Public Information and Avoidance Behavior: Do People Respond to Smog Alerts?

Matthew Neidell 1 Columbia University October 2006

Abstract: This paper examines short-run responses to information about risk by examining the effect of smog alerts on daily attendance at three outdoor facilities in Southern California. To identify the effect of smog alerts, I employ a regression discontinuity design by exploiting the deterministic selection rule used for issuing smog alerts. I find significant and robust evidence that people respond to smog alerts: on days when smog alerts are announced, attendance declines 2 to 11 percent. As smog alerts become increasingly frequent, however, people decrease their response, suggesting decreasing returns to substitute activities. (JEL D80, I18, Q53) Keywords: pollution, avoidance behavior, information, risk, regression discontinuity

1

I thank Janet Currie, Michael Greenstone, Ken Chay, Enrico Moretti, Sherry Glied, Helen Levy, Will Manning, Tomas Philipson, Paul Rathouz, Bob Kaestner, Kerry Smith, Sylvia Brandt, Elizabeth Powers, Michael Khoo, Amanda Lang, Pat Bayer, Wes Hartmann and numerous seminar participants for valuable feedback. I am also very grateful to Mei Kwan, E.C. Krupp, Ken Warren, Jim Bauml, Bruce Selik, Patricia Florez and Joe Cassmassi for help with assembling the data used for this analysis, and Sarah Kishinevsky, Mike Kraft, and Sonalini Singh for excellent research assistance. Financial support from the University of Chicago’s Center for Integrating Statistical and Environmental Science is gratefully acknowledged.

Information is a powerful policy tool. A wealth of evidence spanning various disciplines supports the use of information as a public health intervention. Studies on topics as diverse as nutrition (Ippolito and Mathios (1990)), hygiene (Jin and Leslie (2003)), chemical hazards (Viscusi et al. (1986), Smith and Johnson (1988)), disease prevalence (Mullahy (1999), Philipson (1996)), physical activity (Bauman et al. (2001)), and medical procedures (Domenighetti et al. (1988)) consistently document that individuals adjust their behavior in response to information about permanent, long-standing risk. Additionally, this information is often disseminated to shape individuals’ short-run behavior in response to imminent dangers, such as terrorism threats, environmental hazards, and disease outbreaks, but existing studies do not indicate whether people respond to such rapidly changing information. This paper examines short-run responses to public information about risk by examining the impact of air quality episodes, or “smog alerts”, on daily outdoor activities in Southern California. Smog alerts are designed to warn individuals about high levels of ground-level ozone 2 , a pollutant linked to numerous health conditions, such as asthma, pneumonia, and bronchitis. Smog alert status is a dichotomous variable updated on a daily basis. On the days alerts occur, they are widely disseminated throughout the region via various media channels, such as television and radio. Similar warning systems are now commonplace throughout the U.S. and Europe, so it is important to understand whether people respond to this information about risk. To identify the effect of smog alerts, I employ a regression discontinuity design. Smog alerts are only issued when ambient ozone is forecasted to exceed a particular threshold. If days

2

Ground level ozone is distinct from stratospheric ozone (the “ozone layer”), which affects the amount of ultraviolet radiation that reaches the earth.

1

just above or below this threshold do not vary systematically with outdoor activity decisions, then I can obtain estimates of the causal effect of alerts on avoidance behavior. In support of this approach, other observable characteristics, such as weather and observed pollution, move smoothly around this threshold, suggesting that any change in outdoor activities at this threshold can be directly attributed to smog alerts. Although smog alerts are the main device for conveying information about air quality, the continuous index used for determining smog alert status is also available. But obtaining it involves greater costs. For the time period examined, it is only disseminated in major newspapers 3 , and far fewer people read the newspaper than watch television or listen to the radio. 4 For example, the Los Angeles Times, the most widely distributed newspaper in the region, circulates to only 7% of the population. 5 Therefore, considerably fewer people obtain this information than broadcasted smog alerts. Conditional on the continuous index being available, this paper assesses whether people respond to the more simplistic and readily available information contained in alerts. As a measure of daily activities, I use unique data specifically gathered for this analysis. These are daily attendance from 1989 to 1997 at three major outdoor facilities in Southern California (the Los Angeles Zoo, Griffith Park Observatory, and the Los Angeles County Arboretum) and one major indoor facility (the Natural History Museum). These data consist of administrative records, which are available at a daily level and are not subject to a recall bias that

3

It is now available via the internet or e-mail, so obtaining it may be less costly. In the 1970 Census IPUMS, the rate of households with television and radio in California was 96% and 80%, respectively. The current rate is likely to be higher. 5 The numerator (the average weekly paid total circulation of the Times as of 3/31/2006) is 851,832, according to the Audit Bureau of Circulations (available from http://www.accessabc.com/freereports.htm). The denominator (number of households in Los Angeles and Orange counties) is 12,923,547, according to the 2000 census of population and housing. This is an overestimate of the circulation rate because some households in other counties in the region receiving the Times are counted in the numerator but not the denominator. 4

2

may be present in survey data. Although using data from these sources limits the generalizability of this analysis, conventional survey data do not allow testing of short-run responses. The results provide considerable evidence that people respond to smog alerts. Attendance is significantly lower on days when smog alerts are announced, with declines ranging from 2 to 11 percent across the three facilities considered. These results are generally insensitive to functional form assumptions of the regression discontinuity design and are robust to several specification checks. For example, the estimates are remarkably unaffected by the inclusion of numerous controls for weather conditions and realized air quality, both significant predictors of outdoor activities. Responses are larger for local residents, who are more likely to process this information and have lower costs of substituting activities than non-locals. Attendance at the indoor activity increases when alerts are issued, suggesting people substitute from outdoor to indoor activities. Attendance for children and the elderly, two groups specifically targeted by air quality information, decreases by roughly 20% in response to alerts. These findings indicate that people value the provision of information contained in the warnings. As alerts become increasingly frequent, people may be less likely to respond because of decreasing returns to substitute activities. In support of this, I find that when smog alerts are issued on consecutive days, individuals generally decrease their response on the second day. This implies information decision rules that incorporate long-run conditions might improve the effectiveness of these public good. Therefore, terror alerts that are constantly “code red” may have little effect on public responsiveness. The rest of the paper proceeds as follows. The following section provides background information on air quality. Section 2 describes the theory for this analysis. Section 3 describes

3

the data and section 4 presents the empirical strategy. Section 5 shows the results, and section 6 concludes. 1. Background Information on Air Quality Ground-level ozone, both its 1-hour and 8-hour concentration, is a criteria pollutant regulated under the Clean Air Acts. 6 A long line of evidence links ozone with many short-run health conditions. Ozone is believed to irritate lung airways and increase susceptibility to respiratory related health conditions such as asthma. Symptoms can occur in as quickly as one hour of exposure, with normal lung functioning typically returning within 24 hours (U.S. EPA (2003)). Therefore, altering short-run exposure can substantially reduce the onset of symptoms. The process leading to ozone formation makes it highly predictable and straightforward to avoid. Ozone is not directly emitted into the atmosphere, but is formed from interactions of nitrogen oxides and volatile organic compounds (both of which are directly emitted) in the presence of heat and sunlight. Ozone formation also increases with solar radiation. Because of this process, ozone levels vary considerably both across and within days, tending to peak in the summer and middle of the day when heat, sunlight, and/or solar radiation are at their maximum (U.S. EPA (2003)). Therefore, ozone levels are predicted using weather forecasts and ozone rapidly breaks down indoors because there is less heat and/or sunlight 7 -- so people can avoid exposure by going indoors. To inform the public of local air quality, the U.S. Environmental Protection Agency developed the pollutant standards index (PSI). 8 The PSI ranges from 0-500 and is indexed so that a value of 100 corresponds to the National Ambient Air Quality Standards as set forth in the

