1. Introduction
There is some increasing scientific evidence that our climate is changing and that more adverse climatic conditions will affect human activity over the world in the future. Besides this, social scientists have paid attention to the way individuals can cope with these adverse developments. The Stern report in 2007 looked at the global consequences of climate change, suggesting that many countries could suffer from the adverse developments related to changing climatic conditions. The first United Nations intergovernmental report on climate change emphasized that human migration might be the most important consequence of climate change, especially in developing countries.
A noticeable prediction of climate change on human mobility was of Myers (Reference Myers2002) who predicted that by 2050, climate change would displace more than 200 million individuals. While such a prediction did not rely mainly on scientific calculations, it reflected that human mobility was already seen as one of the most obvious adjustment mechanisms to cope with the adverse consequences of this evolution of climatic conditions.Footnote 1 Early views on the topic started from the idea that a large number of people would be forced to move. In order to refine these dire predictions and give credit to these early views, social scientists have over the subsequent years attempted to look at the possible relationships between climate change and human mobility. These scientific attempts have been facilitated by the growing availability of data needed to investigate the complex nexus between these two phenomena. On the one hand, data on climatic factors have become more available. As reported by Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017), this holds both for data concerning slow onset climatic factors such as warming and rain precipitations and for measures of the climate-related extreme shocks, i.e., the natural disasters. On the other hand, the scientific literature has also taken advantage of the growing availability of data on human migration, both at the macro and the micro levels.
The key question raised in this literature is quite simple: to what extent do climatic conditions lead to the displacement of people from their initial location? While the basic question is quite simple, the answers provided by the set of empirical studies in this area are much more complex. It is not straightforward to summarize the main finding of that extensive literature given the significant degree of heterogeneity, both in terms of results and in terms of approaches. Recent surveys [see Millock (Reference Millock2015), Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017), Cattaneo et al. (Reference Cattaneo, Beine, Frölich, Kniveton, Martinez-Zarzozo, Mastrorillo, Millock, Piguet and Schraven2019), and Piguet et al. (Reference Piguet, Pecoud and de Guchteneire2011) among others] have tried to summarize the whole literature. They all emphasize the diversity in terms of findings and methods used in this literature. While a substantial proportion of papers find some evidence that climatic shocks tend to displace people, a significant number of findings reach different conclusions. A number of papers find that the connection between climate shocks at origin and movement of people initially located in this area is very weak. Also, the recent literature has also emphasized that in a number of cases, the occurrence of adverse climatic developments reduces rather than enhances the mobility of the affected population. These findings are in line with the concept of trapped population put forward by a number of contributions [see among others Black et al. (Reference Black, Arnell, Adger, Thomas and Geddes2012)].Footnote 2 This conclusion in turn drew the attention of important institutions devoted to development issues such as the United Nations and the European Commission and has been advanced as one of major current and future issues related to climate change.
A further complication of this empirical literature is the diversity of outcomes that are considered in the various analyses. While a lot of papers simply look at the mere displacement effect, other studies make a clear distinction between direct and indirect influences of climatic shocks [see Cattaneo et al. (Reference Cattaneo, Beine, Frölich, Kniveton, Martinez-Zarzozo, Mastrorillo, Millock, Piguet and Schraven2019) e.g., for a discussion]. An additional distinction concerns the difference between partial and total effects of climatic factors on mobility.Footnote 3 The surveys of the literature report the significant diversity in terms of outcomes. This diversity is not totally surprising since one can reasonably expect that the effect of climatic shocks on the propensity to move from individuals might be context dependent. The way these shocks operate might depend on the level of development of the area, the type of economic activity, the various possible adaptation mechanisms, and last but not least the availability of external options.
A complementary explanation of the observed heterogeneity is that findings might also depend on the methodology adopted in the empirical strategy of each paper. Recent surveys of the literature also emphasize this important source of heterogeneity. A specific paper in this literature is at the crossroad of a myriad of methodological choices underlying the empirical analysis. These methodological options concern many dimensions, including the type of country covered, the type of the key data that are used, the period of investigation, the measurement of human mobility and of the climatic shocks, or the adopted econometric methodology. While most recent surveys of the literature emphasize the heterogeneity of its results, they do not look, at least explicitly, at the relationship between the results and the methodology adopted. This paper tries to fill this important gap.
In order to highlight the links between the various dimensions of the adopted methodologies and the findings of climate shocks on mobility, we conduct a meta-analysis of the literature. We understand the term methodology in a broad sense since we include not only in that concept the choice of data or statistical methods but also the context in which the studies were conducted. This for instance includes whether the area is located in a developing country or not, whether the type of mobility is within a country or across different countries or the type of climatic conditions under investigation. We also give details about the exact structure of each regression used to generate the findings. Therefore, our paper aims at providing guidelines for the authors in that literature but also answers to more general questions. One of these important questions is whether there is more evidence of a displacement effect of climatic shocks within a country than across countries. Another one is whether the evidence of displacement effects is specific to developing countries.
In terms of details, we include in the meta-analysis 51 empirical papers that are representative of the existing literature on climate change and migration. We include published and unpublished papers, but restrict our attention to papers using econometric methods looking at the potential displacements effects exerted by climate shocks. These 51 papers give rise to 1,355 regression results that we code along a large number of dimensions. These dimensions include the context of the study, the adopted methods, the type of data, and measures of the key variables as well as the exact types of outcomes that the study tries to capture. The coding of these dimensions results in more than 85 different variables that are used subsequently as potential explanatory variables in our regressions. We look in particular at the impact of these dimensions on the probability that a regression concludes in favor or not of a particular effect of climate change on human mobility. To that aim, we adopt a range of econometric models suited to limited dependent variable such as the pooled logit model, the panel logit model with random individual effects, and the (pooled) order logit model.
