Analysis Ofmultivariate Hierarchical data With Missingness- anapplication to in-patient Paediatric Pneumonia Care

Abstract

Routine health data are used to monitor quality of care and to inform interventions to improve patient care. However, statistical analysis of such data presents several challenges related to handling missing data and multiple responses in the presence of complex data structures. In this study we sought to: i) Analyze multilevel clustered data accounting for covariate missingness. ii) Explore appropriate strategies for handling missing data when the outcome is a composite of partially observed components. iii) Examine sensitivity of results to departures from the commonly assumed missing at random (MAR) mechanism. iv) Simultaneously estimate joint covariate effects and association amongst multiple correlated outcomes. We analysed routine data collected during a cluster randomized trial in 12 Kenyan hospitals between March and November 2016. There were 2127 children admitted by 378 clinicians ascross the study sites. The outcomes of interest were 12 pneumonia quality of care indicators spanning assessment, diagnosis and classification and treatment domains of care. For the first three objectives, we constructed Paediatric Admission Quality of Care (PAQC) score, an ordinal composite outcome using 12 pneumonia care indicators. Covariates of interest included : trial arm and follow-up time, hospital, clinician and patient-level variables. Missing data occurred in patient and clinician level variables. Missing data in covariates were imputed using latent normal joint modelling approach assuming MAR mechanism. Random-effects and marginal models were the substantive models of interest. To explore appropriate strategies of handling missing PAQC score subcomponents, we conducted a simulation study. Multiple imputation (MI) at subcomponent level versus the conventional method where missing PAQC score subcomponents were scored with value 0. We assessed departure form MAR assumption within pattern mixture models. Elicited experts’ opinions were incorporated into the imputation models in the form of prior distributions and delta adjustment parameters to create missing not at random imputed values. In the fourth objective, we analyzed 9 binary pneumonia care indicators under the correlated random effects joint model, by applying pairwise fitting and pseudolikelihood methods before and after MI of missing covariates. From results, trial intervention was associated with higher uptake of the paediatric pneumonia guidelines during the trial period. Parameter estimates were precise after MI of covariates compared to complete case analysis. In a range of simulation scenarios, multiple imputation of missing PAQC score elements at item level produced minimally biased estimates compared to the conventional method. Our inferences were insensitive to departures from MAR assumption using either sensitivity analysis approach. Lastly, there was a significant joint interaction effect between intervention arm and follow-up time on pneumonia care indicators. The strength and direction of association amongst outcomes varied within and across domains care. This study demonstrates the practical utility of advanced biostatistical analyses methods with an aim to promote their use while answering substantive health research questions. Uptake of such methods can improve analysis and reportiv ing of health data used to inform policies and in the long run enhance optimal utilization of limited resources while promoting better patients’ outcomes.