2.10.2012

Carefully interpreting climate-conflict results (JPR 2012) - Part 1


The Journal of Peace Research recently published an entire special issue dedicated to climate and conflict which contains papers from a 2010 conference.  Nils Petter Gleditsch is the editor of the issue, and in his introduction he suggests that he's not worried about future climate change (WRT conflict) based on findings presented in the issue.  We had a brief exchange about his interpretation on the Monkey Cage where he suggests that climate may influence conflict (recognizing our 2011 findings), but he is skeptical that anthropogenic climate change will matter.  I suppose this is fair enough, since nobody can say anything about the future with certainty. But his conclusions are based on empirical studies which are also necessarily historical, so it seems a little contradictory to think that they are somehow more informative about the future than other empirical historical studies.

This point aside, I've taken Nils' advice and started reading the papers in the issue.  My first reaction was a second wave of surprise at his conclusions, since most of the empirical papers in the issue seem to actually find a link between climatological parameters and conflict, although I haven't carefully kept score yet.  When I finish reading the issue, I will post my conclusions on this.

The first paper I got to that seemed to have a reasonable identification strategy was this one:

Climate change, rainfall, and social conflict in Africa 
Cullen S Hendrix and Idean Salehyan 
Abstract: Much of the debate over the security implications of climate change revolves around whether changing weather patterns will lead to future conflict. This article addresses whether deviations from normal rainfall patterns affect the propensity for individuals and groups to engage in disruptive activities such as demonstrations, riots, strikes, communal conflict, and anti-government violence. In contrast to much of the environmental security literature, it uses a much broader definition of conflict that includes, but is not limited to, organized rebellion. Using a new database of over 6,000 instances of social conflict over 20 years – the Social Conflict in Africa Database (SCAD) – it examines the effect of deviations from normal rainfall patterns on various types of conflict. The results indicate that rainfall variability has a significant effect on both large-scale and smaller-scale instances of political conflict. Rainfall correlates with civil war and insurgency, although wetter years are more likely to suffer from violent events. Extreme deviations in rainfall – particularly dry and wet years – are associated positively with all types of political conflict, though the relationship is strongest with respect to violent events, which are more responsive to abundant than scarce rainfall. By looking at a broader spectrum of social conflict, rather than limiting the analysis to civil war, we demonstrate a robust relationship between environmental shocks and unrest.

I wasn't sure if these results (HS) contradicted the results in our 2011 paper which showed a link between ENSO and conflict (HMC), since El Nino tends to cause the tropics to dry on average but also causes heavier-than-normal rainfall in some parts of Africa.  So I downloaded the replication data and merged it with the NINO3 time series from our paper.  The first thing I noticed (which isn't substantively important but is concerning from an editorial standpoint) is that the replication code and the published table didn't quite match.  The main conflict table from the paper is this:




But just running the code produced this:



The highlighted coeffs don't match, but they are just "controls" so we may not care much about them being mistabulated.  The underlined variables aren't reported in the first model (although I think they must have been included in model the authors ran because when I dropped them, the rest of the coeffs went haywire).  I mention this not to be a nag, but to explain why my estimates below don't match those in the paper exactly.

Having merged the HS data with our ENSO data (exactly from HMC), I replicated their main result (Column 1), then restricted it to the years which overlap with our paper (Column 2), then include our measure of ENSO {NINO3 averaged May-December} (Column 3), then drop 1997 since its a strong El Nino year that many people have complained about WRT our paper (Column 4):


Two things happen when I try to reconcile HS with HMC.

First, dropping 2005-2008 causes their strong positive correlation between conflict-onset and normalized-standardized-rainfall (GPCP_precip_mm_deviation_sd in the table) to drop by half in magnitude.  Because the standard errors don't change much, this also makes the coeff much less significant.  However, the coeffs are not actually significantly different from one another, so this change is hard to interpret conclusively.  Although it is worth noting that almost none of the other coeffs (except current normalized-standardized-rainfall) move very much.

Second, when I introduce our measure or ENSO (nino3_late in the table), it comes out with a strong positive correlation, consistent with earlier results of HMC.  The coeff on normalized-standardized-rainfall doesn't change much, suggesting these correlations are reasonable orthogonal.  When I drop 1997, the coeff on ENSO doubles in size, but so does the uncertainty.

Overall, I think this tells us that HS doesn't overturn the findings of HMC, suggesting that the effect they may be measuring is not related to ENSO. Although it also suggests that the positive correlation HS observe (between local rainfall and conflict onset) might be a little sensitive to the sample/specification (since notability, they also don't control for temperature which is definitely correlated with rainfall).