Our results show that, in general, results of the literature depend on a large variety of dimensions in terms of methodology and context of the study. The results depend on variables belonging to each broad category in terms of methodology. To illustrate, we find that results tend to depend on the context since there is more evidence of an effect of climatic factors in an area located in a developing country. In contrast, we do not find that studies looking at internal displacement effects find systematically more effects than those looking at international mobility. Results depend on the way mobility is measured in the sense that papers using direct measures of migration find on average more evidence of an effect. Measures of climatic factors tend also to influence the results, albeit in a complex way. For instance, while we find little evidence that a specific type of natural disaster is more associated with a displacement effect, we find that the way slow onset factors and the climate-related extreme events are actually measured has an influence on the results. A final illustration concerns the choice of the econometric method, as we find that papers using panel data and accounting for measurement errors through instrumental variables tend to find more evidence of an effect on mobility. Another important finding is that papers allowing for conditional effects of climate shocks, i.e., effects depending on a specific condition, find also more evidence of an impact.
The paper is organized as follows. Section 2 gives a detailed description of how we code the literature and provides descriptive statistics gaging the representativeness of our sample. Section 3 explains how we carry out meta-analysis and gives the results of our regressions with respect to the various outcomes of the literature. Section 4 summarizes the implications of our results and provides some concluding remarks.
2. Coding the literature
Our dataset consists of 1,355 regression results extracted from 51 papers. These regression results represent our unit of analysis in the subsequent evaluation of the impact of the methodology on the findings. We provide here a short description of all the variables used in the meta-analysis. The codebook of all coded variables, including those which are not used in this paper is provided in Appendix B.
2.1 Coding the evidence of an effect of climate on migration
An important piece of information that we extract from each regression is whether there is evidence of an effect of climate change on migration. The existence of an effect is coded when a coefficient relative to a specific climatic factor is significant at least at the 10% level in the regression involving mobility and these climatic factors. This outcome can be then further decomposed into various categories depending on whether the effect is direct or not and whether the effect is positive or negative.
Direct effect is a binary variable equal to 1 if we find a significant direct effect of climate on migration in the regression. A direct effect is found if there is a direct causal link between a climate variable and migration in the regression, for example, a wave of extreme temperature leading to emigration. Nevertheless, there might be also evidence of an indirect effect of climate on migration, which is captured by the binary variable Indirect effect. We identified two scenarios of an indirect effect in the literature. The first one is when the authors do additional regressions to highlight one specific channel through which climate impacts migration. For example, rainfall variability might impact mobility through its effect on GDP per capita [e.g., Coniglio and Pesce (Reference Coniglio and Pesce2015)]. The second one is when the authors use climate variables as instruments in a two-stage regression. An example is provided by Feng et al. (Reference Feng, Krueger and Oppenheimer2010) who use climate variables to instrument the effect of crop yields on migration.
Direct and indirect effects can be further decomposed either into a positive effect (evidence of a displacement effect) or a negative effect (evidence of increased immobility) using binary variables. To code if an indirect effect is either positive or negative we rely on other results of the paper, such as the results from auxiliary regressions. For instance, if at the same time the authors find in an auxiliary regression that a climate shock decreases crop yields and in their main regression that a decrease in crop yields increases migration, we code the indirect effect as a positive effect.
Since many regressions include several different climate measures at the same time, some choices have to be made in terms of coding. We code a direct/indirect effect if at least one of the climate variables is significant. In the specific case of opposite results of several climate variables, we duplicate the regressions results. Some regressions have indeed negative and positive displacement effects at the same time. This is also the case in the multinomial regressions analyzing the effect of climate on several levels of migration (local, internal, and international). If different results were found in one regression, it was duplicated and coded once as a positive effect and once as a negative effect. The binary variable Split keeps track of the fact that the regression results come from a splitting procedure.
2.2 Coding climate-related measures
While each paper displays some specificity regarding the way the climate variables are included in the regressions, these variables can all be classified into several broad categories. To start with, they can be classified into long-run (slow onset) factors and climate-related extreme events. To control for the fact that some papers include both types of climate variables in the same regression, we define the binary variable Joint Inclusion.
2.2.1 Capturing the slow onset climatic factors
The variable Long Run captures that the regression includes long-run climate measures. Regarding the long-run effects, we identified four measures of temperature and four measures of rainfall that span the whole spectrum of measures used in the empirical literature:
• Temperature or precipitation levels [see e.g., Cattaneo and Peri (Reference Cattaneo and Peri2016)];
• Temperature or precipitation deviations [see e.g., Beine and Parsons (Reference Beine and Parsons2015)];
• Temperature or precipitation anomalies [see e.g., Beine and Parsons (Reference Beine and Parsons2015)];
• Temperature or precipitation variability [see e.g., Coniglio and Pesce (Reference Coniglio and Pesce2015)].
More recently authors have relied on a soil moisture measure which we code separately as Soil Moisture. This measure is usually based on the Standard Precipitation Evapotranspiration Index. It aims at combining rainfall and temperature in a single variable [see e.g., Mastrorillo et al. (Reference Mastrorillo, Licker, Bohra-Mishra, Fagiolo, Estes and Oppenheimer2016)]. We also create a dummy variable labeled joint_temp_rain capturing whether the regression includes jointly factors in terms of rainfall and temperature.
2.2.2 Capturing the climate-related extreme events
The climate-related extreme events considered in the literature usually belong to the category of natural disasters. We create a Natural disasters binary variable taking unity if the regression includes climate-related extreme events. Most papers looking at the effect of natural disasters focus on one particular type of climatic event. Six specific climate-related disasters tend to emerge in the literature. Among these “popular” disasters, earthquakes are definitely worth being investigated but it is unclear and subject to controversy if earthquakes are related to climate change. We therefore disregard earthquakes as specific disasters under investigation. We therefore consider the following specific types of disasters:
• Extreme temperatures [see e.g., Hirvonen (Reference Hirvonen2016)].
• Extreme precipitations [see e.g., Thiede et al. (Reference Thiede, Gray and Mueller2016)].
• Floods [see e.g., Ruiz (Reference Ruiz2017)].
• Hurricanes and storms [see e.g., Koubi et al. (Reference Koubi, Spilker, Schaffer and Bernauer2016a) or Mahajan and Yang (Reference Mahajan and Yang2017)].
