Joan Valls1, Gerard Castellà2, Tadeusz Dyba3, Ramon Clèries4. 1. Biostatistics Unit, Biomedical Research Institute of Lleida (IRBLLEIDA) Hospital Universitari Arnau de Vilanova de Lleida (HUAV) C/ Rovira Roure, 80 Laboratoris d'investigació HUAV, 25198 Lleida, Catalonia, Spain; Department of Mathematics, Autonomous University of Barcelona, Facultat de Ciències, Edifici C, 08193 Bellaterra (Cerdanyola del Vallès), Catalonia, Spain. 2. Biostatistics Unit, Biomedical Research Institute of Lleida (IRBLLEIDA) Hospital Universitari Arnau de Vilanova de Lleida (HUAV) C/ Rovira Roure, 80 Laboratoris d'investigació HUAV, 25198 Lleida, Catalonia, Spain. 3. European Commission, DG Joint Research Centre, Institute for Health and Consumer Protection, Public Health - Cancer Policy Support, Ispra, Italy. 4. Plan for Oncology of the Catalan Government, IDIBELL, Hospital Duran i Reynals, Av. Gran Via de l'Hospitalet, 199-203-1ª planta. 08908 L'Hospitalet de Llobregat, Catalonia, Spain; Dept. of Clinical Sciences, University of Barcelona, Campus de Bellvitge, Edifici Pavelló de Govern, Feixa Llarga s/n, 08907 L'Hospitalet de Llobregat, Barcelona, Catalonia, Spain. Electronic address: r.cleries@iconcologia.net.
Abstract
BACKGROUND: Predicting the future burden of cancer is a key issue for health services planning, where a method for selecting the predictive model and the prediction base is a challenge. A method, named here Goodness-of-Fit optimal (GoF-optimal), is presented to determine the minimum prediction base of historical data to perform 5-year predictions of the number of new cancer cases or deaths. METHODS: An empirical ex-post evaluation exercise for cancer mortality data in Spain and cancer incidence in Finland using simple linear and log-linear Poisson models was performed. Prediction bases were considered within the time periods 1951-2006 in Spain and 1975-2007 in Finland, and then predictions were made for 37 and 33 single years in these periods, respectively. The performance of three fixed different prediction bases (last 5, 10, and 20 years of historical data) was compared to that of the prediction base determined by the GoF-optimal method. The coverage (COV) of the 95% prediction interval and the discrepancy ratio (DR) were calculated to assess the success of the prediction. RESULTS: The results showed that (i) models using the prediction base selected through GoF-optimal method reached the highest COV and the lowest DR and (ii) the best alternative strategy to GoF-optimal was the one using the base of prediction of 5-years. CONCLUSIONS: The GoF-optimal approach can be used as a selection criterion in order to find an adequate base of prediction.
BACKGROUND: Predicting the future burden of cancer is a key issue for health services planning, where a method for selecting the predictive model and the prediction base is a challenge. A method, named here Goodness-of-Fit optimal (GoF-optimal), is presented to determine the minimum prediction base of historical data to perform 5-year predictions of the number of new cancer cases or deaths. METHODS: An empirical ex-post evaluation exercise for cancer mortality data in Spain and cancer incidence in Finland using simple linear and log-linear Poisson models was performed. Prediction bases were considered within the time periods 1951-2006 in Spain and 1975-2007 in Finland, and then predictions were made for 37 and 33 single years in these periods, respectively. The performance of three fixed different prediction bases (last 5, 10, and 20 years of historical data) was compared to that of the prediction base determined by the GoF-optimal method. The coverage (COV) of the 95% prediction interval and the discrepancy ratio (DR) were calculated to assess the success of the prediction. RESULTS: The results showed that (i) models using the prediction base selected through GoF-optimal method reached the highest COV and the lowest DR and (ii) the best alternative strategy to GoF-optimal was the one using the base of prediction of 5-years. CONCLUSIONS: The GoF-optimal approach can be used as a selection criterion in order to find an adequate base of prediction.