Air pollution and internal migration: evidence from an Iranian household survey

The dependent variable of the study was the net outmigration ratio (Gawande et al. 2000), defined as the population leaving two major cities of a province minus new arrivals, divided by their combined total population. We collected data on the inflows and outflows of migration from two major cities in each province of Iran, available from the 2011 and 2016 National Population and Housing Censuses (conducted by the Statistical Center of Iran³). The choice of the 2011 and 2016 censuses for the analysis was based on the availability of inflow and outflow migration data. In addition to the availability of data, we focused on the two major cities because most economic and migration activity in each province occurs in these cities. Aerosol optical depth (“AOD”) also measures the air pollution in the included major cities of each province. We used the total number of migrations in the province’s two major cities as a measure of migration for each province. Figure 1 shows the location of each sample city on Iran’s map.

The net outmigration ratio varied from a minimum of − 0.08 (Semnan in 2016) to a maximum of 0.06 (Lorestan in 2011 and Chaharmahal Bakhtiari in 2011), with an average of − 0.003. In total, 50% of the provinces have a net outmigration ratio of less than zero, while the net outmigration ratio of others exceeds this level. Table 6 in Appendix provides detailed descriptive statistics for the various quartiles of the net outmigration ratio.

The main explanatory variable of interest, AOD, measures air pollution. Aerosols are very small solid and liquid particles that are suspended in the atmosphere. Examples of aerosols include dust, sea salts, volcanic ash, smoke from wildfires, and pollution from factories. The inhalation of aerosols is often associated with increased morbidity and mortality. Long-term excessive exposure to aerosols leads to higher risks of lung cancer, chronic obstructive pulmonary diseases, ischemic heart disease, and strokes (Traboulsi et al. 2017).

We used remote sensing data to measure AOD across the provinces of Iran. In particular, we employed the moderate resolution imaging spectroradiometer (MODIS), which is a key instrument aboard the Terra (originally known as EOS AM-1) and Aqua (originally known as EOS PM-1) satellites. The Terra MODIS and Aqua MODIS constantly monitor the entire Earth’s surface. They collect information on global dynamics and processes occurring on land, in the oceans, and in the lower atmosphere.⁴ They have been measuring global aerosol optical depth and optical properties since 2000 (van Donkelaar et al. 2010).

MODIS aerosol products are freely available and are being used in several studies for different purposes. MODIS AOD has shown a good relationship with ground measurements (Mardi et al. 2018; Shi et al. 2019). The data to measure AOD in our study of Iranian provinces were sourced from NASA Earthdata.⁵ The average AOD values in the study area are illustrated in Fig. 2. It shows that the southern part of Iran has suffered more from air pollution compared to the northern part. Dust storms are the main source of AOD in Iran, except in metropolitan cities (e.g., Tehran and Isfahan), where urban pollutants typically dominate. The deserts in Iraq, Saudi Arabia, and Syria, as well as local soil erosions are the main dust sources in Iran. Therefore, the southern and western provinces of Iran have higher AOD than the northern provinces. Based on Fig. 2, Zahedan, Zabol, Ahvaz, Khorramshahr, and Urmia were the most polluted cities in Iran during the period of our study.⁶

The AOD measure changed from a minimum of 0.16 (Ardebil in 2016) to a maximum of 0.41 (the oil-based province of Khuzestan in 2011), with an average of 0.26. Of the provinces in our sample, 50% have AOD values below 0.25, while the AOD values of other provinces exceed this level. Table 6 in Appendix provides a detailed set of descriptive statistics for the various quartiles of AOD. For 2011, we used the average AOD of the 2007–2011 period. For 2016, we used the average AOD of the 2012–2016 period.⁷

Figure 3 shows the scatter plot of the net outmigration ratio and AOD. The positive slope of the trend line illustrates our main result. To show the positive link, we also provide a numerical example. When the annual mean AOD of Khorramshahr (in the Khuzestan province) and Zabol (in Sistan and Baluchestan province) increased from 0.36 and 0.29 in 2011 to 0.43 and 0.35 in 2016, respectively, their outmigration also increased from 7905 to 9872 and from 16,534 to 21,772, respectively. In contrast, in Nowshahr (in the Mazandaran province), when the annual mean AOD was stable at approximately 0.20 between 2011 and 2016, outmigration changed only slightly from 6445 to 6619, which is consistent with changes in AOD.