• Droughts [see e.g., Dallmann and Millock (Reference Dallmann and Millock2017) or Ruiz (Reference Ruiz2017)].
Additionally, we coded three variables to capture the way each considered disaster were measured and coded in the original paper. The benchmark case is the simple occurrence of at least one disaster of this king over the considered time period. We code additional variables to capture further characteristics of the disasters. The variable Count takes 1 if the authors use the aggregate number of events over a period of time [see e.g., Beine and Parsons (Reference Beine and Parsons2015)]. Intensity takes 1 if the authors use an intensity measure of the disaster such as the number of affected people or the amount of damages [see e.g., Bohra-Mishra et al. (Reference Bohra-Mishra, Oppenheimer and Hsiang2014)], and Duration takes 1 if the authors use a measure of duration such as the number of consecutive months [see e.g., Dallmann and Millock (Reference Dallmann and Millock2017) or Ruiz (Reference Ruiz2017)].
2.3 Coding the dependent variable
We created four categories for the dependent variable that is used in the regressions of the literature. Direct measure is a binary variable capturing if migration is directly observed and measured. Examples for direct measures of migration can be found in survey data in which individuals are directly asked about their migration history [see for instance Gray and Bilsborrow (Reference Gray and Bilsborrow2013)]. It can also be measured directly for instance when the administrative data reports the origin and the timing of the movements of people. In contrast, when mobility is inferred, the variable Direct measure takes the zero value. This is for instance the case when migration flows are built from differences in migration stocks captured from Census data. Researchers also use different dependent variables to infer the impact of climate change on mobility. The dependent variable might capture a flow (Migration Flow [see Coniglio and Pesce (Reference Coniglio and Pesce2015)]), or a rate (Migration Rate [see Beine and Parsons (Reference Beine and Parsons2015)]) or another measure proxying mobility Other. Other can either be an alternative measure of migration or a dependent variable different from migration such as the rate of urbanization which proxies internal migration in the absence of such a measure [see Barrios et al. (Reference Barrios, Bertinelli and Strobl2006)]. It can also be variables used in auxiliary regressions to capture specific channels of influence.
2.4 Coding the channel
As discussed previously, some authors analyze the channels through which the climate variables affect migration. A specific channel is highlighted in different cases. The first one is when the dependent variable is not migration (e.g., Crop yields, or GDP per capita). The second case occurs when there is an interaction term between climate variables and another variable that refers to a specific channel. We code four main channels considered in the literature: the economic channel [Beine and Parsons (Reference Beine and Parsons2015)], the agriculture channel [Cattaneo and Peri (Reference Cattaneo and Peri2016)], other channels [such as the urbanization channel, see Marchiori et al. (Reference Marchiori, Maystadt and Schumacher2012)], and in case no channel is specifically highlighted, an aggregation of channels.
2.5 Coding mobility
We identified three types of mobility in the literature and coded them using three binary variables. Internal migration and international migration make up for the majority of the sample but some authors also analyze local displacement (e.g., migrants moving from one village to another one).
2.6 Coding the data
We code various features of the characteristics of the data used in the regressions. The variable Developing only is a binary variable capturing if the regression is based on a sample of observations involving only developing areas as origins of the potential emigrants. We capture the frequency of the data as well. This variable is expressed in years. In the case of a cross-section, the frequency is set to 0. If the length between each wave of data differs, we take the average frequency [see e.g., Gray and Mueller (Reference Gray and Mueller2012b)]. The variable Cross Country identifies if the regression uses cross-country data. The starting time of the sample as well as the time span of the sample are coded as well.Footnote 4 The time span variable is expressed as the number of years of the period under study. If the data are dyadic, i.e., they capture bilateral movements from a given origin to a given destination, it is captured by the binary variable Dyadic.
2.7 Coding the context of the regression
A number of characteristics of each regression are also coded. The variable Theory based is a binary equal to 1 if the empirical analysis is derived from a theoretical model such as the Random Utility model which is often the underlying framework for the gravity regression models. We coded the binary variable Main as being equal to 1 if the regression belongs to the core of the econometric analysis, as opposed to being involved in a robustness analysis. We code the regression to be an Auxiliary Regression if the dependent variable is not directly related to migration. Many authors run auxiliary regressions to highlight the underlying channel rationalizing their findings. The binary variable Elasticity is equal to 1 if the estimated coefficient is an elasticity or a semi-elasticity. If the regression is restricted to a sub-sample, for example some regressions being run for male migrants only or in a subset of countries, the binary Conditional Sample equals 1. If the climate variable is interacted with another variable to highlight a certain channel [e.g., rainfall level interacted with sub-Saharan countries as in Barrios et al. (Reference Barrios, Bertinelli and Strobl2006)], the variable Conditional Regression takes unity. Finally, we include a binary variable to identify if the regressions control for additional variables not related to climate.
2.8 Coding the context of the study
Several characteristics of the context of the study were coded for each paper. Some of them are self-explanatory, such as the number of authors, the year of publication and whether the paper is published in a peer-reviewed journal or not.Footnote 5 If the paper is published, we capture the impact factor of the journal as reported on the journal's webpage. We also capture the number of citations and the average and maximum h-index of the authors reported by Google Scholar. This information was coded around the same time for all papers.
2.9 Coding the estimation technique
The most popular estimation techniques in the literature are coded as binary variables. Panel takes 1 if the paper uses panel data along with panel specific techniques. IV, OLS, Poisson, and Multinomial logit capture regression techniques using instrumental variables, ordinary least squares, Poisson, and multinomial logit models, respectively.Footnote 6
2.10 Sample selection and representativeness
2.10.1 Sample selection of papers
The aim of our analysis is to uncover the complex links between the methodological approaches and the results. The choice of the papers included in our analysis results from a specific strategy aimed at yielding a sample of empirical studies as representative of the literature as possible. While the inclusion of 51 papers might seem slightly restrictive at first glance, the selection of papers follows a specific strategy. It is worth emphasizing that the total number of papers is constrained by the amount of work and time devoted to the coding (the coding of these papers took more than 2 months of work in total, which means more than 1 day of coding per paper).
