreatment period (before the extension is made) and introducing artificial discontinuities within this period. In some cases we did find significant effects. In the same vein, we estimated equation (1) using the introduction of the Yellow Line, which is characterized by an insignificant jump in ridership (see earlier discussion under "Ridership Discontinuities"). For this expansion, we had expected to find insignificant effects, but instead we found significant declines in both 2 NO and CO at Siri Fort. These perverse results could be due to sporadic and mobile sources of pollution in Delhi. While we admit that this is a potential threat to identification, the fact that we consistently get negative effects for all discontinuities in tables 4 and 5 and that, in most cases, these effects are larger when we reduce the time window, provides strong support that DM did reduce pollution. Below, we provide additional evidence to support this claim.  (Sillman, 1999 Region of Delhi (marked as triangles in figure 2) during our study period.

Effects on Nontransportation Source Pollutants
According to Gurjar et al., 2004, andCalorie, 2012, 68 and55 percent, respectively, of 2 SO in Delhi is generated by these power plants. Figure   S1.3 in the supplementary appendix S1 shows wide variation in the monthly power production of these plants. More importantly, it shows that monthly output was either high or rising when the two extensions were made. Since 2 SO emissions are strongly correlated with power produced, there will be corresponding variations in 2 SO . Thus, it is possible that it is the coal plants that are behind the significant impacts seen in table S1.4. In any case, that the two metro expansions were accompanied by an increase in 2 SO increases the credibility that these expansions led to reductions in some transportations source pollutants (as seen in The variables are similarly defined as in equation (1), and as above, t x now includes season fixed effects. Because we look at the entire series we include all extensions shown in table 3 for ITO and this is represented by the set of discontinuity dummies, {DMi}. We expect that the sporadic sources of pollution are likely to be evenly spread within a long window of three years and would therefore cancel each other. This is therefore our preferred specification. Figure S1.4 in the supplementary appendix S1 (similarly constructed as figures S1.2a through S1.2h), visually presents the effects of multiple extensions on CO at ITO. The impact magnitudes are presented in table 7 which shows that the first extension of the Yellow Line, characterized by the largest increase in ridership, led to a 33.5 percent reduction in CO at ITO, while the other two extensions did not lead to statistically significant reductions. Again, it is reassuring that all point estimates are negative even if some are not statistically significant. 38 We also carry out similar analysis for 2 NO at ITO, which has 18 percent 38. To check whether the result is robust to choice of polynomial order, we ran the regression using fourth through seventh order polynomial in time. The main result does not change: the first extension of the Yellow Line is the only one that is significant, with estimate values of -32.9 and -39.3 for the fourth and seventh order polynomial, respectively. We also estimate equation (2) using the entire series for 2 SO at ITO which has 16 percent missing observations. The only significant discontinuity was the introduction of the Blue Line for which the sign was positive. Thus, the result in table 7 is also robust to the check using nontransportation source pollutant 2 SO . missing observations for the entire series. Figure S1.5a in supplementary appendix S1 (counterpart of figure S1.4 for CO), seems to suggest that the first extension of the Yellow Line and the second extension of the Blue Line led to an increase in 2 NO at ITO. However, looking at the plot, it seems that, unlike CO, the simple time trend does not fully capture the systematic changes for 2 NO , and therefore we interact it with the discontinuity dummies. Figure S1.5b presents the new picture.
The effects seen in the previous plot disappear. If anything, the introduction of the

25
increase was smaller for Delhi compared to Mumbai. This is suggestive of a traffic diversion effect being operative in Delhi after the metro was introduced.

V DISCUSSION OF RESULTS
We start by summarizing our main results and then discuss our estimates in the context of other studies, especially the one by CW.

Summary of Findings
When we look at a short span of nine weeks, we find that the first extension of the Yellow Line, characterized by the largest surge in ridership, led to a 69 percent reduction in CO at ITO. When we extend this span to forty one weeks, the effect size reduces and is no longer statistically significant. However, when we consider our entire study period, 2004-2006, we find that this extension resulted in a 34 percent decline in CO at ITO. The fact that we find a decline when we look at the whole series (our preferred specification) suggests that the effect is not transitory.
The introduction of the Blue Line, the longest extension considered here, led to a 31 percent reduction in 2 NO at ITO when we look at a nine week window, and the effect remains when we consider the entire study period, which once again suggests that this is not a transient effect. Using a five week window, there is some evidence that the introduction of the Blue Line also led to a decline in 2.5

