Although the Indonesian pollution episode in 1997 was primarily driven by meteorological phenomena and such phenomena are usually considered fully exogenous and good sources of identification, we turn our attention to several issues that could potentially impact the validity of our estimates.
Attrition
Attrition is usually a big source of concern when dealing with longitudinal surveys like the IFLS, especially when these surveys span such a long period of time. If attrition is random, or in our case if it does not correlate with pollution, then it is not a problem. However, it is possible that respondents leave the sample systematically. If pollution affects health and the most affected respondents leave the sample, then our results will be biased. In our case, sample attrition can occur when respondents die or retire or when the IFLS loses track of them.
As mentioned in the data section, attrition rates in the IFLS are very low overall, which is encouraging for our study. Furthermore, even if we accept the possibility of a certain attrition bias, this bias would actually strengthen the qualitative implications of our estimates. If pollution were to lead to attrition, then overall we would be left with a slightly healthier sample and our estimates would be upward biased. In other words, the true impact of pollution would be actually greater (more negative) than our estimates might suggest.
We nevertheless proceeded to investigate sample attrition between 1993 and 2007. Approximately 40% of the respondents left the sample between 1993 and 2007. About 46% of these respondents however were over 55 years old in 1997 and about 34% were over 60 years old in 1997. This would suggest that most of the sample attrition comes from old retirees, rather than due to pollution.
We also ran a linear probability and a probit regression to test if the attrition status depends on the pollution exposure. We coded a dummy variable equal to 1 if the respondent was in the sample in 1993 but not in 2007 and 0 if the respondent was present in both years. We regressed this dummy variable on our main pollution variable, age and age squared, a sex dummy, and the presence of an outside kitchen and water source. The coefficient of the pollution variable was statistically insignificant, which strengthened our belief that attrition bias is not a concern for our estimates. We present these two regression in Table 6 in the Appendix.
Economic confounders
While the 1997 fires were undoubtedly a major pollutant in Indonesia, these fires might have had additional effects on the local economy that could have resulted in a reduction of hours worked outside of the pollution channel described in Section 3. Pollution could be for instance correlated with economic conditions in the forestry or agricultural sectors, since arguably, both these sectors were affected by the fires—agricultural land was cleared at the expense of losing the forests. These economic conditions might have affected local demand for sector-specific labor, and it is possible that our estimates pick up some of those effects as well. We tried to address some of these issues by including additional community-level controls, but this is a limitation of our study that needs to be more carefully investigated by further research.
Ideally, this issue should be addressed by considering pollution variation within local economies and very detailed data on local economic conditions, both of which we do not possess. Using community or even regional fixed effects is also not doable since the pollution episode in Indonesia occurred on such a large scale that even regional fixed effects confound with the pollution variable and render it insignificant.
The IFLS does include, however, a number of community-level variables that point to certain economic conditions that might correlate with the fires. These variables are constructed from surveying the community leaders and could suffer from mis-measurement and a dose of subjectivism. Community leaders were asked in 2007 whether their communities were still suffering from the financial crisis7, whether their communities had any wood processing or wood-related factories, whether they experienced drought and the drought frequency, and the percentage of their communities covered by forest and by farming land. We included all these variables together with a community population variable as additional controls and found very little change in our estimates. We present these results in Table 7 in the Appendix.
None of these additional controls are statistically significant, with the exception of the percentage of the community that was farm land. The inclusion of the first four controls leave the estimate of pollution virtually unchanged, while the inclusion of the farm land variable slightly decreases the point estimate of pollution from −0.87 to −1.06. This change in the point estimate is statistically significant and consistent with the economic channel hypothesis. Since fires clear land for agriculture, higher pollution will correlate with more agricultural land which will lead to more hours worked in the absence of the negative effects of pollution. This bias is removed when we control for agricultural land, and the true impact of pollution seems to be more severe than initially estimated.
While the droughts, financial crisis, or the forestry sector do not seem to play roles, there are some spillovers from the fires in the agricultural sector. However, they only seem to strengthen our initial estimates. Arguably, such spillovers can occur in other sectors as well and future research should try to correct for these whenever possible. Studies using finer measures of pollution and more detailed local economic data might be more successful at addressing these concerns.
Migration
Temporary migration and avoidance is another possible cause of concern. If wealthier people decided to temporarily leave their communities to avoid the pollution episode, this could potentially bias our estimates. To tackle this potential problem, we looked at the respondents who migrated between September 1997 and September 1998 and tested whether their migration was determined by the pollution levels or their economic status.
Just like with attrition, we coded a dummy variable equal to 1 if the person migrated during the specified period and 0 otherwise. We then regressed this dummy variable on the 1997 pollution level, the log of household per capita expenditure (PCE), age and age squared, having an outside kitchen and water source, a sex dummy, and the household size. We present these results in column 2 of Table 8 in the Appendix. We found that neither the pollution nor the economic status significantly affects the migratory status. Furthermore, only about 1.7% of respondents were found to have migrated between 1997 and 1998.
Furthermore, we employed another robustness check involving migration, similar in nature to the method described in Jayachandran (2009). We re-estimated our medium-term regression from Table 2, column 4, but instead of matching the pollution data with the community of residence in 1997, we matched it with the community of residence in 2000. If certain people were able to avoid the pollution episode by simply migrating to another area or if the observed migration was due to the pollution episode, we should be able to estimate different coefficient using the pollution exposure at the new location of the migrants. We report these coefficients in the third column of Table 8 in the Appendix. We find virtually identical results with those from our main specification, which further confirms that migration is not a cause of concern.
Current pollution
Another possible confounder is the current pollution level. If certain areas of the country are more predisposed to pollution, it is possible that the past pollution from 1997 is correlated with the current pollution. Since current pollution has already been shown to affect labor supply, omitting this variable could create a bias in our estimates, and furthermore, we might wrongly attribute the effect of current pollution to the past pollution. In that respect, our results would no longer represent long-term effects of pollution, but short-term effects.
To investigate this possibility, we re-estimated our main specifications from Table 2, columns 4 and 5, and included the pollution level from 2000 as an additional regressor. We reported this in Table 9 in the Appendix. We found virtually identical coefficients for the 1997 pollution, while the 2000 pollution was statistically insignificant. While we were not able to replicate this robustness check for our long-term effects since we do not have pollution data in 2007, we believe the medium-term results clearly show that the current pollution and its possible correlation with the past pollution is not a cause of concern.
Most of the issues we covered, such as the financial crisis or migration, were also found to be insignificant in papers like Jayachandran (2009), although for a different outcome variable. While most of these robustness checks leave our estimates unaffected, it is important to acknowledge the possibility of additional channels that could affect hours worked and be correlated with the pollution spike in 1997. The IFLS provides the longitudinal data needed to study the long-term effects of pollution, but it does not go very deeply into recording detailed economic conditions that could serve as additional controls.