The changing climate-migration relationship in China, 1989–2011

Abstract

A persistent concern about the social consequences of climate change is that large, vulnerable populations will be involuntarily displaced. Existing evidence suggests that changes in precipitation and temperature can increase migration in particular contexts, but the potential for this relationship to evolve over time alongside processes of adaptation and development has not been widely explored. To address this issue, we link longitudinal data from 20 thousand Chinese adults from 1989–2011 to external data on climate anomalies, and use this linked dataset to explore how climatic effects on internal migration have changed over time while controlling for potential spatial and temporal confounders. We find that temperature anomalies initially displaced permanent migrants at the beginning of our study period, but that this effect had reversed by the end of the study period. A parallel analysis of income shares suggests that the explanation might lie in climate vulnerability shifting from agricultural to non-agricultural livelihood activities. Taken together with evidence from previous case studies, our results open the door to a potential future in which development and in-situ adaptation allow climate-induced migration to decline over time, even as climate change unfolds.

Introduction

The potential for climate change to result in large-scale involuntary human migration and displacement has been a frequent theme in the climate vulnerability literature. Early scholarly writings and most current popular writings have envisioned a rising tide of “climate refugees” from low and middle-income countries (LMICs) who would make involuntary, permanent and long-distance moves (Myers 1997), but recent scholarship has pushed back against this narrative, suggesting a more nuanced understanding of the relationship between climate and human mobility in which people make decisions to move, stay, or engage in translocal livelihoods within the context of a wide array of social, cultural, and environmental factors and their interlinkages (Faist & Schade 2013; Wiegel et al. 2019). Over the past five years, a significant number of studies using demographic and econometric approaches have directly investigated this issue, revealing that temperature in particular does tend to increase migration across a variety of contexts (Bohra-Mishra et al. 2014; Mueller et al. 2014; Jennings & Gray 2015; Mastrorillo et al. 2016; Nawrotzki & DeWaard 2016; Bohra-Mishra et al. 2017; Call et al. 2017; Riosmena et al. 2018). However these effects are spatially heterogeneous (Gray & Wise 2016; Thiede et al. 2016), potentially strongest for short-distance moves (Gray & Mueller 2012), and occasionally work in reverse to trap potential migrants (Cattaneo & Peri 2016; Nawrotzki & Bakhtsiyarava 2017; Thiede & Gray 2017). The latter findings dispute the older, popular narratives about the nature of climate-induced migration, but are consistent with the contemporary theory and broader scholarly literature documenting the context-specific nature of migration processes (Bell et al. 2015) as well as the many barriers to migration in LMICs (Bryan & Morten 2019).

One issue that has received little attention in the empirical literature, and that might help explain the mixed findings to date, is the possibility that climate-migration relationships are changing over time (McLeman 2018). Improving population well-being under ongoing demographic and development transitions might make individuals more resistant to climate-induced displacement. Alternatively, the compounding effects of increasing exposure to climate extremes could create new vulnerabilities, or the new resources accruing under development could permit more moves than in the past. To date few studies have had access to the longitudinal data necessary to test these hypotheses, and those studies have found conflicting results (Gutmann et al. 2005; Jennings & Gray 2015; Call et al. 2017; Riosmena et al. 2018). The climate and health literature has similarly found no consistent pattern in how temperature effects on health are changing over time (Barreca et al. 2016; Burke et al. 2018). Thus, while we know that exposure to climate extremes will increase globally (Frame et al. 2017), we have little basis on which to predict whether these exposures will become more or less costly to individuals over time on a per exposure basis.

To address these issues, we link a time-series of high-resolution climate data to the China Health and Nutrition Survey (CHNS; Popkin et al. 2009), allowing us to investigate changing climatic effects on internal migration across eastern China during a 22-year period of rapid social and environmental change. China is home to the world’s largest population of internal migrants (Liang et al. 2014), representing the largest migration event in human history, and is highly vulnerable to climate change (Piao et al. 2010; Zhao et al. 2016). However, to date few studies have investigated climate-migration relationships in this context, in part due to data limitations. We create a person-period dataset on 20 thousand Chinese adults for whom we observe both temporary and permanent migration, and then examine climatic effects on these moves using a strategy that accounts for both attrition and potential spatial confounders. This analysis reveals that, prior to the year 2000, temperature anomalies increased permanent moves, and then after 2000 these effects reversed. Using data on household income shares, we show that this change corresponds to changes in the climate responsiveness of particular livelihood strategies. Together this evidence suggests that declining household vulnerability to temperature in the agricultural sector has encouraged a shift to retaining migrants during adverse climate conditions instead of sending them.

Previous Studies

Our research draws on a large previous literature on climate-induced migration in LMICs and climate vulnerability in China, and a small previous literature on how the former relationships change over time and how the latter findings relate to migration. Regarding climate-induced migration, the wide public interest in this phenomenon together with the growing availability of geo-located, longitudinal demographic data has fueled a rapid expansion of this literature, with a focus on LMICs where climate change presents the greatest risks. The most common approach among quantitative social scientists has been to link individual or household-level data on migration to external data on origin-area climate shocks, and then analyze climatic effects on migration while attempting to control for potential confounders (Fussell et al. 2014). A parallel literature has primarily drawn on country-level data on urbanization and international migration to address a similar set of questions (Marchiori et al. 2012; Cai et al. 2016; Cattaneo & Peri 2016). Climate exposures are typically defined by some measure of deviation from a historical baseline, but the literature has not reached a consensus as to which climate measures are the most appropriate nor on the most appropriate way to account for potential confounders (Hsiang et al. 2013). While this set of approaches has been criticized for ignoring the multicausal and context-specific nature of migration (Nicholson 2014; Parsons 2019; Wiegel et al. 2019), we argue that most of these critiques can be addressed through a judicious application of econometric methods, which can account for the role of macro-level socio-economic and geographic contexts within which individuals experience climate and risk, as well as the individual and household risk factors impacting migration decision-making.