PM at
ITO, but we could not carry out the analysis for the entire study period due to a large number of missing observations. Finally, we do not find any significant effects at Siri Fort which is mainly a residential area, and relative to ITO, is further away from the extensions considered here.
In supplementary appendix S2, we compare our estimate of a 34 percent reduction in CO with the one in Doll and Balaban, 2013. Doll and Balaban also study the effect of the DM on air pollution, but they use a different methodology to 26 do so. The comparison reveals that our impact estimate is many times higher than theirs. We refrain from commenting on the reasons for this difference beyond what we have stated in this appendix.
Comparison with Chen and Whalley (2012) CW estimate the impact of introduction of the Taipei Metro (TM), in 1996, on pollution in Taipei City. Using a two-year window of very good quality data (they had 1 missing observation in a two year window), they find that the opening of the TM resulted in a 15 percent decline in CO. Using a three year window and spanning multiple extensions, we find a much larger impact of 34 percent for Delhi. 40 One reason for why our impact is larger could be that CO measurements in our study are from a single monitoring station located at a major traffic intersection, whereas CW use the average CO across ten monitoring stations and they exclude the few stations located at traffic intersections (see footnote 20 on page 15 in CW). If traffic diversion is the main mechanism via which the metro impacts pollution, then one might expect to see bigger effects at traffic intersections.
Next, we compare the implied ridership-pollution elasticities across the two studies. From not having any metro in the city, CW report an average daily ridership of 40,410 in year following the TM introduction. Using data from the DMRC, we note that average daily ridership before and after the first extension of

Benefits from Reduced CO Pollution
Drawing from literature that looks at health impacts of air pollution, we quantify the benefits of reduced CO levels in Delhi in terms of infant lives saved. Currie, Neidell, and Schmieder (2009)  the extension. 43 Viscusi (2008) argues that less developed countries tend to have a lower value of statistical life (VSL), consistent with their lower levels of income.
He calculates a VSL of 1.2-1.5 million USD (2000 prices) for India, and 0.2-0.9 for Taiwan, as against 7 for the U.S. Using the lower bounds of VSL estimates for India and Taiwan the infant lives saved may be valued at 73.2 and 0.34 million USD (2000 prices), respectively. Given that baseline pollution levels in Delhi are higher than in Taipei, and perhaps, avoidance behavior is much less in Delhi due to lack of pollution related warnings, we conjecture that this figure is a lower bound for Delhi.

How Viable is the Delhi Metro?
Winston and Maheshri (2007)  population and high rates of economic growth, and this should alleviate concerns around low ridership demand. 44 As of 2012, the total route length of the DM was 190 kilometers, and its ridership was about 2 million per day (DMRC 2012). In contrast, the San Francisco/Oakland area, has a population density that is slightly more than one fourth that of Delhi's, and in 2014, the BART's route length and average daily ridership was 167 kilometers and 0.4 million, respectively. 45 Moreover, Delhi is a polycentric city and the route design of the DM is ideal to serve daily commuters. It has a radial track layout, with major north-south and east-west corridors connecting different parts of Delhi to the border cities such as Gurgaon and Noida, which are the new centers of employment. Also, ridership is likely to increase further as new routes are completed. According to Litman (2014) rail transit is more appropriate in areas where development is more compact and noise and air pollution are serious considerations, while buses are more appropriate where travel is more dispersed. It seems to us that Delhi meets the criteria that   Murty et al. (2006) carry out a social cost benefit analysis of phases I and II of the DM, covering a track length of 108 kilometers. Considering the estimates of financial flows during the period 1995-2041, they estimate the financial benefit-cost ratio to be between 1.92 and 2.30. They also estimate the capital costs of phases I and II to be 64,060 million and 80,260 million, respectively, and the net present social benefit from both phases to be 419.98 billion INR (in 2004-05 prices). Their calculations account for differences in shadow and market prices of unskilled labor; premiums for importing fuel; benefits accruing from reduced road congestion, accidents, and air pollution; and effects of redistribution of income among stakeholders. Their estimates are based on several assumptions regarding annual flows of costs and benefits during the entire lifetime of the DM. Evaluating the accuracy of these assumptions is beyond the scope of our study. the DM, there was a significant reduction in carbon monoxide at a major traffic intersection in central Delhi. There is also suggestive evidence that the introduction of the Blue Line led to a decline in nitrogen dioxide at the same intersection.
Although we find a favorable impact of the DM on Delhi's pollution soon after some of the larger metro expansions, the overall impact of the metro system on air pollution crucially depends on how the electricity needed to drive the metro is generated. If it is not cleanly generated, then some or all of the benefits from reduced pollution in Delhi may be offset by increased pollution elsewhere. 47 Also, our impact estimates are for the short run, and quantifying the net social benefit of the metro network in the long run is beyond the scope of our analysis.
Much of our analysis was restricted by the poor quality of data on pollution.
A longer time series, with fewer missing observations, would have allowed us to draw more definitive conclusions. Given the severity of Delhi's pollution problem, we would urge the environment ministries at the state and the center to invest in better technology and equipment in order to record pollution levels more accurately and completely. As our analysis reveals, it is difficult to conduct rigorous impact assessments without good quality data.
Our paper also highlighted the nature of pollution in Delhi by citing several inventory studies. We believe that a multipronged strategy needs to be adopted if Delhi wants to shed its distinction of being the most polluted city in the world (based on WHO's Pollution Database: Ambient Air Pollution 2014). As potential pollution abatement measures, Delhi's planners should consider the following: increasing accessibility to metro stations by improving feeder systems, promoting 47. The DM has a regenerative braking system on its rolling stock, which generates electricity when brakes are applied and then feeds it back into the system. This way, almost 35 percent of the electricity consumed is regenerated by the system (Sreedharan 2009). Doll and Balaban (2013) provide an excellent analysis of the overall carbon footprint of the DM. For the year 2011, they estimate that the DM saved 232,162 tons of 2 CO because of its regenerative braking technology.

