In late 1997, Southeast Asia experienced the worst forest fires on record. Large tracts of forest and arable land on the islands of Kalimantan and Sumatra were destroyed. Parts of these islands were blanketed with thick haze of smoke for several months, forcing President Suharto to declare a state of emergency in September 1997. An analysis of SPOT images suggested that about 5 million hectares of land burned, of which 20% was forested (liew et al. 1998).

The effects of the fires were felt throughout the region-in Brunei, Singapore, Malaysia, and as far afield as Thailand and Vietnam. Early estimates suggested that the fires caused at least $4.5 billion in damage in the region (Schweithelm and Glover 1999). Most of these costs were attributed to the health consequences of the fires. Because the haze was the heaviest and persisted the longest over parts of Indonesia, several studies have concluded that the Indonesian population suffered the greatest health costs. Ruitenbeck (1999) based his conclusions on data from hospitalizations and self-treatments in Sarawak, Malaysia. Sastry (2002) examined the effects of the smoke on daily mortality in Kuala Lumpur and reported an increase in mortality for 65- to 74 year olds that lasted several weeks. He concluded that the results are "suggestive of wider short-term health impacts, particularly with respect to acute morbidity," and speculated that the effects of the smoke haze "in Indonesia itself are likely to have been tremendous" (p. 20).

These conclusions are potentially flawed for two reasons. First, the degree of exposure to haze was substantially lower in Malaysia than in Kalimantan and Sumatra. It is unclear how to interpret the extrapolation to the Indonesian population unless one assumes that the effects of haze on health are linear. Second, in poor economies, analyses of those who use health care likely miss the poorest and potentially most vulnerable part of the population.

We used uniquely rich data from Indonesia to measure directly the short-term effects of the fires on the health status of the adults who were exposed to the most severe haze. Rather than rely on inpatient, outpatient, or mortality data to infer effects on health outcomes, we used longitudinal data from a population-based household survey, the second wave of which was collected at the time of the Indonesian fires. We combined these survey data with satellite-based aerosol measures to produce a rich data source with which to examine the immediate effects of the fires.

In the following sections, we describe the Indonesian fires and discuss what is known about the health consequences of exposure to particulates. Our data, from the Indonesia Family Life Survey (IFLS), are described in conjunction with two important methodological issues. First, the health indicators that are examined are introduced and their interpretation is discussed. Second, comparisons of the health status of adults who were living in areas that were exposed to haze in 1997 with the health status of adults who were living in areas that were not exposed is shown to be contaminated by unobserved heterogeneity and to overstate the impact of the fires on health. Exploiting the longitudinal dimension of IFLS, we developed a difference-in-difference approach to estimation. Specifically, we compared health status reported several years prior to the fires with health status reported by the same person at the time of the fires. We then contrasted changes among individuals who were exposed to the haze in 1997 with changes among individuals who were not exposed. The results indicate that exposure to fires has a negative and significant impact on the health of older adults and prime-age women, but that much of the impact appears to be transitory.

THE INDONESIAN FIRES

The Indonesian fires have their origins in the ecologies and economies of Sumatra and Kalimantan-the two major islands on which the 1997 fires occurred. Population densities on these islands are low, and tropical rain forest covers considerable portions of the land area. In some areas, the forest floor is covered with a thick and exceptionally flammable layer of dried organic material.

On both Sumatra and Kalimantan, small-scale farmers have traditionally used controlled burns as a method of land clearing. Recent excavations have revealed sites where swidden (slash-and-burn) agriculture has been practiced for some 200 years (Lawrence and Schlesinger 2001; United Nations Development Program, UNDP, 1998). Used correctly, fire plays a valuable ecological role in swidden agriculture because it makes the nutrients that are bound up in plant material that has been cleared available to future crops (UNDP 1998).

Over the past quarter-century, the magnitude of fires on Sumatra and Kalimantan has increased dramatically, despite an ostensible ban on the practice put in place in 1984 and reaffirmed in 1997 (Ketterings et al. 1999; UNDP 1998). The increase in fires is a function of several phenomena. First, the amount of land under commercial control has risen as a result of timber and plantation concessions that have been granted in the past three decades. These industries create more flammable debris and use fire for clearing larger areas than do small-scale farmers. Second, when fires burn out of control, logged areas sustain more damage than does primary forest (Siegert et al. 2001). Third, the Indonesian government's efforts to move people from the densely populated islands of Java and Bali to less-settled areas have increased the number of small-scale farmers-so much so that on Sumatra, cultivation can no longer be called "shifting," although these farmers continue to use fire to clear land (Ketterings et al. 1999). Finally, conflicts over claims to land have increased, and fire is sometimes used as a weapon in such disputes (Glover and Jessup 1999).

In recent years, the most damaging fires have occurred in 1982-1983, 1987, 1991, 1994, and 1997-1998. In all these years, the fires were exacerbated by drought brought on by the El Nio Southern Oscillation (ENSO; Jim 1999). In ENSO years, the delay of the monsoon means that fires burn for several months longer than usual. Also, because the land is unusually dry, fires burn out of control more easily, sometimes escaping into peat forests, where they burn underground and may ignite shallow coal seams.

