Heat Stress Increases Long-term Human Migration in Rural Pakistan

Abstract

Human migration attributable to climate events has recently received significant attention from the academic and policy communities (1-2). Quantitative evidence on the relationship between individual, permanent migration and natural disasters is limited (3-9). A 21-year longitudinal survey conducted in rural Pakistan (1991-2012) provides a unique opportunity to understand the relationship between weather and long-term migration. We link individual-level information from this survey to satellite-derived measures of climate variability and control for potential confounders using a multivariate approach. We find that flooding—a climate shock associated with large relief efforts—has modest to insignificant impacts on migration. Heat stress, however—which has attracted relatively little relief—consistently increases the long-term migration of men, driven by a negative effect on farm and non-farm income. Addressing weather-related displacement will require policies that both enhance resilience to climate shocks and lower barriers to welfare-enhancing population movements.

Donors spend approximately 4.6 billion dollars per year in emergency relief for natural disasters (1). Astonishing forecasts of the number of environmentally displaced persons are broadly based on measures of population exposure and ignore individual adaptation (2). Recent quantitative evidence suggests that individual, permanent migration increases with natural disasters and climate shocks, but not uniformly (3-9). Empirical work on the causes of migration has typically been limited to analysis of data covering only a few years, and can therefore conclude little about migration in the longer-term. Using a unique, 21-year longitudinal survey (1991-2012), we examine the long-term migration of household members in response to states of extreme temperature and rainfall in rural Pakistan. Significantly expanding on previous studies of climate-induced migration, we allow climate effects to be time-varying, multidimensional, interactive, nonlinear, and heterogeneous, all while accounting for various spatial and temporal confounders. This approach reveals a complex migratory response that is not fully consistent with common narratives of climate-induced migration.

Pakistan is highly vulnerable to climate change and involuntary displacement. In 2010 alone, floods in Pakistan affected 20 million people, destroying an estimated crop value of 1 billion US dollars (10). Some 14 million people relocated temporarily, and 200,000 moved to internal displacement camps funded by international donors (11). Uncharacteristically high temperatures (heat stress) also reduce population well-being by lowering agricultural yields. For example, the early maturity of wheat grains due to heat stress reduced Pakistani wheat yields by 13 percent in 2010 (12). However, Pakistan’s social protection programs and international relief efforts have been far more responsive to flood victims than heat stress victims, as in other parts of the developing world.

This study aims to answer three unresolved questions in this literature. First, which weather patterns explain the long-term mobility patterns of men and women in Pakistan? Second, is there evidence that extreme rainfall and heat affect agricultural income—indicating a possible channel through which they impact migration? The channels through which disasters affect migration have rarely been addressed due to data limitations (13). Third, are there barriers to weather-induced movement? Knowledge of what motivates migration and the barriers to adaptation through migration is important for designing appropriate policies that respond to natural disasters, migration, and displacement.

To answer all of these questions, we construct a longitudinal survey based on the Pakistan Panel Survey (PPS) collected in 1986-1991 (14) and two tracking studies (Supplementary Methods). The heads of the 1991 PPS households or proxy respondents were resurveyed in 2001 and 2012 to track the movement of original, 1991 household members. The data collected from the PPS and the two tracking studies are used to build an individual-level panel of migrating and non-migrating household members over a 21-year period. We create a person-year dataset following previous work (5-6, 15-18). As migration rates are very low for individuals younger than 15 or older than 39, individuals are included in the dataset starting from baseline or when they reach age 15, and excluded after migrating or when they turn 40. This sample consists of 44,791 person-years, where 4,428 individuals are represented from 583 households.

