{"id":"72779c00-1103-4863-b14e-07ced11861a8","arxiv_id":"1908.07100","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Using generalized joint regression models that model alliance formation and conflict together, the study finds that defensive alliances have an average treatment effect indistinguishable from zero.","lead":"A political science paper tests whether military alliances deter conflict after accounting for the fact that alliances may themselves be caused by the expectation of conflict. Using a joint statistical model with infant mortality as an instrument, it finds that defensive alliances have no measurable deterrent or provoking effect.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The null ATE rests on an instrument exclusion restriction that the paper does not validate; infant mortality plausibly affects conflict through target state fragility, not alliances alone.","rationale":"The reader's weakest-assumption identification matches the most load-bearing threat to the central claim: the exclusion restriction for the infant-mortality instruments. The headline claim is explicitly causal ('once correctly accounting for the endogeneity of alliances... challengers are neither deterred from nor induced to attack'), and that causal interpretation is available only if the instruments are valid. The paper's support for validity is weak—absence of published evidence is not evidence of absence—and the substantive mechanism linking infant mortality to state fragility makes a direct path to conflict plausible even after conditioning on GDP per capita and civil conflict history. Other concerns, such as the predictive comparison using different covariate sets and the single-split copula selection, affect the paper's supporting narrative but do not bear as directly on the causal identification. The proposed overidentification test is a feasible check because the paper already has two instruments and a well-defined specification. Since the reader's verdict was already CONDITIONAL and this stress-test reinforces rather than overturns that verdict, I recommend no change; the paper should add the instrument-validity check, ideally with a same-covariate exogenous baseline and replication code, before the null causal claim is accepted.","tokens_in":17711,"tokens_out":8809,"duration_ms":110290,"concrete_test":"Compute the Sargan/Hansen overidentifying-restrictions statistic in a linear 2SLS analog of the specification using both infant-mortality instruments and the same controls. Under the null that both instruments are excluded from the conflict equation, the statistic is chi-square with one degree of freedom; a rejection at the 5% level indicates that at least one instrument affects conflict through a channel other than alliances, directly invalidating the identification assumption. If the statistic is far from significant and the IV alliance coefficient remains near -0.016, the concern is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is identification of the alliance ATE in the recursive bivariate probit, which requires that the target's infant mortality rate and its change affect the challenger's dispute initiation only through the target's defensive alliance commitments (Section 4.1). The paper's stated defense—'I find no published evidence of a direct relationship'—is an appeal to absence of documentation, not a statistical or design-based justification. Infant mortality is a broad proxy for state capacity, public health, demographic structure, and fragility; even after conditioning on target GDP per capita and civil conflict history, these dimensions are not fully blocked, and GDP and civil conflict are themselves potentially endogenous to conflict. If high infant mortality directly raises the probability that A attacks B (a fragile target) while also increasing B's propensity to seek alliances, the estimated alliance coefficient is contaminated by the direct instrument→conflict path. The observed null is then compatible with a genuinely deterrent alliance effect, or with a range of effects, rather than establishing no effect. Because the two equations are identified by this instrument alone, no amount of copula flexibility or out-of-sample PR-AUC addresses failure of the exclusion restriction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper challenges the conventional finding that defensive alliances deter militarized conflict. The author argues that alliance formation is endogenous to the security environment and conflict expectations, and models this with a recursive bivariate probit 'generalized joint regression model' using the target's infant mortality rate and its change as instruments for the target's relevant defensive alliance commitment. The central empirical result is an average treatment effect of -0.016 with a 95% credible interval [-0.053, 0.016], which is interpreted as evidence that alliances neither deter nor provoke aggression. The paper also reports that the joint model outperforms a standard Johnson-Leeds logistic regression in precision-recall AUC and that the null result is robust to alternative copula specifications.","tokens_in":17911,"tokens_out":5629,"duration_ms":59991,"significance":"If the causal identification were credible, the paper would be a significant challenge to a prominent finding in international relations: that explicitly relevant defensive commitments deter attack. The manuscript has genuine