The forest fires of 1997-1998 were by far the largest in Indonesia's history, burning some 5 million hectares before they were eventually quenched by the rains in mid-to-late November (Ruitenbeck 1999). No sooner had the rains stopped than fires sprang up again on Kalimantan (but not on Sumatra) in early 1998.

One way to characterize the extent and consequences of the fires is through satellite imagery. Figure 1 displays the locations of fires that occurred during the second half of 1997, identified by light emissions recorded by the DMSP-OLS satellite (Fuller and Fulk 2000).

The consequences of the fires, however, were not limited to the areas of burning. The fires produced visibility-limiting haze that caused transportation slowdowns and accidents, shutdowns of schools and businesses, and health problems. Easterly and southeasterly winds spread haze from the fires over an area far larger than where the fires occurred.

This situation is reflected in Figure 2, which displays levels of haze on a particular day every two weeks from the start of the fires in early September to mid-November 1997. Haze is measured using an aerosol index calculated with data from NASA's Total Ozone Mapping Spectrometer (TOMS). A value of 0 indicates that the air is crystal clear, and a value of 4 corresponds to barely being able to see the midday sun; aerosol index levels that are 4 and higher are depicted by a black cloud in the figure. The figure highlights two important facts. First, the haze spread right across the southern part of Kalimantan (Borneo), covered all but the most northern part of the island of Sumatra, and spread up to Singapore and Malaysia for a short period. Estimates suggest that the haze covered over 300 million square kilometers. Second, the figure clearly demonstrates that the areas with high levels of haze varied during this period, with the haze building up in the last half of September and dissipating in early October; then, as the atmospheric pressure changed, both the area covered by haze and the intensity of the haze increased again until mid-November, when the fires started to die out. Whereas Java, Bali, Lombok, and Sulawesi were not affected by the haze, across Kalimantan and Sumatra, there was substantial variation in the timing, duration, and intensity of exposure to haze. This variation is important, since we exploited it in our analyses of the impact of the haze on health.

The magnitude of the 1997 fires is placed in a longer-term context in Figure 3, which reports the daily TOMS aerosol index from 1996 (when the satellite started recording data) through 2002. The capital cities of three provinces were selected because the provinces were blanketed with haze, and they are included in the survey data we used. The figures illustrate a key point: 1997 was nothing short of a catastrophe. The TOMS aerosol index peaked close to 6 in Sumatra and slightly over 5 in South Kalimantan. These levels are unprecedented in recent history. The figures also underscore the spatial and temporal heterogeneity in the haze.

Both Figures 2 and 3 indicate that levels of haze in Kalimantan and Sumatra were significantly worse than the levels in Singapore and Malaysia. Other data corroborate this evidence. The most general ground-based measure of the hazard that smoke presents for health, referred to as the PM^sub 10^ measurement, reflects the number of particles with a diameter of less than 10 micrometers (a size that can enter the respiratory tract) per cubic meter of air (m/m^sup 3^). Measurements in Jambi (on the east coast of Sumatra) in early October documented particulate concentrations of 1,864 m/m^sup 3^-a level three times higher than that at which the U.S. Environmental Protection Agency (EPA) warns people to use respirators during unavoidable outdoor activities (Kunii et al. 2002). Measurements taken a month later in Palembang, Sumatra, indicated PM^sub 10^ levels of 402, which generate EPA warnings that healthy people should curtail vigorous activities (Pinto and Grant 1999). In contrast, in Singapore and Malaysia, PM^sub 10^ levels in September averaged below 200 m/m^sup 3^ (Emmanuel 2000; Sastry 2002).

HEALTH CONSEQUENCES OF EXPOSURE TO HAZE

Ambient air quality, which reflects the presence of both particulate matter and gaseous compounds, has been associated with increased risks of mortality and respiratory morbidity in numerous studies. Although the associations are strong, the precise biological pathways through which exposure to poor-quality air affects health have not been fully delineated. Both the size of the particles and the chemical composition of the particles and the gasses appear to be relevant (Churg and Brauer 2000; Harrison and Yin 2000). With respect to size, all particles with a diameter of less than 10 micrometers (m) can enter the respiratory tract, but those with a diameter of less than 2.5 m ("fine" particles) are a particular concern because they are small enough to penetrate deeply into the lungs, enter the bloodstream, and be transported to other tissues (Malilay 1999; McClellan 2002). With respect to composition, attention has focused particularly on carbon monoxide, nitrogen dioxide, lead, sulfur dioxide, and polycyclic aromatic hydrocarbons. Less is known about the roles of aldehydes, free radicals, and volatile organic compounds.

