This analysis employed longitudinal data from 193 countries for the period 2000–2023. This study period was selected to coincide with the modern era of global immunization expansion following the launch of Gavi, the Vaccine Alliance, in 2000. This period ensures data availability and standardized reporting for newer-generation vaccines (e.g., HepB3, Hib3, and RCV1) and provides the highest temporal alignment with the UCDP and UNHCR displacement datasets. We utilized regression analysis, two-way fixed-effects (TWFE) and Callaway–Sant’Anna (CS) difference-in-differences (DiD) models, event-study models with HonestDiD sensitivity analysis, and dose–response analyses to quantify the effect of forced displacement, armed conflict, and natural disasters on national childhood vaccination coverage.
Ethics statement
This research complies with all relevant ethical regulations. The study is a secondary analysis of publicly available, aggregated country-year data containing no individually identifiable information, and involved no human participants, human tissue or animals. Institutional review board approval and informed consent were therefore not required and no ethical approval was sought. No participants were recruited or compensated.
Sex and gender
The unit of analysis is the country-year-vaccine observation and the study involved no human participants. WUENIC national immunization coverage estimates are not published disaggregated by sex or gender, and many other data inputs were likewise reported as aggregate totals. No sex- or gender-based analysis was therefore possible and sex and gender were not considered in the study design; because no individual-level data were collected, no disaggregated source data can be provided.
Data
National vaccination coverage rates, our dependent variable, were obtained from the WHO/UNICEF Joint Estimates of National Immunization Coverage (WUENIC)42. This dataset includes annual vaccination coverage estimates for 14 essential childhood vaccines. For our analysis, we dropped vaccines that were imbalanced across country-year observations, resulting in the national average of 10 included vaccines: Bacille Calmette-Guérin (BCG); the first and third doses of diphtheria, tetanus toxoid, and pertussis containing vaccine (DTP1 and DTP3); the third dose of hepatitis B vaccine (HEPB3); the third dose of Haemophilus influenzae type B containing vaccine (HIB3); the first and second doses of measles containing vaccine (MCV1 and MCV2); the third dose of pneumococcal conjugate vaccine (PCV3); the third dose of polio containing vaccine (POL3); and the first dose of rubella containing vaccine (RCV1). More details on the selection of these vaccines and sensitivity analyses incorporating WHO national introduction dates are included in the Supplementary Information (Supplementary Tables 1–3, Supplementary Fig. 1a–d).
Our primary exposure variable was derived from national totals of forcibly displaced persons, defined as individuals displaced from their homes due to conflict or disaster. These data were obtained from the United Nations High Commissioner for Refugees (UNHCR), based on inputs from the Expert Group on Refugee and Internally Displaced Persons Statistics (EGRIS) and endorsed by the Bureau of the United Nations Statistical Commission43. The estimates draw from multiple sources, including censuses, surveys, and administrative records. For this analysis, we defined exposure as the annual country-level total of forcibly displaced persons, aggregating both cross-border and internal displacement, given that disaggregated data are frequently inconsistent or incomplete across contexts. This origin-and-internal definition means that, for a given country, the exposure captures demand-side pressure on that country’s own system and the broader systemic disruption that accompanies large-scale out-migration, rather than the receiving-country burden borne elsewhere. We note that these data primarily capture conflict-associated displacement, so displacement counts may be underestimated due to missing IDPs; we re-estimate our main analysis in the Supplementary Information (see sensitivity to Alternate Displacement Data in the Supplementary Information) using displacement data (2008–2023) from the Internal Displacement Monitoring Centre, which explicitly delineates conflict-attributable and disaster-attributable displacement flows and stocks44.
We also incorporated data on armed conflict from the Uppsala Conflict Data Program (UCDP) Country-Year dataset on organized violence within country borders (version 25.1)45,46. The UCDP dataset captures all armed conflicts in which the conflict results in a minimum of 25 battle-related deaths (BRDs) within at least one calendar year. To operationalize conflict exposure, we followed standards in the literature and categorized conflict intensity into three levels based on annual BRD counts that took place in each country: no conflict (< 25 BRDs), low-intensity conflict (25–999 BRDs), and war (≥1000 BRDs)2.
Data on natural disasters were derived from the Emergency Events Database (EM-DAT), maintained by the Centre for Research on the Epidemiology of Disasters (CRED) at the Université catholique de Louvain (UCLouvain)47. We subset EM-DAT data to only natural disasters—comprising hydrological, meteorological, geophysical, biological, and climatological disasters—and removed infectious disease outbreaks to mitigate possible correlation between our independent and dependent variables.
Finally, our dataset also included GDP per capita data from the World Bank as a control variable48. We included lagged GDP per capita as the sole time-varying covariate given its consistent availability across countries and years, its role as a parsimonious proxy for socioeconomic capacity and health system resources, and to avoid introducing post-treatment bias from controls that may themselves be influenced by conflict or displacement. Country and year fixed effects further absorb unobserved heterogeneity and global shocks.