The criteria of inclusion of the papers are the following ones. First, we include only empirical papers making use of econometric techniques to identify any possible link between climatic factors and mobility. In that respect, we exclude papers studying the possible effects based on quantitative models of migration [see for instance Burzynski et al. (Reference Burzynski, Deuster, Docquier and De Melo2018)] even though they estimate a subset of the parameters used to simulate expected impacts of projected climatic conditions. We also exclude papers that make use of migration and climatic data but that do not make use of econometric regressions, such as Dun (Reference Dun2011), Findley (Reference Findley1994), or Noy (Reference Noy2017). Second, among the econometric analyses, we exclude papers whose main focus is not the connection between mobility and climate change.Footnote 7 Third, among the econometric papers, we favor those quoted in the recent surveys of the literature mentioned before. We complement this selection by recent papers quoted by other work. This allows to set an implicit minimal level of perceived quality of the included papers and avoid to include more “esoteric” studies. Finally, in order to be able to pin down the methodologies that could explain the variation in the obtained findings, we make sure that we have a more or less balanced sample of regression results in terms of the main methodological approaches. Tables 1 and 2 provide evidence that our sample of regressions is more or less balanced in terms of these main methodological approaches.
Summary statistics of categorical variables—part 1
Summary statistics of categorical variables—part 2
Tables 1–3 provide a set of descriptive statistics computed from our unit of analysis (regressions) on the main methodological dimensions. Table 1 reports the proportion of regressions with a specific characteristics captured through a dummy variable. Tables 2 and 3 provide the mean of the continuous covariates of our regressions (such as the specific measures of the climatic conditions). Table 20 in the appendix reports the list of 51 empirical papers from which these regression results are extracted.
Summary statistics of continuous variables
For the main dimensions, our sample of regressions generates enough variability in order to identify the impact of these methodological choices on the outcomes. By main dimensions, we mean the fact that (1) the paper is published or not, the fact that (2) the regressions use conditional regression models or not, (3) use conditional samples or not, the fact that they look at (4) international mobility or not, (5) internal mobility or not, (6) the fact they are based on some theory, (7) they cover developing countries only, (8) they use cross-country data, or (9) different types of individual data (households, individual agents, etc.). Another important dimension is (10) whether mobility is directly measured or not and (11) the way the dependent variable capturing mobility is expressed (migration flows or rate or other proxies).
Tables 1 and 2 report the proportion of categorical variables taking the value 1. For some of these, we should not expect to have an equal allocation across the possible values. For instance, the fact that about a fifth of the regressions in our sample finds a negative effect of climatic factors on mobility suggests that this result is far from being anecdotical, even though a majority of regressions do not conclude in favor of such a result. Likewise, the fact that 44% are based on a theoretical framework shows that the literature pays some decent attention to the underlying theories, even though the majority of regressions are mostly data driven.
2.10.2 Representativeness of the included papers
In spite of the selection criteria, we might face the legitimate concern that our sample is not fully representative of the existing literature. While the existing literature is an elusive object, we carry out two preliminary analyses to document to what extent our sample of papers is an unbiased subset of the relevant population of papers.
First, we look at whether our included papers are quoted in recent surveys of the literature. The idea is that these surveys span more or less the existing relevant literature. Therefore, the majority of the papers we included should be quoted by these surveys. We should also make sure that the unquoted papers that we have included display more or less the same observable characteristics of the quoted papers.Footnote 8 We select four recent surveys: Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017), Millock (Reference Millock2015), Cattaneo et al. (Reference Cattaneo, Beine, Frölich, Kniveton, Martinez-Zarzozo, Mastrorillo, Millock, Piguet and Schraven2019), and Piguet et al. (Reference Piguet, Kaenzig and Guélat2018).Footnote 9
Table 4 suggests that the majority of our included papers (about 71%) are quoted by at least one survey. “Only” 15 papers out of 51 are not reported by any of the chosen survey. A subset of these are not included for obvious reasons, for instance because they are recent and still unpublished [see for instance Ruiz (Reference Ruiz2017)].
Quoted and unquoted papers in recent surveys
Column (2) gives the number of papers that are quoted by the reported number of surveys given in column (1). Column (3) gives the corresponding proportion of papers.
Second, we can compare our sample of papers with a list of quoted papers in the most relevant survey. We choose the one of Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017) since it is the most recent published survey with a focus on the economic literature. We identify in Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017) 34 papers that match our selection criteria. Out of these 34 papers, we include 28 papers, which means that more than four papers out of five reported by Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017) are treated in our meta-analysis. All in all, we are confident that our 51 papers form a representative sample of the empirical economic literature on climate change and mobility.
3. Results
3.1 Econometric approach
In order to investigate the impact of methodology on the results in the academic literature, we carry out a meta-analysis relating these results to a set of key characteristics of these approaches. It is important to understand that the unit of the analysis is at the level of regression. We have basically 45 coded papers, leading to 1,307 regression results. In order to get robust results, we conduct two different econometric procedures. We focus on the common results generated by both econometric approaches. The first one relies on a logit specification, pooling all regressions together:
where Prob(y ij=1) is the probability that regression j in paper i gets a specific outcome y, x ij is a vector of characteristics of regression j in paper i, β is a vector of parameters of interest to be estimated, Φ() is the logistic function, and ɛ ij is an i.i.d. error term. In order to account for the underlying correlation between regressions of the same paper, we cluster the standard errors of the β coefficients at the paper level. A panel approach with paper fixed effect is obviously unfeasible, given the low variability across regressions of the same paper. Nevertheless, we can account for unobserved heterogeneity using a panel regression model with (paper) random effects (RE). The RE logit model takes the following form:
with c i the set of random effects at the paper level, with no particular assumption about how these are related to the x ij [see Wooldridge (Reference Wooldridge2010), chap. 15 for a discussion].
