Data-driven methods for imputing national-level incidence in global burden of disease studies

Abstract

ObjectiveTo develop transparent and reproducible methods for imputing missing data on disease incidence at national-level for the year 2005.MethodsWe compared several models for imputing missing country-level incidence rates for two foodborne diseases – congenital toxoplasmosis and aflatoxin-related hepatocellular carcinoma. Missing values were assumed to be missing at random. Predictor variables were selected using least absolute shrinkage and selection operator regression. We compared the predictive performance of naive extrapolation approaches and Bayesian random and mixed-effects regression models. Leave-one-out cross-validation was used to evaluate model accuracy.FindingsThe predictive accuracy of the Bayesian mixed-effects models was significantly better than that of the naive extrapolation method for one of the two disease models. However, Bayesian mixed-effects models produced wider prediction intervals for both data sets.ConclusionSeveral approaches are available for imputing missing data at national level. Strengths of a hierarchical regression approach for this type of task are the ability to derive estimates from other similar countries, transparency, computational efficiency and ease of interpretation. The inclusion of informative covariates may improve model performance, but results should be appraised carefully.

Methodological Framework for World Health Organization Estimates of the Global Burden of Foodborne Disease

“…To overcome the inevitable problem of missing data, we imputed missing data using a hierarchical random effects model as a default. Although the use of explanatory covariates such as eating habits or income levels is often considered in these exercises, we decided not to pursue such models driven by our earlier model evaluations and comparisons [ 29 ]. The choice of our default imputation model was further motivated by a strive for parsimony and transparency, while recognizing that other approaches could be used for stand-alone studies.…”

“…This is a strong assumption, and led to the exclusion of five hazards for which the assumption was clearly violated, i.e., Bacillus cereus , Clostridium perfringens , Clostridium botulinum , Staphylococcus aureus , and peanut allergens. For the remaining hazards, it is difficult to evaluate the validity of this assumption, as this would require a comparison of incidence data in countries with data versus countries without data, which per definition is not possible [ 29 ]. As a result, this assumption is made in all global burden of disease studies, even though this is not always explicitly mentioned.…”

“…The external data used must thus be representative of the selected population, region and time. The CTF developed, tested and evaluated several possible approaches to impute missing incidence data at the country level [ 29 ]. This exercise identified several pitfalls in the use of explanatory covariates, such as the potential for overfitting and the arbitrariness in the selection of covariates.…”

World Health Organization Estimates of the Global and Regional Disease Burden of 11 Foodborne Parasitic Diseases, 2010: A Data Synthesis

“…Incidence estimates and clinical sequelae, for diseases caused by foodborne trematodes were mainly based on the results of two review articles [ 16 , 17 ] ( S1 Table ). We also imputed incidence rates for countries without reported national prevalence, but with reports of at least one autochthonous human infection, by using a hierarchical random-effects models and incidence information from other countries as input data [ 18 ]. In highly endemic zones adult subjects either maintain the parasites acquired when young or can be newly infected as the consequence of inhabiting a zone of high infection risk.…”

World Health Organization Global Estimates and Regional Comparisons of the Burden of Foodborne Disease in 2010

“…When incidence data were not available for all countries in a subregion, incidence rates were extrapolated from available data using a Bayesian log-normal random effects model, specifying the subregion as random effect. The predictive value of this model was explored and compared with other possible imputation models by McDonald et al [ 17 ]. Detailed descriptions of methods to estimate incidence and mortality of all 31 hazards can be found in papers in the collection on enteric diseases [ 11 ], parasitic diseases [ 18 ], and chemicals and toxins [ 19 ].…”