Appendix A
I begin this appendix by presenting further information on the explanatory and control variables included in the main analysis. The information on temperature comes from the NOAA’s NCEP/NCAR Reanalysis Monthly Means Dataset 1948–2011 (in degrees Celsius) [
42]. These data provide surface or near surface air temperatures (at a 0.995 sigma level) with spatial coverage of a 2.5 × 2.5-degree longitude native resolution (144 × 72). Specifically, as described in detail in Landis [
24] (p. 608), the temperature shock measure “uses the monthly deviation from a country’s long-term monthly mean, indicated by (X
itz − X
it-bar)/a
it where X
itz is the mean temperature of country
i in month
t in year
z, X
it-bar is the panel mean of country
i’s long-term monthly (
t-bar) mean temperature for the period 1948–2011, and a
it is the standard deviation of that panel.” According to Landis [
24], this approach is adopted from Hendrix and Salehyan’s [
25] (pp. 40–41) measure of rainfall deviation, as the latter study argues that deviations from the panel mean are an optimal operationalization of the “eco-shock” mechanism. Additionally, I employ standardized precipitation deviations (see also [
25]) using monthly precipitation data (mm/month) from the Global Precipitation Climatology Project Version 2.2. These data have a spatial coverage of 2.5 × 2.5-degrees with a longitude resolution (144 × 72) for 1979–2011.
Regarding the control variables, first, there is some evidence linking climate to conflict and, thus, migration [
43,
44]. Hence, I include a civil war onset indicator (based on at least 1000 battle deaths). This information is taken from the Armed Conflict Database [
31]. Second, a variable for a state’s (i.e., the state sending/receiving migrants) regime captures whether people’s choice of leaving their country is also affected by domestic politics in the source location. Additionally, the regime the destination controls for people’s choices based on political-related factors. This factor also captures influences like state repression or human rights violations. I include the polity 2 item taken from the Polity IV data set, which covers basically all countries in my sample over the entire period [
32]. The polity 2 measure ranges between −10 and +10, with higher values standing for more democratic countries. High unemployment might also push people to leave their countries looking for better life conditions. At the same time, low levels of unemployment should be attracting more migrants. I thus include an indicator for unemployment from the World Bank Development Indicators. The measure for unemployment refers to the share of the labor force that is without work but available for and seeking employment. For example, in 2000, Greece was affected by 11.1% of unemployment. The original variable of unemployment suffers from missing values. To address this issue, I linearly interpolate these missings. This interpolation explains why some states in the sample then have an unemployment rate of 0. To address any concerns stemming from this treatment, I also present models that omit the unemployment variable. I also control for population size and GDP per capita using data from Gleditsch [
33]. These measures are log-transformed to reduce their distributions’ skewness, because some countries are much wealthier and larger than others. Finally, in light of the dyadic nature of this analysis, I also add a measure on the cultural distance between states. The rationale behind this item is to capture a truly dyadic influence on emigration, i.e., whether cultural similarities impact on emigrants’ choice of the destination country. To this end, I adopt Kandogan’s [
29] revised variable of Kogut and Singh’s [
30] standardized measure of cultural differences.
In
Table A1 of this appendix, I examine the impact of climate variability on transnational migration, as discussed in the main analysis (
Table 3), while excluding the variable on precipitation shocks. When including this item as done in
Table 2 of the main text, I only capture the period between 1979 and 2009 because of the limited data availability for the precipitation variable. Omitting the precipitation item, and increasing the number of observations as a consequence, does not change the main finding.
Table A1.
The impact of climate variability on emigration—Omitting precipitation.
Table A1.
The impact of climate variability on emigration—Omitting precipitation.
| Model 1 | Model 2 |
---|
Temperature shock | 0.14 *** | 0.10 *** |
(lag)—Origin | (0.02) | (0.02) |
Temperature shock | −0.23 *** | −0.20 *** |
(lag)—Destination | (0.02) | (0.02) |
Precipitation (lag)—Origin | | |
Precipitation (lag)—Destination | | |
Constant | 3.42 *** | 16.29 *** |
| (0.00) | (1.20) |
Obs. | 1,218,626 | 446,279 |
Lagged dependent variable | Yes | Yes |
Country fixed effects | Yes | Yes |
Year fixed effects | Yes | Yes |
Clustered standard errors (dyad) | Yes | Yes |
R2 | 0.96 | 0.95 |
RMSE | 3.90 | 3.73 |
Table A2 shows the short-term effects of climate variability on transnational migration (based on
Figure 3 in the main text). For example, the results indicate that the coefficient estimate of Temperature shock (lag)-origin is 0.12 (based on specifications of Model 2 in
Table 3 of the main analysis). While this is a rather small impact, it does not come unexpectedly; the fact remains that the variable is highly statistically significant. In addition, due to the temporally lagged dependent variable included as a predictor, the coefficient estimates of all explanatory variables only reflect the short-term effect, i.e., the impact in a current year (i.e., the short-term effects). Hence, I have calculated the long-term effects of climate variability on transnational migration.
Table A3 shows that the coefficient estimate of Temperature shock (lag)-origin is 2.61 (based on specifications of Model 2 in
Table 3 of the main analysis).
Table A2.
Short-term effects of climate variability.
Table A2.
Short-term effects of climate variability.
| Marginal Effect Estimate | Lower Bound | Upper Bound |
---|
Temperature shock (lag)—Origin | 0.11 | 0.04 | 0.17 |
Temperature shock (lag)—Destination | −0.19 | −0.25 | −0.13 |
Precipitation shock (lag)—Origin | −0.17 | −0.23 | −0.11 |
Precipitation shock (lag)—Destination | 0.01 | −0.05 | 0.07 |
Table A3.
Asymptotic long-term effects of climate variability.