This literature makes clear that climate has persistent effects on human migration, and that, despite an early focus on precipitation, temperature effects tend to dominate when the two are compared (Carleton & Hsiang 2016). This finding is consistent with studies of climatic effects on agriculture (Zhao et al. 2017), labor productivity (Zhang et al. 2018) and health (Ma et al. 2015) in these settings, which have also found large temperature effects. Building on livelihood approaches to migration (Ellis 2000), agriculture is typically assumed to be the conduit between climate and population mobility (Cai et al. 2016), but these other potential pathways cannot be excluded. The story thus far is generally consistent with “climate refugees” narrative described above, but other aspects of this literature have often diverged from common assumptions about the nature of climate-induced migration. Most future climate migrants are likely to be internal or intra-regional migrants given that international moves are quite rare (Gray & Wise 2016) and also that climatic influences on migration are at times larger on shorter moves (Gray & Mueller 2012). Also, studies which have compared climatic effects across space have found considerable spatial heterogeneity, including between adjacent countries and regions (Gray & Wise 2016; Thiede et al. 2016). Additionally, these studies and others have also documented several cases in which adverse climate conditions reduce migration and thus trap potential migrants in place (Cattaneo & Peri 2016; Nawrotzki & Bakhtsiyarava 2017; Thiede & Gray 2017), consistent with the high costs of migration in these contexts (Bryan & Morten 2019). Up to this point, it should be noted that these studies have typically examined only one type of population mobility, and have assumed that climatic effects on migration are constant over space and time.

In contrast, a small number of studies have examined whether climate-migration relationships can change over time. In one of the first quantitative studies in the climate-migration literature, Gutmann et al. (2005) examined climatic influences on county-level net migration in the US Great Plains from 1930–1990, revealing the effects of temperature and precipitation declined over time while the effects of environmental amenities such as water bodies increased. Jennings and Gray (2015) subsequently used individual-level population registry data from 1865–1937 in the Netherlands to show that the climate-responsiveness of long-distance moves similarly declined over time. Using a similar approach, Call et al. (2017) drew on 17 years of monthly registry data from Matlab, Bangladesh to show that, over time, households shifted from sending migrants during precipitation extremes to retaining them during these same periods. Most recently, Riosmena et al. (2018) used household-level data from the 2000 and 2010 Mexican censuses to measure climatic effects on international migration during the preceding five-year periods, finding significant climate effects but little change over time. Thus three out of four case studies, covering the longest time periods, suggest that climate-migration relationships might be evolving over time away from involuntary moves imposed by environmental hardships. However this remains a small evidentiary base.

Also unclear is the extent to which previous results, which have focused on long-distance permanent moves, extend to other types of moves such as temporary migration. Most previous studies have also ignored processes of sample attrition that are also potentially driven by climate. Because temporary moves are typically lower cost, they might be particularly responsive to climate shocks, a process that has been previously described for natural disasters (Fussell et al. 2010; Gray et al. 2014). On the other hand, as climate change gradually alters mean climatic conditions this would be expected to influence permanent moves as well. Previous studies comparing permanent moves across distance categories, another proxy for cost, have tended to find important differences in the effects of climate (Gray & Mueller 2012; Gray & Bilsborrow 2013; Jennings & Gray 2015; Gray & Wise 2016; Leyk et al. 2017). Differential climatic effects on permanent versus temporary migration have received much less attention, but were explored by Henry et al. (2004) in one of the earliest papers on this topic. In that study, climatic effects on temporary migration were stronger in some cases and on permanent migration for other cases, leaving unclear what lesson to extract.

China is a uniquely relevant study site to investigate these issues given its high level of exposure to climate shocks as well as its rapid social and economic transformation in recent decades. China has among the world’s highest per capita exposure to flooding (Jongman et al. 2012) and is also highly exposed to temperature shocks (Zhao et al. 2016). Since 1990, it has also rapidly transitioned from being a predominantly rural, poor and agrarian country into one that is majority urban and middle-income (Yang et al. 2013; Liang et al. 2014), representing the world’s largest such transition to date. During this period, China’s complex household registration system (hukou) has also placed significant constraints on mobility by tying access to state resources (including schools, hospitals, and public sector jobs) to the place of registration. Up until recently a critical component of one’s household registration was the designation of the household as either rural or urban, with the latter enabling individuals to more easily migrate and access services in urban destinations (Chan 2012). As a result, this two-tiered system is likely to reinforce climate vulnerability in rural areas where many households are both directly dependent on the natural environment and have additional constraints placed on their mobility.

Whether population-level climate vulnerability has increased or decreased during this period is unclear. Evidence suggests that recent changes in temperature, precipitation and flooding have slowed the growth of agricultural production (Chen et al. 2016) as well as urban economic activity (Elliott et al. 2015), but there is also evidence of household adaptation to both of these changes (Chen et al. 2014; Ma & Maystadt 2017). For example, farmers have shifted sowing and harvesting dates of maize and wheat to correspond to changes in seasonality (Wang et al. 2012), and may be re-allocating effort to non-agricultural activities in order to reduce exposure to flooding (Tian & Lemos 2018). Other studies from China have shown that agricultural vulnerability to shocks declines with exposure to external labor markets (Giles 2006) and that cropping intensity has also declined nationally (Qiu et al. 2017), both suggesting the potential for declining vulnerability to climate in agriculture.