In addition to our variable of interest (AOD), we also controlled for other key determinants of internal net outmigration. The choice of control variables was based on our literature review of the determinants of internal migration and the availability of data for the provinces of Iran. Our control variables are the level of economic activity in the source province, unemployment rate, amenities, environmental quality indicators, and income inequality. Table 1 provides a detailed description of the variables and data sources. Descriptive statistics are presented in Table 2. In the following section, we briefly explain the relationship between each control variable and migration.

We apply a panel fixed effects model for the estimations.⁸ Equation (1) presents the empirical model. The fixed effects model accounts for all time-invariant differences across provinces, such as cultural, traditional, or geographical factors that may affect both air pollution and net outmigration. Provinces of Iran differ not only because of their stage of economic development, but also because of the high degree of ethno-linguistic fractionalization in Iran. Different cultural backgrounds may shape various aspects of the social and economic life of households. For example, the issues of energy consumption, attention to natural resources, and environmental protection also have cultural drivers (e.g., Soyez 2012; Horlings 2015). Cultural norms and differences in risk attitudes may also explain and form migration behavior within and across provinces. In addition, geography may shape the pattern of energy consumption, transportation, and urban life, which may also influence AOD. At the same time, geography may influence the allocation of the budget by the state and investment rate. Attractive places may accommodate a larger number of investors in various economic sectors (e.g., real estate and tourism) than less geographically attractive locations. The latter may then influence an individual’s decision to migrate. For these reasons, we include province fixed effects in our estimations. In fixed effects regressions, we estimate the effect of within-province changes in our explanatory variables on within-province changes in our outcome variable, removing the possible effects of any province-specific time-invariant characteristics. This approach reduces the risk of spurious correlation between net outmigration and AOD. It is noteworthy that several studies on determinants of international and regional immigration have also applied fixed effects estimators (e.g., Clark et al. 2007; Mayda 2010; Arif 2020).where EMIGOUT_RATIO is outgoing migration minus incoming migration in two major cities of province divided by their combined population; AOD is the aerosol optical depth (as a proxy for air pollution); X is a vector that includes the control variables; v_i captures the province-specific effects; i = 1,…; n denotes the province; t = 1, …; t denotes the time period; βs are coefficients; and u_it is an error term.⁹ Following de Chaisemartin and D’Haultfoeuille (2020), Imai and Kim (2021), and Kropko and Kubinec (2020), we decided not to include time fixed effects because using both province and time fixed effects can obscure interpretation.¹⁰ We use robust standard errors in reporting t-statistics. These standard errors are adjusted for 31 clusters at the province level. They are robust against heteroskedasticity and/or serial correlation in the idiosyncratic error. We also test for panel cross-sectional dependence. It is commonly assumed that disturbances in panel data models are cross-sectionally independent, especially in samples with large number of cross-sectional dimension. However, the existence of cross-sectional dependence maybe present in panel regression. Residual dependence may result in estimator efficiency loss and invalid test statistics. We conducted residual cross-sectional dependence test based on Baltagi, Feng, and Kao Bias-corrected Scaled LM (Baltagi et al., 2012) and Pesaran CD (Pesaran, 2021), which is appropriate for the short time and large cross-sectional dimensions as in our study, after estimating the general model (Model 7 in Table 3). The results show that we cannot reject the null hypothesis of “no cross-sectional dependence (correlation) in residuals” with p-values of 0.80 and 0.47, respectively.

The above findings are based on the panel fixed effects estimation method. As a robustness check, we re-estimated Eq. 1 using a two-stage least squares (2SLS) estimator, which is one of the most common instrumental variable estimators. The reason for employing the 2SLS estimator is to account for possible endogeneity in the model due to the likely reverse causality issue (e.g., bidirectional relationship between air pollution and outmigration). For example, we argue that local air pollution increases outmigration; however, it can be the case that increasing outmigration reduces air pollution via a decrease in the labor force and economic activity. In such cases, we may experience a simultaneous bias. To address this concern, we treated our measure of local air pollution (AOD index) as an endogenous variable and used an instrumental variable procedure to estimate the exogenous impact of AOD on the net outmigration ratio.