6

Criteria pollutants are considered those most responsible for urban pollution. For empirical evidence on the low correlation between indoor and outdoor ozone, see Chang et al. (2000). 8 The PSI, which was replaced by the Air Quality Index in 1999, is also available for other criteria pollutants believed to affect health, such as particulate matter and carbon monoxide. 7

4

Clean Air Acts. The PSI as reported in the Los Angeles Times contains a brief legend to summarize the health effects: 0-50 good; 51-100 moderate; 101-200 unhealthful; 201-275 very unhealthful; and 275+ hazardous. In order to provide ample notification for the public to react, the PSI is typically forecast one day in advance. Major newspapers are required to report this information, usually in the weather section (U.S. EPA (1999)). In addition to providing the ozone forecast, California state law requires the announcement of an air quality episode when that forecast equals or exceeds 200 on the PSI. 9 These episodes are more widely publicized than the ozone forecast; they are announced on both television and radio. When an episode occurs, susceptible members of the population – those with a history of respiratory illness or part of a more vulnerable segment of the population, such as children or the elderly – are encouraged to remain indoors and shift outdoor activities to night time. All other members of the population are encouraged to avoid rigorous outdoor activity during the day. Furthermore, schools are directly contacted and instructed to reschedule or cancel outdoor activities, such as physical education classes, recess, and sports practices. The public is also encouraged to minimize their contribution to pollution by ride sharing, for example, although no financial incentives are offered for doing so. Although air quality episodes can be issued for any criteria pollutant, they have only been issued for ozone. Furthermore, because ozone is a major component of urban smog, this has given rise to the term “smog alerts.” 10 The agency responsible for providing air quality forecasts and issuing smog alerts for Southern California is the South Coast Air Quality Management District (SCAQMD). Because

9

200 PSI corresponds with 0.20 parts per million (ppm). Additionally, a stage II air quality episode is issued when the ozone forecast exceeds 250 PSI or 0.30 ppm, but this only occurred once over the time period studied. 10 While these alerts are offered on a statewide basis, they typically only occur in Southern California because of its exceptionally high levels of ozone.

5

SCAQMD covers all of Orange county and the most populated parts of Los Angeles, Riverside, and San Bernardino counties, an area with considerable spatial variation in ozone, a separate forecast is provided for each of the 38 source receptor areas (SRAs) within SCAQMD. They produce an air quality forecast by noon the day before in order to leave enough time to disseminate the information. When an alert is issued, the staff at SCAQMD directly contacts a set list of recipients, including local schools and newspapers. The media then further circulate the information to the public, but in a greatly condensed form. For example, the Los Angeles Times provides air quality forecasts for only 10 air monitoring areas (AMAs) in SCAQMD by taking the maximum forecasted value of the SRAs within an AMA. Given the reporting process and the factors believed to affect ozone formation, the model used for issuing an alert can be summarized as: (1)

alat = 1{maxat(ozfst = f(wfst,ozst-1,srt)}≥0.20}

where the subscripts a, s and t indicate AMA, SRA, and date, respectively, al is an alert, ozf is the forecasted 1-hour level of ozone, w f is the weather forecast, oz is observed 1-hour ozone, sr is solar radiation, and 1{•} is an indicator function equal to 1 when the forecasted ozone exceeds 0.20 ppm and 0 otherwise. Alerts for ozone are only issued from March through October, compatible with the seasonal patterns of ozone. Despite the fact that these air quality forecasts and alerts have been around since the late 1970s, there is no published evidence on whether people respond to them. More broadly, there is limited empirical evidence on the existence of avoidance behavior with respect to air pollution. Bresnahan et al. (1997) found that people spent less time outdoors when air pollution levels rose. Their study relied on survey data, which is potentially subject to recall bias. Furthermore, it looked at responses to actual pollution levels rather than information about pollution. Therefore,

6

it is unclear whether people reduced their time outside because of avoidance behavior or because of health symptoms from exposure to the elevated pollution levels. 11 To overcome these concerns, this paper uses daily administrative data on attendance at various localities to directly test whether people respond to information about pollution. 2. Theory Individuals may substitute between indoor and outdoor activities because they believe exposure to outdoor pollution affects health and because of the direct utility they receive from engaging in these activities. A decrease in outdoor activities in response to information about pollution is an increase in avoidance behavior. To understand the implications of air quality information on this behavior, I use a simplified version of the model developed by Bresnahan et al. (1997) and extend it to include information about pollution. 12 Assume individuals maximize utility defined over consumption (C) and outdoor activities (O). People process information about ozone according to: (2)

Ik = ω1k·al + ω2k ·ozf(Q)

where I is information about ozone (i.e., expected ozone levels), al is smog alerts, and ozf is the continuous ozone forecast. Q is the effort involved in acquiring information. Since smog alerts are widely disseminated, assume no effort is required to obtain that information. Acquiring information on the continuous ozone forecast requires effort since it is less widely disseminated. 13 Consistent with standard Bayesian learning models (Viscusi et al. (1986), Smith and Johnson (1988)), the ω’s are the weights people place on alerts and forecasted ozone (ω1k,ω2k ≥0, ω1k+ω2k=1). The subscript k indicates heterogeneity in individual’s knowledge of 11

In related work (Neidell (2004)), I found that smog alerts lowered hospital admissions for asthma. That study used a monthly measure of smog alerts, which is potentially correlated with other factors related to ozone and health, and did not provide direct evidence that people responded to the alerts. 12 The main simplifications are that pollution and health do not enter directly into the utility function.

7

pollution levels. To remain consistent with the EPA’s targeting of two distinct groups, I assume two types of people: susceptible (s) and unsusceptible (u). Accordingly, the health effects, and therefore health care expenditures, are greater for susceptible than unsusceptible people for a given level of expected ozone (δH/δIs > δH/δIu). Short-term health is affected by O and I, with heterogeneity of effects determined by susceptibility. 14 Individuals’ exogenously determined incomes are spent on consumption at price pC, outdoor activities at price pO, medical expenses on health ((G(H)), and the cost of efforts to acquire information about pollution at price pQ. People maximize utility by choosing levels of C, O, and Q. The ratio of first order conditions for optimization for C and O yield: (3)

∂U / ∂C pC = O . ∂U / ∂O p + ∂G / ∂H ⋅ ∂H / ∂O

People spend less time outside (increase avoidance behavior) when expected ozone increases if avoidance behavior is more productive when expected pollution increases (δ2H/δOδI < 0). In other words, more time outside is worse for health when ozone is high. This condition seems likely to hold because it is precisely what the PSI and smog alerts convey (and because indoor ozone levels are typically uncorrelated with outdoor levels). An additional insight from this model is that if people are rational Bayesian updaters, then susceptible people are less likely to respond to an alert than unsusceptible people. This arises because the ozone forecast that determines smog alert status is potentially observable. If susceptible people obtain the ozone forecast, then smog alerts offer no additional information (ω1s=0,ω2s=1). For example, if an individual discovers over time that he is susceptible to the effects of ozone when it reaches 0.18 ppm, then he obtains the continuous ozone forecast Alternatively, Q can represent the effort obtaining the continuous forecast relative to alerts, with Q > 1. Another distinction in this model from Bresnahan et al. (1997) is that I only focus on short-run health effects from pollution exposure, such as respiratory illnesses. 13 14

