Count Data– Generalized Linear Models Department of 计数数据的广义线性模型–系.docVIP

Count Data – Generalized Linear Models –Poisson Regression Poisson Regression Data: Units: NASCAR Winston Cup Races (1975-1979 = n=151 races) Dependent Variable: Y = # of Caution Flags (e.g. crashes) (CAUTIONS) Independent Variables: X1 = # of Drivers in the Race (DRIVERS) X2 = Circumference of the Track (TRKLENGTH) X3 = # of Laps in the Race (LAPS) Random Component: Poisson Distribution for the # of accidents Density Function: Link Function: g(?) = log(?) Systematic Component: Testing for the overall model: H0: ?1 = ?2 = ?3 = 0 (# of crashes is independent of all predictors) HA: Not all ?i = 0 Test Statistic: Xobs2 = -2(L0-L1) Rejection Region: Xobs2 ( Where L0 is the maximized log likelihood function under H0 and L1 is the maximized log likelihood under HA. H0: g(?????? L0 = 410.8784 (SAS Code and output below) HA: g(?) = ?????1X1 + ?2X2 + ?3X3 L1 = 433.0160 Xobs2 = -2(410.8784-433.0160) = 44.28 Strong evidence that number of crashes is associated with predictors Testing for individual (partial) regression coefficients: H0: ?j = 0 HA: ?j ( 0 (can clearly do or as well) Test Statistic: (some software packages report z2) Rejection Region: |zobs| ( z?/2 (obvious adjustment for 1-sided tests) Predictor bj SE(bj) Zobs Drivers 0.0365 0.0125 2.92 Trklength 0.1145 0.1684 0.68 Laps 0.0026 0.0008 3.25 Note that as the number of drivers increases, number of crashes increases, similarly, as the number of laps increases, crashes increase. Controlling for Drivers and Laps, Trklength is not associated with crashes. Simpler Model: Testing for Goodness of Fit for the Poisson Model If only a few categories (combinations of levels are such that exceeds 5 in most cells) of the independent variables, the goodness of fit can be computed directly from the and likelihood ratio goodness of fit statistics as: These are approximately chi-square for fixed number of categories under the restriction on fitted values above. The degrees of freedom a