strengths: it takes the endogeneity of alliances seriously, uses instruments borrowed from prior published work, reports an ATE with a credible interval, and includes a sensitivity analysis over nineteen copulas. The out-of-sample predictive comparison is useful descriptive evidence. However, the central claim rests on an untestable exclusion restriction, and the predictive comparison does not isolate the role of endogenizing alliances. The significance of the paper is therefore conditional on whether the identification strategy can be defended more convincingly or the claims appropriately weakened.","major_comments":[{"comment":"The identification of the alliance ATE rests entirely on the exclusion restriction that the target's infant mortality rate and its change affect the challenger's conflict initiation only through the target's defensive alliance commitments, after conditioning on target GDP per capita and civil conflict history. The support offered in the text — 'I find no published evidence of a direct relationship' — is an appeal to an absence of documentation rather than a design-based or statistical justification. Infant mortality is a broad proxy for state capacity, public health, demographic structure, and fragility; these dimensions are not fully blocked by GDP per capita and civil conflict history, and those two controls are themselves potentially endogenous. If infant mortality directly raises the probability that a challenger attacks a fragile target while also increasing the target's propensity to ally, the estimated alliance coefficient is contaminated by a direct instrument-to-outcome path. Because the recursive bivariate probit is identified by these instruments alone, this concern is load-bearing for the null ATE.","section":"Section 4.1, 'Modeling Alliance Formation'"},{"comment":"The claim that endogenizing alliances improves predictive performance is not supported by the reported comparison. The Johnson and Leeds (2011) baseline and the GJRM differ in their covariate sets: the GJRM adds target GDP per capita, target civil-conflict history, challenger trade dependence, and a peace-years spline, among other differences. The 57–59% PR-AUC improvement in Figure 4 therefore cannot be attributed to modeling alliance endogeneity. The conclusion in Section 6 that 'endogenizing... improves our ability to predict' requires either a nested comparison with identical covariate sets or a decomposition of the performance gain.","section":"Section 5, 'Predictive performance' and Figure 4"},{"comment":"The conclusion that alliances 'neither deter nor provoke aggression' is stronger than the evidence. The estimated ATE is -0.016 with 95% credible interval [-0.053, 0.016]; a null result with this interval cannot establish equivalence or the absence of substantively meaningful effects. The paper itself notes in Section 5 that the result is 'not definitive,' but the abstract and conclusion use language of a definitive null. Equivalence testing, a sensitivity analysis to unobserved confounding, or bounds on plausible effect sizes would be needed to support the strong conclusion.","section":"Section 5, 'Average treatment effect' and Section 6"}],"minor_comments":[{"comment":"The phrase 'Infant morality rate' should read 'infant mortality rate.'","section":"Section 4.1"},{"comment":"The definitions of precision and recall are garbled: precision is TP/(TP+FP) and recall is TP/(TP+FN), but the text appears to reverse or confuse these quantities. This should be corrected because the PR-AUC comparisons rely on these definitions.","section":"Section 5, 'Copula selection'"},{"comment":"The instruments are described as strong based on chi-square and likelihood-ratio tests, but no first-stage coefficient table is presented. Including a table with the first-stage estimates and standard errors would clarify the strength and sign of the instrument-alliance relationship.","section":"Section 4.1 and Section 5"},{"comment":"The predictive comparison between the Johnson-Leeds model and the GJRM may also be affected by differences in estimation samples (e.g., due to missing GDP or trade data). The paper should report the number of observations used in each model and any sample restrictions.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious empirical application, but the central identification assumption is fragile. I would not recommend rejection solely because the instruments are borrowed from prior published work; however, the manuscript needs either to substantially strengthen the exclusion-restriction defense or to temper the causal conclusions. The predictive performance claim also needs to be reframed as a model comparison rather than evidence about endogeneity. If the editor sees the paper primarily as a methodological demonstration, the novelty relative to the cited GJRM literature should be made clearer."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a serious, well-written attempt to endogenize alliance formation in models of interstate conflict, and if the identification held it would overturn a well-known result. But the identifying assumption is doing essentially all the work, and the paper does not validate it. I would still send it out—the field should confront this—but I would not take the null as established.