8

provided in the newspaper. 15 Therefore, issuing a smog alert when ozone is expected to reach 0.20 ppm offers no additional information, so he does not respond to an alert. On the other hand, unsusceptible individuals do not obtain the ozone forecast because there are no benefits from avoiding ozone levels less than 0.20 ppm and the costs of acquiring the information are positive. Because the acquisition costs for smog alerts are zero, smog alerts are the only piece of information they possess about air quality. Since unsusceptible people are potentially affected by ozone exposure above 0.20 ppm, as the warning accompanying smog alerts implies, then they respond to alerts. Therefore, susceptible people are less likely to respond to an alert than unsusceptible people. 16 To derive this prediction, start with the first order condition for Q, which is pQ – δG/δH·δH/δQ ≥ 0. If the expected health benefits from reducing exposure when forecasted ozone is less than 0.20 ppm are zero (unsusceptible), then this condition is not binding, and Qu=0. Therefore, smog alerts are the only piece of information they possess about ozone levels (ω1u=1,ω2u=0), so Iu = al. Based on the first order conditions in equation (3), avoidance behavior is increasing in smog alerts (δOu/δal ≥ 0). On the other hand, the susceptible benefit from information about pollution when forecasted ozone is less than 0.20 ppm, so they are more likely to exert effort in obtaining information (Qs>0). Therefore, they obtain the ozone forecast from the newspaper, so smog alerts offer no additional information (ω1s=0,ω2s=1) and Is = ozf. Therefore, avoidance behavior is unaffected by smog alerts (δOs/δal = 0). 3. Data A. Outdoor activities

15 16

This assumes the cost of acquiring the PSI is less than the health costs from exposure. Note that this may not hold for individuals who do not know their susceptibility.

9

For a measure of time spent outdoors, the dependent variable of interest, accurately recorded individual level time diaries would be an ideal source of data. Because such data are generally unavailable at a daily level over a sufficient period of time, I instead use daily aggregate measures of attendance at three outdoor facilities within the boundaries of the SCAQMD. If time spent outdoors and attendance at these facilities are positively correlated, then aggregate attendance is a valid construct for the dependent variable. The three distinct outdoor attractions from which data are collected are the Los Angeles Zoo and Botanical Gardens (“Zoo”), Griffith Park Observatory (“Observatory”), and the Los Angeles County Arboretum and Botanic Gardens (“Arboretum”). 17 Descriptive statistics for each are shown in Table 1. 18 Although focusing on three specific facilities limits the generalizability of this analysis, these data provide several advantages over time use surveys. One, because they are administrative data, they are likely to be free of recall errors that arise in survey data. Two, these data are available over a long period of time in which there is substantial variation in ozone levels, forecasts, and smog alerts. Three, the exact dates are available in the attendance data. This allows me to explicitly test short-run responses to information and to use the regression discontinuity design. Therefore, this approach improves upon measurement, precision, and causality at the expense of generalizability. Total attendance data, available from 1989-1997 for the Zoo and Observatory and 19901997 for the Arboretum, are collected using different techniques at each place. The Zoo charges an admission fee, and the register is linked to an automated system that tracks attendance. The 17

The Zoo and Observatory are owned and operated by the City of Los Angeles, and the Arboretum is jointly operated by the Los Angeles Arboretum Foundation and Los Angeles County. 18 I also obtained attendance for the Los Angeles Dodgers and California (Anaheim) Angels, both major league baseball teams, but chose not to include them in the analysis because admissions reflect advance ticket purchases

10

Observatory does not charge an admission fee, and attendance is recorded from two turnstiles that people use to enter the grounds, with the numbers entered by hand in daily log files. The Arboretum charges a nominal entrance fee for all customers, and attendance is calculated by dividing the daily cash deposit by the admission price, and is hand-recorded in a log book. 19 In terms of the accuracy of these data, the more sophisticated record keeping system used by the Zoo suggests attendance there is more likely to be accurately measured. Data from both the Observatory and Arboretum were hand-recorded, suggesting greater potential measurement error. Furthermore, because use of the Observatory is free of charge, people can leave and reenter multiple times on the same visit. 20 Also, a turnstile at the Arboretum records attendance, but this value is only recorded in the log at a monthly level. If the measurement error in attendance is uncorrelated with smog alerts, then it will not induce bias in estimates, but will reduce their efficiency. Therefore, more precise estimates are expected for the Zoo, and less precise estimates for the Observatory and Arboretum. The Zoo, because it charges varying admission fees, also offers a breakdown of attendance by age, enabling me to test the Bayesian updating model. Separate attendance is available for adults, children under 2, children aged 2-12, and seniors aged 62 and up. Children and the elderly are two groups considered susceptible to the effects of ozone, so they are more likely to obtain the ozone forecast and therefore not respond to alerts. 21 Although this definition

and involve sedentary activities. In accord with this, I found no statistically significant effect of the alerts on attendance, though sample sizes were quite small. 19 The Arboretum offers a “Free Tuesday” once a month in which attendance is not recorded. 20 An employee of the Observatory noted this occurs because of the interest in immediately adjacent areas that require passing through one set of turnstiles. If customers had originally entered through a different set of turnstiles and seek to exit through the original set, they must re-enter the Observatory and hence will be double counted in the attendance figures. 21 A concern with using children is their responses to alerts could be a function of the state-wide rule requiring schools to reschedule outdoor activities. To address this, I perform this analysis using only days during the summer months when children are unlikely to attend school.

11

of susceptibility is not exhaustive 22 , both groups are specifically targeted by air quality information. If Bayesian updating holds, then children and elderly are less likely to respond to alerts than adults. The Zoo also offers attendance for Greater Los Angeles Zoo Association (GLAZA) members. 23 While the Zoo is both a tourist and local attraction, GLAZA members are typically only local residents. They may be more likely to be aware of alerts or may find it easier to switch activities. Therefore, they are more likely to respond to alerts, providing a robustness check of the model. Because ozone levels vary throughout the day, it is important to point out difference in hours of operation. The Zoo and Arboretum are only day time activities, while the Observatory is both a day and night activity. 24 Many people frequent the Observatory for night time activities, such as stargazing, when ozone levels are considerably lower. Therefore, because it is possible that people shift their outdoor activities to the night on days when alerts are announced, there may be less of a response to alerts at the Observatory. If smog alerts encourage people to reduce time spent outdoors, in order to avoid exposure to ozone they can increase time spent indoors. To test this, I examine whether attendance at an indoor facility increases when alerts are issued. I use data on total attendance at the Natural History Museum of Los Angeles (“Museum”), available from 1991 to 1997. It is located in

22

Unfortunately I am unable to identify customers with a history of respiratory illnesses. These individuals are, however, more likely to be children or elderly than adults, so most should be captured in this definition of susceptibility. 23 GLAZA members do not pay an admission fee per visit. 24 The Zoo is open everyday from 10 a.m. to 5 p.m., with the closing time extended to 6 p.m. from July 1 to Labor Day. The Arboretum is open from 9 a.m. to 5 p.m. everyday. The Observatory is open from 2 p.m. to 10 p.m. Tuesday through Friday and 12:30 p.m. to 10 p.m. on Saturday and Sunday. When school lets out, it is open from 12:30 p.m. to 10 p.m. everyday.