In the subsequent tables reporting the estimations results, we report the marginal effects of the variables on the probability to find a specific outcome (such as the displacement effect, a direct effect, or a negative effect on the mobility of people). The marginal effects are computed at the mean of the explanatory variables.
3.2 Explaining outcomes of studies
In this section, we look at the impact of the methodological choices in terms of the results. We consider a large variety of outcomes: evidence of any effect, direct vs. indirect effect, positive, negative, or no effect on mobility. We consider the main features of the analyses, such as the type of data, the econometric methods, or the way they capture and measure mobility. Given the large set of potential factors, we do not consider here the issues of the specific modeling choices of climatic factors, which is also an important issue. This will be investigated in more detail in the subsequent section.
3.2.1 Probability of any effect
Tables 5 through 8 report the estimation results concerning the impact of methodology on any type of effect of climatic factors. By any type, we mean that the effect can be positive (displacement effect) or negative (increase in immobility), direct or indirect. Tables 5 and 6 report the results using all regression results while Tables 7 and 8 are based only on a sample that excludes auxiliary regressions. Auxiliary regressions in that literature are usually conducted to provide a refined assessment of a previous piece of analysis, such as uncovering indirect effects in the case of little evidence of a direct one [see for instance Beine and Parsons (Reference Beine and Parsons2017), Cattaneo and Peri (Reference Cattaneo and Peri2016), or Feng et al. (Reference Feng, Krueger and Oppenheimer2010)]. Tables 5 and 7 use pooled logit estimates while results in Tables 6 and 8 rely on the random effect panel estimation.
Probability of any effect: pooled logit estimates
Marginal effects are reported in the table.
Standard errors computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Probability of any effect: panel logit estimates
Marginal effects are reported in the table. Panel logit estimation with paper random effects.
Standard errors are computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Probability of any effect (excluding auxiliary regressions): pooled logit estimates
Marginal effects are reported in the table. Sample excludes auxiliary regressions.
Standard errors are computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Probability of any effect (excluding auxiliary regressions): panel logit estimates
Marginal effects are reported in the table. Sample excludes auxiliary regressions.
Standard errors are computed by the Delta method. Panel logit estimation with paper random effects.
*p < 0.10, **p < 0.05, ***p < 0.01.
Our results bring support in favor of some influence of the various methodological categories. First, the way mobility is measured matters. In particular, we find that the use of migration flow rather than other measures (such as migration rates or proxies of mobility) increases the probability of finding an effect. This effect amounts to about 20% depending on the specifications and the estimation method of our paper. This effect is even more important when focusing on the main effect on mobility, i.e., excluding auxiliary regressions. The type of data used in the econometric analysis is definitely an important factor. Frequency plays a clear role: data sampled at higher frequencies tend to support more the case of an effect on mobility. Such a result is consistent with the fact that migration measures spread over several years might be less able to capture short-term movements of individuals in response to climatic conditions. Results from Tables 5 and 6 suggest that using proxies to capture mobility is also more likely to deliver significant results. This effect is quite substantial since the probability of finding an effect increases by about 20%.
Second, the way regressions are conducted is also an important factor. The use of conditional regressions allowing the impact on mobility to be dependent on a specific condition (such as the importance of the agricultural sector) tends to generate more evidence of an effect. The increase in the probability due to the use of conditional regressions is above 20%, both when including or excluding auxiliary regressions. This finding tends to be quite robust across a lot of investigations that we carry out throughout this meta-analysis. Relying on instrumental variable estimation—although its use remains quite limited in the literature—tends to generate more evidence of an effect, especially when excluding auxiliary regressions. This effect is important since the increase in the probability is about 25%. This might reflect that attenuation bias associated with measurement errors of the climatic factors is an important issue in this literature.
Third, the context of the study seems to play some role. The results clearly support a role for more recent analyses (see the negative effect of the starting period of the sample). This might reflect that migration plays a more important role over time as an adjustment mechanism but also that climatic shocks have tended to become more adverse over time. The context in terms of geographical coverage also matters. In the pooled logit estimations, analyses involving mainly developing countries tend to find a more important role of mobility, which is in line with the perception that developing countries face a double risk in terms of climate change [Cattaneo et al. (Reference Cattaneo, Beine, Frölich, Kniveton, Martinez-Zarzozo, Mastrorillo, Millock, Piguet and Schraven2019)]. Finally, while we do not really find a role for authors' reputation (through the use of some measures involving the h index), we find some evidence in favor of a small publication bias. We find that the fact that the paper is published in a journal with a high impact factor tends to increase the probability of finding an effect by 2–3% (see Tables 7 and 8). This suggests that the publication bias in this literature tends to be rather modest, as assessed also by the insignificant effect of a simple variable capturing whether the paper has been published or not.
3.2.2 Probability of a direct effect
We carry out a similar analysis, but restrict our attention to evidence in favor of a direct effect of climatic factors. We therefore look at the probability of a direct effect as opposed to either no effect or an indirect effect. Tables 9 and 10 report the results. In each table, columns (1) and (2) include the results using all regressions, while results in columns (3) and (4) are obtained excluding auxiliary regressions as before.
Probability of a direct effect: pooled logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Probability of a direct effect: panel logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method. Panel logit estimation with paper random effects.
*p < 0.10, **p < 0.05, ***p < 0.01.
The results are quite similar to those of Tables 5 through 8, which is not totally surprising given the fact that most regressions focus on direct effects. The identification of indirect effects has been considered in a few papers (10 out of 51 in our sample) and is often provided as a complementary piece of analysis to the core of the paper.Footnote 10 The results of Tables 9 and 10 emphasize the role of the starting period of time, of using conditional effects, of the frequency of the data and of measuring directly mobility. In addition, the results show that the use of dyadic data that are for instance involved in gravity models produces more evidence in favor of a direct effect. The use of dyadic data tends to increase the probability of a direct effect by about 15%. The use of multinomial logit frameworks allowing for the occurrence of several alternatives in terms of mobility also increases the probability of a direct effect of about 17%.
