Joan Valls Marsal 1 Ramon Clèries Soler 2

Transcription

1 Joan Valls Marsal 1 Ramon Clèries Soler 2 1 Biostatistics Unit. Biomedical Research Institute of Lleida 2 Catalan Cancer Registry jvalls@irblleida.cat / r.cleries@iconcologia.net October 25th 2013

2 Continguts / 27

3 Cancer incidence and mortality Incidence and mortality cancer : usually defined as the incident cases or deaths in one single year, per 100,000 persons. recorded by the Incidence and Mortality Cancer Registries, regional or population-based. conventional epidemiological measures for assessing cancer. rate trends and their modelling: key issue to evaluate the impact of new treatments, health programs or other epidemiological related issues. 2 / 27

4 2008 World wide cancer incidence and mortality Jemal et al (2001) CANCER J CLIN ;61:6990) 3 / 27

5 2008 World wide cancer incidence and mortality Jemal et al (2001) CANCER J CLIN ;61:6990) 4 / 27

6 2008 Catalan cancer incidence Borras et al (2008) Med clin 5 / 27

7 Catalan and European mortality Gispert et al (2008) Med clin 6 / 27

8 Predictions Prediction of incidence or mortality in the future. Important both from the epidemiological and health planing point of view and to assess the burden of cancer. Used to evaluate the effect of the new health programs or treatments. Most used models for prediction: Age-period or age-period-cohort models (Poisson distribution) Historical data to feed predictions. Prediction out of the range of observed values. 7 / 27

9 To evaluate the optimal a prediction in the future f or mortality using age-period models. 8 / 27

10 Models for rate trends c it and n it incident cases or deaths and persons at risk in the population, respectively, for ith age group in the t period (year). Models: Model AAAP (Additive Age-specific Age-Period) ( ) cit E = α i + β i t n it Model MAAP (Multiplicative Age-specific Age-Period) ( ( )) cit Model MAAP log E = α i + β i t Model MAP (Multiplicative Age-specific Period) ( ( )) cit Model MAP log E = α i + βt n it n it c it Poisson distributed and fixed n it. Dyba i Hakulinen (2000) Stat Med, Dyba et al (1997) Stat Med 9 / 27

11 Predictions from a model Taxa ajustada per persones inici fi Període d ajust del model 10 / 27

12 Predictions from a model Taxa ajustada per persones inici fi Període d ajust del model projecció 10 / 27

13 Predictions from a model Taxa ajustada per persones inici fi Període d ajust del model? projecció 10 / 27

14 Predictions from a model Taxa ajustada per persones inici fi Període d ajust del model?? projecció 10 / 27

15 Predictions from a model Taxa ajustada per persones inici fi Període d ajust del model??? projecció 10 / 27

16 Predictions from a model Taxa ajustada per persones ??? inici fi Període d ajust del model projecció 10 / 27

17 Model selection Taxa ajustada per APC 2% APC 5% APC 10% Taxa ajustada per APC 2% APC 5% APC 10% Dades observades 11 / 27

18 Model selection Taxa ajustada per APC 2% APC 5% APC 10% Taxa ajustada per APC 2% APC 5% APC 10% Dades observades Model multiplicatiu 11 / 27

19 Model selection Taxa ajustada per APC 2% APC 5% APC 10% Taxa ajustada per APC 2% APC 5% APC 10% Dades observades Model multiplicatiu Model additiu 11 / 27

20 Pearson goodness of the fit Pearson goodness-of-fit statistic used to assess the model X 2 = I T i=1 t=1 (c it ĉ it ) 2 Var(ĉ it ) Distributed as χ 2 on the residual degrees of freedom for the model 12 / 27

21 Analyses performed Data of mortality cancer in Spain from 1951 to 2011 downloaded from WHO. Three cancer sites for women and men: Women: breast, colorectal and lung cancer. Men: lung, colorectal and prostate cancer. AAAP model when increasing trends and MAP or MAAP models when decreasing trends. 13 / 27

22 Analyses performed Prediction made for each year in 1975 up to 2011 Prediction using data historical data 5 years before the prediction. Basis period for prediction: 1, 2,..., 20 last years time basis period for prediction 1, 2,..., 20 last years 5 years prediction 14 / 27

23 Analyses performed Different basis period time for the prediction Last year Last 5 years Last 10 years Last 20 years GoF-optimal years GoF-optimal years: smallest period of time where the model fits, according to Pearson s GoF statistis 15 / 27

24 Analyses performed Comparison of the different methods (1,5,10,20 last years or GoF-optimal): For each year predicted (1975 up to 2011) obtention of 95% confidence interval for the predicted deaths. Boolean evaluation of the inclusion of the true observed value respect to the interval. Computation of a whole coverage (% years where the true value is inside the interval). 16 / 27

25 Breast women p-value GoF selection Pearson's GoF p value Pearson's GoF p value Number of years for the basis time period of prediction Number of years for the basis time period of prediction Pearson's GoF p value Pearson's GoF p value Number of years for the basis time period of prediction Number of years for the basis time period of prediction 17 / 27

26 Breast women comparison methods correct.cases.all prediction basis: last year * * 2 out of 36: coverage= 5.56 % correct.cases.all prediction basis: last 5 years * * * * * * * * * * * * * * * * * * * * * * * 23 out of 36: coverage= % correct.cases.all year.predicted prediction basis: last 10 years * * * * * * * * * * * 11 out of 36: coverage= % year.predicted correct.cases.all year.predicted prediction basis: last 20 years * * * * * * 6 out of 36: coverage= % year.predicted prediction basis: last optimal years correct.cases.all * * * * * * * * * * * * * * * * * * * * * * * * out of 36: coverage= 75 % year.predicted 18 / 27

27 Breast women: last year prediction basis: last year * * correct.cases.all out of 36: coverage= 5.56 % year.predicted 19 / 27

28 Breast women: last 5 years prediction basis: last 5 years * * * * * * * * * * * * * * * * * * * * * * * correct.cases.all out of 36: coverage= % year.predicted 20 / 27

29 Breast women: last 10 years prediction basis: last 10 years * * * * * * * * * * * correct.cases.all out of 36: coverage= % year.predicted 21 / 27

30 Breast women: last 20 years prediction basis: last 20 years correct.cases.all * * * * * * 6 out of 36: coverage= % year.predicted 22 / 27

31 Breast women: GoF-optimal prediction basis: last optimal years correct.cases.all * * * * * * * * * * * * * * * * * * * * * * * * out of 36: coverage= 75 % year.predicted 23 / 27

32 Lung women correct.cases.all prediction basis: last year * * * * * * * * * * * * * * 14 out of 36: coverage= % correct.cases.all prediction basis: last 5 years * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 34 out of 36: coverage= % year.predicted prediction basis: last 10 years year.predicted prediction basis: last 20 years correct.cases.all * * * * * * * * * * * * * * * * 17 out of 36: coverage= % year.predicted correct.cases.all * * 2 out of 36: coverage= 5.56 % year.predicted prediction basis: last optimal years correct.cases.all * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * out of 36: coverage= % year.predicted 24 / 27

33 All tumor sites Method Tumor site 1 year 5 years 10 years 20 years GoF-optimal Women Breast 5.56 % % % % 75 % Colorectal % 50 % % 2.78 % % Lung % % % 5.56 % 97.22% 25 / 27

34 Simple method with reasonable results and useful at practice. Good alternative to choosing a subjective years. supporting that large historical data are not needed for. Limitations: Ongoing results: more results with new tumor sites and countries. Not possible to establish general statistical properties from particular data sets. 26 / 27

35 Thank you Thank you for your attention! 27 / 27