12

central Los Angeles and open during daylight hours everyday except Monday. Attendance is also recorded automatically in a fashion similar to the Zoo. 25 B. Air Quality Information To assign smog alert status and forecasted ozone to each of the facilities, I obtain the 1hour ozone forecast from the Los Angeles Times, thus making the AMA the smallest geographic unit for which the smog alert data is available. The resulting measurement error in alert status, if random, will attenuate the estimated coefficients. The ozone forecast is reported in PSI units, rather than parts per million (ppm) that observed ozone levels are recorded. Because the PSI is a nonlinear function of ppm, I convert the ozone forecast to ppm. In accordance with the availability of attendance data, I collect air quality information from 1989-1997. Although I do not have information on the AMA in which an alert is issued, I use the selection rule in equation (1) to assign alert status. To verify the appropriateness of this approach, I use an administrative file from SCAQMD that contains the dates alerts were issued, though not the AMA they were issued for. I then compute the daily maximum ozone forecast within SCAQMD using the ozone forecast from the Los Angeles Times, and assign alert status based on the maximum ozone forecast. Comparing the assigned alert status to the administrative file (shown in Figure 1) indicates the selection rule is strictly followed: there are only 7 inconsistencies in the 2138 data points available. To assess the accuracy of alerts, I compare predicted alert status to realized alert status using the maximum 1-hour ozone level in each AMA, shown in Table 1. Accuracy is quite low. Of the 104 alerts issued in the AMA for the Zoo, less than 20% were correctly issued; conversely, there were 35 days where ozone surpassed 0.20 ppm but no alert was issued. Accuracy improves for the AMA for the Arboretum, with 25% correctly issued and only 17 days 25

Except on the first Tuesday of each month when admission is free. 13

missed. 26 Furthermore, the R-squared from a regression of observed ozone on forecasted ozone is roughly 0.5 for each venue, suggesting much of ozone formation is not captured in the ozone prediction model. If people track alert accuracy, the inaccuracy of alerts may be problematic for encouraging people to respond. It is, however, potentially useful for the research design of this study: if scientists and meteorologists cannot distinguish between days above and below the threshold, then it is likely that individuals cannot either. C. Other sources Observed air quality and weather are important factors affecting time spent outside. In particular, individuals may adjust their response to alerts depending on observed levels of ozone. Furthermore, heat and sunlight are related to ozone formation. To account for these factors, I include measures of daily 1-hour ozone, 8-hour carbon monoxide (CO), 1-hour nitrogen dioxide (NO2), maximum temperature, precipitation, maximum relative humidity, resultant wind speed, and average sky cover from sunrise to sunset in percent (0=clear, 1=overcast). Pollution data are readily available from the California Air Resources Board air pollution monitoring network. 27 Data on weather come from surface summary of the day (TD3200), maintained by the National Climatic Data Center. 28 Each outdoor venue is assigned to the closest pollution and weather station, and all data are linked at the daily level. Using the date, I assign year-month dummies to control for seasonal patterns in ozone formation and behavior, and day of week, holidays, and summer schedule dummies to account for changes in leisure time. 26

The Arboretum has higher levels of ozone, and therefore alerts, than the Zoo and Observatory because it is on the north side of the Hollywood Hills, where it is warmer and pollutants are trapped by surrounding mountains. The Zoo and Observatory are located near downtown Los Angeles and experience cooler weather. 27 This data is available at http://www.arb.ca.gov/aqd/aqdcd/aqdcddld.htm. 28 This data is available at http://www4.ncdc.noaa.gov/cgi-win/wwcgi.dll?wwAW~MP~AP. Data on humidity and sky cover are more frequently missing than the other weather variables. To prevent a loss in sample size, these

14

4. Empirical Strategy The main objective is to estimate the demand for outdoor time separately for each location, thereby allowing for differential responses to alerts. For example, as noted earlier, the Observatory is open during the evening when ozone levels are lower. Assuming a linear model for demand yields: (4)

ot = β0 + β1 · alt + β2 · ozft + β3 · xt + β4 · xt* + εt

where ot is the log of aggregate attendance 29 at day t (as a measure of outdoor time), alt is a dummy variable indicating whether a smog alert is issued in the AMA of the outdoor facility, xt are observed covariates affecting outdoor time, xt* are unobserved covariates, and εt is an i.i.d. error term. Based on the prediction from the avoidance behavior model, the hypothesis is that β1 < 0: outdoor attendance at the specific facility decreases when alerts are announced. The main limitation in estimating (4) is that unobserved variables, such as forecasted weather, may be correlated with both the decision to issue an alert and engage in outdoor activities. Since alerts are a deterministic function of the forecasted ozone, as indicated in equation (1), forecasted ozone fully governs the alert selection rule and makes it possible to leverage a regression discontinuity (RD) design. To do this, specify (4) as: (5)

ot = β0 + β1 · alt + f(ozft, β2) + β3 · xt + vt

where f is a function that relates the ozone forecast to attendance and vt is the composite error term (vt = β4 · xt* + εt). If days just below forecast ozone levels of 0.20 ppm are identical to days

missing values are imputed using a best-subset regression with the other weather variables and observed ozone as covariates. Regression results are, however, insensitive to excluding the imputed values. 29 Specifying the dependent variable in levels yields comparable results.

15

just above 0.20, then the discontinuity in attendance that occurs at 0.20 ppm represents the causal effect of alerts. 30 While it is impossible to know whether the unobservables are identical across such days, I examine how well the observable covariates balance across alert status for days with ozone forecasts near 0.20 ppm. If the observed factors balance, then it is more reasonable to believe the unobserved factors do as well. Figure 2 shows a plot of three likely influential covariates (temperature, humidity, and carbon monoxide (CO)) by ozone forecast levels. All three covariates evolve smoothly throughout this plot, suggesting they are unaffected by smog alerts. As preliminary evidence that people respond to alerts, average attendance at the Zoo is also plotted in Figure 2. Focusing on the observations near 0.20 ppm, attendance is generally increasing in forecasted ozone prior to 0.20 ppm. At 0.20 ppm, the point at which an alert is issued, attendance sharply drops. After that, attendance levels remain increasing in forecasted ozone, but are generally lower than the attendance levels below 0.20 ppm. This drop in attendance at the alert threshold provides the first piece of evidence that people respond to smog alerts. Specification of f is crucial to the RD design. It enables the use of points far from the alert threshold to improve efficiency and generalizability, but misspecification can render biased estimates of β1 (Dinardo and Lee (2004)). Figure 2 suggests that attendance at the Zoo is roughly linear in forecasted ozone, so I begin by estimating that specification. Additionally, I estimate equation (5) using only observations centered near the alert threshold, but omit forecasted ozone from the equation. Furthermore, I also control for f non-parametrically by specifying a dummy variable for each value of the ozone forecast, but omit the smog alert variable (also using

30

If people directly respond to the ozone forecast, this does not invalidate the empirical strategy. It implies there will be no effect of alerts.

16

observations centered near the alert threshold). This method does not restrict the discontinuity to occur at 0.20 ppm; instead, it is estimated from the data. I display results from all specifications to assess the sensitivity of estimates of β1 to the functional form of f. 31 Two additional assumptions are necessary to obtain unbiased estimates of β1. The first is that alert status is not “corrected” once actual levels of ozone are realized; i.e., alert status is correctly dictated by equation (1). Despite the temptation to continually update the forecast, officials at SCAQMD indicate this rule is strictly followed because of flaws inherent in detecting and disseminating an alert as it occurs. For example, ozone typically peaks in the late afternoon, around 3:00 pm. The data is not received until an hour later, and once a violation is detected, it must be double-checked to ensure accuracy. At this point, which can be up to 2 hours after the violation first occurs, the media are first made aware. By the time this information is received by the public, sunlight has decreased and ozone levels have typically fallen to safer levels. Therefore, this assumption is likely to be satisfied. 32 The second assumption is no supply-side effects. For example, facilities do not lower their price to entice customers on alert days, keep animals inside to protect their health on alert days, or reach maximum capacity on non-alert days such that they turn customers away. Of the venues considered, none violate these possibilities. It is possible, however, that a more crowded atmosphere, although under capacity, provides less enjoyment because of longer waiting times, for example. In this case, if attendance drops in response to an alert being issued, then there is less crowding, which may induce other people to go to these outdoor attractions. Therefore, this

31

In this application, the covariate (ozone forecast) that determines the treatment (smog alert) is discrete. If the deviations from the continuous measure are random, this can be modeled econometrically by accounting for the group structure of the ozone forecast (Card and Lee (2004)). This involves computing standard errors clustered on each value of forecasted ozone. 32 Furthermore, I estimated models that included an indicator for whether alerts are issued correctly, and found no significant differences in the estimates.