I was also interested in the SCAD data HS use, since my coauthors and I have played with the idea of using it to look at smaller scale conflicts.  Unfortunately, the data doesn't go back in time very far, so it's pretty tough to look at the effects of ENSO since ENSO varies annually.  Nonetheless, I did the quick and dirty thing of collapsing all SCAD events in Africa to yearly observations (following the ACR approach of HMC), linearly de-trending the series, and plotting the scatter against ENSO:


Since the panel is so short, 1997 is a monstrous outlier. But if we ignore it (as the critics who don't like the results of HMC keep arguing for) we see a pretty large positive correlation (although, unsurprisingly it's not quite significant with N = 13).  I think this is pretty interesting, since the SCAD database contains less-organized, less-deadly or less-political conflict than the PRIO data analyzed in HMC and it still exhibits a positive correlation with ENSO, similar to organized, deadly, political violence presented in HMC.

[Code to run the models in the table are below the fold.]

// This runs with replication code from Hendrix & Salehyan JPR 2012 (http://jpr.sagepub.com/content/49/1/35/suppl/DC1) and replication code from Hsiang, Meng & Cane Nature 2011 (http://www.nature.com/nature/journal/v476/n7361/full/nature10311.html#/supplementary-information)

// To run, put the file enso_conflict_TS.dta from HMC into the replication folder of HS. Then run this as a .do file in Stata, where the replication folder for HS is the working directory.





//load in "Time series data for replication for Hsiang, Meng, and Cane. Nature, 2011"
clear
use enso_conflict_TS.dta 
drop if region_dummy == 0
keep year nino3_late
save ENSO, replace


use Hendrix_Salehyan_JPR491_Replication.dta
merge m:1 year using ENSO.dta


//cut and pasted regression for HS2012
logit conflict_onset incidence_l GPCP_precip_mm_deviation_sd GPCP_precip_mm_deviation_sd_sq GPCP_precip_mm_deviation_sd_l GPCP_precip_mm_deviation_sd_l_sq polity2_l polity2_sq_l log_pop_pwt_l log_pop_pwt_fd_100_l log_rgdpch_pwt_l grgdpch_pwt_l peaceyears*, cluster(ccode)


outreg2 using HG_JPR_2012_ENSO_table2, excel br dec(3) ct("logit", " ", "1991-2008", " ") replace


//restricting to sample that overlaps with HMC2011
logit conflict_onset incidence_l GPCP_precip_mm_deviation_sd GPCP_precip_mm_deviation_sd_sq GPCP_precip_mm_deviation_sd_l GPCP_precip_mm_deviation_sd_l_sq polity2_l polity2_sq_l log_pop_pwt_l log_pop_pwt_fd_100_l log_rgdpch_pwt_l grgdpch_pwt_l peaceyears* if nino3_late ~= ., cluster(ccode)


outreg2 using HG_JPR_2012_ENSO_table2, excel br dec(3) ct("logit", " ", "1991-2004", " ") 


//modeling ENSO and normalized rainfall together
logit conflict_onset nino3_late incidence_l GPCP_precip_mm_deviation_sd GPCP_precip_mm_deviation_sd_sq GPCP_precip_mm_deviation_sd_l GPCP_precip_mm_deviation_sd_l_sq polity2_l polity2_sq_l log_pop_pwt_l log_pop_pwt_fd_100_l log_rgdpch_pwt_l grgdpch_pwt_l peaceyears*, cluster(ccode)


outreg2 using HG_JPR_2012_ENSO_table2, excel br dec(3) ctitle("logit", "ENSO", "1991-2004", " ")


//modeling ENSO and normalized rainfall together (dropping 1997)
logit conflict_onset nino3_late incidence_l GPCP_precip_mm_deviation_sd GPCP_precip_mm_deviation_sd_sq GPCP_precip_mm_deviation_sd_l GPCP_precip_mm_deviation_sd_l_sq polity2_l polity2_sq_l log_pop_pwt_l log_pop_pwt_fd_100_l log_rgdpch_pwt_l grgdpch_pwt_l peaceyears* if year ~= 1997, cluster(ccode)


outreg2 using HG_JPR_2012_ENSO_table2, excel br dec(3) ctitle("logit", "ENSO", "1991-2004", "drop 97")

3 comments:

  1. Dear Sol-

    The blogosphere is not to be used for careful empirical analysis or actual replication. It is to be used for fawning over the titles of articles you haven't read in support an as-yet-unsubstantiated worldview. Please see here:

    http://chrisblattman.com/2012/02/09/climate-change-and-conflict-thank-goodness-the-skeptics-have-not-given-up/

    ReplyDelete
  2. That's a nice little addition to our main findings, thanks. I think perhaps the population variable we included in the replication data was not scaled properly--otherwise the results are the same. We'll look into it.
    --Idean

    ReplyDelete
  3. Yeah, and we did not report the cubic polynomial of time in the table. That's noted in the text.

    ReplyDelete