Poster De Conférence Année : 2022

Linked police and health data: how to apply capture-recapture to correct for under-reporting and bias (tutorial)

Résumé

Police data are often used to estimate the number of injured road users. However, these data suffer from under-reporting and more importantly from bias. When police data are linked with health data, one can correct for under-reporting and bias with the capture-recapture approach. We describe how. Capture-recapture is based on the condition of homogeneity of capture/recording by a given source. This is usually not satisfied, as the probability of being captured /registered in a given database (police especially) usually depends on severity, mode of transport, etc. One way to account for this is to stratify on these characteristics, and in each strata, calculate the simple Petersen estimate, and then sum over the strata. However, frequencies in some strata may be too small and/or additional variables may influence the probability of recording, at least in the police data: single /multi vehicle, driver/passenger, type of road… In this case, one should use a multivariate model and include these characteristics as covariates. The response variable is whether the injured road user is recorded in police data only, health data only, or in both. We can then estimate the number of injured road users not recorded in any database or directly estimate the total number of injured road users. Formulas will be provided; implementation with the SAS and R software will be described. Examples will be given with the French Rhône county, over 2006-2016. The police recorded 2700 injured road users on average by year, the road trauma registry 7400; 1700 were linked. This means 8400 altogether. Around 11 800 are estimated by capture-recapture, stratified on injury severity or by a multivariate capture-recapture model: it accounts for five variables strongly associated with recording probability (and 3 more). Only the multivariate model can provide unbiased estimates by injury severity, mode of transport, age, and gender.
Presentation_IRTAD_recapture new.pdf (723.66 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04297617 , version 1 (24-11-2023)

Identifiants

  • HAL Id : hal-04297617 , version 1

Citer

Emmanuelle Amoros, Cebrail Aksoy, Amina Ndiaye, Blandine Gadegbeku, Bernard Laumon, et al.. Linked police and health data: how to apply capture-recapture to correct for under-reporting and bias (tutorial). 7th IRTAD conference : Better road safety data for better safety outcomes, Sep 2022, Vaulx-en-velin, France. 21 p., 2022. ⟨hal-04297617⟩
25 Consultations
4 Téléchargements

Partager

More