To answer the first question, we employ discrete-time event history models to measure individual responsiveness to weather variables, controlling for baseline (1991) household wealth and demographic characteristics, and for village and time fixed effects. (Explanatory variables are summarized in Supplementary Table 1.) We estimate the event history model as a logit model, analyzing migration as a binary dependent variable. A household member is considered a migrant in year t if he was permanently not present in year t for reasons other than death. The individual is considered a within-village migrant if they moved elsewhere in the village, and an out-of-village migrant if they moved outside of the village (including abroad). The multinomial event history model, estimated as a multinomial logit, differentiates the impacts of weather anomalies on local (within-village) vs. long-distance (out-of-village) moves, and gender-differentiated migration (5-6):

$$\text{log}\left(\frac{{\pi }_{rit}}{{\pi }_{sit}}\right)={\alpha }_{rt}+{\alpha }_{r\nu }+{\beta }_r{X}_{it},$$

where π_rit is the odds of moving distance r for individual i in year t, π_sit is the odds of not moving, and parameters α_rt and α_rv are the baseline hazard of mobility in village ν and year t, respectively, for the specific types of mobility r. X is our vector of controls. Inverse probability weights are used to correct estimates for individual attrition (Methods).

From various secondary data sources (Supplementary Methods), we construct the key weather variables included in the analysis: cumulative rainfall over the monsoon period (June – September), average temperature over the Rabi season (November – April) when wheat is grown, a measure of flood intensity (the annual number of deaths caused by flooding), and a 12-month moisture index—the Standardized Precipitation Evapotranspiration Index (SPEI) (19). All weather variables are measured at the village level, with the exception of flood intensity which varies by province. Our preferred specifications use average weather values from year t and t-1 to capture the weather preceding period t migration decisions. Trends appear stationary and migration corresponds with peaks in temperature (Supplementary Figure 1).

Table 1 presents results of migration responses to weather by gender using an event history model, and further by within and out of village moves using a multinomial event history model. We focus on the estimates of the weather parameters (though estimates of all coefficients are presented in Supplementary Table 2). Specifications A-C present results from linear and non-linear specifications of the rainfall and temperature variables (Methods).

Table 1.

Migration Responses to Climate. Q abbreviates quartile; the omitted category in non-linear models is the interquartile range. All coefficients reflect odds ratios. Inverse probability weights account for individual attrition. Standard errors are clustered at the village level. Statistical significance of parameters based on t tests, where ***, **, and * indicate p<0.01, p<0.05, p<.10. Joint tests of statistical significance based on Chi-squared tests.

			Men						Women
	Logit Move		Multinomial logit Distance of move				Logit Move		Multinomial logit Distance of move
			In village		Out of village				In village		Out of village
Specification A
Rainfall	1.28		0.94		1.93	**	1.19		1.24		1.17
Temperature	2.69	***	2.42	***	2.90	**	1.87	***	2.03	**	1.69	*
Joint test of variables	17 92	***			21.96	***	11.60	***	13 92	***
Specification B
Rainfall	1.05		1.75		0.53		1.04		1.59		0.62
Temperature	2.62	***	2.64		2.42	**	1.85	***	2.06	**	1.53
Rainfall × Temperature	1.01		0.97		1.07	*	1.01		0.99		1.03	*
Joint test of variables	17 92	***			26.32	***	14.56	***	21.80	***
Specification C
Rainfall in 1Q	1.47		1.51		1.57		1.13		0.99		1.36
Rainfall in 4Q	0.82		0.84		0.81		1.20		1.20		1.30
Temperature 1Q	0.84		1.02		0.68		0.83		0.80		0.84
Temperature 4Q	5.09	***	2 83	***	11.16	***	1.85	***	1.82	***	2.19	**
Joint test of variables	25.53	***	41.83	***			15.45	***	21.87	***
Specification D
Flood	0.96	*	0.96	*	0.96		0.97	**	0.95	***	0.99
Temperature	3.00	***	2.76	***	3.35	***	2.00	***	2.22	***	1.74	*
Joint test of variables	18.98	***			22.45		13.11	***	17.01	***
Specification E
Moisture index	0.71	*	0.70		0.75		0.75	**	0.64	**	0.85
Individuals	2,125		2,147				2,303		2,303

Source: Pakistan Panel Survey 1991; Pakistan Panel Tracking Surveys 2001, 2012

Overall, we observe no robust effect of rainfall on the mobility of men or women. Men are slightly more likely to move out of the village in response to greater rainfall levels, but only when temperature is also sufficiently high (Specification B). When we flexibly allow rainfall and temperature to have a non-linear impact on migration by including dummy variables for rainfall and temperature in the first (Q1) and fourth (Q4) quartiles (the omitted groups being the second and third quartiles), we find that it is only temperature in Q4 that significantly affects migration. Specification D, (controlling for flood intensity instead of rainfall), further corroborates that flooding has no effect on out-of-village moves and indeed causes a modest decline in the within-village migration of men and women.