3.2.3 Probability of a displacement effect
We turn now to the investigation on the probability to find some displacement effect of climatic shocks, i.e., the fact that these shocks lead to an increase in human mobility. This is probably the effect that has received the most important attention from researchers in that literature. This is also the main type of effect that most people have in mind when they think about the connection between climatic factors and mobility. Tables 11 and 12 report the results. In each table, columns (1) and (2) report estimates aiming at looking at the probability of a positive effect. The alternative to a positive effect is therefore either no effect or a decrease in mobility triggered by climate shocks. In columns (3) and (4), the analysis looks at the impact on the probability of finding a direct displacement effect.
Probability of displacement effect: pooled logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Probability of displacement effect: panel logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method. Panel logit estimation with paper random effects.
*p < 0.10, **p < 0.05, ***p < 0.01.
Most determinants that explained the probability of having an effect of any type or a direct effect tend also to explain the evidence in favor of a displacement effect. From the estimates in columns (1) and (2) of Tables 11 and 12, variables relative to the starting time of the analysis, the publication in a good journal, the coverage of developing countries, the use of dyadic data, the use of a direct measure of mobility, the use of flows as the dependent variable in the regressions, and the adoption of IV estimation tend to provide more evidence of a displacement effect. The use of a direct measure of mobility tends to increase the probability of finding a displacement by a substantial effect of about 30%. The use of migration flows increases that probability by 20%. Consistent with this picture, the use of proxies to measure mobility decreases the probability of finding a direct displacement effect by 25%.
In contrast, using conditional regressions does not seem to have some explanatory power, which suggests that this approach is relevant for both positive and negative effects of climate shocks in terms of mobility. The same holds for the use of panel data estimation techniques. These methodological choices turn out therefore to be as important for uncovering displacement effects and the trapped population phenomenon recently emphasized in the literature. The context of the study plays a moderate role for the evidence of a displacement effect. The choice of developing countries tends to increase such a piece of evidence by about 5% to 10%.Footnote 11
When talking about the impact of climate change on migration, the most immediate effect people have in mind is probably the combination of a positive and a direct effect of climatic shocks. Still, such an effect results from the combination of two separate effects, which in turn might be influenced by different methodological choices. Columns (3) and (4) of Tables 11 and 12 show that a subset of the variables explaining the previous considered effect tend to explain the occurrence of this type of result. This is the case of the period of the analysis, the publication in a good journal, the coverage of developing countries and the use of direct measures of mobility. Besides this, additional methodological choices tend to lead to more evidence of a direct displacement effect. This is basically the case for regressions using cross-country data. Once again, measuring mobility through other variables than migration flows (including proxies such as urbanization rates) tends to lower the probability of such an effect. This result confirms that the measure of mobility is of tremendous important in this literature.
3.2.4 Probability of a negative effect
The issue of adverse climatic events such as natural disasters as shocks trapping the affected population has been increasingly emphasized by some authors, such as Black et al. (Reference Black, Arnell, Adger, Thomas and Geddes2012). Empirically, the case for finding negative effects of climate shocks on mobility, i.e., cases in which individuals tend to stay more on average in the affected area is not an isolated econometric fact. An increasing number of papers have documented such cases in a variety of contexts. Our sample of regressions reflects this trend. About 20% of our regressions report such an effect associated at least with one of the included climatic shock. In fact, a large majority of the papers (38 out of 51) included in our analysis tend to display at least one effect of this type, albeit some time in an isolated or subtle way.
In general, as reported by Table 13, we do not find many obvious patterns for generating negative effects of climate shocks. The most convincing feature is the use of conditional regressions. The possibility of conditioning the mobility effect on specific conditions (such as the level of income) increases the possibility of capturing the immobility effect of climate shocks by about 10%.Footnote 12 This result is in line with the findings of some papers in favor of a liquidity conditioning effect of natural disasters for instance. The use of alternative proxies to migration measures to capture mobility also favors the findings of some negative impact (with an effect of about 15%). Direct measures of mobility also seem, to a certain extent, to favor the results of some increased immobility, but this finding is not supported by the panel regressions. The results suggest that studies documenting the so-called trapped population effect come from quite heterogeneous backgrounds in terms of the adopted methodology and that the existence of such effects might be very context specific.
Probability of a negative effect
Marginal effects are reported in the table. Panel logit estimation with paper random effects [cols (3)–(4)].
Standard errors are computed by the Delta method and clustered at the paper level [cols (1)–(2)].
*p < 0.10, **p < 0.05, ***p < 0.01.
3.2.5 Ordered logit estimates
In this section, we combine in a single regression the three possible outcomes on mobility in order to overcome the limitations of using binary alternatives. We carry out an ordered logit estimation with outcomes ordered with respect to an increase in mobility. The three modalities are therefore (i) evidence of a negative effect, (ii) no effect, and (iii) evidence of a positive effect of climatic shocks on the propensity of people to move. This means that we look at the features in the analysis favoring either a displacement effect or a reduction in immobility of people. Of course, the combination of both effects in the same regression rests on a potentially strong assumption that implicitly considers the reduction of immobility and the increase of mobility as comparable effects. Nevertheless, such an analysis has the advantage of including all regressions in the same estimation on the one hand, and to overcome the limitation of binary alternatives (with alternatives capturing heterogeneous situations) on the other hand.Footnote 13
The results allow us to identify the methodological choices that favor a displacement effect of climate shocks. These include the fact that the paper is published in a journal with a high impact factor, the use of a conditional sample, the use of a direct measure of mobility, the use of dyadic data and the adoption of IV estimation. We should nevertheless emphasize that the fit of the ordered logit model remains quite low. This might reflect that mixing up the three possible outcomes in a single analysis relies on some strong assumptions that are not fully supported by the data (Table 14).
Probability of an increase in mobility: ordered logit estimates
The three modalities in the order are negative effect, no effect, and positive effect.