cycle-rickshaws to deliver last mile connections, extending the Delhi Bus Rapid
Transit System (BRTS) and ensuring strict enforcement of existing bus corridors, using congestion pricing and citywide parking policies to dissuade use of private vehicles, adopting uniform emissions standards throughout the country, constructing a bypass road around Delhi to eliminate interstate freight traffic, designing efficient waste disposal systems to prevent sporadic burning of garbage and foliage, shutting down the remaining coal based power plants in the city, and educating Delhi's residents of the severity of the problem to enable them to take more informed decisions.
Finally, before investing in a capital intensive rail network, it is imperative that a proper cost benefit analysis be undertaken. A specific consideration that needs to be made is its desirability vis-à-vis a bus transit system (BTS). In the context of the United States, Winston andMaheshri (2007) andO'Toole (2010) claim that the less capital intensive BTS is more suitable for most US cities where there isn't enough demand for the metro to be able to recoup its high capital costs.
Litman (2014) makes contentions in favour of the metro, stating that while it is more capital intensive, it has lower operating costs per passenger mile. Both sides, however, agree that metro systems are best suited in areas with high population density, characterized by high demand for travel. Today, governments at various levels in India are planning to build metro rail systems in several tier II cities.
These cities have lower population densities compared to tier I cities such as Delhi.
Governments, at times, invest in capital intensive projects like the metro to emulate more advanced economies or to pander to private firms who build and design these systems. We caution them against these temptations. As many cities in the United           (1) (1) (13) Source: Authors' analysis based on data described in the text.
Missing observations in each included week do not exceed 20 percent. Each estimate is calculated from a separate regression (equation (1)  Std. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level.   Source: Authors' analysis based on data described in the text.
1. Estimates for NO2 and CO pertain to a nine week window, and for PM 2.5 to a 59 seven week window.
Std. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. Source: Authors' analysis based on data described in the text. Source: Authors' analysis based on data described in the text.
Estimates for NO2 and CO pertain to a nine week window, and for PM 2.5 to a seven week window. Std. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. Source: Authors' analysis based on data described in the text.
For both pollutants results pertain to a nine week window.
Std. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level.

Observations 1622
Source: Authors' analysis based on data described in the text.
Results pertain to a nine week window.
Std. errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. Source: Authors' analysis based on data provided by CPCB.
Shares less than 10 percent highlighted in bold Source: Authors' analysis based on data described in the text.
The dependent variable is an indicator of whether the observation is missing (=1 if missing and =0 otherwise). The explanatory variables are the same as in Table 4 with the addition of season fixed effects. Std. Errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. Source: Authors' analysis based on data described in the text.
The dependent variable is an indicator of whether the observation is missing (=1 if missing and =0 otherwise). The explanatory variables are the same as in Table 4 with the addition of season fixed effects. Std. Errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. Source: Authors' analysis based on data described in the text.
The dependent variable is an indicator of whether PM2.5 observation is missing, (=1 if missing and =0 otherwise). The explanatory variables are the same as in Table 4, except that we drop all discontinuity dummies and the time polynomial, and include an indicator for whether CO measure is high (=1 if high and =0 otherwise) and also include season fixed effects. Std. Errors are clustered at one week. * indicates significantly different from zero at 10 percent level, ** at 5 percent level, and *** at 1 percent level. ii) The two methodologies are very different and each has some limitations.
While our RD identification strategy is not robust to the presence of sporadic and mobile sources of pollution, our data are actual measurements on pollution and weather obtained from monitoring stations located in Delhi. Their method relies on estimating travel sector parameters using data from multiple studies, and sometimes relying on estimates for other cities (e.g. their occupancy rates for cars is taken from average vehicle occupancy for Asian countries