Prevalence and factors associated with self-medication in Pabna, Bangladesh: An interpretable statistical and machine learning approach

Background : Self-medication (SM), the use of medicines without a doctor’s supervision, is common in Bangladesh and other low- and middle-income countries and may cause adverse effects and contribute to antibiotic resistance. Factors associated with SM have mostly been identified with a single analytical method, leaving their consistency across methods underexplored. This study aimed to estimate the prevalence of SM and identify associated factors by comparing several methods within one interpretable framework. Methods : A cross-sectional survey of 421 adults was conducted in Pabna Sadar, Bangladesh. Four methods were applied in parallel to 13 variables and compared at the variable level: chi-square tests, multivariable logistic regression, LASSO stability selection, and SHAP-based importance from machine learning. Nine models were compared by nested cross-validation (CV) and evaluated on a held-out test set. Factors supported by at least two methods were considered associated. Antibiotic SM was examined with multivariable logistic regression. Results : The prevalence of SM in the previous month was 54.2%, and pharmacies were the main source of medicine information (55.7%). Among self-medicators, 39.0% took antibiotics, more often those aged 48 to 58 years than those aged 18 to 28 years. A chronic condition, gender, and education were supported by all four methods, and occupation and awareness of the antibiotic-dairy interaction by two. Respondents with a chronic condition had over four times the adjusted odds of SM (adjusted odds ratio [AOR] 4.36; 95% confidence interval [CI] 2.61 to 7.27), and males had about twice the odds of females (AOR 2.13; 95% CI 1.17 to 3.87). Random forest achieved the highest discrimination (CV AUC 0.818 ± 0.052; test AUC 0.827). Conclusions : SM was common in this setting. The identified factors depended partly on the analytical method, and agreement across methods distinguished robust from method-dependent factors. People with chronic conditions may need particular attention, and the pharmacy counter may be a practical point for intervention, including antibiotic sales. The framework may be applied to other survey-based health studies.