The extent to which these findings on climate vulnerability translate to migration is also unclear, and relatively few studies have investigated climate-migration relationships in China more broadly. Fu et al. (2012) estimated a model of province-province migration flows using the 1995 One-percent Population Survey, and found that migrants tended to move towards destinations with less seasonal variation in temperature. More closely related to our own approach, Giles and Yoo (2007) used household panel data from 1995–2000 to show that past rainfall shocks predicted the size of the current migrant network as part of an instrumental variables approach. Building on this contribution, Minale (2018) used three rounds of another household panel to examine climate effects on temporary migration, finding that negative rainfall anomalies increased participation in temporary migration. Finally and most relevant to our own contribution, Ward and Shively (2015) used the absolute value of county-level rainfall anomalies to predict the number of household migrants in the CHNS, revealing that spring and summer rainfall shocks had positive and negative effects on migration respectively. These studies make clear that climate does play a role in internal migration in China, but what role changing temperatures may play and how these relationships may evolve over time have not been addressed.

Our research extends these studies by examining temperature as well as rainfall effects, decomposing temporary and permanent moves, and carefully accounting for attrition as well as the potential endogeneity of control variables to same-period climate. We directly respond to criticisms that this literature has ignored the complexity of population mobility (Faist & Schade 2013; Wiegel et al. 2019) by (1) separately examining four distinct types of mobility, (2) allowing the rate of each type of mobility to vary across locations and across time, (3) including a large set of social and economic control variables which account for non-climatic influences on mobility and (4) allowing the climatic effects to vary across time and subsequently across household characteristics. Our core hypothesis for this analysis is that hot and dry conditions increase temporary and permanent migration, but that these effects have declined over time towards zero. This expectation is driven by observations that population-level reliance on agriculture, the most climate-dependent sector, has declined over time (below), and that adaptations are occurring within agriculture that should make it more resilient per unit of climate shock (Wang et al. 2012; Chen et al. 2014). Hot and dry conditions are costly for agriculture in China (Chen et al. 2016; Ma & Maystadt 2017), and, while flooding is costly, we cannot directly measure it with our data (below). Increasing household assets and access to migrant networks as well as remittances over the study period (below) should also lower the relative costs of migration, provide buffers against climate shocks, and also routinize migration so that a large fraction of prospective migrants depart with or without the occurrence of climate shocks (Nawrotzki et al. 2015). Below, we describe how we test the core hypothesis as well as the potential role of these potential pathways for the effect.

Data and Methods

To investigate changing climate-migration relationships in China, we (1) use the CHNS to observe 20,471 adults over inter-survey periods of 2–4 years, (2) link these data to climate exposures at the county level, and (3) use fixed effects regressions to measure climatic effects on migration and attrition while controlling for potential spatial and temporal confounders. CHNS is one the longest-running longitudinal surveys in a LMIC, having conducted interviews in 1989, 1993, 1997, 2000, 2004, 2009 and 2011, with an additional 2015 round released while our analysis was ongoing. An initial sample of 3795 households was drawn in 1989 using a multistage process from 32 urban neighborhoods, 30 suburban neighborhoods, 32 towns and 96 rural villages in eight provinces (Liaoning, Shandong, Henan, Jiangsu, Hubei, Hunan, Guizhou and Guangxi). We refer to these sampling clusters as communities. Sampling weights are not available, but the CHNS provinces are diverse and the characteristics of sample households are similar to the Chinese population as a whole (Chen 2005). Following the 1989 baseline interview, follow-up rounds have targeted the same households. Over this 20-year period, approximately one third of the 1989 households have been lost to death, refusal or departure, but the sample size has been maintained by the addition of (1) new households formed by baseline household members, (2) newly sampled households, (3) newly sampled communities, and (4) the addition of Heilongjiang province in 1997 (Popkin et al. 2009). With these additions, the sample on average contains 14% new households in each post-1989 round, and this appears to mitigate against the sample aging over time (Table 1).

Table 1.

Descriptive statistics from the person-period dataset (n = 64,829).

Variable/Statistic	Units	Level	Mean	Std. Dev.	Min	Max
Outcomes
Temporary migration	0/1	Individual	0.124	-	0	1
Permanent migration	0/1	Individual	0.093	-	0	1
Individiual attrition	0/1	Individual	0.048	-	0	1
Household attrition	0/1	Individual	0.086	-	0	1
Predictors
Temperature anomaly	SD	County	0.022	0.896	−2.04	1.91
Precipitation anomaly	SD	County	−0.018	0.906	−2.06	3.06
Year	years	National	9.63	6.87	0	20
Age 15–19	0/1	Individual	0.158	-	0	1
Age 20–24	0/1	Individual	0.156	-	0	1
Age 25–29	0/1	Individual	0.141	-	0	1
Age 30–34	0/1	Individual	0.133	-	0	1
Age 35–39	0/1	Individual	0.146	-	0	1
Age 40–44	0/1	Individual	0.142	-	0	1
Age 45–49	0/1	Individual	0.124	-	0	1
Female	0/1	Individual	0.492	-	0	1
Head/Spouse	0/1	Individual	0.502	-	0	1
Child of head	0/1	Individual	0.383	-	0	1
Other relation to head	0/1	Individual	0.102	-	0	1
Relation missing	0/1	Individual	0.012	-	0	1
Ever married	0/1	Individual	0.620	-	0	1
Married missing	0/1	Individual	0.154	-	0	1
Less than primary edu.	0/1	Individual	0.123	-	0	1
Primary edu.	0/1	Individual	0.207	-	0	1
Lower middle edu.	0/1	Individual	0.412	-	0	1
Upper middle edu.	0/1	Individual	0.153	-	0	1
Technical school	0/1	Individual	0.049	-	0	1
University edu.	0/1	Individual	0.032	-	0	1
Education missing	0/1	Individual	0.023	-	0	1
Household size	#	Household	4.41	1.47	1	14
Migrant network	#	Household	0.50	0.93	0	10
Consumer assets	0–10	Household	5.12	1.80	0.00	9.71
Business assets	0–10	Household	3.16	1.24	0.00	10.00
Assets missing	0/1	Household	0.038	-	0	1
Urbanicity index	scale	Community	53.4	19.4	14.3	106.5

