{"id":"a9fb4619-9a06-4982-81ca-e2a2065ee202","arxiv_id":"2505.03405","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Risk-averse individuals are more likely to flee conflict areas and risk-tolerant individuals to stay, consistent with a quantile maximization model, in Nigerian panel data.","lead":"This paper applies quantile maximization theory to explain why some people flee conflict zones while others stay. Using Nigerian household and conflict data, it finds that risk-averse people are more likely to flee and risk-tolerant people are more likely to stay, contrary to findings for economic migrants.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The imputed leaver lottery (Eqs. 8-9) is the linchpin of the QM prediction; the panel's 329 actual migrants provide observed leaver outcomes that the paper never compares against the imputation.","rationale":"The reader's CONDITIONAL verdict is appropriate. The most load-bearing step in the paper is the construction of the counterfactual 'leaver' distribution in Eqs. (8)-(9). The QM model does not by itself predict who flees; the prediction is read off the CDFs in Figure 7. If those CDFs are wrong, the claimed match between the model and the observed risk preferences in Table 7 is not evidence for the model—it is simply a correlation between migration and a risk question measured after the move. The paper flags the constant-coefficient assumption as strong, but it does not exploit the panel's actual leavers to validate the imputation. The proposed test would do exactly that. A secondary concern, not chosen here, is that risk aversion is measured post-migration and the DiD in Table 8 lacks a pre-treatment baseline, so the regression alone cannot distinguish self-selection from reverse causality; this further supports the CONDITIONAL verdict but is downstream of the imputation issue. If the imputation test fails, the central claim should be downgraded; if it passes, the QM mechanism gains credibility.","tokens_in":19912,"tokens_out":10818,"duration_ms":106456,"concrete_test":"Using the GHS panel, extract the post-migration per-capita expenditure for the 329 households that changed LGA, measured in the first wave after the move. Compare the CDF of these observed leaver outcomes with the imputed leaver distribution used in Figure 7, using a Kolmogorov-Smirnov test and a test for first-order stochastic dominance. If the observed leaver CDF differs significantly from the imputed one, or crosses at a different location, the imputation is unsupported and the CDF-crossing-based prediction is not credible. As a secondary check, re-estimate Equation (9) using only the characteristics of actual leavers rather than all conflict households, and see whether the crossing persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central prediction that risk-averse individuals flee and risk-tolerant stay depends on the CDF comparison in Figure 7, in which the 'leaving' lottery is not observed but imputed. Equation (8) estimates a welfare equation on non-conflict households only; Equation (9) applies those coefficients to conflict households and Duan-smears to predict expenditure if they left. The paper itself concedes in Section 3 that 'the Betas of the prediction equation are constant if one migrates or stays' is 'a strong assumption.' If the true coefficients for leavers differ—because of unobserved kin networks, labor-market discrimination at destination, or selection into who actually leaves—the imputed leaver distribution is biased. The CDF crossing, and hence the maxmin-leaves/maxmax-stays prediction, could be an artifact of the imputation rather than a property of the actual lotteries. Notably, the GHS panel contains 329 households that actually changed LGA and whose post-migration expenditure is observed in later waves; these data are not used to validate the imputed leaver distribution, despite being directly relevant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantile-maximization (QM) model of the decision to flee conflict, in which individuals compare two lotteries—staying and leaving—and are distinguished by which quantile of the outcome distributions they focus on: maxmin (risk-averse) individuals choose the lottery with the best lower-tail outcome, while maxmax (risk-tolerant) individuals choose the lottery with the best upper-tail outcome. Using Nigerian General Household Survey panel data (2010-2016) matched to ACLED conflict fatalities by Local Government Area, the authors estimate a welfare equation on non-conflict households (Eq. 8), apply the coefficients to conflict households to impute their counterfactual expenditure if they left (Eq. 9), and compare the resulting 'leaver' distribution with the observed 'stayer' distribution. They report that the CDFs cross once: the leaver lottery dominates at the lower tail and the stayer lottery dominates at the upper tail, implying that risk-averse individuals should flee and risk-tolerant individuals