Although the role of gaseous compounds has sometimes been considered, particularly in studies of outdoor air pollution, we focus our discussion on the health implications of exposure to particulates because most of the evidence from the Indonesian fires suggests that fine particulates were elevated to far more dangerous levels than were gaseous compounds (Kunii et al. 2002; Pinto and Grant 1999). Studies that have considered the health impacts of exposure to particulates can be divided into three general types with respect to the source of exposure: outdoor air pollution resulting from routine activities that involve the combustion of fossil fuels, such as the operation and manufacturing of vehicles; indoor exposure from routine activities that involve the combustion of biomass fuels, such as cooking or heating; and exposure from a catastrophic event, such as a forest fire, building fire, volcanic eruption, or explosion. Next, we discuss the literature on mortality and respiratory morbidity-the health outcomes on which most analyses have focused.

Mortality

Exposure to particulates appears to elevate the risk of mortality. Most of the evidence for this relationship has come from studies of air pollution, which often has a chemical makeup similar to biomass smoke (because both involve combustion processes), although exposure to air pollution is generally at a lower level over a longer period.

Using data from the Harvard Six Cities Study, Dockery et al. (1993) found that total suspended particulate (TSP) levels were significantly associated with increased mortality in each of six U.S. cities. Chay and Greenstone (2003) used variations in air quality attributable to the 1981-1982 recession to identify the effects of pollution on infant mortality in the United States. They found that a 1 mg/m^sup 3^ reduction in particulates results in about 4-8 fewer infant deaths per 100,000 live births at the county level. Cropper et al. (1997) studied the health effects of air pollution in Delhi, India. They presented evidence of a positive relationship between particulate air pollution and daily nontraumatic deaths, but the impacts were smaller than those estimated for other countries. They attributed the lower impact to differences in distributions of age and cause of death. Most deaths in Delhi occur before age 65 from causes that are not strongly associated with air pollution.

Evidence of the impact of the Indonesian fires on mortality has been mixed. Sastry (2002) found that daily increases in 1997 haze levels in two urban areas of Malaysia (Kuala Lumpur and Kuching) were associated with increased mortality rates for older individuals in these areas. For Singapore, however, Emmanuel (2000) found no significant increase in mortality during the 1997 haze.

Respiratory Morbidity

Breathing air with a high level of particulates can damage both the upper and lower respiratory tract, resulting in inflammation of the airways and conditions such as coughs, bronchitis, difficulty breathing, reduced lung function, and ultimately more severe obstructive breathing disorders (Chretien and Nebut 1996). Most of the evidence about the respiratory consequences of exposure to intense levels of particulates has come from studies that have been conducted during or shortly after short-term exposure. Little is known about the long-term effects of exposure to smoke.

Nor is much known about respiratory morbidity in Indonesia during the 1997 fires. Comparisons of routine data collected from governmental health facilities revealed an increase in cases of acute respiratory infection (ARI) and bronchial asthma between September 1997 and June 1998, relative to the same period in 1995-1996 (Aditama 2000). A study based on a convenience sample of some 600 Indonesians in Jambi in September 1997 (when particulate levels far exceeded the EPA's "hazardous" rating) found high reported levels of respiratory problems (91% of the interviewed respondents), shortness of breath when walking (44%), and shortness of breath with hard physical work (36%; Kunii et al. 2002). Another study that compared 127 high school students in two areas of Central Kalimantan (one with twice the levels of particulates of the other) found that the male students in the area of poorer air quality performed significantly worse than did the male students in the area of better air quality on physical assessments of lung function, although no difference emerged in the prevalence of bronchitis or bronchial asthma (Santoso 1998, cited in Aditama 2000).

Other studies have considered respiratory health in Singapore and Thailand, where levels of particulates were also elevated by the Indonesian fires, but to a considerably lesser degree. In Singapore, increases in PM^sub 10^ levels (from 50 to 150 m/m^sup 3^) were associated with a 30% increase in haze-related health conditions (upper-respiratory-tract illness, asthma, and rhinitis; Emmanuel 2000). In Thailand, a comparison at the time of the haze of outpatient visits in southern Thailand (where the air quality deteriorated) and northern Thailand (where it did not) revealed a relative increase in southern Thailand in both outpatient visits and inpatient admissions for respiratory conditions (Phonboon et al. 1999). In Malaysia, data from outpatient visits in Kuala Lumpur and Kuching indicated a rise in respiratory-related visits during the haze (Brauer and Hisham-Hashim 1998; World Health Organization 1998).

A number of studies in other settings have considered the respiratory consequences of exposure to smoke from fires. Two studies that were conducted in California suggested that exposure is associated with immediate increases in respiratory morbidity (Lipsett et al. 1994; Shusterman, Kaplan, and Canabarro 1993). Another study of fire victims immediately and three months after exposure indicated that whereas airway reactivity diminished as the duration since the fire increased, other aspects of lung function showed no improvement (Kinsella et al. 1991).