The longitudinal design of CHNS allows us to observe migration over time for a known sample of baseline individuals as well as sample attrition. In each follow-up round, the household roster collected information on the permanent and temporary migration status of all household members from the previous round, as well as the distance moved by permanent migrants. The wording of these questions, which ask about exits from the households and absences of current household members respectively, matches the structure of commonly used survey questions on these topics and also closely matches Chinese definitions of migrants with and without household registration in the destination. Several previous studies have used these data to observe out-migration over time from the CHNS sample (Giles & Mu 2007; Chang & MacPhail 2011; Tong & Piotrowski 2012; Chen 2013; Ward & Shively 2015), but they have largely ignored attrition and to our knowledge have not combined the measures of permanent and temporary migration into a multinomial framework as we do here. CHNS also measures a large set of other individual, household, and community characteristics at each round, which we draw on as controls and supplementary outcomes as described below.

To enable this approach, we construct a person-period dataset (n = 64,829) in which the migration/attrition status of each individual aged 15–49 in the baseline round is measured at the time of the follow-up round 2–4 years later (Table 1). This dataset includes all person-periods where the individual was reported to be a member of a sample household in the baseline round (including temporary migrants), with the exception of cases where individuals died in the inter-survey period and cases where the entire community was lost to follow-up, which were excluded. Using information from both migration measures, we define a four-category multinomial migration/attrition outcome where the reference category is individuals who remained in the community, either as members of the baseline household or another household. The first outcome captures temporary migration and is defined as individuals who remained members of the baseline household but were temporarily absent since the previous round (12% of person-periods). Temporary migrants who moved to attend school were defined as non-migrants as our focus is on livelihood dimensions of migration. The second outcome captures permanent migration and is defined as individuals who exited the household and joined another household outside of the community (9% of person-periods). Note that this outcome excludes within-community moves but includes intra-county moves (68% of permanent moves) that would often be missed in census or administrative data sources. We then decompose attrition into individuals who were not observed at follow-up but whose household was observed (5% of person-periods) and individuals whose entire household was lost at follow-up (9% of person-periods), reasoning that these outcomes are likely to reflect different processes of unreported migration and/or refusal. By decomposing attrition in this way, we are able to examine which outcomes resemble migration in their determinants, and to capture whether climate might be contributing to otherwise-unobserved departures (Thomas et al. 2012).

To this person-period dataset we also attach a variety of control variables measured at baseline, including age, gender, relationship to the household head, marital status, educational attainment, household size, a measure of the migration network, measures of consumer and business assets, a measure of urbanicity, and indicators for missing data on specific variables (Table 1). By measuring all controls in the baseline survey, we avoid potential endogeneity to climate exposures during the inter-survey period. Potential measures of the migration network at baseline are limited by the questionnaire used in 1989, which measured temporary migration only. Thus we define the migration network as the number of household temporary migrants recorded at baseline. To create the asset variables, we applied polychoric principal components analysis to a large set of asset variables using the full sample of households (Kolenikov & Angeles 2009). The first and second components placed large weights on consumer and business assets respectively, and collectively explained 42% of the variation in assets. We rescale these components to vary from 0–10 to create our measures of consumer and business assets. The time-varying urbanicity measure was developed by the CHNS team as a composite index of population density, access to infrastructure, non-agricultural livelihoods, and mean educational attainment (Jones-Smith & Popkin 2010). Finally, missing data on relation to head, marital status, education and assets were accounted for, while retaining the full sample size, by including an indicator for missingness on each variable in the regression, with missing assets interpolated to the median. Descriptive statistics from this dataset on change over time are displayed over time in Table 2, which also brings in additional data on inflation-adjusted sources of household income as described below. Consistent with previous accounts (Yang et al. 2013; Liang et al. 2014), the study period saw a rapid increase in urbanicity, consumer assets, household income and temporary migration, as well as generally increasing temperature anomalies.

Table 2.

Time trends in individual and contextual characteristics of adults age 15–49 (n = 64,829 person-periods).

Characteristic/Year	Baseline Year
	1989	1991	1993	1997	2000	2004	2006	2009
Individual characteristics
Temporary migration	7.0%	7.6%	6.7%	8.7%	13.1%	17.9%	16.7%	20.8%
Permanent migration	7.2%	7.1%	11.6%	8.8%	10.1%	8.8%	11.3%	9.7%
Individual attrition	4.5%	4.5%	5.3%	4.1%	6.1%	3.7%	5.5%	5.2%
Household attrition	5.5%	8.0%	5.3%	8.5%	11.2%	8.2%	13.1%	8.6%
Mean age	29.6	30.3	30.6	31.5	32.1	31.9	32.0	32.8
Less than primary edu.	22%	20%	18%	13%	9%	6%	6%	6%
Primary edu.	24%	24%	24%	22%	19%	20%	17%	17%
Lower middle edu.	35%	38%	39%	40%	43%	45%	44%	44%
Upper middle edu.	13%	14%	15%	16%	18%	17%	17%	14%
Technical school edu.	3%	2%	3%	5%	6%	6%	7%	7%
University edu.	2%	2%	1%	2%	4%	4%	5%	6%
Education missing	2%	1%	1%	1%	1%	3%	4%	6%
Household characteristics
Migrant network	0.32	0.34	0.40	0.28	0.38	0.63	0.85	0.83
Consumer assets	3.95	4.13	4.29	4.94	5.26	5.78	6.05	6.49
Business assets	2.87	2.92	3.00	3.14	3.25	3.36	3.36	3.37
Crop income	2411	1942	2752	2105	3401	3536	3484	3963
Transfer income	599	1074	1288	1700	4227	3434	3900	3877
Other income	9811	11355	13415	17447	15850	20742	33615	37090
Crop income share	19%	14%	16%	10%	14%	13%	8%	9%
Other income share	77%	79%	77%	82%	68%	75%	82%	83%
Transfer income share	5%	7%	7%	8%	18%	12%	10%	9%
Community characteristics
Urbanicity index	43.3	44.8	44.8	51.1	56.9	59.7	61.4	64.7
County characteristics
Temperature anomaly	−0.17	−0.79	−0.24	0.46	0.77	0.44	0.28	−0.69
Precipitation anomaly	0.11	−0.15	−0.16	0.09	0.28	−0.21	0.09	−0.28