17

offsetting behavior will understate the amount of avoidance behavior, leading to a downward bias in β1. 5. Results A. Main results The main set of regression results, shown in Table 2, provides further evidence that people respond to smog alerts by decreasing outdoor activities. There is a separate panel for each outdoor facility, and each panel includes results from six specifications. These specifications are reported to gauge the sensitivity of estimates to both potential confounding and functional form assumptions of the RD. The odd numbered columns omit several variables likely to affect the demand for outdoor activities (precipitation, wind, humidity, observed ozone, carbon monoxide, and nitrogen dioxide), while the even numbered columns include them. 33 Columns (1) and (2) use a linear specification for the RD using all observations (“full sample”). Columns (3) and (4) restrict the sample to forecasted ozone values between 0.15 and 0.24 ppm but omit forecasted ozone (“restricted sample”). Columns (5) and (6) are non-parametric estimates on the restricted sample that include separate dummy variables for all values of the ozone forecast. For the non-parametric results, the coefficients on the dummy variables are not directly interpretable because they index for the unobservables in the model. To infer the impact of alerts, I use the change in coefficients around the alert threshold in two ways. One, I look at the change in coefficients at the point of the discontinuity by comparing the dummy variable on ozf = 0.19 to ozf = 0.20 ppm. Because this approach is demanding on the data and individual coefficients on the dummy variables may be imprecise, I also infer the impact by subtracting the mean of the point estimates for the dummy variables from 0.20 to 0.24 ppm from the mean of the dummy variables from 0.16 to 0.19 ppm (0.15 is the omitted category). This second approach is

18

less likely to be sensitive to outliers. If the dummy variable(s) above the alert threshold are lower than those above, this implies a decrease in attendance in response to alerts. Results for the Zoo are shown in Panel A. Columns (1) and (2) show a statistically significant drop in attendance of 15% both with and without the additional weather and air quality covariates. When limiting the sample to points centered on the alert threshold, shown in columns (3) and (4), attendance shows a statistically significant drop of 10% and 11%. The nonparametric results in columns (5) and (6) indicate that smog alerts reduce attendance between 11 and 12% according to both methods of inferring the impact. These are remarkably similar to the parametric estimates on the restricted sample. In fact, nearly all of the individual coefficients on the dummy variables below the alert threshold are larger than all of those above the threshold (not shown). This robustness to the functional form of ozone forecast is impressive given the demands of this technique. Furthermore, the results are almost completely unaffected by excluding several important covariates. This is notable given that weather and air quality are important predictors of outdoor activities. This suggests that these variables evolve smoothly around the discontinuity, so the change in attendance at the discontinuity represents the causal effect of smog alerts. Results for the Observatory, shown in Panel B, also show a decline in attendance from smog alerts, and the decline is smaller in magnitude than for the Zoo, as expected. Attendance declines by roughly 3% in the full and restricted sample, though statistically insignificant. In the non-parametric specification, at the point of the discontinuity there is a statistically significant drop in attendance of 6%, which is larger than the parametric estimates. The effect using all of the dummy variables reveals an effect comparable in magnitude to the parametric results, and is statistically significant. All estimates are also insensitive to including the weather and pollution 33

I do not omit temperature and sun cover because these variables are part of the ozone prediction model. 19

variables. These results suggest a response to alerts, with the smaller magnitude than the Zoo consistent with expectations because the Observatory includes night time hours when ozone levels are typically lower. Attendance at the Arboretum also declines when alerts are issued, as shown in Panel C, though the results are more sensitive to covariates and functional form. In the full sample, attendance declines by roughly 11%. In the restricted samples, attendance declines by 6 to 10% depending on the included covariates. The non-parametric results are less consistent than the parametric results. There is a surprising positive jump in coefficients at the point of the discontinuity, indicating alerts increase attendance. Given the small samples available at each ozone forecast, this may be due to unusual occurrences on those specific days, such as an unobserved special event 34 or a recording error. Despite the purported anomaly, the coefficients below the threshold are typically larger than those above. Using all coefficients to compute the effect, which is less sensitive to outliers, indicates a statistically significant drop in attendance that is close in magnitude to the parametric estimates. Though these results are less robust than the results for the other two facilities, they generally support evidence of a response to alerts. Based on parametric results from the restricted samples, which are more likely to provide unbiased estimates, estimates suggest a drop in attendance of 2 to 11 percent from an alert at all three facilities. The results for the Zoo and Observatory are extremely insensitive to functional form assumptions and potential confounding, but are less so for the Arboretum. In only one instance are there unexpected results, but this occurs at the Arboretum, where estimates are generally less robust. The results from all three places are of the same order of magnitude despite coming from three distinct sources, making it unlikely these results are due to sampling

34

The Arboretum has numerous scheduled special events throughout the year, such as garden sales, plant fairs, and environmental education fairs.

20

variability or general misspecification. Overall, the results suggest people respond to smog alerts by reducing attendance at these outdoor attractions. B. Sensitivity analyses 35 If the costs of avoiding these outdoor activities are lower for local residents, either because they are more aware of information or have lower costs of substitution, then they should display larger responses to alerts. Using the attendance breakdown at the Zoo, I explore how GLAZA members respond to alerts. Shown in column (1) of Table 3, GLAZA members reduce attendance by 18%, which is larger than the total attendance response. Given that GLAZA members are more likely to be local residents, this conforms to expectations that locals show a greater response to alerts than other attendees. Responses to alerts may vary by the amount of leisure time available. For example, people have greater discretion over their time on weekends and may find it easier to switch activities, suggesting a potentially larger effect of an alert on the weekend. Columns (2)-(4) in Table 3 show results for each place including an interaction between alerts and a weekend dummy variable. The joint effect of alerts and alerts*weekend is still statistically significant at all facilities. Results indicate responses are larger on the weekend for all three facilities, with nearly all of the response at the Observatory coming on the weekend. Though weekend responses are not statistically different from the overall response, they suggest the greater leisure time available on weekends allows a larger response to alerts. Total leisure time is divided between outdoor and indoor activities. Therefore, if individuals decrease time spent outside, they must increase time spent inside. To assess this, I use attendance at the Museum as the dependent variable in (5). The results, shown in column (5)

35

In all sensitivity analyses I only include results using the full set of covariates and omit the non-parametric results because the estimates are generally similar to the parametric estimates on the restricted sample.