Some analyses have concentrated particularly on firefighters, who are regularly exposed to smoke. Generally, studies of wildland firefighters (for whom exposure tends to be seasonal) have pointed to an association between exposure to smoke and acute respiratory health and to the persistence of some symptoms after the firefighting season ends (Brauer 1999). Studies are now under way of firefighters at the World Trade Center (WTC) site, a number of whom developed persistent coughs during their work at the site (Prezant et al. 2002). Banauch et al. (2003) used data from a sample of firefighters who were exposed to the WTC disaster to show that for about 55% of those who were highly exposed, lung dysfunction that was documented one or three months after the event was still present at six months.

Exposure to indoor smoke from the combustion of cooking fuels has also been found to be strongly associated with both acute respiratory infections and acute lower-respiratory-tract infections (Bruce, Perez-Padilla, and Albalak 2000). Women and young children are particularly likely to be exposed because they tend to be indoors, especially when food is being cooked. In a carefully executed observational study conducted in Kenya, Ezatti and Kammen (2001a, 2001b) constructed measures of exposure to particulates from biomass fuels, based on longitudinal data collected over a two-year period. They showed that time spent with both ARI and acute lower respiratory infection (ALRI) is an increasing concave function of daily exposure to PM^sub 10^. In an effort to address directly whether indoor air pollution affects respiratory illness, a controlled experiment is now under way in Guatemala, in which the respiratory health of individuals in treatment households who have received cookstoves that have been shown to reduce the presence of particulates will be compared to that of individuals in control households in which open fires are used for cooking (Albalak et al. 2001; Smith 2004).

A substantial body of other work has considered the consequences for respiratory health of exposure to particulate air pollution (for an extensive review, see Pope, Dockery, and Schwartz 1995). Exposure has been associated with increased hospitalization for respiratory disease, exacerbation of asthma, the increased incidence and duration of respiratory symptoms, declines in lung function, and restricted lung activity.

Several difficulties complicate the interpretation of the results just described. A well-documented issue with respect to the studies of mortality is the question of whether elevated particulate levels simply hasten death among frail individuals for whom the end is already near.

With respect to the consequences of biomass smoke for respiratory morbidity, most of the studies have analyzed groups that have been selected in various ways. Two of the studies of Indonesia relied on small samples chosen by convenience. Several studies have analyzed administrative data from health facilities. In contexts where access to health care is limited, however, the group who chooses to seek medical care may be different from the group who does not. Firefighters, of course, are likely to be a particularly fit group of individuals. Although the studies have documented increases in respiratory morbidity that accompany exposure to haze, the results do not necessarily generalize to broader populations.

DATA

We combined data from a longitudinal household survey, collected in Indonesia in 1993, when there were no fires of note, and data collected in 1997, while the fires were burning, with information on the intensity of the smoke haze derived from satellite data. This combination provided unique opportunities to measure the effects of the fires on the health and well-being of the Indonesian population. We begin with a description of the household survey data and then turn to the measurement of smoke haze.

IFLS is an ongoing longitudinal survey of individuals, households, communities, and facilities. The first wave, conducted between August and December 1993, interviewed over 7,200 households in 321 enumeration areas on the islands of Sumatra, Java, Kalimantan, Sulawesi, Bali, and West Nusa Tenggarra, representing 83% of the Indonesian population (Frankenberg and Karoly 1995). Individual interviews were conducted with the household head, spouse, up to two children, and up to two other adults.

IFLS2, the first follow-up survey, was conducted in 1997. Between August and December 1997, the 321 enumeration areas were visited, the original household was located, and all household members were reinterviewed. If a household or an individual who was interviewed in 1993 had moved nearby (within 30 minutes by public transportation), the interviewer would attempt to conduct the interviews at the new location. Longer-distance movers were interviewed in late 1997 and early 1998 as long as their new locations were in one of the 13 IFLS provinces included in IFLS. Over 94% of the IFLS1 households were successfully reinterviewed in IFLS2 (Frankenberg and Thomas 2000).

In addition to basic demographic and economic characteristics of respondents, the IFLS collects detailed information on health. The respondents are asked to report their general health status (GHS), whether they have difficulty with activities of daily living (ADLs), and whether they experienced various symptoms in the month before the survey. Height and weight are measured as well. In 1997, additional physical assessments were conducted by trained health workers. Because these assessments are not available in the 1993 data and data from two rounds are necessary to assess the effects of the fires accurately, we did not analyze the effect of the fires on the physical health assessments.

As we discussed earlier, Indonesia's fires dramatically reduced air quality. Ground-based pollution monitors are one method of measuring air quality, but in Indonesia only a few cities have pollution monitors. Because of the limited coverage of ground-based monitors, we measured air quality with the aerosol index developed by NASA from the TOMS data.