\n\nWhat is actually new is the application: a recursive bivariate probit GJRM where a target's relevant defensive alliance commitment is endogenous, instrumented by the target's infant mortality rate and its year-to-year change, and the conflict equation then estimates an ATE on challenger dispute initiation. The paper engages both the alliance-formation and alliance-conflict literatures, includes controls from both strands, checks every copula in the GJRM package, and reports an ATE of -0.016 with a 95% interval [-0.053, 0.016]. That is a dramatically different conclusion from Johnson and Leeds's 20% deterrent effect. The theoretical story—unobserved threat, ally reliability, and burden sharing drive both alliance formation and attack—is coherent, and the null is not imposed by construction. Citation-wise, it engages the main exchange fairly and credits Braumoeller et al. for the method.\n\nThe soft spot, and it is load-bearing, is the exclusion restriction in Section 4.1. The paper argues that infant mortality should influence whether A attacks B only through alliance formation, civil conflict, or capabilities, with civil conflict history and GDP per capita conditioned out. But infant mortality is a broad proxy for state fragility, public health, and demographic stress; a fragile target is plausibly a more attractive target for reasons not captured by GDP and civil conflict history. And those two conditioning variables are themselves potentially endogenous to conflict. The paper's stated defense—'I find no published evidence of a direct relationship'—is an appeal to absence of documentation, not a design-based justification. No amount of copula flexibility or out-of-sample PR-AUC fixes a failed exclusion restriction.\n\nThe second issue is the predictive improvement claim. The endogenous model includes extra covariates and a different functional form, so the PR-AUC gain over the Johnson and Leeds baseline does not isolate the effect of endogenizing alliances. A same-covariate non-endogenous baseline would be needed. Minor points: no replication code or data is provided, and the instruments are asserted to be strong in a footnote without reported test statistics.\n\nWho is this for: IR scholars working on alliances and deterrence, and applied methodologists interested in binary endogenous treatments. It deserves a serious referee, but the referee should push hard on the instruments and on the model-comparison design. My recommendation: peer review, with the expectation of major revisions or a substantial weakening of the causal claims.","headline":"A serious, readable empirical challenge to the alliance-deterrence consensus, but the null result leans entirely on an unvalidated infant-mortality exclusion restriction and a predictive comparison that does not isolate the endogeneity correction.","tokens_in":18413,"tokens_out":2570,"would_cite":true,"duration_ms":31139,"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":"The paper claims that once alliance formation is modeled jointly with conflict, a target's relevant defensive alliance commitment neither deters nor provokes a challenger, with an estimated average treatment effect of -0.016 (95% CI…","keywords":["military alliances","deterrence","endogeneity","generalized joint regression models","militarized interstate disputes","instrumental variables","causal inference","alliance formation"],"falsifier":"A credible demonstration that infant mortality or its short-run change has a direct effect on interstate dispute initiation—through, say, state weakness or diversionary incentives—after conditioning on GDP per capita and civil conflict would break the exclusion restriction and leave the null average treatment effect unidentified.","tokens_in":17526,"feed_emoji":"🛡️","tokens_out":9403,"duration_ms":89022,"temperature":0.7,"pith_summary":"Military alliances are usually studied as if they were randomly assigned, but states form alliances because they expect conflict. This paper re-estimates the effect of a target's defensive alliance commitments on whether a challenger initiates a militarized interstate dispute (a threat, display, or use of force), allowing alliance formation and conflict to be driven by the same unobserved security concerns. When alliance formation is modeled jointly with conflict, the targeted state's defensive commitment has essentially no effect: an average treatment effect of -0.016 with a 95% interval crossing zero. The paper argues that this null means alliances neither deter nor provoke aggression, and that conventional estimates showing deterrence mostly captured selection into alliances. The finding matters because it challenges a central premise of deterrence theory and redirects research toward the conditions under which alliances might work.","feed_headline":"Defense pacts neither deter nor provoke once modeled jointly","feed_subtitle":"Jointly modeling alliance formation and conflict cuts the apparent 20% deterrent effect to a null 1.6%.","key_machinery":"The carrying mechanism is the generalized joint regression model (GJRM): a recursive bivariate probit in which the first equation predicts whether the target has a relevant defensive commitment and the second predicts whether a challenger initiates a dispute. A copula—a function linking the two equations' unobserved errors, chosen here as the 180-degree rotated Clayton—lets the same unmeasured factors influence both alliance formation and conflict, which is the channel that the paper says accounts for the apparent deterrent effect. The alliance