21

of Table 3, indicate an increase in attendance at the Museum of nearly 17%, though it is statistically insignificant. It is larger in magnitude than the response for any of the outdoor facilities. However, children are a high fraction of attendees at the Museum, and this estimate is the same magnitude for children attendees at the Zoo (shown below). The estimates are too imprecise to suggest full offsetting behavior, but the lack of a negative effect on attendance – which would reflect unobserved heterogeneity – supports the above evidence that people take protective actions in response to alerts. One concern with the evidence presented thus far is that people may respond to alerts out of altruistic rather than health concerns. When alerts are issued people may not participate in certain activities because it involves driving, and they do not want to contribute to pollution on days already considered highly polluted. 36 Therefore, people may reduce visits to specific outdoor attractions but not limit total time spent outdoors, implying a decrease in outdoor attendance is not evidence of avoidance behavior. If people limit their contribution to pollution, then they reduce attendance at any facility, indoor or outdoor. The evidence of no negative effect at the Museum rejects the altruism hypothesis. Furthermore, if people drive less in response to alerts, then CO levels -- a proxy for traffic because it primarily comes from automobiles -- should fall at the alert threshold. As shown in Figure 2, however, CO evolves smoothly through the alert threshold, rejecting the altruism effect. These findings help to substantiate the overall evidence that people are displaying avoidance behavior in response to information about pollution. 37

36

Although individuals are price-takers with respect to ozone levels, the model could be accommodated to include altruism by entering efforts towards environmental goods directly in the utility function. 37 As an additional test of altruism, I also examined whether the traffic fatalities – a proxy for traffic volume – responds to alerts. If people drive less in response to an alert, then there should be an accompanying decrease in traffic fatalities, all else equal. The results from this specification indicated no effect of alerts on fatalities.

22

I provide a general specification test for this model by including future alerts in equation (5). People cannot respond to a forecasted alert before it occurs, so finding an effect would suggest misspecification. If people use a naïve version of equation (1) to forecast ozone on their own, however, then they may anticipate future smog alerts by increasing current outdoor activities; this suggests a positive coefficient on future alerts. Therefore, only a negative coefficient on future alerts would suggest misspecification. The results, shown in Table 4, indicate a positive effect from future alerts for all three facilities, though imprecisely estimated, suggesting people may anticipate future smog alerts. Furthermore, the effects from contemporaneous alerts are unaffected by the inclusion of future alerts. C. Bayesian Updating and Intertemporal Substitution If more susceptible people obtain the continuous ozone forecasts, they are less likely to respond to alerts than unsusceptible people. To test this hypothesis, I estimate equation (5) for the Zoo separately for children and the elderly, two susceptible groups. The results, shown in Table 5, indicate the opposite: responses are larger for children and seniors. These groups decrease attendance between 18 and 20% as compared to an overall response of around 11%. 38 Although this basic classification of susceptibility is not sufficient for capturing those with a history of respiratory illness, it captures two major groups targeted by air quality information. In general, these results reject a model of rational Bayesian updaters and suggest information content in the alerts for potentially susceptible populations. As alerts become increasingly common it may become more costly to switch activities; switching activities after one alert may be easy, but people may tire of responding as they become more frequent. To assess intertemporal responses, I also include in equation (5) an

38

Although responses by children may reflect administration laws by schools, the results using the restricted sample only include the summer period when school is out of session.

23

interaction between contemporaneous and lagged alert status (alt*alt-1). 39 This variable is an indicator function for whether an alert is issued on two consecutive days. The coefficient on alt measures the response to an alert if it gets issued on one day only. The coefficient on alt*alt-1 measures the response on the second day relative to the first day. If this coefficient is positive, people are less likely to respond on the second day. 40 Although individuals are unlikely to go to a specific outdoor venue on successive days, using aggregate responses still provides a valid test. To see this, imagine individuals have a sequence of outdoor activities for the next two days {O1, O2}, and do not substitute outdoor activities across days. Individuals do not have the same ordering of activities across the two days, so some have {Zoo, Arboretum} and others have {Arboretum, Zoo}, for example. The alternative choice on each day is an indoor activity. If people respond to an alert issued on the first day, O1 will decrease, so both the Zoo and the Arboretum experience a decline in attendance. If people do not respond to an alert issued on the second day, O2 will not change, so both the Zoo and the Arboretum experience no change in attendance. Therefore, with enough variation in the ordering of {O1, O2} across individuals, using aggregate attendance is sufficient to perform this test. Table 6 shows that when an alert gets issued on one day only, responses are comparable to results in Table 2. Estimates for the Zoo and Observatory remain statistically significant, but not for the Arboretum. When an alert is issued on two consecutive days, though, the response on the second day decreases at the Zoo and Observatory. These estimates are imprecise because alerts occurring on two consecutive days are rare events, but are significant at the 10% and 12% level for the Zoo and Observatory, respectively. Furthermore, they are statistically different than 39

I also include lagged smog alert in levels.

24

the first day response. The decreases are comparable in magnitude to the first day response, suggesting that avoidance behavior almost completely disappears on the second day. The same pattern does not occur for the Arboretum. Responses on the second day are comparable to those on the first, though t-statistics are less than 1, so it is difficult to draw strong conclusions. Although responses are not consistent across the three facilities, in instances where the estimates are borderline significant they suggest people are less likely to respond to alerts as they become more frequent. 41 6. Conclusion Providing information about pollution is an integral part of environmental policy. The EPA mandates that large cities provide the PSI on a daily basis (U.S. EPA (1999)). Numerous local air quality districts also provide pollution warning systems, such as “ozone action days” in Arkansas, Texas and Baltimore-Washington Metropolitan Area, “ozone alerts” in New York, Oklahoma, and Kansas, and “ozone warnings” in Chicago, Illinois. If people respond to these warnings, this can improve the efficiency of air quality regulations. Traditional regulations rely on reducing overall emissions and improving air quality for all people – whether they are susceptible or not to the health effects of pollution. Alternatively, regulations can provide information to allow susceptible individuals to reduce their exposure on high pollution days. Depending on the welfare costs from responding and fraction of susceptible people in the population, the latter may be a more efficient policy. This paper examines whether people in Southern California respond to smog alerts by decreasing attendance at three outdoor facilities. I find a significant decrease in time spent 40

An alternative explanation is that the media may not report an alert as vigorously on the second day. It is unfortunately not possible to distinguish this explanation from the above explanation. 41 Responses on the second day may vary depending upon the accuracy of the first day’s forecast if people track errors. The occurrence of two consecutive alerts where the first is issued correctly is an extremely rare event,

25

outdoors in response to smog alerts, and results are extremely insensitive to numerous specification checks for two of the facilities. Furthermore, susceptible individuals are more likely to respond to alerts than unsusceptible individuals. People value the information contained in the warnings, though their patience for responding wanes as more warnings are supplied. If people change their behavior in response to smog alerts, this action must have some cost to them. It is unfortunately not possible with the given data to measure the loss in consumer welfare from responding to alerts. Measuring these costs is an essential part of environmental regulation, 42 so it should be a focus of future research. Given that individuals adjust their behavior in response to information about risk, it is essential to account for this compensatory behavior when estimating the effect of ozone on health. Several recent analyses using quasi-experimental techniques often provide convincing empirical evidence of exogenous variation in ambient pollution, but typically assume inelastic behavioral responses to ambient pollution. 43 Even if ambient pollution levels are randomly assigned, this study suggests individuals re-optimize in light of their assigned ambient pollution level so their exposure to pollution remains endogenously determined. If unaccounted for, this compensatory behavior may lead to biased estimates of the biological effect of pollution on health. This is explored in more detail in Neidell (2006). There are also potential health costs from responding to the alerts that cannot be detected in this analysis. If people spend more time inside as a result of the alerts, then this can influence people to adopt a more sedentary lifestyle. Although one day alone may not induce changes in behavior, a longer string of alerts might. For example, an alert was issued every single day in making it unlikely to detect an effect. I estimated such models, and results were generally comparable to Table 6, though less precise as expected. 42 For a more detailed description of the role of avoidance behavior in environmental regulation, see Harrington and Portney (1987).