The TOMS data offer several key advantages for this study. First, recent work has shown that the aerosol index is linearly correlated with ground-level aerosol optical thickness (AOT), which, in turn, is highly correlated with levels of TSP (Brimblecombe 1995; Hsu et al. 1996; Hsu et al. 1999; Torres et al. 2002). Second, aerosol levels have been measured on a daily basis since 1978 (although the failure of instruments resulted in the lack of data for Indonesia between mid-1993 and mid-1996). Third, the geographic coverage of the TOMS data includes all the locations of the households who were interviewed in the IFLS. We matched the TOMS data to the IFLS data on the basis of the latitude and longitude of each IFLS enumeration area, which was recorded with a handheld global positioning system. Because the TOMS data are available over time, we could precisely capture each individual's exposure to the smoke on several dimensions: level of smoke inundation, duration of exposure, and timing of the exposure relative to the IFLS interview.

Figure 4 displays the location of IFLS enumeration areas and whether these areas were exposed to smoke. Following standard practice, we defined an area to have experienced smoke if the TOMS aerosol index exceeded 1.5 for at least three days between July 1, 1997, and the interview date. On the basis of these criteria, exposure occurred in all the enumeration areas in Southern Kalimantan; Northern, Western, and Southern Sumatra; and some areas in Lampung and West Nusa Tenggarra. About 25% of the IFLS respondents lived in "haze areas."

It is possible that the fires could have elevated mortality or migration just before the survey. In fact, the respondents in "haze areas" were no more likely either to die or to move in the three months preceding the interview date than were the respondents in the "nonhaze" areas. The fires do not appear to have induced differential attrition.

MEASUREMENT AND INTERPRETATION OF HEALTH

Health status is difficult to measure. It is multidimensional, and an individual's perception of each domain of his or her own health reflects a complex combination of physical, psychosocial, phenotype, and genotype influences over the life course in conjunction with the individual's expectations and information about health. Moreover, self-reported health is conditioned by levels of and knowledge about the health of the reference group used by the respondent. These complexities are inherent in all interview-based survey questions about health and have been subjected to extensive inquiry. (See Murray et al. 2002 for a state-of-the-art discussion of the issues and King et al. 2004, e.g., for some recent proposals for anchoring self-reported health in surveys.)

We examined three indicators of adult health status, each of which is potentially affected by exposure to haze from the fires and was measured in both the first and second waves of IFLS. The indicators are whether the respondent had difficulty carrying out strenuous tasks, a specific morbidity related to respiratory function, and a general measure of overall health.

Our first health indicator is whether the respondent had difficulty carrying a heavy load, one of a battery of questions about difficulties the respondent had with ADLs. It has been argued that questions about ADLs are easy for a respondent to answer, since they ask about specific activities, such as walking a certain distance, climbing stairs, or carrying a heavy load, that are well defined and capture important dimensions of functional health. Moreover, ADLs have been shown to be predictive of later mortality (Reuben, Siu, and Kimapu 1992; Scott et al. 1997). Many of the standard ADLs are of the greatest salience for the elderly because they concern activities as basic as bathing. We focused on an item that is also relevant for prime-age adults and that provides information about the respondent's capacity to perform physically strenuous activities. If haze affects respiratory functioning, strenuous activities like carrying a heavy load are likely to be more difficult.

Adult respondents in IFLS were also asked whether they experienced a series of specific symptoms during the four weeks prior to the interview. Self-reported incidence of coughing, our second health measure, is indicative of respiratory problems that are affected by exposure to haze. The temporal framing of the questions likely points the respondent to compare the incidence at the time of the interview with the incidence more than a month earlier (for an insightful discussion of the importance of the temporal frame when interpreting self-reported morbidities, see Sanchez-Paramo and Das 2003).

Self-reported GHS, an indicator of overall health, is our third measure of health. Each adult respondent was asked to rate his or her own health as very good, good, fair, or poor. Whereas there is a clear mechanism through which exposure to haze will affect coughing and difficulty carrying a heavy load, an examination of the impact on GHS is intended to capture the effects on a broader set of health domains. We focused on whether the respondent reported being in poor GHS, which has been shown to be a powerful predictor of subsequent mortality in a wide array of settings (see Idler and Benyamini 1997). This is true even after physicians' reports of health problems are controlled, suggesting that GHS contains information that may not be readily observed by a physician, such as an individual's health-related behaviors, own health history, and family's health history.

The percentage of adult respondents aged 30 and older who reported having each health problem in the 1997 wave of IFLS is presented in the first column of Table 1. All three are common: 1 in 5 adults reported difficulty carrying a heavy load, 1 in 6 reported being in poor general health, and over one third reported coughing in the previous month. The interpretation of any self-reported health indicator is complicated because the meaning of each question may vary across the respondents. Several studies have shown that higher-income (and arguably healthier) people report themselves as being in poorer health than do lower-income (less-healthy) people (see, e.g., Murray and Chen 1992; Sadana et al. 2002). To provide some insights into the quality and nature of the information contained in the three self-reported health indicators used in this study, we relate them to two physical health assessments: lung capacity (measured by a puff test, in which the respondent blows into a plastic tube) and the timed sit-to-stand-that is, the time it takes for the respondent to stand from a sitting position (repeated five times, as fast as he or she can). Lung capacity has been shown to be diminished by exposure to air pollution, and respiratory problems are likely to affect the timed sit-to-stand.