Standard errors clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
3.3 Modeling choices of climatic factors
In this section, we focus on the influence of specific choices in terms of modeling and/or measuring climatic factors. This dimension is important since there are extensive discussions in the literature about the consequences of adopting particular measures of the climatic conditions.Footnote 14
3.3.1 Joint inclusion of slow onset factors and natural disasters
In this section, we raise a simple question: to what extent is it important to include jointly the slow onset factors such as gradual warming or variations in rainfall and the extreme climatic shocks such as the natural disasters? There is to a certain extent a dichotomy in the literature. A first portion of the literature focuses specifically on the impact of gradual changes in the climatic conditions such as the increase in temperature, the decrease in average rainfall, or the changes in the cyclical patterns of annual rainfalls. Other papers investigate the role of climate-related extreme events. These papers look at the impact of various types of natural disasters. More recently, some papers have tried to integrate both types of factors in order to isolate their respective influence in a more convincing way. Nevertheless, it is unknown whether the joint inclusion of both factors exerts an actual influence on the findings.Footnote 15
Table 15 reports the results. The key variable is the joint inclusion of the long-run and short-run factors (“joint inclusion LR-SR” variable in Table 15). In our sample, 14 papers and about a quarter of the regressions (22.7%) jointly include both types of factors. Columns (1) and (2) look at the impact on finding any type of effect while columns (3) and (4) look at direct effects. As before, we use pooled and panel logit estimation. In order to keep the model parsimonious, we include only controls for which we found significant effects in the previous estimations of subsection 3.2.
Impact of joint inclusion of LR and SR factors
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level [cols (1) and (3)]. Panel logit estimation with paper random effects [cols (2) and (4)].
*p < 0.10, **p < 0.05, ***p < 0.01.
The results dismiss the idea that failure to include both measures might significantly affect the results. In some sense it is some good news as it suggests that the findings emerging from the majority of papers that focus only on one type of factor are not subject to some bias due to the omission of the other factors. It is also a positive result to the extent that collecting appropriate measures of one type of factor is already an important and often tedious task.
3.3.2 Influence of modeling slow onset climatic factors
In Tables 16 and 17, we look at the role of modeling choices of slow onset factors. The literature does not display any strong consensus about the way researchers should model long-run climatic factors. As reported by Berlemann and Steinhardt (Reference Berlemann and Steinhardt2017), most papers consider temperature and/or rainfall, but diverge on their specific measures. We therefore code basically these two main dimensions and capture the diversity of measures which are used in the literature. In that respect, authors use either levels of these factors, deviations from some long-run average (with or without scaling with variability measures) or measures pertaining to variability of these factors. Another issue is whether the papers include jointly temperature and rainfall in their analysis. Auffhammer et al. (Reference Auffhammer, Hsiang, Schlenker and Sobel2013) claim that failure to include both factors can lead to an issue of omitted variables in the estimation of the economic impact of climate change. We also test whether this might have an impact on the mobility outcome. Related to that, in some recent papers, researchers use measures of soil moisture which combine both types of long-run factors.
Impact of modeling slow onset factors: pooled logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level.
*p < 0.10, **p < 0.05, ***p < 0.01.
Impact of modeling slow onset factors: panel logit estimates
Marginal effects are reported in the table.
Standard errors are computed by the Delta method. Panel logit estimation with paper random effects.
*p < 0.10, **p < 0.05, ***p < 0.01.
The results of Tables 16 and 17 suggest that the way these long-run factors are modeled definitely plays a role. Studies relying on measures that capture the variability of precipitation seem more inclined to find an effect on mobility, with an increase in the probability of about 20%.Footnote 16 The same holds for papers using rainfall levels, albeit with a more moderate effect (around 10%). The use of soil moisture as a measure of long-run climate change increases the probability of finding a direct effect. The effect is quite substantial, with an increase of the probability of around 25%. We do not find that failure to account jointly for both long-run factors (i.e., measures of rainfall and temperature) tend to influence the outcomes on mobility: the joint_temp_rain variable capturing the joint presence of rainfall and temperature measures is never significant in any regression.
3.3.3 Influence of modeling natural disasters
We look at the role of the way extreme events, namely natural disasters, are captured in empirical studies on the impact of climate change. In this section, we look at the way these natural disasters are measured and disregard the type of natural disaster. The benchmark reference level is the simple occurrence of at least one natural disaster in each period of time. To overcome the simplicity of the simple occurrence measure, papers use different alternative measures: the number of these events over the period of investigation (aggregate count), a measure of intensity such as the number of casualties of affected individuals (intensity) or a measure of duration (such as the proportion of the period subject to such a natural disaster).Footnote 17
Table 18 reports the estimation results regarding these features. We look at the three possible outcomes (any effect, direct effect, and direct displacement effect), using as before pooled and panel logit regressions. The use of simple measures such as the mere occurrence has often led to failure to find a direct effect of natural disasters on the displacement of people. This might be explained in two different ways. First, self-reported measures of natural disasters such as the EM-DAT data can be subject to large measurement errors. This is especially the case for such measures covering older periods of time such as the 1980s and 1990s. This has led some researchers to deal with this issue.Footnote 18
Impact of modeling and measuring natural disasters
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level [cols (1), (3), and (5)].
Panel logit estimation with paper random effects [cols (2), (4), and (6)].
*p < 0.10, **p < 0.05, ***p < 0.01.
A second explanation is that measures of magnitude of these natural disasters are needed to capture their effect beyond the mere occurrence of an extreme event. The estimation results of Table 18 test the relevance of this hypothesis. While the way natural disasters are measured is definitely an important issue, the results of Table 18 do not support that this matters for bringing more (or less) evidence of a direct displacement effect. Nevertheless, measures matter for other types of effect. The use of count measures rather than the simple occurrence tends to give rise to more evidence of any effect of climate change, with an increase of the probability of about 10%. The use of intensity or duration measures lowers the probability of finding direct effects of climate change on mobility by about 10%.
