Submission ID 117098
| Issue/Objective | Vaccination effectively protects children from diseases, disabilities, and early death, yet estimating immunization coverage in India is challenging due to inconsistent records. The National Family Health Survey (NFHS) often relies on maternal recall when vaccination cards are unavailable, introducing potential biases. Though critical, maternal recall's accuracy varies by region and demographic. This study uses Bayesian spatial modeling to quantify maternal recall's role in estimating routine immunization rates nationally and sub-nationally, assessing its reliability, identifying biases, and improving vaccination coverage and dropout rate estimates. |
|---|---|
| Methodology/Approach | We analyzed measles-containing vaccine (MCV) coverage and DPT1 (diphtheria-tetanus-pertussis-containing vaccine dose 1)-DPT3 dropout (received DTP dose1 but not DTP dose3) among children under five years in India using data and GPS positions from NFHS-4 (2015-2016) and NFHS-5 (2019-2021). Bayesian spatial modeling, via Integrated Nested Laplace Approximation (INLA) and Spatial Generalized Linear Models (spGLM), estimated vaccination probabilities for hyper-local mapping. Vaccination status was classified as Card Marked or Maternal Recall. We compared INLA and spGLM performance, built baseline spatial models with geographic coordinates, and evaluated maternal recall's impact by training models with and without recall data. Model performance was assessed using coverage and Root Mean Squared Error (RMSE). |
| Results | We compared spGLM and INLA performance across Indian states. Models using only card data (spGLM(CARD), INLA(CARD)) were less accurate than those including maternal recall (spGLM(CARD+MR), INLA(CARD+MR)). Example state: Bihar, RMSE dropped from 0.41 (NFHS-4, MR=0) to 0.29 (NFHS-4, MR=1) and from 0.29 (NFHS-5, MR=0) to 0.23 (NFHS-5, MR=1), showing improved accuracy. INLA outperformed spGLM in efficiency, completing analyses in under an hour versus spGLM's hours, though predictive accuracy was similar. Maternal recall significantly enhanced model performance. |
| Discussion/Conclusion | Using Bayesian methods with maternal recall improved vaccine coverage estimates in unobserved areas, aiding dropout prediction and identifying cold-spots for targeted interventions. Regional disparities, influenced by socioeconomic and healthcare factors, highlight the need for localized strategies. Maternal recall's impact varied, with minimal effect in states like Punjab with robust healthcare. MCV1 coverage rose from 2016-2021, but dropout rates persist. Hyper-local mapping via spGLM and INLA reveals disparities, guiding precise interventions. |
| Presenters and affiliations | Ritika Singh Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi Ritika Singh Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi Sumeet Agarwal Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi Alexandre de Figueiredo Department of Infectious Disease Epidemiology, LSHTM |