The second column of Table 1 presents the results from a regression relating lung capacity to the three health indicators (and a control for gender of the respondent). Both lung capacity and timed sit-to-stand were converted to z statistics so that the regression coefficients can be interpreted as standard deviations of change in the dependent variable. All three self-reported health indicators are significantly negatively correlated with lung capacity, and over one third of the variation in measured lung capacity is explained by these three health indicators (and gender). The lung capacity of a respondent who has difficulty carrying a heavy load is 0.4 standard deviation lower than that of a respondent who has no such difficulty. This difference is significantly larger than the decline associated with being in poor GHS, which seems reasonable since lung capacity is likely to be more closely related to difficulty with strenuous activities than to overall health status. Having a cough has the smallest effect on lung capacity, which suggests that coughing in the previous month was a transitory problem for most respondents. The results for the timed sit-to-stand, in the third column of the table, are qualitatively similar (a longer time indicates poorer health).

Both lung capacity and sit-to-stand assessments were conducted in the respondent's home by a trained health worker, who was usually a nurse, and so are not contaminated by individual-specific self-reporting biases. However, both involve participation by the respondent, and people who are inclined to report that they are in poor health may also be inclined not to try as hard on these assessments. These linked propensities would result in spurious correlations between the self-reported and physically assessed health indicators. In the final column, we turn to a health measure that involved no participation by the respondent at the interview and that is indicative of overall health status: whether or not the respondent was alive three years after the 1997 interview. All three self-reported health indicators are significant predictors of three-year mortality, with difficulty carrying a heavy load continuing to be the best predictor. Coughing has only a modest effect, again suggesting that it was a transitory problem for most respondents. Clearly, all three self-reports of health contain information about the GHS of the respondents and about specific domains that are related to exposure to haze.

The fact that difficulty carrying a heavy load is a better predictor not only of the physical health assessments but also of three-year mortality suggests that it may be less subject to respondent bias than is GHS. There is evidence in support of this interpretation. First, many studies have demonstrated that socioeconomic status (SES) is a powerful predictor of multiple dimensions of health status, including physical assessments and mortality. SES is also strongly predictive of ADLs, including difficulty carrying a heavy load. However, SES is not as highly correlated with GHS, and, in some studies, higher-SES individuals have reported being in worse general health than have lower-SES respondents (for discussions, see Strauss et al. 1993 and Sadana et al. 2002; for evidence from the IFLS, see Thomas and Frankenberg 2002). One reason suggested for this observation is that relative to the concrete task of carrying a heavy load, GHS is less well defined, and the meaning of "poor" health likely depends on the reference population against which a respondent compares his or her own health. If the reference population is the entire population, the meaning of "poor" should be the same for everyone. However, if it depends on the people with whom one has contact, then higher-SES respondents will tend to expect a higher standard of health.

The importance of information about health was highlighted by Dow et al. (1997), who presented evidence from two randomized experiments in which the user fees for health care services were changed for persons in the treatment groups relative to those in the control groups. One experiment was conducted in the United States, the other in Indonesia. In both cases, among those for whom the price of care was lower, the use of health care services increased, ADLs improved, but GHS worsened. Dow et al. suggested that seeing a health professional likely changes information about one's own health and may affect one's reference level of health, which is reflected in GHS. ADLs are apparently less prone to these effects.

The role of changes in the health of a reference population is discussed in Thomas et al. (2004), who reported the results from another treatment-control experiment in Indonesia. Households were randomly assigned to receive either iron supplements or identical-looking placebos. After a year of supplementation, adults in the treated group were in better health than were those in the control group. They had higher levels of iron in the blood and reported lower levels of fatigue and less difficulty with ADLs, including carrying a heavy load. Self-reported GHS, however, did not differ between the treatment group and the control group, suggesting that as the health of others in one's household (and community) changes, so does one's own reference health level.

In sum, all three self-reported health indicators-difficulty carrying a heavy load, coughing, and poor GHS-are predictive of physical health and subsequent mortality. They clearly provide valuable information about the health of respondents. That their interpretation is not straightforward, however, will be discussed later in the article. We turn now to a discussion of the empirical methods and results.

METHODS AND RESULTS

We examined the effect of exposure to haze from the fires on the health of adults aged 30 and older. The literature suggests that older people are more susceptible to the deleterious effects of smoke haze. Thus, we examined adults aged 56 and older separately from prime-age adults (aged 30-55). Among prime-age adults, women are more likely to have been exposed to indoor pollution (from cooking, for example), which may affect their susceptibility, whereas men are more likely to work outdoors and on physically arduous tasks and so may be more exposed to smoke haze. The analyses of these adults are stratified by gender.