Note: Variables are as defined in Table 1, plus income which is defined as net household income from this source in inflation-adjusted 2011 RMB.

Climate data were extracted from Climate Research Unit’s Time Series (CRUTS), which is a gridded climate dataset preferred by climatologists for many applications. CRUTS uses a geostatistical approach to combine data on climate from over 4000 weather stations, including a large number in China, producing a monthly global dataset at 0.5° resolution (~50km at the equator) (Harris et al. 2014). Using a set of digital maps maintained by the CHNS project for all counties in the sample provinces, we extracted CRUTS mean daily temperature and mean daily precipitation rate as yearly, county-level spatial means for the period 1981–2011 using an approach that weighted grid cells by their overlap with the target county¹ CHNS project staff then linked these data to the recoded CHNS sample counties (n = 54, the location of which is not publicly released) using a secure approach that used rounding to prevent disclosure of the county locations. We then transformed these linked temperature and precipitation values into standardized climate anomalies for the two-year period encompassing the year of each follow-up survey and the year prior, defined as the z-score of temperature and precipitation relative to all other two year periods in that county during 1981–2011. These temperature and precipitation anomalies thus capture the extent to which climate conditions in the period preceding and overlapping each follow-up survey deviated from historical conditions in that location.

In previous work, we have shown standardized climate anomalies to be stronger predictors of migration than raw climate values and that climatic effects on migration can extend across multiple-years lags (Mueller et al. 2014; Gray & Wise 2016; see also Nawrotzki & DeWaard 2016; Henderson et al. 2017). Anomalies also have the advantage of being uncorrelated (on average) with historical climate and other baseline characteristics, meaning that they can be interpreted as natural experiments (Nordkvelle 2017). These climate values were then linked to the person-period data at the county-period scale (Table 1). To account for shared measurement of climate at the county level, all regression results reported below account for clustering at this scale (Williams 2000).

To examine how climatic influences on migration change over this period, we estimate multivariate models of migration and attrition that allow the underlying risk of migration to change over time in different ways, while also allowing the climatic effects on migration to change over time in the same way. Specifically, we use the person-period dataset described above to estimate multinomial logistic regressions of migration/attrition as a function of climate exposures, baseline controls, county fixed effects, various specifications of the time trend, and interactions between climate and the time trend. The multinomial approach allows us to simultaneously examine all four streams of migration and attrition while also allowing the effects of climate and controls to be stream-specific. Climate exposures and baseline controls are included as described above, capturing controls at the beginning of the interval and climate exposures during the interval. The inclusion of county fixed effects accounts for all time-invariant county characteristics such as baseline climate that might confound the effects of climate anomalies. To account for the direct effects of the changing national context in increasingly sophisticated ways, time is defined as years from 1989 to the baseline round, and we alternately include three different specifications of this variable: (1) a linear term for the year, (2) a quadratic transformation, and (3) a stepwise time trend (i.e., year fixed effects). By starting with a simple linear trend and proceeding to a stepwise trend, we gradually eliminate potential time-varying sources of confounding to the climate-migration relationship, but also increasingly exclude potentially relevant sources of climate variation such as national-level climate shocks. To each of these three models we add interactions between the time trend and the climate variables, thus allowing the climate effects to vary across time. For the case of the stepwise time trend, we allow the climate effects to differ before and after the year 2000, rather than year to year, in order to take better advantage of our sample size. If, in all specifications, (1) temperature anomalies increase temporary and permanent migration while precipitation anomalies decrease them, and (2) these effects shrink over time towards zero, then our core hypothesis will be strongly supported.

Expressed mathematically, the log-odds of making a move of type r relative to no move (event s) are given by

$$\text{log}\left(\frac{{\pi }_{rip}}{{\pi }_{sip}}\right)={\alpha }_r+{\alpha }_{rc}+{\beta }_r{X}_{ip}+{\beta }_r{A}_{cp}+{\beta }_{r}t+{\beta }_r{A}_{cp}t$$

where π_rip is the odds of making a move of type r for individual i in period p, π_sip is the odds of no move, α_r is the intercept for move r, α_rc is a vector of county fixed effects for move r, X_ip is a vector of controls for individual i in period p, A_cp is a vector of climate anomalies for county c in period p, t is a linear term for the year (subsequently modified as described below), and β_rA_cp t captures interactions between the time trend and climate anomalies. The linear term for the year subsequently becomes a quadratic transformation, making the final two terms into β_rt + β_rt² + β_rA_cpt + β_rA_cpt². Finally, we replace the linear term with period fixed effects and allow the climate effects to vary before and after the baseline year 2000, creating α_p + β_rA_cpα_p<2000 where α_p<2000 is an indicator for periods beginning before the year 2000 that is included in the period fixed effects. With the inclusion of county fixed effects and the specified time trend, these models are identified by within-county variation in climate over time, as well as national-level variation in climate that deviates from the time trend (in the linear and quadratic cases). We subsequently extend this analysis with several supplementary specifications as described in detail below, including a quadratic model (1) without socio-demographic controls, (2) with added interactions between the time trend and the household and community controls, (3) with permanent moves divided into within-county and out-of-county moves, (4) stratified by gender, (5) stratified by education (no secondary versus any secondary), (6) by whether the community was designated rural or urban (as distinct from rural/urban hukou status, which we do not observe), and (7) stratified by North versus South China. All multinomial results are displayed in exponentiated form as relative risk ratios and subsequently as marginal effects.

