Effect of sea surface temperature in El Niño regions on dengue dynamics in Colombia: Evidence from causal machine learning

Dengue fever is among the most rapidly expanding vector-borne diseases globally, with Colombia ranking among the most affected countries in the Americas. Although previous research has linked climate variability and El Niño–Southern Oscillation (ENSO) episodes to dengue dynamics, the direct causal effect of sea surface temperature (SST) in El Niño regions remains insufficiently explored. We conducted a retrospective ecological analysis using monthly laboratory-confirmed dengue cases from 1,044 Colombian municipalities (2013–2023), combined with atmospheric, oceanic, and socioeconomic data. We emulated an experimental design to estimate the effect of SST in El Niño regions 1–2, 3, 3–4, and 4 on excess dengue cases. Confounder adjustment was guided by a Directed Acyclic Graph (DAG), and causal effects were estimated using Double Machine Learning (DML) with XGBoost learners. We estimated the Average Treatment Effect (ATE) and the Conditional Average Treatment Effect (CATE) conditioned on altitude. Robustness was evaluated with refutation tests introducing random confounders, subset replacement, and placebo exposures. A total of 455,329 confirmed dengue cases were reported during the study period, peaking in 2023. The strongest association was observed for El Niño region 3, where a standard deviation increase in SST (1.24 °C) raised the probability of excess dengue cases by 6.9 percentage points (ATE = 0.069, 95% CI = 0.056 – 0.083). El Niño regions 3–4 and 4 showed slightly weaker ATE yet significant effects (6.4% and 6.2%), while El Niño region 1–2 had the lowest effect (4.6%). The CATE analysis revealed that the effects of El Niño regions 1–2 and 3 were stronger at higher altitudes; meanwhile, for El Niño regions 3–4 and 4, the effects showed a slightly negative trend, suggesting a heterogeneous effect of El Niño regions on dengue incidence in Colombia based on altitude. Robustness checks indicated the presence of residual bias, particularly when applying the subset replacement test. These findings highlight the importance of integrating oceanic monitoring into early warning systems of the disease and tailoring vector-control strategies to local ecological contexts.