Other
Zero-Dose and Dropout Children in Nigeria: A Comparative Machine Learning Analysis of Childhood Immunization Cascade Failure Using the 2023–24 Nigeria Demographic and Health Survey
Eloghosa Aisosa Nosa-Ihaza, Michael Chiebuka Ogidi, Stephen Okello Odeke, Uyioghosa Nosayise Nosa-Ihaza ·
Summary
Written by AI, not yet AICS's view · claude-sonnet-5This study uses Nigeria's 2023 to 2024 national health survey to look at children who received no vaccines at all, called zero-dose, and children who started but did not finish their vaccine series, called dropout. It finds that these two groups have different causes and should be treated as separate problems. About 37 percent of children aged 12 to 23 months were zero-dose, and vaccination coverage was lower than official global estimates suggest. The study also tested whether complex computer prediction models work better than standard statistics for identifying these children, and found they did not.
Key points
- Zero-dose prevalence was 37.3%, meaning implied Penta1 coverage of 62.7%, lower than the WHO/UNICEF estimate of 71%
- Full vaccination coverage was 53.6%, lower than the global administrative estimate of 67%
- Maternal education, facility delivery, ANC attendance, and household wealth strongly predicted zero-dose status but were weak or not significant predictors of dropout
- Muslim religious affiliation remained an independent predictor of zero-dose status even after adjusting for state-level clustering
- Standard survey-weighted logistic regression performed as well as or better than machine learning methods like XGBoost and random forest for predicting both outcomes
Limitations
The study only covers children aged 12 to 23 months from one survey round and cannot show how or why the identified factors cause zero-dose or dropout status, only that they are statistically associated.
Abstract
Abstract Background There are two distinct failure points in childhood immunization: children who receive no immunization (“zero-dose”); and children who start but do not complete the immunization series (“dropout”). Two failure points to address are possible: ‘zero-dose’, where children receive no first doses; and ‘dropout’, where children begin but do not complete the immunization series. These are plausibly different failure points that require different interventions. Nigeria has the largest number of children with zero doses in absolute numbers in the world. The newly released Nigeria Demographic and Health Survey (NDHS) 2023–24 is used to describe failure cases and to compare the performance of machine learning techniques with standard regression in predicting them. Methods We analyzed 4,937 children aged 12-23 months (the age range covered by the WHO/UNICEF children’s recode) in the NDHS. Survey-weighted logistic regression was used to model ZDS and dropout (Penta1 received vs. Penta3 not received) and to compare them with the WHO/UNICEF national coverage estimates (WUENIC). With the same held-out test data, each of three additional algorithms - elastic net, random forest, and XGBoost - was compared to logistic regression, and sampling weights were added to each. Geographic clustering, after adjusting for individual-level effects, was quantified using a mixed-effects model with a state-level random intercept. Results Zero-dose prevalence was 37.3% (implied Penta1 coverage 62.7%), compared with the WUENIC national estimate of 71% Penta1 coverage; incomplete series coverage was at 9.1% (14.5% of Penta1 recipients), and full vaccination coverage was at 53.6%, which is lower than the administratively weighted global coverage estimates (67% Penta3). The associations of zero-dose status were strong and independent for maternal education, facility delivery, attending ANC, and household wealth, whereas these factors were surprisingly weak, albeit statistically significant for Maternal education and Facility delivery, and not significant for the other factors in the dropout group. After full adjustment, religious affiliation (among Muslims) remained an independent predictor of zero dose, even though clustering by state was adjusted for. Survey-weighted logistic regression performed best for both outcomes (zero-dose AUC=0.825; dropout AUC=0.613), whereas XGBoost performed worst for both. Only 4.7% of the residual variation in dropout could be accounted for by state of residence, but 11.2% could be accounted for by zero-dose status. Conclusions Zero-dose status as well as dropout are characterized by very different predictor profiles and are, therefore, separate policy problems and should not be treated as a single “coverage gap”. There was no improvement in prediction performance in terms of algorithmic complexity compared to a correctly specified logistic regression for either outcome.