3.3.4 Influence of type of natural disasters
Finally, we look at whether some specific natural disasters tend to be more associated with evidence of a mobility effect of these events. While there are numerous types of various disasters, we restrict our attention to the most used ones in the literature, namely floods, storms and hurricanes, droughts, extreme temperatures, and extreme precipitations. Table 19 reports the results of this investigation. The structure of this table mimics the one of Table 18, with the exception that the last two columns (5) and (6) which report the impact for a displacement effect instead of a direct displacement effect. To sum up the main findings, we do not find some robust evidence that a particular type of disaster is more (or less) associated with an effect in terms of mobility, with the exception of disasters classified as extreme temperatures and floods. Studies focusing on extreme temperatures and floods find less evidence that such a shock has an impact on the mobility of people, by a respective impact on the probability of 15% and 10%. The occurrence of floods is likely to decrease the finding of displacement of people by about 20% with respect to the base case. Nevertheless, the whole picture still remains, in that no other particular disaster seems to be more prone to induce some displacement of populations.
Impact of types of natural disasters
Marginal effects are reported in the table.
Standard errors are computed by the Delta method and clustered at the paper level [cols (1), (3) and (5)].
Panel logit estimation with paper random effects [cols (2), (4) and (6)].
*p < 0.10, **p < 0.05, ***p < 0.01.
4. Conclusion
This paper provides a meta-analysis of the empirical literature devoted to the identification of the complex link between climatic factors and mobility of people. This literature has reached very different results in terms of the effect of climatic shock on the propensity of people to relocate elsewhere. This diversity of results is reflected by the fact that a significant subset of papers conclude in favor of the three possible outcomes with respect to this basic relationship. While some papers find evidence of a displacement effect, others find either no evidence of an effect of climate shocks or even an opposite effect, i.e., adverse climatic developments increasing the immobility of the affected individuals. As such, this diversity is not surprising given the large range of different contexts that are covered by the literature. Our meta-analysis allows to investigate the specific role of the adopted methodology in explaining specific results obtained in the empirical studies, in addition to the context dependent findings.
In our analysis, the term methodology encompasses many dimensions. We start from the fact that each regression in each paper is at the crossroad of a large set of various methodological choices that can potentially affect the findings. We indeed code a large set of characteristics in terms of methodological approaches adopted in each paper and each regression used to assess the impact of climate change. We look at the type of mobility that is considered in each regression. We code the type of data and measures used to capture mobility of individuals. The same applies to the way climatic variables are modeled in the various papers in the literature. Our analysis takes into account the context of the study (e.g., whether it concerns a developing country or not). We also code the context of each regression (e.g., use of a conditional sample). We also pay attention to the econometric methods. Finally, we look at the characteristics of the authors such as their previous citations and the reputation of the journals where the paper is published. One of the goals of our analysis is to identify the most important features of the empirical studies in this literature that can lead to more (or less) evidence of a displacement effect of climate change.
Our results emphasize the importance of the various broad categories of methodological choices. First, the adoption of particular measures matters a lot in this literature. Using high frequency data such as annual data allows us to capture the short-run mobility of people, which obviously increases the probability of finding an effect. The use of direct measures of mobility is also important. Measures of mobility that are computed or that are derived from proxies tend to generate less evidence in favor of an effect. The use of migration flows as the dependent variable in the regressions also increases the probability of finding an impact. The way climatic factors are modeled and measured turns out to play some role, too. The way slow-onset factors such as temperature and rainfall variations are measured is also important. Using variability and rainfall levels tend to deliver more evidence of an effect. The recent use of soil moisture which aims at combining warming and rainfall in a single indicator has the opposite effect. Regarding climate-related extreme events, while we do not find any systematic pattern for a stronger role of a specific type of natural disaster, we do find that the way they are measured matters. In general, relying on indicators simply capturing the occurrence of a disaster over a certain period of time tends to lower the probability of finding an effect. Indicators capturing the intensity of the disaster such as the number of affected people, or reflecting the duration of a climatic event deliver less evidence in favor of a direct effect of this disaster.
Second, the context of the paper and the specificity of each regression play a role. Investigating the occurrence of mobility in the developing world increases the probability of finding some effect. This confirms that the issue of climate change and migration concerns primarily developing countries. This might be explained by the fact that these countries face a double issue related to climate change, namely a higher exposition to adverse climatic developments and a lower capacity to cope with these developments. Another interesting finding is that papers covering more recent investigation periods tend to find more often some effect. At the level of the regressions, analyses allowing a conditional effect of climate shocks tend to find more evidence of an effect—positive or negative—of climatic variables on the propensity to move. This confirms the importance of identifying the channels or the mechanisms through which climate shocks affect the movement of people. Related to that, a significant proportion (15.5% in our sample) has attempted to document indirect effects of climate change, i.e., evidence that climatic events can affect determinants of migration of people. We think that this is a valuable development of the literature as it obviously helps explaining the diversity of situations and contexts in which climate plays a role.
Third, the statistical approach of the analyses plays a role, too. Using dyadic data or panel data tends to favor evidence of an effect. Related to that, the use of the relevant estimation techniques for these data structures, i.e., panel estimation techniques accounting for unobserved heterogeneity and Poisson regressions in gravity models also tend to provide more evidence of an effect. Instrumental variable estimation, albeit not so widespread in the literature, also increases the probability of an effect. This might be due to the fact that IV estimations mitigate the attenuation bias related to the error of measures on climatic factors. Finally, we find only some modest publication bias. Papers published in relatively good journals tend to find more effects, including displacement ones exerted by climatic variables. In contrast, we do not find any bias related to the simple fact that the paper is published or not, or any bias related to the reputation of the authors. In general, the literature seems to have been quite honest in terms of the reported findings.
Acknowledgments
This work has benefitted from suggestions and comments from the participants. We are in particular grateful to R. Abramitzky, M. Berlemann, G. Brunello, M. Burke, B. Cockx, J. De Melo, F. Docquier, E. Lodigiani, K. Millock, I. Noy, I. Ruyssen, M. Steinhardt, I. Schumacher, and J. Voorheis for useful suggestions. Of course, the usual disclaimer applies.
Appendix A List of coded papers
Coded papers
Appendix B Complete codebook