Statistical analysis
We first grouped countries into quintiles based on the total number of forcibly displaced persons to explore patterns in vaccination coverage across displacement burden. Using a fixed-effects ordinary least squares (OLS) regression with vaccine and year fixed effects, we estimated the association between displacement burden and immunization coverage. To quantify associations, we estimated a two-way fixed-effects (TWFE) difference-in-differences (DiD) specification on a country-vaccine-year panel, evaluating the average effect of displacement shocks on coverage outcomes in both the shock year alone (Year 1) and across a 3-year post-treatment window (Years 1–3). Displacement shocks were defined by combining a relative-increase criterion (e.g., year-over-year growth of ≥50% in displaced persons) with an absolute-volume criterion (e.g., thresholds of ≥100,000 displaced persons), alongside parallel specifications based on the share of national population displaced. Requiring both percentage growth and absolute volume reflects the scale and intensity of the displacement event, and is methodologically necessary to avoid misclassifying small proportional changes in low-displacement settings as shocks while still capturing substantial increases in comparatively high-displacement settings. We varied both criteria across scenarios to assess the robustness of our findings.
Our primary specification is enumerated in Eq. (1), wherein the binary Post equals one for country-years in which a country had experienced a displacement shock meeting the scenario-specific treatment criteria. In the Year-1 specification, only the shock year was coded as post; in the 3-year specification, the shock year and two subsequent years were coded as post. For every treated country, the 5 years preceding the shock were coded as pre. Post was interacted with conflict intensity, allowing treatment effects to vary across conflict contexts. Conflict intensity (ConflictCat) was a time-varying, three-level categorical variable derived from annual battle-related deaths (BRDs): no conflict (< 25 BRDs), low-intensity conflict (25–999 BRDs), and war (≥1000 BRDs), with no conflict as the reference category; 1(·) denotes an indicator function equal to one when its argument holds. Lagged log of GDP per capita (GDPpcc,t−1) was included as a covariate. Models included country-vaccine unit fixed effects (αcv), year fixed effects (\({\lambda }_{t}\)), with standard errors clustered by country (\({\varepsilon }_{{{{\rm{cv}}}}t}\)). Identification derives from within-unit deviations in the Post indicator and its interactions with conflict intensity; no separate static treatment dummy is estimable. In Eq. (1), β1 is the displacement-shock effect in the no-conflict tier, and the tier-specific effects in the minor-conflict and war tiers are recovered as β1 + δminor and β1 + δwar, respectively, via clustered linear combinations of the model coefficients.
$${Coverag}{e}_{{cvt}}={\beta }_{0}+{\beta }_{1}{{Post}}_{{ct}}+\mathop{\sum }_{k}{\gamma }_{k}\,1({{ConflictCat}}_{{ct}}=k)\,+\,\mathop{\sum }_{k}{\delta }_{k}\,[{{Post}}_{{ct}\cdot}{{{\rm{\cdot }}}}1({{ConflictCat}}_{{ct}}=k)]\,+\,{\beta }_{2}{{GDPpc}}_{c,t – 1}+{\alpha }_{{cv}}+{\lambda }_{t}+{\varepsilon }_{{cvt}}$$
(1)
Because treatment timing was staggered across countries, conventional TWFE estimators can be biased by comparisons that use already-treated units as controls for newly treated units49. To address this possible bias, we estimated the average treatment effect on the treated (ATT) using two heterogeneity-robust estimators. First, we applied the Callaway–Sant’Anna50 estimator with never-treated countries as the control group, lagged log GDP per capita as a covariate (estimated via outcome regression, with an unadjusted specification used as a fallback in cases of singularity) and standard errors clustered by country. Second, as a cross-check, we applied the Sun-Abraham51 interaction-weighted estimator, specified as cohort-by-relative-time interactions within the same country-vaccine and year fixed-effects structure, aggregated to an overall post-treatment ATT. We assessed robustness to violations of the parallel-trends assumption using the HonestDiD relative-magnitudes sensitivity analysis52, reporting the breakdown value M̄, the largest violation of parallel trends, expressed as a multiple of the maximum pre-treatment deviation, under which the estimated post-treatment effect remains significantly negative.