should stay. A probit regression (Table 7) using a single binary investment question from the 2016 survey finds that risk tolerance is negatively associated with staying, and a difference-in-differences specification (Table 8) is used to argue that conflict does not confound this association.","tokens_in":20101,"tokens_out":5998,"duration_ms":55224,"significance":"If the results hold, the paper makes a novel contribution: it is, to my knowledge, the first empirical application of quantile maximization to forced displacement, and it proposes a sharp behavioral contrast between risk-averse forced migrants and risk-tolerant economic migrants, with potentially important policy implications. The paper is also commendable for its transparency: it explicitly acknowledges the strong coefficient-stability assumption in the imputation step, provides attrition diagnostics (Tables 3-5), and separates the CDF-prediction step from the observed risk-preference regression, so the central test is not simply a restatement of the model's assumptions. However, two load-bearing gaps prevent the evidence from being fully convincing: the imputed leaver lottery is never validated against the observed post-migration outcomes of the 329 actual migrant households in the panel, and risk preferences are measured once, after migration decisions were made, with a single binary investment question.","major_comments":[{"comment":"The imputed leaver lottery is never validated against observed outcomes of the 329 households that actually changed LGA. The GHS panel contains their post-migration expenditure in later waves, but Figure 7 relies entirely on the counterfactual predicted by applying non-conflict-household coefficients to conflict households. The paper itself calls the coefficient-stability assumption 'a strong assumption'; if the true returns to characteristics differ for leavers (e.g., through kin networks, labor-market discrimination at destination, or selection into who actually leaves), the single crossing in Figure 7 and the resulting maxmin-leaves/maxmax-stays prediction could be an artifact of the imputation. Please add a validation exercise comparing predicted vs. observed expenditure for actual migrants, and report robustness of the CDF crossing to alternative imputation models (e.g., including destination effects or relaxing the constant-coefficient assumption).","section":"Section 3, Eqs. (8)-(9); Figure 7"},{"comment":"Risk preferences are measured in 2016 with a single binary investment question, after the migration decisions (2010-2016) have already been made. The DiD in Table 8 does not resolve the timing problem because risk aversion is observed only at the endline; the interaction term is a cross-sectional contrast, not a pre/post comparison. The negative association between staying and risk tolerance in Table 7 could therefore reflect the effect of migration or conflict experience on measured risk preferences rather than selection on stable preferences. The paper should either use a pre-migration measure of risk preferences, or provide a convincing exogeneity/bounding argument (for example, a placebo test using 2010 characteristics to predict 2016 risk attitudes, or a discussion of how measurement error in the binary question would affect the probit estimates).","section":"Section 4, Tables 7 and 8"},{"comment":"The claim that the CDFs 'cross in one point' and that the differences are statistically significant needs formal support. The text reports no test statistic for the crossing or for stochastic dominance; a Kolmogorov-Smirnov-type test or confidence bands for the difference in CDFs would strengthen the prediction step. In addition, please clarify whether the observed staying distribution uses all conflict-area households or only those that never migrated during the panel, since the two definitions could shift the location of the crossing and hence the predicted direction of selection.","section":"Section 5, Figure 7"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and garbled references (e.g., 'Khanemann', 'C ¸ ha˘ glar¨Ozden', 'Tizeorie', 'acaderniae') and needs careful proofreading before publication.","section":"Throughout"},{"comment":"Footnote 1 contains a local file path ('file:///C:/Users/...') that should be replaced with a proper citation to the source study.","section":"Footnote 1"},{"comment":"The description of dwelling-quality variables repeats '1=Low quality, 1=Medium quality'; the codes should be 1, 2, and 3.","section":"Table 2"},{"comment":"The Mironova and Whitt work is cited as 'Mironova and Whitt [2017]' in the text and as 'Whitt and Mironova, 2017' in the conclusion, while the bibliography gives 'Under Review, 2017'; please standardize the citation and update the status.","section":"References"},{"comment":"The statement that 'only relatively more risk-tolerant individuals would leave their homes' refers to the illustrative example but is easily read as a general model prediction; since