However, the results consistently show that men move out of the village in response to extreme temperatures in the Rabi season (Specifications A–D). Lastly, when the weather variables are substituted by a moisture index (Specification E), we see that periods of high moisture in general are associated with the retention (as opposed to migration) of household members. Thus, we are left with an overall picture that heat stress—not high rainfall, flooding, or moisture—is most strongly associated with migration. The risk of a male, non-migrant moving out of the village is 11 times more likely when exposed to temperature values in the fourth quartile (Specification C). Male migration responses are robust to accounting for spatial autocorrelation (Methods).

Figure 1 provides the predicted rates of out-of-village migration for both men and women, based on the preferred model (Specification C, which flexibly allows for non-linear effects). Men and women consistently migrate the most under scenarios with extreme temperature. The baseline migration of men (women) moves from 0.01 (0.02) to 0.12 (0.05) under the scenario of temperature in Q4 (extreme heat) and rainfall in Q1 (extreme scarcity).

Figure 1. Predicted Probabilities of Out-of-Village Migration by Gender.

The bubble size reflects the predicted probabilities obtained using Specification C under different temperature and rainfall extreme scenarios (Supplementary Table 5). Solid teal green bubbles indicate the probability of men moving out of the village in a given scenario. Black dashed bubbles indicate the probabilities of women moving out of the village in a given scenario. Predicted probabilities are specified for the scenario where the temperature and rainfall lie in the interquartile range and extreme hot scenario (low rainfall, high temperature) for reference and differentiated by color for the gender of migrants.

Is there evidence that extreme rainfall and heat affect agricultural income—indicating a possible channel through which they impact migration? To answer this second question, we examine how fluctuations in temperature have affected the annual farm, farm wage, and non-farm income of the PPS households during 1986-1991. We estimate a linear regression including rainfall and temperature extreme variables (as in Specification C), along with household and time fixed effects. Table 2 displays the marginal effects of temperature and rainfall on various sources of income, with 90% confidence intervals. Agricultural income suffers tremendously when temperatures are extremely hot (in Q4)—wiping out over a third of farming income. Non-farm income also experiences losses from heat stress, but to a lesser extent (16%). Interestingly, high rainfall increases all sources of income substantially. This analysis suggests one possible reason that heat stress drives migration, while extreme rainfall does not: heat stress (unlike rainfall) provides a negative income shock.

Table 2.

Marginal Effects of Rainfall and Temperature Extremes on Annual Income, with 90% Confidence Intervals. The marginal effects are computed using the point estimates from a linear regression which includes household and time fixed effects. Confidence intervals are based on village-clustered standard errors.

	Net farm income	90% CI	Farm wage income	90% CI	Non-farm income	90% CI
Variable Mean (1000s 2000 Rupees)	44.15		0.75		31.45
Rainfall in 1Q	−9.25	[−20, 1]	−0.12	[−0.5, 0.3]	3.93	[0.4, 7.5]
Rainfall in 4Q	13.92	[2, 26]	1.31	[0.4, 2]	15.38	[10, 20]
Temperature 1Q	−10.20	[−28, 8]	0.32	[−0.0, 0.6]	−4.70	[−9, −0.2]
Temperature 4Q	−15.89	[−31, −0.6]	0.59	[−0.1, 1]	−4.90	[−10, −0.1]
Households	648

Source: Pakistan Panel Survey (1986-1991)