To characterize the dynamics of displacement shocks over time, we estimated event study models that compare outcome trajectories in treated versus untreated units relative to the timing of the shock. We estimated three specifications. The first was a treated-only TWFE specification (Eq. (2)), which retained the lagged log GDP per capita covariate, the conflict intensity controls and fixed-effects structure of Eq. (1), and added event-time indicators spanning the 5 years before and 5 years after the displacement shock, with the year immediately preceding the shock as the reference period. The second was a corrected TWFE event study that incorporated never-treated countries as the comparison group by assigning them to the reference period. The third was the dynamic ATT obtained from the Callaway–Sant’Anna50 estimator. Across all specifications, standard errors were clustered by country.
$${{\mbox{Coverage}}}_{{cvt}}=\mathop{\sum }_{k= – 5,k\ne – 1}^{5}{{{{\rm{\beta }}}}}_{k}\cdot 1\left({{EventTime}}_{{ct}}=k\right)+{\beta }_{2}\cdot {{\mbox{GDPpc}}}_{{ct} – 1}+\mathop{\sum }_{j}{\gamma }_{j}1({{ConflictCat}}_{{ct}}=\,j)+{\alpha }_{{cv}}+{\lambda }_{t}+{\varepsilon }_{{cvt}}$$
(2)
Dose–response modeling
To quantify the relationship between population displacement, conflict, natural disasters, and vaccination coverage, we employed generalized additive models (GAMs) with penalized regression splines. Coverage outcomes were modeled as a function of total displacement, BRDs, and disaster-related fatalities, allowing for nonlinear dose–response relationships. Models included fixed effects for country, year, and vaccine type to account for unobserved heterogeneity across space, time, and immunization program characteristics. Log transformations were applied to highly skewed predictors to stabilize variance and improve model fit. We fit separate models for each exposure (proportional displacement, absolute displacement, battle-related deaths, and disaster-related fatalities), with the proportional displacement specification serving as the primary model for the main text. As a supplementary analysis, we refit the absolute-displacement model within population terciles (small, medium, large countries) defined by each country’s median population over the study period, to assess whether dose–response patterns vary with country size. Sensitivity to the choice of basis dimension was assessed by refitting the primary model across k ∈ [4, 6, 8, 12, 18, 24, 36, 54]; results were stable across this range. Smoothing parameters were selected via restricted maximum likelihood (REML) to balance model flexibility and overfitting.
Model diagnostics included residual checks and evaluation of concurvity to ensure appropriate spline behavior and robustness. Predicted coverage estimates were derived over a fine grid of exposure values, holding covariates such as GDP per capita at reference levels. Uncertainty around predictions was quantified using standard errors, which were used to generate 95% confidence intervals. Heatmaps and dose–response curves were produced from these predictions to visualize expected coverage across the range of displacement and conflict intensities, highlighting nonlinear and synergistic effects on immunization coverage.
Robustness and sensitivity analyses
Placebo modeling was used to introduce fake shocks at random years into the DiD models at different treatment criteria; no significant values were observed, increasing confidence in our model results. Missingness was evaluated for systemic correlations and none was observed (see Supplementary Materials). Alternative DiD models were run with 2024 Social and Economic Rights Fulfillment Index (SERF) Right to Health scores (ranging from 2000–2021) instead of logged GDP per capita (see Supplementary Table 12), identifying similar results in both substantive effect sizes and statistical significance.
We also ran DiD models with the continuous version of V-Dem’s health equality scores (see Supplementary Table 13); when accounting for health equality, forcible displacement was only statistically significant in the most extreme displacement thresholds (e.g., 100% year-over-year increase with more than 500,000 persons displaced). This attenuation may reflect the scope of V-Dem’s health equality metric, which captures whether high-quality basic healthcare is broadly guaranteed across the population. In contexts where health systems are relatively equitable, the additional strain from moderate displacement inflows may be absorbed without measurable declines in vaccination coverage. By contrast, when displacement exceeds the absorptive capacity of even well-functioning systems, the negative impact becomes detectable. These results highlight the importance of considering underlying health system equity as a moderating factor in the relationship between displacement and vaccination outcomes.
The reporting of this study conformed with the STROBE guidelines. All analyses were conducted using R software (Version 4.6.1) using the packages dplyr (version 1.2.1), tidyr (version 1.3.2), purrr (version 1.2.2), tibble (version 3.3.1), stringr (version 1.6.0), readr (version 2.2.0), readxl (version 1.5.0), countrycode (version 1.8.0), ggplot2 (version 4.0.3), patchwork (version 1.3.2), scales (version 1.4.0), fixest (version 0.14.2), did (version 2.5.0), HonestDiD (version 0.2.8), mgcv (version 1.9.4), sensemakr (version 0.1.6), broom (version 1.0.13), marginaleffects (version 0.32.0), zoo (version 1.8.15), and states (version 0.3.3).
Case studies
To contextualize the statistical estimates we conducted two descriptive case analyses. At the national level we examined Ukraine (2014), Syria (2012), and Venezuela (2016), tracing annual displacement, WUENIC coverage, and battle-related deaths around each country’s first qualifying shock together with contemporaneous reports of vaccine-preventable disease outbreaks (Supplementary Case Study 1). At the subnational level we examined state-level displacement and zero-dose vaccination in Nigeria (Supplementary Case Study 2). Both analyses are descriptive and are not used to estimate treatment effects.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.