As a complementary analysis to provide insight into the potential mechanisms of these effects, we also create a parallel household-period dataset (n = 21,824) in which household and community-level controls are observed at baseline, climate exposures are again measured during the inter-survey period, and the proportion of income received from various sources (i.e., income shares) is observed at follow-up (Supplement Table 1). Using income shares accounts for the highly skewed distribution of income, and allows us to directly examine dependence on agriculture, potentially the most vulnerable economic sector. We exclude the 5% of household-periods with one or more negative income shares. Specifically, we examine the share of crop income (19% on average), transfer income (11%, including migrant remittances), and other sources of income which make up the majority (71%, of which wage and business income are the largest components). Crop and transfer income are of particular interest as agricultural production is directly dependent on climate (Piao et al. 2010), and transfers have also been shown to respond to climate shocks (Gröger & Zylberberg 2016). We analyze this dataset using linear regression with the same set of predictors and clustering corrections as in the analysis of migration described above.

Results

The results for the primary multinomial specifications are displayed in Table 3. We first examine the results of the control variables (displayed with the linear time trend) to ensure they are consistent with expectations and previous studies. As expected, temporary migration is most common for young adults who are male, unmarried, educated, living in rural areas, not heads of household, and connected to migrant networks. Interestingly, there is negative selection on consumer assets. Selection for permanent migrants is mostly in the same direction but generally stronger, as expected given the higher costs associated with these moves. However, permanent moves are also distinguished by being more common for women (likely reflecting patrilocal marriage practices), peaking at ages 25–29, and being negatively selected on business assets. The determinants of individual-level attrition are most similar to those of permanent migrants, suggesting that some fraction of this flow may be composed of unobserved departures from the household. Attrition of whole households, in contrast, is only weakly predicted by individual-level demographics. Instead, it increases with education, migrant networks and consumer assets and decreases with household size, suggesting that small, affluent households might be more likely to refuse or to make unobserved moves. Overall, the results for the controls are consistent with expectations and suggest that attrition might be capturing otherwise unmeasured aspects of migration, reinforcing the importance of considering these outcomes in tandem (Thomas et al. 2012). Nonetheless, the effects of climate are never jointly significant for either form of attrition (below), suggesting that these flows do not contain significant numbers of climate-induced migrants.

Table 3.

Multinomial logistic regression of migration/attrition on climate variables and controls with alternative time trends (relative risk ratios and significance tests, n = 64,829 person-periods).

Specification/Outcome	Temporary migration		Permanent migration		Individual attrition		Household attrition
Linear time trend
Temp anomaly	1.06		1.26	***	1.09		1.01
Temp X year	1.00		0.98	**	1.00		1.01
Precip anomaly	0.96		0.98		1.03		1.01
Precip X year	1.01	+	1.01	*	1.00		1.00
Year	1.07	***	1.00		0.99		1.01
Age 20–24	1.24	***	3.11	***	3.01	***	1.26	***
Age 25–29	1.08		3.63	***	3.75	***	1.14
Age 30–34	0.84	*	2.78	***	3.18	***	1.20	+
Age 35–39	0.72	***	2.80	***	2.70	***	1.07
Age 40–44	0.60	***	2.56	***	2.63	***	0.87
Age 45–49	0.43	***	2.63	***	2.77	***	0.79	*
Female	0.61	***	2.06	***	1.03		1.03
Child of head	1.97	***	17.09	***	7.50	***	1.02
Other relation to head	1.90	***	12.07	***	8.06	***	1.17
Ever married	0.61	***	0.40	***	0.77	***	0.87	+
Primary edu.	1.69	***	1.47	***	1.25	*	1.24	**
Lower middle edu.	1.87	***	1.62	***	1.21	+	1.42	***
Upper middle edu.	1.97	***	1.97	***	1.20	+	1.69	***
Technical school	2.48	***	2.28	***	1.19		1.47	**
University edu.	2.44	***	3.15	***	1.31		2.19	***
Household size	1.00		1.09	***	1.21	***	0.76	***
Migrant network	1.45	***	1.27	***	1.17	***	1.41	***
Consumer assets	0.92	***	0.99		0.95		1.08	**
Business assets	0.98		0.94	**	1.00		0.94	*
Urbanicity index	0.99	***	1.00		0.99	+	1.00
Joint: Climate X year	3.5		12.6	**	0.3		0.6
Quadratic time trend (controls not shown)
Temp anomaly	0.89		1.25	*	0.96		0.92
Temp X year	1.04		0.98		1.04		1.02
Temp X year2	1.00	+	1.00		1.00		1.00
Precip anomaly	0.89	+	1.04		0.95		0.95
Precip X year	1.03	*	0.98		1.04	+	1.03
Precip X year2	1.00		1.00		1.00		1.00
Year	1.11	***	1.09	***	1.00		1.09	**
Year2	1.00		1.00	***	1.00		1.00	*
Joint: Climate X time	7.2		20.3	***	5.4		4.1
Joint: Temp X time	2.8		17.0	***	2.2		0.6
Joint: Precip X time	5.5	+	3.1		2.8		3.3
Stepwise time trend (controls not shown)
Temp anomaly	0.96		0.83	**	1.16	*	0.97
Temp X pre-2000	0.98		1.43	**	0.83		1.10
Precip anomaly	1.08	*	1.00		1.02		0.93
Precip X pre-2000	0.87	*	1.03		1.00		1.05
Pre-2000	0.31	***	0.51	***	0.74	+	0.48	***
Year is 1989	-		-		-		-
Year is 1991	1.06		1.11		0.91		1.58	*
Year is 1993	0.92		1.61	***	0.96		1.09
Year is 1997	1.68	***	1.13		1.03		1.53	*
Year is 2000	-		-		-		-
Year is 2004	0.90		0.55	***	0.40	***	0.53	***
Year is 2006	0.82	+	0.66	***	0.65	**	0.96
Year is 2009	1.17		0.45	***	0.65	*	0.55	*
Joint: Climate X pre-2000	5.4	+	9.8	**	1.9		0.4

