CK030 - Poisson Regression

Question 1

What is ‘non-convergence’ ?

Answer 1

The likelihood function is not maximized by any value of the coefficients

Question 2

What is ‘complete seperation’ ?

Answer 2

The outcome completely seperates a covariate (so for a specific covariate, all subjects are events/non-events)

Question 3

What is ‘quasi-complete seperation’ ?

Answer 3

The outcome almost completely seperates a covariate (so for a specific covariate, almost all subjects are events/non-events)

Question 4

How to detect (quasi-)complete seperation?

Answer 4

• Estimated OR is 0 or very large
• Confidence Intervals are very wide

Question 5

What are key assumptions behind the Poisson regression model?

Answer 5

• The event rate (hazard rate) is constant within a category
• The event rate (hazard rate) does not differ across inidividuals

Question 6

What is the interpretation of the coefficients in Poisson regression?

Answer 6

The exp(b) are the Incidence Rate Ratios (IRR)

Question 7

What is ‘overdispersion’ ?

Answer 7

If the variance is larger than the mean

Question 8

What are causes of ‘overdispersion’ ?

Answer 8

• Dependence of events
• ‘Zero-inflated data’ (so events cannot happen for some categories/patients)
• Misspecification of the model

Question 9

What are the main consequences of ‘overdispersion’ ?

Answer 9

• Standard errors are too small
• P-values are too small (inflated type-1 error rate)

Question 10

What are solutions for ‘overdispersion’ ?

Answer 10

• ‘Quasi-Poisson regression’
• ‘Negative binomial regression’

Question 11

What are the main differences between ‘quasi-Poisson regression’ and ‘negative binomial regression’ ?

Answer 11

• ‘Quasi-Poisson’ can model both under- & overdispersion, ‘Negative binomial’ can only model overdispersion
• Coefficients in ‘Quasi-Poisson’ are the same as in ‘Poisson’, but they are different in ‘Negative binomial’
• Interpretation of the coefficients as the log(IRR) is the same in all models