Are there barriers to weather-induced movement? To answer this final question, we examine the relationship between mobility and weather anomalies by land ownership and asset wealth to see whether financial constraints influence migration decisions (Table 3). Interestingly, extreme heat stress is associated with more migration for both land-owners and non-land owners, and for those in the first and third terciles of asset wealth. However, the magnitude and statistical significance of the estimates are most pronounced for the land- and asset-poor, and their moves are most likely to be out-of-village moves. It seems that for the poor, the migration benefits following heat stress outweigh the moving costs, spurring migration of all forms. The poor may have more locational flexibility as they are not tied to the land or assets which can be hard to sell, and at risk of loss if untended. Furthermore, given that the poor often provide goods and services to land cultivators, this is consistent with our find in Table 2, where we showed that heat stress especially reduces rural non-farm income. When farmers are hit by a shock, the livelihoods of those dependent on providing goods and services to them will also be affected.

Table 3.

Migration Responses to Rainfall and Temperature Extremes by Land Ownership and Asset Wealth. All coefficients reflect odds ratios. Inverse probability weights used in all models. Statistical significance parameters based on t tests, where ***, **, and * indicate p<0.01, p<0.05, p<.10. Joint tests of statistical significance based on Chi-squared tests.

	Owned land							Asset Wealth
	None Distance of move			Some Distance of move				1st Tercile Distance of move				3rd Tercile Distance of move
	In village	Out of village		In village		Out of village		In village		Out of village		In village		Out of village
Rainfall in 1Q	1.40	1.07		1.09		1.70	*	1.41		1.11		1.04		1.56
Rainfall in 4Q	1.41	1.37		0.89		0.91		1.23		1.45		0.83		0.75
Temperature 1Q	0.97	1.01		0.74		0.65		0.67	*	0.90		1.81	**	0.81
Temperature 4Q	1.69	4.89	***	2.55	***	2.67	**	2.66	***	2.98	**	1.41		2.31	*
Joint test of variables	13.26			40.30	***			28.28	***			14.77	*
Individuals	1,592			2,858				2,204				2,246

Source: Pakistan Panel Survey 1991; Pakistan Panel Tracking Surveys 2001, 2012

Our empirical work offers the first quantitative evidence of how long-term migration decisions in Pakistan are affected by weather extremes. Both women and men respond to heat stress by moving, but men mostly move long-distances. Our results are consistent with earlier evidence of risk diversification through the marriage migration of women in India (20). Long-distance moves also coincide with farm income losses, yet men appear more responsive to temperature fluctuations and historically are inclined to migrate for employment in this setting. While all individuals use migration to cope with heat stress, the poor are more likely than the rich to relocate outside of the village.

Our study has broader policy relevance for development strategy in Pakistan. Existing flood relief programs may potentially crowd out private coping mechanisms like migration, particularly for the poor and risk-averse living in flood-prone areas. Our results also show the important role of heat stress—a climate shock which has attracted relatively less relief—in lowering farm and non-farm income and spurring migration. Sustainable development will require policies that enhance adaptation to weather-related risks for farmers and for enterprises tied to the rural economy. Shifting relief towards investments in heat-resistant varieties, producing and disseminating better weather forecasting data and weather insurance, and policies that encourage welfare-enhancing migratory responses might improve individual abilities to adapt to an array of weather-related risks (21,22).

METHODS

To account for individual attrition, all of our statistical models use inverse probability weights constructed from the ratio of predicted probabilities, of remaining in the sample between 1991 and 2012, from a restricted and unrestricted probit model (Supplementary Table 4) (23,24). The F statistic testing the joint significance of the rainfall variable and its interaction with temperature (p<0.05) suggests Specification B is preferred to Specification A for the sample of men under the multinomial logit model. Conclusions are similar when including five-year (rather than one-year) fixed effects (Supplementary Table 5) and without averaging values from year t and t-1 (Supplementary Table 6); the latter being imprecise due to the collinearity between weather variables. We test for the robustness of the results of Specification C under spatial correlation (25) using a grouped bootstrap (where years are resampled and replaced) for the logit model (Supplementary Table 7). Male migration responses remain responsive when facing temperature in the fourth quartile.