Notes: County fixed effects, missing indicators and constant are included but not shown. Reference categories are age 15–19, head/spouse, and no education.

⁺p<0.10,

^*p<0.05,

^**p<0.01,

^***p<0.001

Table 3 also displays the effects of climate anomalies, the time trend, and their interactions for all three specifications of the trend. As expected, the effect of time across the four migration/attrition streams (net of climate effects and interactions) is mostly positive and often highly significant, particularly for temporary migration. To assess whether the effects of climate are modified by time, we examine the joint significance of the climate-by-time interactions for each outcome and each specification of time. The overall finding is that our core hypothesis is strongly supported for temperature effects on permanent migration, weakly supported for precipitation effects on temporary migration, and not supported for other climate-to-migration pathways. More specifically, Table 3 reveals that the climate-by-time interactions are jointly significant solely for permanent migration, with the exception of temporary migration in the stepwise model where they are also marginally significant. The main climate effects, which capture the effects of climate on moves during the 1988–1993 period, are also statistically significant for these same cases, confirming that climate is primarily influencing permanent moves. Below we discuss these results in detail.

In the model with the linear time trend, the climate-by-time interactions are jointly significant for permanent migration only (p = 0.002) and this is driven by temperature-by-time (p 0.006). In the baseline period, the odds of making a permanent move increase by 26% with each standard deviation of two-year temperature (p = 0.001). However the interaction with time is negative, and by the end of the study period a unit temperature increase decreases permanent migration by 11% (p = 0.08). There is also a weaker precipitation story for temporary migration. The interaction with time is positive and marginally significant, which changes the effect from negative and non-significant in 1989 to positive and significant in 2009.

In the model with the quadratic time trend the story is similar: the climate-by-time interactions are highly significant for permanent migration (p < 0.001) and non-significant for the other streams, and again these effects are driven by temperature (p < 0.001). The marginal effects of temperature on permanent migration in this model are visualized in the first panel of Figure 1. Relative to an unadjusted baseline rate of permanent migration of 9%, a one standard deviation increase in temperature increased permanent migration by 1.0 percentage points in 1989 (p = 0.020), but by 1999 this effect became negative, decreasing permanent migration by 0.8 percentage points in the final period (p = 0.002). In this model, there is also a weaker story for precipitation on temporary migration where the interaction terms are marginally significant (p = 0.06). In this case, precipitation has a marginally significant negative effect on temporary migration in the initial period, which increases to a non-significant positive effect by the final period.

Figure 1.

The marginal effects of temperature on permanent migration

In the model with the stepwise time trend the story is again similar: the climate-by-time interactions are highly significant for permanent migration (p = 0.008) and driven by the effects of temperature (p = 0.002). Examining the main temperature effect and its interaction with time reveals that, before 2000, a one standard deviation increase in temperature increased the odds of permanent moves by 19% (p = 0.027), whereas for the 2000 baseline round and later it decreased by them 17% (p = 0.004). In this model, the precipitation-by-time term also becomes significant for temporary migration. Consistent with the story above for the quadratic time trend, precipitation has a non-significant negative effect on temporary migration before the year 2000, which increases to a significant positive effect after the year 2000.

To provide insight into the mechanisms of these effects we also estimate a number of supplementary specifications of the migration model, and we focus the presentation on temperature effects for permanent migration in order to provide insight into the results described above. We use the quadratic time trend, which allows smooth changes in the background context of migration without completely discarding national-level climate shocks. To supplement the main specifications above, we estimate the quadratic model (1) without socio-demographic controls, (2) with added interactions between the time trend and the household and community controls, (3) with permanent moves divided into within-county and out-of-county moves, (4) stratified by gender, (5) stratified by education (no secondary versus any secondary), (6) stratified by rural/urban, (7) and stratified by North versus South China.

The supplementary results for permanent migration are displayed in Supplement Table 2, and visualized for temperature by education in Figure 1. All of the supplementary specifications reveal declining temperature effects on permanent migration over time, though in some cases these results are stronger than others. Examination of the temperature-by-time interactions reveals that they remain jointly significant in most specifications but not all: the effects become non-significant for the less-educated, urban and southern subsamples, but remain so for the more-educated, rural and northern subsamples. They also become non-significant when within-county and out-of-county permanent moves are divided, but retain the same shape, suggesting a sample size constraint. The climate-by-time interactions remain significant when control variables are excluded as well as when all household and community controls are interacted with time, indicating that the observed trends in temperature effects are not driven by observed changes in the composition of the sample and specifically cannot be explained by changing effects of wealth, household size, networks or urbanicity on migration. The strongest contrast revealed in Supplement Table 2, between the more and less-educated samples, is also visualized in the second and third panels of Figure 1. The effects for the sample with secondary education mimic those of the full population, but for individuals with no secondary education the effects are much weaker and depart from zero for only three of eight time periods.