The relationship between exposure to haze and difficulty carrying a heavy load is reported in Table 2 for older adults. In 1997, over 50% of older adults who were exposed to haze reported this difficulty, which affected less than 40% of those who were living in areas that were not affected by the fires. The difference between these groups is one potential measure of the effect of the haze on this dimension of health. It is both large (15%) and significant. However, this estimate of the effect of the fires is predicated on the strong assumption that exposure to haze was spread randomly across the population. The validity of this assumption can be tested. If differences in health status between the haze and nonhaze areas reflect the impact of the haze, rather than other differences, then the incidence of difficulty carrying a heavy load, as reported before the haze in 1993, should be the same for respondents in the haze and nonhaze areas. As is shown in the second row of Table 2, the assumption is false. Older people who were living in areas that were affected by haze in 1997 reported that they were in worse health in 1993 than did the rest of the older population.

The change in health in haze areas between 1993 and 1997, in the first column of the third row of the table, reflects the combined effect of aging of the respondents, exposure to haze, and any other changes that occurred during this period. The aging of the sample is common to both the haze and nonhaze areas and, to the extent that other changes are similar, the difference between the change in health in haze areas and the change in health in nonhaze areas yields an estimate of the effect of the fires on health that controls for unobserved differences between the haze and nonhaze areas. This "difference-in-difference" estimate is reported in the third column of the third row and indicates that 5.6% of the older population had more difficulty carrying a heavy load because of the fires.

That estimate may be contaminated by differences between respondents who were and were not exposed to haze. If all these differences are observed in the data, they can be controlled in a multivariate regression context. However, these estimates will also be biased if there are unobserved differences between respondents who were exposed and those who were not. If the differences are fixed over time and affect health outcomes in a linear and additive way, then the inclusion of a person-specific fixed effect in the regression model will absorb their influence, and the estimates will not be contaminated by this form of observed or unobserved heterogeneity. Intuitively, the change in health of an individual who was exposed to haze in 1997 is compared with the change in health of an individual who was not exposed, controlling for all fixed observed and unobserved differences between these two individuals.

The fixed-effects estimate of the effect of haze on difficulty carrying a heavy load is in the fourth row of the table. It is 5.3% and significant at the 5% level. The fact that this estimate is close to the difference-in-difference estimate suggests that our identification strategy is robust to several sources of potential bias that are due to unobserved heterogeneity. In addition to controlling for unobserved heterogeneity, the fixed-effects difference-in-difference estimates have two important advantages. First, they are more efficient (as demonstrated in Table 2). Second, biases in self-reported health that arise from differences in the propensities of respondents to report themselves as being in poor health will be absorbed by the fixed effect as long as the individual's reporting propensity does not change over time. The rest of this article focuses on fixed-effects estimates of the effect of the haze; the difference-in-difference estimates are similar in all cases.

Difficulty Carrying a Heavy Load

Table 3 presents fixed-effects estimates of the effects of haze for each of the three self-reported health indicators and for the three demographic groups. In each block, the results for older adults are in the first column, for prime-age women in the second column, and for prime-age men in the third column. We begin with difficulty carrying a heavy load.

The regression in Panel A measures the effect of exposure to haze, controlling for individual fixed effects and observed differences in household resources and the location of the respondent. The first column in the first row repeats the estimate for older adults discussed earlier. For older adults and prime-age women, exposure to haze results in worse health, as indicated by higher levels of reported difficulty carrying a heavy load.

Figure 2 highlighted the fact that the smoke and haze spread across Indonesia in two major waves. Some of the respondents who had been exposed to haze prior to the 1997 interview were no longer exposed by the time of the interview. The regression in Panel B distinguishes among respondents who were exposed at the time of the interview, those who were interviewed at least a month after the haze and smoke had cleared, and those who were not exposed (the excluded group). Whereas older adults had more difficulty carrying a heavy load while exposed to the haze, there is no evidence of longer-term effects. In contrast, the effects on prime-age women persisted even a month after the haze had cleared.

The regression in Panel C further disaggregates the timing of exposure to haze. We separately identify respondents whose exposure to haze began no more than 30 days prior to the interview, those who were exposed at the time of the interview who had been exposed for at least a month, those who were not exposed at the time of the interview but had been exposed within the previous 30 days, and those who had been exposed prior to the interview but more than 30 days earlier. The key novel result from this specification is for prime-age men, who reported more difficulty carrying a heavy load at the onset of haze, but the effect quickly disappeared. There are at least two plausible interpretations. Either the prime-age men became accustomed to the haze after a month of exposure, or they adjusted what they thought of as being "difficult" if the exposure persisted.

In sum, haze has a deleterious impact on the ability of all adults to carry out strenuous tasks. Among prime-age men, the effect is short lived; among older adults, it persists until the haze has cleared; and among prime-age women, it persists for at least a month after the haze has cleared.

Respiratory Problems

The second block of Table 3 reports the fixed-effects difference-in-difference estimates of the effect of exposure to haze on the incidence of coughing in the previous 30 days. The first row indicates that older adults who were exposed to haze prior to the 1997 interview were 9% less likely to report coughing, and prime-age adults were 5% less likely to do so. If the respondents coughed more when they were blanketed by haze, relative to before the haze or after it cleared, then it is important to distinguish those who were exposed at the time of the interview. This is a particularly relevant point for this health indicator, given that the temporal framing of the question suggests that the respondents compared coughing at the time of the interview with coughing a month before the interview.

