Pertussis notifications in the Netherlands: workload versus benefits

Abstract

Purpose: Model-based geostatistics (MBG) is a branch of spatial statistics that is increasingly used to support disease control programmes in low-resource settings. MBG methods allow modelling of the spatial variation in disease prevalence and can thus be used to inform intervention policy. In this study, our focus is to identify the best practice for MBG when the goal is to estimate the likelihood of exceeding relevant thresholds for intervention policy implementation. Continuous data on disease indicators, collected from surveillance programmes, are typically dichotomised to positive or negative for the disease of interest before MBG analysis is conducted. Our conjecture is that this dichotomisation reduces the accuracy of spatial mapping of disease prevalence due to loss of data.Methods & Materials: Consequently, we carry out a geostatistical analysis of data on both simulated data, and on malnutrition and anaemia data from a cross-sectional survey conducted in Uganda and Ethiopia. We then compare two possible approaches for disease mapping: the first defines a model for a continuous measurement and uses this to predict the probability of exceeding a predefined threshold. The second approach first dichotomises the continuous measurement into a binary outcome indicating the exceedance or not of the threshold, and then develops a geostatistical logistic regression based on this binary outcome.Results: Our results indicate that these two approaches lead to a substantially different quantification of uncertainty associated with the estimates of disease prevalence.Conclusion: Dichotomisation of the data leads to higher uncertainty in the spatial predictions of prevalence. Modelling with continuous data should therefore be applied whenever this is feasible.