the empirical finding is the opposite, please clarify that the sign of the QM prediction is data-dependent and depends on which lottery dominates at the relevant quantiles.","section":"Section 2.2, final paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid and interesting core, and the authors are transparent about their main assumption. The key issue is that the central prediction rests on an imputed counterfactual distribution that is never checked against the observed outcomes of actual migrants in the same panel; this is fixable within the scope of the manuscript and should be the primary focus of the revision. The timing of the risk-preference measure is a second concern that also needs to be addressed explicitly. I recommend major revision rather than rejection because the questions are concrete and the data to answer them appear to be available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The first thing to know: this paper gives the first application of quantile maximization (Rostek) to conflict displacement and reports that risk-averse individuals are more likely to leave conflict zones while risk-tolerant stay. That is a genuine reversal of the standard economic-migrant selection result, and if it holds up, it matters for policy. The second thing: the empirical support is real but conditional, weaker than the authors’ tone suggests.\n\nWhat's new and what works: the QM framing is clean. Instead of mean-variance EU, the authors compare CDFs of staying vs leaving and let different risk types focus on different parts of the distributions. The Nigeria panel (GHS 2010–2016) is well matched to the question, and the paper is careful about attrition. The direct risk question in 2016 and the dominance analysis give a simple, transparent test.\n\nSoft spots, in proportion: the linchpin is the imputed leaving lottery. Eq. 8 estimates a welfare equation on non-conflict households and Eq. 9 applies those betas to conflict households to predict expenditure if they left. The authors acknowledge the constant-beta assumption is strong. If leavers have different coefficients—kin networks, labor-market discrimination—the CDF crossing in Fig. 7 could be an artifact. There are 329 households that actually changed LGA with observed post-migration expenditure; the paper never compares the imputed distribution to those actual leavers. That is the most serious gap, and it is fixable. Second, risk preferences are measured with a single binary investment question asked in 2016, after the migration decision. The authors treat risk preference as time-invariant but don't establish that; the DiD only tests whether conflict exposure changes risk preferences, not whether migration changes them. So reverse causality remains on the table. Third, the DiD has no pre-treatment baseline, which makes it a weak check on the confounding story.\n\nNone of these are fatal to the idea. The association in Table 7 is significant and the direction is consistent with the Mironova-Whitt experiment. But the verification step is not yet strong enough to call the reversal a robust fact.\n\nWho this is for: migration economists, conflict researchers, and anyone designing assistance for IDPs. Worth a serious referee: the question is important, the method is novel in this domain, and the flaws are addressable. I'd send it out with a request for the imputation validation and more sensitivity around the risk measure. Recommendation: send to peer review, with the expectation of major revision.","headline":"First quantile-maximization account of forced migration finds risk-averse flee and risk-tolerant stay; the reversal is new, but the imputed leaving lottery and post-migration risk measure keep the finding conditional.","tokens_in":20630,"tokens_out":3260,"would_cite":true,"duration_ms":32427,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"People who flee conflict are more risk-averse than those who stay, the paper argues, reversing the usual migrant profile.","keywords":["conflict","forced displacement","quantile maximization","risk aversion","migration self-selection","expected utility","stochastic dominance","Nigeria"],"falsifier":"Take the 329 Nigerian households who actually changed local government areas, use the living standards they actually had after moving as the leaving lottery, compare with stayers' observed standards, and check whether the CDFs still cross exactly once with leaving dominant at the low end. If the crossing disappears or reverses, the predicted direction of risk selection is an artifact of the imputation.","tokens_in":19665,"feed_emoji":"🏃","tokens_out":8858,"duration_ms":82060,"temperature":0.7,"pith_summary":"This paper aims to establish that the people who flee conflict are not the same kind of people who make economic migrations: under quantile maximization, the risk-averse choose to leave and the risk-tolerant choose to stay. It models staying and leaving as two lotteries over household living