Finally, to examine the potential for a livelihood pathway from climate to migration, we examine climate-by-time interactions for income shares. As displayed in Supplement Table 3, this reveals that the climate-by-time interactions are jointly significant for the “other” income share (non-crop, non-transfer income) using all three specifications of the time trend, and that this is again driven by temperature. In all three cases, the temperature effect on other income is positive in the first period and then becomes negative by the final period. The crop and transfer shares generally move in the opposite direction, but the climate-by-time interactions are jointly significant only with the quadratic time trend. To visualize these effects, we use the quadratic time trend to plot the results for the other income share in the fourth panel of Figure 1. This reveals that temperature increased the other income share at the beginning of the study period, but this effect becomes negative by the middle of the study period. Below we discuss how these results might inform the interpretation of the climate-by-time effects on migration.

Discussion

We present a new case study of changing climate-migration relationships over time in China, adding to previous examples from the US (Gutmann et al. 2005), the Netherlands (Jennings & Gray 2015), and Bangladesh (Call et al. 2017). We document that over our 1989–2011 study period the effect of temperature on permanent migration transitioned from positive to negative, and that these findings are robust to various specifications of the time trend and to interactions between time and a subset of socio-demographic controls. This pattern is strongest for individuals with at least a secondary education and who live in rural areas or in northern China. There is also a weaker but parallel story in which drought (i.e., low precipitation) initially increased temporary migration, but over time began to decrease it. Over this same period, the temperature effects on income shares also shifted, from decreasing the share of agricultural and transfer income to increasing it.

What to make of these changes? Given that changing relationships between socio-demographic characteristics and migration cannot explain these results, we must look to the rapidly changing social and economic context in which migration decisions were made. As described above, China rapidly transitioned during the study period from being a predominately rural and agrarian society to one that is increasingly urban and employed in manufacturing and the service sector. Plausibly, this transition has increased vulnerability to climate shocks in the urban and manufacturing sectors, where total factor productivity has been shown to decline at higher temperatures (Zhang et al. 2018) and labor costs to increase with high temperature days (T_MAX < 35°C) (Zhao et al. 2016). As firms respond to adverse climate exposures, there may be fewer opportunities for work in the urban sectors that attract permanent climate migrants. This interpretation is supported by a rapidly growing literature that identifies urban and manufacturing activities in China and other LMICs as specifically vulnerable to increasing temperatures (Sudarshan & Tewari 2014; Li et al. 2016; Cai et al. 2018; Zhang et al. 2018).

At the same time, agricultural households have gained more tools for crop adaptation and livelihood diversification in the face of environmental shocks (Ma & Maystadt 2017). Increased access to off-farm work, short-term labor migration, and remittances from family members in urban areas have increasingly insulated agricultural households from shocks to their income (Giles 2006). At the beginning of our study period, when agricultural households had fewer tools to deal with climate shocks, migrants moved in response to heat and drought, but as households gained more access to adaptive tools, agricultural incomes became more robust to climate shocks. This, coupled with the increasing vulnerability of other income sectors to climate, allowed people to respond to heat and drought by instead remaining at home. Thus climate shocks might have declined in significance as “push” factors for migration, while at the same time (spatially-correlated) climate shocks likely reduced the “pull” towards destinations.

This study has important implications for the questions asked and the methods used by future research on climate-induced migration. First, we need to know more about how climate-induced migration, and climate vulnerability more broadly, are changing over time (Barreca et al. 2016; McLeman 2018; Burke et al. 2018). The number of long-term longitudinal datasets from LMICs such as CHNS that could enable this research remains relatively small (Thiede & Gray 2017), but these questions can also potentially be addressed using increasingly available micro-data from censuses as well as from demographic and health surveys (Thiede et al. 2016, Eissler et al. 2019). Second, urban and non-agricultural populations in LMICs have been neglected in the study of climate-induced migration and climate vulnerability. Global poverty is transitioning to urban areas (Ravallion et al. 2007), and urban populations face unique exposures to food prices as well as urban heat islands (Hertel et al. 2010; Tan et al. 2010). Climate shocks in urban destinations that are correlated with rural origins might also explain why many studies in low-resource settings have found only muted effects of climate on migration (Carleton & Hsiang 2016).

Finally, these results have important implications for how we view climate-induced migration and how it is likely to proceed under climate change. These results directly challenge the common view that increasing temperatures under climate change will universally displace migrants, and that these effects will intensify over time (Myers 1997). Instead, as LMICs broadly follow the path of China towards urbanization and de-agrarianization (albeit much more slowly and with considerable local variation) the potential for climate to displace migrants may also decline, paralleling what has been observed in the US, the Netherlands and Bangladesh (Gutmann et al. 2005; Jennings & Gray 2015; Call et al. 2017). These results also challenge another common view that population “trapping”, when adverse climate conditions constrain migration, should be viewed as negative and maladaptive. In our case, climate-induced immobility occurs as China develops and the rural sector becomes less vulnerable to climate change. That being said, these results likely have limited relevance to the world’s poorest countries, including those in East, Central and West Africa, which are still very far from experiencing a China-like transition to prosperity (Fox 2012). In these places, it is very possible to envision an initial transition towards increased propensity to climate-induced displacement as the poverty and low-mobility trap is escaped (Nawrotzki & Bakhtsiyarava 2017), followed by a long-term decline in vulnerability.