Non-Linear Models Flashcards
What is the maximized log-likelihood, l(b)?
b is the maximum likelihood estimate of coefficient B. The maximized log-likelihood, l(b), is the likelihood that the estimate, b, matches the data.
How high do we want l(b) to be?
l(b) should be reasonably high, constrained by the low bound of l(0) of the null model and the high bound of l(sat) of the saturated model. However, if l(b) is too high, i.e. approaches too quickly to l(sat), then there may be cause for concern of overfitting.
What value should the Pearson chi-square statistic approach if the GLM fits well?
n-p-1 if the GLM fits well.
What is the main disadvantage of using the Poisson distribution for the response variable? And how can we combat this?
The restriction of variance equaling the mean. We can use alternative count models that incorporate a Poisson distribution while letting the mean of the response differ from the variance of the response.
What is another name for the Tweedie Distribution?
Compound Poisson-Gamma Distribution
What is the difference between a zero-inflated model vs. a hurdle model?
A zero-inflated model allows its domain to start at 0, which allows its variance to be greater than its mean. A hurdle model’s domain starts at 1, which allows its variance to be greater or less than its mean.
What is the null and alternative hypotheses of a goodness of fit test?
The null hypothesis states that the expected values are close to the observed data. The alternative hypothesis assumes otherwise.
Why is the actual response, Y, not important in a generalized logit regression for a nominal response?
Y is meaningless in this type of regression. Rather, we are more concerned about the probability of each category.