The regression in Panel B indicates that the lower levels of reported coughing among those who were exposed to haze arose because people who were exposed to haze that cleared more than a month before the interview were much more likely to report less coughing than were those who were not exposed to haze and those who were not exposed at the time of the 1997 interview.

The regression in Panel C demonstrates that there was an 8%-9% higher level of reported coughing among prime-age adults at the onset of haze. There is no difference between the effects on men and women, and when the groups are combined, the effect is significant at the 5% level. If respondents who were currently exposed to haze and had been exposed for at least a month compared their coughing in the past 30 days with coughing prior to that time, they were not likely to be different from those who were not exposed to haze. This likelihood is reflected in the small and insignificant coefficients in the second row of Panel C. However, when the haze cleared, the respondents apparently noticed that they were coughing less and reported this fact in the interview. We conclude that exposure to haze results in elevated levels of coughing, but the effects are short lived.

General Health Status

The effect of haze on whether the respondent reported that he or she was in poor general health are displayed in the third block of Table 3. For older adults and prime-age women, exposure to haze resulted in fewer respondents reporting they were in poor general health. Haze had no impact on the GHS of prime-age men.

Panel C reports the time path of the estimated effects of the haze. Among prime-age women, reported GHS is no different for those who were exposed to haze at the time of the interview than for those who were never exposed. However, after the exposure ended, prime-age women were 11% less likely to report they were in poor general health, and the effect persisted for at least a month after the haze cleared. The results for older adults are similar except that they were also less likely to report that they were in poor general health after they had been exposed for at least a month.

The lower rate of reported poor health, after the haze cleared, is consistent with the evidence for coughing discussed earlier. The results suggest that these respondents compared their health at the time of the 1997 interview, when the haze had cleared, with their health prior to the interview, when they were blanketed by haze and were in poorer health. Therefore, at the interview, they were less likely to report that they were in poor health.

This interpretation suggests that people who have recently been exposed to haze should be more likely to report they are in poor general health. They do not. However, recall the discussion of self-reported health indicators in the previous section, which suggested that GHS is likely to be influenced by both a respondent's own prior health experiences and by the health of those around him or her and that the relative salience of these references is likely to shift as circumstances change.

At the onset of the haze, a good deal of discussion probably occurs in the community about its effects on health. Thus, when the haze is a relatively new phenomenon, the health of others in the community is likely to be particularly salient in one's assessment of one's own health. But because everyone in the community is likely to have been affected by the haze, on average, reported GHS will be no better or worse among individuals in communities that are affected by haze than among individuals in communities that are not affected by haze. In other words, within communities, individuals' relative positions with respect to general health have not changed, although absolute levels of health are lower in the haze communities than in the nonhaze communities, as indicated by increased difficulty carrying a heavy load.

The simple comparisons of those who were exposed to haze and those who were not, in Panel A of the table, suggest that the haze resulted in improved general health. A more nuanced examination of how reported GHS varies with the timing of exposure to haze indicates that the haze had a deleterious impact on GHS, which is consistent with our evidence on coughing and carrying a heavy load.

CONCLUSIONS

The fires in Indonesia in late 1997 were an environmental disaster. The effects of those fires on health have been difficult to quantify because of a paucity of survey data on the health status of individuals. We combined information collected in health interviews in the IFLS with satellite measures of aerosol levels and examined the effects of exposure to haze on three domains of health status.

We found that comparisons of the health of the population living in haze areas with the health of those in other areas substantially overestimated the "effect" of the fires because of time-varying location-specific unobserved heterogeneity in health status. Consequently, we exploited the repeat-observation nature of IFLS and compared changes in the health of the two groups.

The haze had an immediate deleterious impact on physical functioning, as measured by self-reported difficulty carrying a heavy load. The effect dissipated quickly for primeage men, persisted until the haze cleared for older adults, and lasted several months after the haze cleared among prime-age women.

The incidence of reported coughing was higher at the onset of the haze and much lower a month after the haze cleared, indicating that the haze had a substantial negative effect on respiratory health. Noting that the interpretation of GHS is likely to be influenced by both a respondent's own prior health and the health of those around him or her, we interpreted the evidence on GHS as indicating that the fires resulted in substantially poorer general health among prime-age women and older adults.

In addition to shedding light on the health consequences of exposure to haze, the evidence presented here provides insights into the nature of data that are necessary to measure the effects of changes in the environment-whether they are economic, social, ecological, or political. We have shown that comparisons of groups that are based on cross-sectional data are fraught with difficulties and can be seriously misleading. High-quality longitudinal survey data that can be matched with administrative or other data sources are of tremendous value in this context. It is also clear that the interpretation of self-reported health status is not straightforward and that the collection of measures and biomarkers of physical health is likely to have substantial benefits.