Missing Data Analysis Using Statistical and Machine Learning Methods in Facility-Based Maternal Health Records

Missing data are a rule rather than an exception in quantitative research. The questionable aspect however is the extent, pattern, mechanism, and treatment of missingness in facility-based paper maternal health records. We utilized data from maternal health records at Kawempe National Referral Hospital, Uganda. Only records of women who had given birth at the Hospital during January 2017 to January 2021 were considered. The analysis was done using R-Studio using frequency distributions, Pearson χ2 Test. Treatment of missingness was done using Listwise deletion (LD), Mode Imputation, Multiple Imputation by chained equations (MICE), Imputation using K-Nearest Neighbors (KNN) and Random Forest (RF) Imputation. Performance of methods was investigated using prediction accuracy and the Kruskal–Wallis Test on Standard Errors (SEs) derived following a Logistic Regression. Overall, 5% of the data was missing. The proportion of missingness ranged from 1.4 to 20.7% in variables. Case-wise missingness was established where 2498 out of the 4626 cases (54%) had at-least one variable with missing value. The pattern of missingness was arbitrary. The data suggest either missing at random or missing completely at random. With the exception of LD, no difference in SEs following Logistic Regression was noted in the imputation methods for treatment of missingness (p > 0.05). Further, LD yielded the lowest prediction accuracy after Logistic Regression. No major variations were noted in the prediction accuracy following a Logistic Regression after imputation using MICE, mode imputation, KNN and RF. Missingness in facility-based health records should not be ignored. Researchers need to pay attention to both overall and case-wise missingness. © 2022, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.