26

August of 1986 in the San Bernardino and Riverside areas. A sedentary lifestyle is unlikely to effect short-term health, but may affect longer-term health outcomes, such as obesity. Therefore, although policies with announcements at lower pollution levels, which are becoming increasingly popular, may improve short-term health outcomes, they may also induce longerterm negative effects on health as well.

43

For examples of such studies, see Ransom and Pope (1995), Chay and Greenstone (2003a,b), Neidell (2004), Currie and Neidell (2005), Jayachandran (2005), and Lleras-Muney (2005).

27

References Bauman, Adrian E., Bill Bellew, Neville Owen, Philip Vita, (2001). “Impact of an Australian Mass Media Campaign Targeting Physical Activity in 1998.” American Journal of Preventive Medicine 21(1): 41-47. Bresnahan, B., M. Dickie, and S. Gerking (1997). “Averting Behavior and Urban Air Pollution.” Land Economics 73: 340-357. Card, David and David Lee (2004). “Regression Discontinuity Inference with Specification Error.” Center for Labor Economics Working Paper #74. Chang LT, Koutrakis P, Catalano PJ, Suh HH (2000). “Hourly personal exposures to fine particles and gaseous pollutants--results from Baltimore, Maryland.” J Air Waste Manag Assoc. 50(7): 1223-35. Chay, Kenneth and Michael Greenstone (2003a). “Air Quality, Infant Mortality, and the Clean Air Act of 1970.” NBER Working Paper #10053. Chay, Kenneth and Michael Greenstone (2003b). “The Impact of Air Pollution on Infant Mortality: Evidence from Geographic Variation in Pollution Shocks Induced by a Recession.” Quarterly Journal of Economics 118(3): 1121-1167. Currie, Janet and Matthew Neidell (2005). “Air Pollution and Infant Health: What Can We Learn from California’s Recent Experience?” Quarterly Journal of Economics 120(3): 10031030. Dinardo, John and David Lee (2004) “Economic Impacts of New Unionization on Private Sector Employers: 1984-2001.” Quarterly Journal of Economics 119(4). Domenighetti G, Luraschi P, Casabianca A, Gutzwiller F, Spinelli A, Pedrinis E, and Repetto F. (1988). “Effect of information campaign by the mass media on hysterectomy rates.” Lancet Dec 24-31;2(8626-8627):1470-3. Harrington, Winston and Paul Portney (1987). “Valuing the Benefits of Health and Safety Regulation.” Journal of Urban Economics 22. Ippolito, Pauline M. and Alan D. Mathios (1990). “Information, Advertising and Health Choice: A Study of the Cereal Market.” Rand Journal of Economics 21:459-480. Jayachandran, Seema (2005). “Air Quality and Infant Mortality during Indonesia’s Massive Wildfires in 1997.” Mimeograph, UCLA. Jin, Ginger Zhe and Phillip Leslie (2003). “The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards.” Quarterly Journal of Economics 118(2): 409451. Lleras-Muney, Adriana (2005). “The needs of the Army: Using compulsory relocation in the military to estimate the effect of environmental pollutants on children's health.” Mimeograph, Princeton University. Mullahy, John (1999). “It'll Only Hurt a Second? Microeconomic Determinants of Who Gets Flu Shots.” Health Economics 8(1): 9-24. Neidell, Matthew (2004). “Air Pollution, Health, and Socio-Economic Status: The Effect of Outdoor Air Quality on Childhood Asthma.” Journal of Health Economics 23(6): 1209-1236. _______ (2006). “Compensatory Behavior and the Health Effects of Ozone.” Mimeograph, Columbia University. Philipson, Tomas (1996). “Private Vaccination and Public Health: An Empirical Examination for U.S. Measles.” Journal of Human Resources 31(3): 611-30 Ransom, M., and C.A. Pope III, (1995). “External Health Costs of a Steel Mill.”

28

Contemporary Economic Policy 13, 86-97. Smith, V. Kerry, and Johnson, F. Reed (1988). “How Do Risk Perceptions Respond to Information? The Case of Radon.” Review of Economics and Statistics 70(1): 1-8. U.S. Environmental Protection Agency (1999). “Guidelines for Reporting of Daily Air Quality – Air Quality Index (AQI).” EPA Document #454-R-99-010, Research Triangle Park, NC. _______ (2003). “Air Quality Criteria Document for Ozone”, First External Review Draft, available at http://www.epa.gov/ncea/ozone.htm. Viscusi, W. Kip, Magat, Wesley A., and Huber, Joel C. (1986). “Informational Regulation of Consumer Health Risks: An Empirical Evaluation of Hazard Warnings.” RAND Journal of Economics 17(3): 351-65.

29

0

.2

.4

alert

.6

.8

1

Figure 1. Adherence to Smog Alert Selection Rule

.05

.1

.15

.2

.25

.3

ppm

Figure 2. Zoo Attendance and Covariates by Ozone Forecast

10 9 8 7 6 5 4 3 2 1 0 0.05 0.06 0.07 0.08 0.09

0.1

0.11

0.12

attendance/1000

0.13

0.14

0.15

humidity

30

0.16

0.17

CO

0.18

0.19

0.2

0.21 0.22 0.23

max. temp.

Table 1. Summary Statistics A. Outdoor and Indoor Facilities

attendance alert ozone forecast (ppm) ozone (ppm) max.temp./10 precip. (in.) relative humidity/10 resultant wind speed (mph) sky cover (tenths) carob monoxide (ppm) nitrogen dioxide (ppm) children <2 children 2-12 seniors adult GLAZA members

Zoo (n=1949) mean std. dev. 4766 3160 0.05 0.22 0.12 0.05 0.08 0.04 8.21 0.97 0.22 1.62 8.97 0.71 56.59 23.70 0.45 0.30 2.31 1.24 0.07 0.03 299 362 766 774 81 51 1530 1655 718 538

Observatory (n=1770) mean std. dev. 5629 2368 0.06 0.23 0.12 0.05 0.08 0.04 8.25 0.96 0.19 1.36 8.98 0.71 57.02 23.63 0.45 0.30 2.30 1.23 0.07 0.03 -

B. Alerts accuracy Arboretum year Zoo Observ. 1989 16 / 6 / 9 16 / 6 / 8 1990 12 / 2 / 5 10 / 2 / 5 54 / 18 / 5 1991 20 / 4 / 8 20 / 4 / 8 37 / 10 / 1 1992 24 / 2 / 8 22 / 2 / 8 48 / 18 / 4 1993 10 / 2 / 1 10 / 2 / 1 36 / 11 / 3 1994 15 / 3 / 2 15 / 3 / 2 37 / 5 / 2 1995 7/0/1 7/0/1 27 / 3 / 1 1996 0/0/1 0/0/1 14 / 0 / 1 1997 0/0/0 0/0/0 1/0/0 Total 104 / 19 / 35 99 / 19 / 34 254 / 65 / 17 number of alerts issued / number of alerts correct / number of alerts missed

31

Arboretum (n=1683) mean std. dev. 449 476 0.15 0.36 0.13 0.05 0.10 0.05 8.34 0.88 0.21 1.48 8.99 0.70 55.81 23.41 0.44 0.30 1.57 0.76 0.06 0.03 -

Museum (n=1136) mean std. dev. 915 592 0.01 0.10 0.09 0.03 0.07 0.03 8.04 0.81 0.19 1.38 9.00 0.70 55.20 22.88 0.47 0.29 2.31 1.25 0.07 0.04 -

