I’m happy to announce the release of new BESH Stat version 0.18. The main change is the implementation of Zero-Inflated Poisson regression as described by Diane Lambert in [1]. BESH Stat results on default settings match on 4 decimal place to the R pscl package zeroinfl function parameter estimates and standard errors (by decreasing the convergence criterion on the Settings tab you can increase precision of the estimates)
What’s new?
- Zero-Inflated Poisson regression model that required few more changes internally mentioned below
- Poisson regression NULL deviance formula update to match [R] GLM function result in case of response variable [y] containing zero values. I that case contribution to the deviance is equal to = 2*(y * ln(1/mu) – (y-mu)). Similar update was made to deviance residuals computation.
- Logistic regression routine updated to accept response variable values in [0, 1] interval it does not need to be 0’s 1’s output as before. This update was needed to implement zero-inflated-poisson (ZIP) regression model estimation using the expectation-maximization (EM) algorithm.
- Complete redesign of Poisson and Logistic regression IRLS algorithm (internally) to accept externally provided starting parameter values and weights. This update was needed to implement ZIP regression model estimation using the (EM algorithm).
- If cancel button is clicked while (Poisson, Logistic, Multinomial Logistic, Ordinal Logistic, or ZIP regression) then calculation if aborted and userform closed. In previous version just the userform was closed but computation continued in the background. (Note that however, the Cox regression still behaves the “old” way as it require more changes in the code to implement. I put it on my TODO list.)
References:
- Diane Lambert. Zero-Inflated Poisson Regression, With an Application to Defects in Manufacturing. Technometrics, Feb1992, 34.1