standards and argues that each person judges the lotteries by a single part of their outcome distribution rather than by the mean. Using Nigerian household panel data matched with local conflict deaths from 2010 to 2016, the paper finds that the cumulative distributions of staying and leaving cross once, with leaving better at the low end and staying better at the high end, exactly the pattern that predicts risk-averse flight. A direct survey measure of risk preferences lines up with that prediction, and a difference-in-differences test finds no evidence that conflict itself created the risk-preference gap. If this is right, forced migrants and economic migrants are selected in opposite directions, so policies that bundle them together are likely to miss.","feed_headline":"Risk-averse people flee; risk-tolerant people stay","feed_subtitle":"In Nigeria, forced migrants are risk-averse and stayers risk-tolerant, the opposite of economic migrants.","key_machinery":"The engine is the $\\tau$-quantile maximization rule. For a lottery with cumulative distribution $F_x$ and $\\tau \\in [0,1]$, the $\\tau$-quantile is $Q_\\tau(F_x)=\\inf\\{x_i : F_x \\ge \\tau\\}$, and a decision maker prefers the lottery with the larger $Q_\\tau$; $\\tau=0$ gives maxmin behavior and $\\tau=1$ gives maxmax behavior, with intermediate $\\tau$ covering intermediate risk aversion. The empirical machinery builds the two lotteries as distributions of household expenditure per capita: the staying distribution is observed, and the leaving distribution is predicted by fitting a welfare equation on non-conflict households and applying those coefficients to conflict households, with Duan's smearing correction for the log-to-level retransformation. The statistical object that carries the argument is the once-crossing cumulative distribution functions: below the crossing the leaving lottery dominates at the low end, which the model maps to risk-averse leavers, and above it the staying lottery dominates at the high end, which the model maps to risk-tolerant stayers.","core_discovery":"The central claim is that in a conflict environment people are quantile maximizers: a risk-averse (maxmin) person focuses on the lowest possible outcomes of staying and leaving and picks the lottery with the better worst outcome, which in the Nigerian data is leaving, while a risk-tolerant (maxmax) person focuses on the highest outcomes and picks staying. The paper constructs the leaving lottery by imputing the expenditure that conflict-affected households would have had in non-conflict areas, compares its cumulative distribution with the observed distribution of expenditure among stayers, and finds a single crossing: leaving dominates for low quantiles and staying dominates for high quantiles. It then tests the model's selection prediction with a direct risk-preference question and finds that migrants are significantly more risk-averse than stayers, with the association stronger for migration from areas with persistent conflict. A difference-in-differences test shows no significant effect of conflict exposure on risk aversion, which the paper reads as ruling out the main confounder. The paper's conclusion is that forced migrants self-select for risk aversion, in sharp contrast with economic migrants, and that the two groups therefore need separate policies.","pith_inferences":["The paper stops at conflict, but the same two-lottery logic transfers to displacement by disasters or economic collapse, where the 'leave' lottery is also chosen for downside protection; a testable extension would apply the CDF-crossing method to climate-displacement panel data.","Because the crossing point of the two CDFs is the threshold that separates predicted leavers from stayers, conflict intensity that shifts the crossing rightward should raise flight even with unchanged risk preferences; comparing crossing points across Nigeria's panel waves would test this.","The paper leaves implicit that its imputation assumes everyone in the conflict zone faces the same leaving lottery; adding person-specific information such as kin networks and information access would change the quantiles individuals compare and could either sharpen or weaken the risk-selection result.","A sharper falsification is available inside the data: the 329 observed migrant households have post-move expenditures recorded when they were followed, so the leaving lottery need not be imputed at all."],"forward_implications":["If the selection result is correct, forced migrants are systematically more risk-averse than the populations they leave behind, which means return decisions hinge on restoring basic safety and minimum living standards rather than on improving opportunity.","The same model predicts that the risk-preference gap between leavers and stayers should widen as conflict becomes more persistent, which the Nigerian data show and which can be checked in other conflict panels.","Host-country policies that