Table 2. Effect of Smog Alerts on Outdoor Attendance A. Dependent variable = total attendandance at Los Angeles Zoo 1

alert oz f =.20 - oz f =.19 ⎛ .24 ⎜ ∑ oz ⎝ k =.20

f

⎞ ⎛ .19 = k ⎟ / 5 − ⎜ ∑ oz ⎠ ⎝ k =.16

f

⎞ = k⎟/4 ⎠

max. temp. max. temp. sq. sun cover precip. wind humidity Observations R-squared

2 3 4 parametric full sample restricted sample -0.148** -0.150** -0.102** -0.110** [0.036] [0.040] [0.025] [0.030] 2.121** 1.658** 1.009* 0.263 [0.388] [0.274] [0.371] [0.581] -0.132** -0.106** -0.070** -0.029 [0.023] [0.017] [0.021] [0.032] -0.166 -0.117 -0.003 -0.021 [0.111] [0.081] [0.032] [0.030] -0.067** -0.019 [0.011] [0.014] -0.001 -0.001 [0.001] [0.001] -0.039 0.118 [0.032] [0.064] 1949 1949 520 520 0.74 0.76 0.81 0.82

5 6 non-parametric restricted sample -0.113** -0.122** [0.015] [0.014] -0.106** -0.120** [0.012] [0.017] 1.350** 0.680 [0.387] [0.574] -0.089** -0.052 [0.021] [0.032] -0.051 -0.060 [0.030] [0.029] 0.006 [0.012] -0.002 [0.001] 0.135 [0.062] 520 520 0.99 0.99

B. Dependent variable = total attendandance at Griffith Park Observatory 1

alert oz f =.20 - oz f =.19 ⎛ .24 ⎜ ∑ oz ⎝ k =.20

f

⎞ ⎛ .19 = k ⎟ / 5 − ⎜ ∑ oz ⎠ ⎝ k =.16

max. temp. max. temp. sq. sun cover precip. wind humidity Observations R-squared

f

⎞ = k⎟/4 ⎠

2 3 4 parametric full sample restricted sample -0.032 -0.031 -0.024 -0.024 [0.023] [0.022] [0.020] [0.018] 0.357** 0.377** 0.032 -0.110 [0.098] [0.073] [0.320] [0.402] -0.022** -0.023** -0.005 0.003 [0.006] [0.004] [0.017] [0.022] -0.079* -0.085** 0.008 0.003 [0.029] [0.027] [0.023] [0.028] -0.004 0.054** [0.002] [0.011] 0.000 0.001** [0.000] [0.000] -0.003 0.017 [0.008] [0.051] 1770 1770 497 497 0.68 0.68 0.71 0.72

32

5 6 non-parametric restricted sample -0.066** -0.063** [0.012] [0.011] -0.028** -0.021** [0.006] [0.005] -0.019 -0.211 [0.274] [0.291] -0.002 0.009 [0.015] [0.016] -0.001 -0.019 [0.035] [0.036] 0.064** [0.007] 0.001* [0.000] 0.037 [0.048] 497 497 0.99 0.99

Table 2. Effect of Smog Alerts on Outdoor Attendance (continued) C. Dependent variable = total attendandance at Los Angeles County Arboretum 1

2 3 4 5 6 parametric non-parametric full sample restricted sample restricted sample alert -0.114* -0.117* -0.070 -0.094* [0.052] [0.051] [0.033] [0.034] oz f =.20 - oz f =.19 0.056** 0.040* [0.012] [0.016] .19 -0.063** -0.101** ⎛ .24 ⎞ ⎛ ⎞ ⎜ ∑ oz f = k ⎟ / 5 − ⎜ ∑ oz f = k ⎟ / 4 [0.012] [0.019] ⎝ k =.20 ⎠ ⎠ ⎝ k =.16 max. temp. 3.139** 2.488** 1.548 1.630 1.952* 1.751 [0.250] [0.208] [0.800] [0.876] [0.785] [0.819] max. temp. sq. -0.189** -0.153** -0.101 -0.104 -0.123* -0.111* [0.016] [0.013] [0.045] [0.049] [0.045] [0.047] sun cover -0.195* -0.131* -0.069 -0.053 -0.107** -0.088** [0.076] [0.053] [0.042] [0.042] [0.022] [0.025] precip. -0.077** -0.365 -0.248 [0.013] [0.306] [0.357] wind -0.001 -0.001 -0.002 [0.000] [0.001] [0.001] humidity -0.047* -0.004 0.023 [0.022] [0.023] [0.032] Observations 1683 1683 579 579 579 579 R-squared 0.81 0.82 0.83 0.84 0.99 0.99 * significant at 5%; ** significant at 1%. Standard errors clustered on ozone forecast that account for heteroskedasticity in brackets. Columns 1 and 2 include a linear term for ozone forecast, columns 3 and 4 limit the sample to ozone forecasts between 0.15 and 0.24 ppm, columns 5 and 6 include a separate dummy variable for each ozone forecast limiting the sample to ozone forecasts between 0.15 and 0.24 ppm. All regressions include an indicator for holiday, an indicator for summer schedule, day of week dummies and yearmonth dummies. The odd numbered columns include controls for observed ozone, carbon monoxide, and nitrogren dioxide. The regressions for the Arboretum also include an indicator if the day was an insect or environmental fair.

33

Table 3. Sensitivity Estimates of Effect of Alerts on Outdoor Attendance GLAZA members Zoo Observatory Arboretum Museum 1 2 3 4 5 alert -0.182** -0.059 0.012 -0.079 0.167 [0.025] [0.046] [0.033] [0.040] [0.242] alert*weekend -0.180 -0.121* -0.043 [0.094] [0.052] [0.037] P(alert=alert*weekend) > F 0.382 0.143 0.607 P(alert=alert*weekend=0) > F 0.006 0.006 0.016 Observations 520 520 497 579 70 R-squared 0.64 0.83 0.72 0.84 0.88 See notes to Table 2. Specifications are comparable to column (4) in Table 2. P(alert=alert*weekend) > F is the p-value from an F-test that the coefficient on alert equals the coefficient on alert*weekend. P(alert=alert*weekend=0) > F is the p-value from an F-test that the coefficients on alert and alert*weekend are jointly zero.

Table 4. Include Future Alerts Arboretum Zoo Observatory 1 2 3 alertt -0.108** -0.024 -0.096* [0.028] [0.018] [0.037] alertt+1 0.053 0.033 0.018 [0.055] [0.032] [0.047] Observations 506 485 560 R-squared 0.82 0.72 0.84 See notes to Table 2. Specifications are comparable to column (4) in Table 2.

34

Table 5. Effect of Alerts on Zoo Attendance by Susceptibility Children <2 Children 2-12 Seniors 1 2 3 alert -0.195* -0.178** -0.175** [0.060] [0.051] [0.029] Observations 520 520 520 R-squared 0.71 0.86 0.77 See notes to Table 2. Specifications are comparable to column (4) in Table 2.

Table 6. Effect of Consecutive Alerts on Outdoor Attendance

alertt alertt*alertt-1 P(alertt=alertt*alertt-1) > F Observations R-squared

Zoo 1 -0.141** [0.021] 0.121 [0.065] 0.005 508 0.82

Observatory 2 -0.060* [0.021] 0.078 [0.046] 0.055 485 0.71

Arboretum 3 -0.045 [0.046] -0.070 [0.070] 0.820 563 0.84

See notes to Table 2. Specifications are comparable to column (4) in Table 2 but also included lagged smog alert. P(alertt=alertt*alertt-1) > F is the p-value from an F-test that the coefficient on alertt equals the coefficient on alertt*alertt-1.

35