Data Collection

To generate the most complete universe of studies, we combine three approaches: (1) we run keyword searches on Google Scholar, (2) include studies citing prominent early works, and (3) publicly solicit recent and unpublished work. This yields 3,346 studies (see Table 2).

Table 2.

Stages of Filtering and Number of Studies Selected

Note: * Second filter also requires that the study uses a fixed effects panel model; $ \dagger $ two working papers were abandoned by the authors.

Our topical filter requires that studies include a quantitative analysis where armed conflict is the dependent variable and commodity prices are an independent variable. Among 376 relevant studies, our research design filter retains 46 natural experiments that leverage plausibly exogenous price variation. These studies represent 201 countries and 10,926 unique country-years.Footnote ⁹ Included countries are on average 40% as wealthy, somewhat more unequal, two-thirds more prone to conflict, and somewhat less democratic than the world at large; they more closely resemble those countries that experienced an intrastate conflict in the postwar period (see Appendix B.3). Identification relies on the inclusion of unit and time fixed effects to absorb time-invariant confounds and global shocks.Footnote ¹⁰ This second filter increases the internal validity of included studies. We retain one estimate per paper for every commodity and conflict type (onset, incidence, and intensity) following prespecified rules (see Appendix A.5).

Together, these two filters ensure the conceptual comparability of study estimates. We take two additional steps to ensure that the estimates are numerically comparable. First, we standardize all estimates to address potential differences in the scales of the conflict outcomes (e.g., binary or count) and price variables (e.g., in different currencies). Our standardized effects are expressed in terms of standard deviation changes in the prices and conflict variables:

(1) $$ {\hat{\beta}}_{std}=\hat{\beta}\times \frac{sd\hskip0.18em \left(\mathrm{Price}\right)}{sd\hskip0.18em \left(\mathrm{Conflict}\right)} $$

Following Mummolo and Peterson (Reference Mummolo and Peterson2018), we residualize the variables using the unit and time fixed effects before computing the standard deviations. More commonly reported pooled standard deviations often overstate the variation used to estimate $ \hat{\beta} $ in a two-way fixed effects model. We compute these statistics (or receive them from authors) for 37 studies (see Appendix Table A.3).

Second, we ensure that all studies use a common functional form:

(2) $$ {\mathrm{Conflict}}_{it}={\delta}_i+{\gamma}_t+\beta \hskip0.5em {\mathrm{Prices}}_{it}+\kappa {X}_{it}+{\varepsilon}_{it}, $$

where $ i $ indexes the authors’ cross-sectional unit (which we use to cluster the standard errors) and $ t $ indexes their temporal unit; $ {X}_{it} $ includes the other time-varying controls included in the authors’ original specification. This overcomes noncomparability that arises from the use of models with nonlinear link functions (e.g., logistic regression) or the choice of fixed effects (e.g., using year fixed effects where the temporal unit in the panel is month). We acquire replication data for 32 studies and estimate this model; we confirm the remaining five estimate a similar linear model.Footnote ¹¹

These standardization steps exclude nine papers for which we lack the necessary statistics (see Appendix Table A.6). We can, however, compute an alternate measure of effect size, the partial r ( $ {\rho}_p $ ), which requires only the t-statistic ( $ t $ ) and degrees of freedom ( $ df $ ): $ {\rho}_p=t/\sqrt{t^2+ df} $ (see Appendix B.4). Reassuringly, the distribution of $ {\rho}_p $ does not change with the inclusion of these nine studies.

Meta-analysis

We first estimate the fixed effects meta-analysis model (Rosenthal and Rubin Reference Rosenthal and Rubin1982), which is a precision-weighted average of the standardized estimates ( $ \hat{\beta} $ _std from Equation 1, with weights equal to the inverse of the standardized variance).Footnote ¹² Under minimal assumptions, this model consistently estimates the average effect for the studies in our sample (Rice, Higgins, and Lumley Reference Rice, Higgins and Lumley2018). We also compute the random effects meta-analysis model (DerSimonian and Laird Reference DerSimonian and Laird1986).Footnote ¹³ This model assumes that the true effects differ across studies, but that these are drawn from a common (normal) distribution.

The fixed and random effects models both recover quantities of interest: the former provides an efficient estimator for the average effect within our sample of studies, while the latter provides both an estimated mean and variance of true effects, which permits generalization to out-of-sample studies. For numerical reasons, the standard errors from the random effects model will always be weakly larger. We estimate both models for each type of commodity. We also present a pooled effect, which averages our estimates across commodity types, giving equal weight to each commodity type.Footnote ¹⁴

Our estimates pool across conflict types (incidence, onset, and intensity).Footnote ¹⁵ In Appendix C, we show that coefficient estimates are stable when re-estimating our models while leaving out each conflict type and, additionally, that conflict type is nonsignificant when entered as a moderator.Footnote ¹⁶

We take steps to mitigate publication bias and assess whether it skews our estimates. We include working papers. We also perform several diagnostic tests: p-curves, funnel plots, and meta-regression analysis (see Appendix H). These find no evidence of publication bias. Prominent papers in this literature have published null results (e.g., Bazzi and Blattman Reference Bazzi and Blattman2014), ameliorating concern that only positive findings escape the file drawer. We also find that our results are not driven by outliers in effect size or precision (Appendix G).