treat economic and forced migrants as one group will fit neither: forced migrants are less likely to take entrepreneurial risks and more likely to need subsistence and security support near their places of origin.","Risk preferences, measured cheaply by survey questions, become a usable early indicator of who is likely to flee a conflict zone."],"supporting_citations":[{"why":"Supplies the theoretical characterization of quantile maximization that the model applies to conflict mobility.","marker":"Rostek [2010]"},{"why":"Establishes quantiles as an essentially unique ordinal decision criterion consistent with weak first-order stochastic dominance, justifying the choice rule.","marker":"Chambers [2009]"},{"why":"Provides the behavioral critique of expected utility that motivates replacing mean evaluation with quantile evaluation.","marker":"Kahneman and Tversky [1979]"},{"why":"Supplies the smearing retransformation used to convert predicted log expenditure into the leaving lottery's outcome levels.","marker":"Duan (1983)"},{"why":"Baseline evidence that economic migrants are risk-tolerant, the contrast that makes the paper's risk-averse leavers a puzzle.","marker":"Jaeger et al. [2010]"},{"why":"Theoretical model showing risk-taking can trigger migration, the economic-migration benchmark the paper reverses.","marker":"Katz and Stark [1986]"},{"why":"Evidence that violence can increase risk tolerance; the confounding hypothesis the DiD test is designed to rule out.","marker":"Voors et al. [2012]"},{"why":"Evidence on violence and the certainty premium; second confounding-factor reference.","marker":"Callen et al. [2014]"},{"why":"The only prior empirical study of risk preferences and forced migration, whose findings are consistent with the paper's results.","marker":"Mironova and Whitt [2017]"}],"fun_headline_variants":["Risk-averse flee, risk-tolerant stay in conflict zones","Conflict migrants are risk-averse, stayers risk-tolerant","Quantile maximization explains who flees and who stays in Nigeria","Forced migrants are risk-averse; economic migrants are risk-tolerant"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, which the paper itself flags as strong, is that the way household characteristics translate into living standards is the same for people who stay and people who leave, so the gains from leaving can be predicted by applying non-conflict-area coefficients to conflict households.","fun_headline_variants_meta":{"raw":{"variants":["Risk-averse flee, risk-tolerant stay in conflict zones","Conflict migrants are risk-averse, stayers risk-tolerant","Quantile maximization explains who flees and who stays in Nigeria","Forced migrants are risk-averse; economic migrants are risk-tolerant"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2662,"prompt_tokens":922,"completion_tokens":1740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1669}},"tokens_in":538,"tokens_out":1740,"duration_ms":14299,"temperature":1.0,"reasoning_tokens":1669,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:52:42.336839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 329 Nigerian households who actually changed local government areas, use the living standards they actually had after moving as the leaving lottery, compare with stayers' observed standards, and check whether the CDFs still cross exactly once with leaving dominant at the low end. If the crossing disappears or reverses, the predicted direction of risk selection is an artifact of the imputation.","supporting_citations":[{"cited_title":"A Primer on Motion Capture with Deep Learning: Principles, Pitfalls and Perspectives","cited_arxiv_id":"2009.00564","evidence_quote":"Supplies the theoretical characterization of quantile maximization that the model applies to conflict mobility."},{"cited_title":"An axiomatization of quantiles on the domain of distribution functions","cited_arxiv_id":null,"evidence_quote":"Establishes quantiles as an essentially unique ordinal decision criterion consistent with weak first-order stochastic dominance, justifying the choice rule."},{"cited_title":"Prospect theory: An analysis of decision under risk","cited_arxiv_id":null,"evidence_quote":"Provides the behavioral critique of expected utility that motivates replacing mean evaluation with quantile evaluation."},{"cited_title":"Labor migration and risk aversion in less developed countries","cited_arxiv_id":null,"evidence_quote":"Theoretical model showing risk-taking can trigger migration, the economic-migration benchmark the paper reverses."},{"cited_title":"Long, and Charles Sprenger","cited_arxiv_id":null,"evidence_quote":"Evidence on violence and the certainty premium; second confounding-factor reference."},{"cited_title":"Risk tolerance during conflict: Evidence from aleppo, syria","cited_arxiv_id":null,"evidence_quote":"The only prior empirical study of risk preferences and forced migration, whose findings are consistent with the paper's results."}],"review_version":1}