Decision Tree
Transcript: Poisson Model Categorical var. Continuous var. Does it violate the POA? If so use multinominal logit 3rd Choice Negative Binominal Check for linearity, normality, heteroskedaticity, multicolliniarity. Considere one of the following 3 and if interaction terms is relevant Panel Data Random vs. Fixed Effects Count var. When we expect a non-linear relationship btw. dep.var. and ind.var. Multinominal Logit (no natural order) or Ordered Logit (Natural order) Endogeneity 2nd Choice Non-linear regression Test for serial correlation. If yes (severe)=Use 1st D. If no=Use FE. (If serial corr. is not severe, may report both FE & 1stD. If there is selection bias, we should use a Heckman Selection Model to account for this Look at the data: Panel data, use PD approach). Cross-sectional, use these. Then look at characteristics of dependent variable If over-dispertion Check for excess zeros - if there are, used zero-inflation model Dummy Variable (0;1 variables) OLS If unsure of Multinom. vs. Ordered: Do POA to test. Logit - remember to look at interactions Fixed Effects vs. 1st Difference Selection Bias Ordered Logit Does it violate the IIA? If so: Nested Logit Mixed Logit Multinom. probit Pooled OLS? Test with the Hausman. h0=Use Random Effects Accept= Use RE Reject= 3rd choice Multinominal Logit 1st Choice If we suspect endogeneity we should use instrumental variable regression. If no over-dispertion Is a simpler model, so should be first choice Check for excess zeros - if there are, used zero-inflation model Use Breuch-Pagan test to see if biased: No - use OLS Yes - 2nd choice Discrete (More than 2 levels)