equation also includes two instruments: the target's infant mortality rate and its one-year change, both used to satisfy the exclusion restriction.","core_discovery":"The paper's central claim is that, once the endogeneity of alliance formation is accounted for, a target's relevant defensive alliance commitment neither deters nor provokes a challenger. In a bivariate recursive probit model estimated on directed-dyad years 1816-2000, the average treatment effect of the defensive commitment on dispute initiation is -0.016 with a 95% credible interval [-0.053, 0.016], materially smaller than the roughly 20% reduction reported by prior exogenous models. The paper attributes the difference to unobserved confounders—offensive intention, ally reliability, and burden sharing—that make states with high conflict risk more likely to form alliances, so apparent deterrence is selection. It interprets the null as the causal effect of alliance commitments once these confounders are accounted for.","pith_inferences":["If the null is correct, past correlational evidence for deterrence mostly documented selection: threatened states form alliances, and those same threats predict attack; policy claims that alliances reliably buy security need re-examination.","The exclusion-restriction strategy could be stress-tested with different instruments tapping public-goods pressure, such as disease burden or schooling enrollment; a second independent instrument producing the same null would strengthen the case.","A null average effect does not mean alliances are useless; their benefits may appear in non-conflict domains such as autonomy, domestic survival, or postwar management, which are not measured here.","The copula results imply a common unobserved factor raises both alliance formation and conflict; identifying that factor empirically could replace instrumental-variable reliance with direct measurement."],"forward_implications":["The conventional finding that defensive alliances cut the probability of being attacked by about 20 percent is not reproduced when alliance formation is endogenous; the estimated effect is 1.6 percent and statistically indistinguishable from zero.","The null result is robust across copula choices: 18 of 19 specifications yield null or provocation effects, and only the poorly fitting Student-t copula recovers deterrence.","Models that endogenize alliance formation predict militarized disputes better than the exogenous benchmark, improving out-of-sample precision-recall by 57.4 percent.","Alliance formation and conflict initiation should be modeled as one joint process, since conflict expectations shape alliance choices and alliances may then shape conflict.","Research should shift from asking whether alliances deter to identifying moderating conditions under which particular commitments deter or provoke."],"supporting_citations":[{"why":"Provides the benchmark exogenous model showing a 20% deterrent effect, which the paper re-estimates and contrasts with the endogenous null result.","marker":"Johnson and Leeds (2011)"},{"why":"Supplies the generalized joint regression modeling framework and the argument that unobserved confounders bias alliance-conflict estimates.","marker":"Braumoeller et al. (2018)"},{"why":"Introduces infant mortality rate and its change as instruments for alliance formation through public-goods outsourcing.","marker":"Kimball (2010)"},{"why":"Provides estimation of the recursive bivariate probit model with endogeneity, used here to compute the average treatment effect.","marker":"Marra and Radice (2011)"},{"why":"Supplies the alliance treaty data used to operationalize a target's relevant defensive commitment.","marker":"Leeds et al. (2002)"},{"why":"One of the recent deterrence critiques whose variables and debate frame the model specification.","marker":"Kenwick, Vasquez and Powers (2015)"}],"fun_headline_variants":["Alliances neither deter nor provoke once endogeneity is modeled","Joint model finds no causal effect of defense pacts on conflict","Accounting for alliance selection erases deterrent effect on conflict","Defense commitments have no significant causal impact on disputes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result stands or falls on the assumption that a target's infant mortality rate affects whether it is attacked only by changing its alliance choices, after taking GDP per capita and civil conflict into account.","fun_headline_variants_meta":{"raw":{"variants":["Alliances neither deter nor provoke once endogeneity is modeled","Joint model finds no causal effect of defense pacts on conflict","Accounting for alliance selection erases deterrent effect on conflict","Defense commitments have no significant causal impact on disputes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1332,"prompt_tokens":858,"completion_tokens":474,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":474,"tokens_out":474,"duration_ms":5872,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:26:46.355915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A credible demonstration that infant mortality or its short-run change has a direct effect on interstate dispute initiation—through, say, state weakness or diversionary incentives—after conditioning on GDP per capita and civil conflict would break the exclusion restriction and leave the null average treatment effect unidentified.","supporting_citations":[],"review_version":1}