To instrument the AOD variable, we used province average annual temperature and the standardized precipitation-evapotranspiration index (SPEI)¹² (with larger values indicating higher degrees of moisture). We assumed that these climatic variables are exogenous to net outmigration and can affect our outcome of interest via their effects on AOD. There is evidence on weather conditions (e.g., temperature, sunlight, precipitation, wind speed, and humidity) affecting air quality (Liu et al. 2020; UCAR 2020), which supports the relevance of these instruments for the AOD index.¹³

The results of the 2SLS estimator are presented in column 8 of Table 3. In line with our findings from the fixed effect estimations, AOD is positively and significantly related to EMIGOUT_RATIO. The estimated size of the coefficient increased even after controlling for possible endogeneity of AOD. A shift from the 1st to the 3rd quartile of the AOD index increased the net outmigration ratio (controlling for other factors) by an estimated size of 0.63 * 0.06 = 0.0378. Considering the standard deviation of the net outmigration ratio in our sample (0.0287), higher air pollution based on Model 8 in Table 3 leads to an increase in net outmigration from the province by 1.3 standard deviation.

The diagnostics also show both the relevance and validity of the employed instruments. To test the null hypothesis that excluded instruments are irrelevant (under-identification of the model), we relied on the Kleibergen–Paap rk LM statistic with a value of 8.72 and a p value of 0.01. Therefore, we can strongly reject the null hypothesis of under-identification. To examine the weak instruments issue, we used the Kleibergen–Paap rk Wald F statistic, which tests the null hypothesis that excluded instruments are weak (weak identification). The Kleibergen–Paap Wald rk F statistic (12.59) was above the 15% critical value for maximal IV size (11.50). Overall, we reject the weakness of the instruments. After gaining confidence in the relevancy condition of our instruments for AOD, we examined the exogeneity of the instruments, which is important for the validity condition. We used the Hansen J-test (Hansen 1982). The null hypothesis that the instruments are valid (i.e., they are not correlated with the error term and are correctly excluded from the equation) cannot be rejected at a 5% level of significance (p value: 0.19).

Are the results presented in Table 3 due to outliers or influential observations? To check the possible outliers or influential observations, we excluded the capital city, Tehran, from estimations and re-estimated our model, as this province has a specific position in Iran as the center of Iranian political and economic administration. As can be seen from Table 4, excluding Tehran from our sample has no significant impact on our initial observation of the positive relationship between the AOD index and EMIGOUT_RATIO.

To check the impact of outliers or influential observations (outside of Tehran) that may affect our main general specification (Model 7 of Table 3) further, we examined the residuals to identify observations with very large leverage or very large squared residuals.

We followed the procedure in Bjorvatn and Farzanegan (2013) and Farzanegan and Witthuhn (2017). Following the identification of the large residuals, we applied robust regression, which assigns lower weights to possible outliers (Hamilton 1991). Outliers are those cases that show high levels of residuals and leverage. We re-estimated Model 7 in Table 3 by using pooled ordinary least squares (OLS) with province dummy variables. Then, we used the "leverage-versus-squared-residual plot,” which produces a figure that illustrates the leverage versus the squared residuals (see Fig. 4).

Following this step, we calculated Cook’s D¹⁴ and its cutoff, which is 4/62.¹⁵ Finally, to address the possible impact of outliers, we used a robust estimator (Li 1985). Robust regression is an appropriate compromise between excluding points with higher leverages or residuals entirely from the analysis and including all data points, treated equally, in OLS regression. Robust regression will assign a missing weight to observations with Cook’s D larger than 1. Cases with large residuals tend to be down-weighted in robust regressions. The more cases there are in the robust regression that have weights close to one, the closer the OLS results are to the robust regressions.

The results are shown in Table 5. The robust regressions in Table 5 assign lower weights to cases with higher Cook’s D indicators (those that exceed the 4/62 threshold). Since Cook’s D did not exceed 1 in any of the cases, none were excluded from the robust regressions and the total number of observations remains at 62. In our sample, robust regressions assigned the lowest weights to Sistan and Baluchestan (0.075), Chaharmahal and Bakhtiari (0.076), Kohgiluyeh and Boyer-Ahmad (0.091), Hamedan (0.24) and Kermanshah (0.70).

Our main results from the robust regressions are qualitatively and quantitatively similar to the previous results.¹⁶