Evaluating the impact of unadjusted confounding and study design on estimated pathogen-attributable incidence of diarrhoea among children in low and middle-income countries: a sensitivity analysis of an attribution algorithm in the MAL-ED cohort

BackgroundUnderstanding diarrhoea aetiology is critical for understanding vaccine impact, but causal attribution is difficult. We evaluated the sensitivity of an existing attributable fraction-based method to unadjusted confounding and study design. MethodsWe used MAL-ED data to estimate attributable incidence (IRattr) using an algorithm regressing diarrhoea incidence on pathogen quantity. The reference model used mixed-effects logistic regression adjusted for co-infection, sex, test batch, and age. We evaluated the magnitude of confounding by prior immunity, antibiotics, socioeconomic status, and breastfeeding in incrementally-adjusted models. To understand the impact of post-diarrheal shedding, we excluded stools collected {+/-}7, 14, or 28 days from an episode. We conducted matched nested case-control and case-crossover studies and used Poisson regression to calculate risk-based estimates of the IRattr. Results40 406 stools samples from 1715 children were included (6625 diarrhoeal, 33 781 control). Prior immunity had the greatest confounding impact, with adjusted models underestimating IRattr compared to the reference; overall, confidence intervals overlapped substantially. Excluding samples proximal episodes showed no consistent pattern of bias, and most estimates were {+/-}10% of the reference except Cryptosporidium, tEPEC, and norovirus. Nested case-control and case-crossover study designs produced similar IRattr estimates, and risk-based estimates were lower than odds-based estimates, ranging from 33% (Shigella) to 49% (tEPEC) fewer episodes per 100 child-years compared to the reference. ConclusionThis algorithm was robust to unadjusted confounding, but risk- and odds-based estimates differed. We recommend adjusting for breastfeeding, antibiotics use, and prior infection using risk-based model. If case-control design is necessary, matched incidence density sampling should be used.