Table A3.
Asymptotic long-term effects of climate variability.
| Marginal Effect Estimate | Lower Bound | Upper Bound |
---|
Temperature shock (lag)—Origin | 2.45 | 1.01 | 3.89 |
Temperature shock (lag)—Destination | −4.31 | −5.67 | −2.96 |
Precipitation shock (lag)—Origin | −3.83 | −5.22 | −2.59 |
Precipitation shock (lag)—Destination | 0.20 | −1.27 | 1.67 |
I linearly interpolated missing values in the outcome variable, which only reports values per decade. While this addresses the issue of missing values, it may increase the risk of inducing another problem: cointegration, particularly since temperature rises on average more or less linearly as well (but note: temperature shocks do not). Cointegration may lead to spurious findings. As described by Toll [
45]:
“a regression analysis seeks to explain as much as possible of the observed variation in the dependent variable by the variations in the independent variables. The variance of a trending variable is dominated by its trend. If an independent variable has a trend as well, then its variance too is dominated by the trend. More importantly, the trend in any independent variable can explain a large share of the trend in the dependent variable. This implies that, in a regression analysis, the confidence in the parameter estimates is overstated. That is, a regression analysis will find a statistically significant relationship even when there is none.”
For examining whether cointegration might be an issue, I re-run the analysis only with the actually observed data, i.e., I drop the linearly interpolated values (
Table A4). The results remain qualitatively the same as in the main analysis (Model 2 in
Table 3): I still obtain evidence for a significantly positive relationship between temperature shocks and transnational migration. An increase in temperature shocks at home increases the amount of emigrants whilst and a decrease in temperature shocks in the destination country attracts more migrants. Additionally, precipitation shocks in the destination country decrease the number of emigrants whereas this was insignificant in the main analysis.
Table A4.
The impact of climate variability on emigration—without linearly interpolated data.
Table A4.
The impact of climate variability on emigration—without linearly interpolated data.
| Model 1 |
---|
Temperature shock | 0.57 ** |
(lag)—Origin | (0.25) |
Temperature shock | −5.01 *** |
(lag)—Destination | (0.27) |
Precipitation (lag)—Origin | −0.37 |
| (0.27) |
Precipitation (lag)—Destination | −1.03 *** |
| (0.24) |
Constant | 64.32 *** |
| (14.66) |
Obs. | 36,369 |
Lagged dependent variable | No |
Country fixed effects | Yes |
Year fixed effects | Yes |
Clustered standard errors (dyad) | Yes |
R2 | 0.52 |
RMSE | 12.89 |
I also examined the predictive power of the main explanatory variables of interest via in-sample predictions techniques. That is, I analyze how accurate “conditional statements about a phenomenon for which the researcher actually has data, i.e., the outcome variable has been observed” [
46] (p. 311) are. I rely on one measure for assessing the in-sample prediction power: Theil’s U According to Böhmelt and Bove [
47] (p. 3), “Theil’s U is the square root of the ratio between the sum of squared prediction errors of a baseline model and the sum of squared prediction errors of a naïve model; that is, a no-change prediction. If Theil’s U is larger than 1, the model actually performs worse than the naïve model; values for Theil’s U smaller than 1 indicate that the “theoretically informed model” performs better than the naïve specification.”
For my baseline model (Model 1 in
Table 3 in the main analysis), Theil’s U is at 0.83798313.
Table A5 below gives an overview of the model’s in-sample prediction power and the individual contribution each of the variables employed in Model 2 makes. The contributions of each variable is measured by calculating the difference between the value of the baseline model’s Theil’s U values on one hand and, on the other hand, the corresponding goodness-of-fit measure’s value calculated for a model that discards that particular item. For example, excluding Temperature shock-origin (lag) from the baseline model leads to an increase in Theil’s U from 0.83680681 to 0.83682049. Therefore, Temperature shock at the origin country does contribute to the model’s overall prediction power by 0.001368 units according to Theil’s U. Finally, note that none of these predictors included in Model 2 diminishes the predictive power. In other words, Theil’s U n decrease when leaving out an item from the model specification. Ultimately, the specifications used in the main analysis perform well in predicting transnational migration and clarifying the robustness of this empirical analysis.
Table A5.
In-sample prediction power.
Table A5.
In-sample prediction power.
Excluded Variables | Mean U | ΔU |
---|
None (baseline model) | 0.83680681 | – |
Temperature shock (lag)—Origin | 0.83682049 | 0.001368 |
Temperature shock (lag)—Destination | 0.83684651 | 0.00397 |
Precipitation shock (lag)—Origin | 0.83684564 | 0.003883 |
Precipitation shock (lag)—Destination | 0.8368069 | 0.000009 |
To further examine whether unemployment rates play a significant role for the choice of the destination country, I interact temperature shock at the origin country and unemployment rate at the destination country (
Figure A1 in this appendix). The information on unemployment comes from the World Bank indicators and it show the percentage of total labour force that is unemployed. The results show that emigration increases at high levels of temperature shock at the origin country and low levels of unemployment in the destination country. This means that emigrants do consider the economic characteristics of the country they are migrating, also while taking into account the climatic influences I focus on.
Figure A1.
Average marginal effects of temperature shock in origin country and unemployment in destination country. Vertical bars are 90 percent confidence intervals; horizontal red line represents a marginal effect of 0; estimates are based on Model 2 (
Table 3 in the main analysis). The values of the unemployment variable are linearly interpolated due to missing values.
Figure A1.
Average marginal effects of temperature shock in origin country and unemployment in destination country. Vertical bars are 90 percent confidence intervals; horizontal red line represents a marginal effect of 0; estimates are based on Model 2 (
Table 3 in the main analysis). The values of the unemployment variable are linearly interpolated due to missing values.