Logit Model Flashcards
Outcome variable
Dichotonomous 0 or 1
Outcome based on….
Utility, if utility of buying (event) is greater then not buying. Model will predict that the product will be bought / the event happens
Probability of event (buying etc)
P[Y=1] = P[U > 0] = 1 - F(-a-xiB) or 1 - (CDF at that point)
Prediction
Y=0 when (u <= 0)
Y=1 when (u>0)
CDF probability is…..
Probability function for a consumer that a event will take place or not. At a certain utility level. At u = 0, probability is 50/50.
Estimation via…
Maximum likelihood = for what value of B is it most likely that the model matches the data.
Ways of assessing the impact of the IV’s
1) Coefficients
2) Odds ratio
3) Marginal effects
Coefficient interpretation
Only possible to derive the direction of the effect if the estimate is significant.
Odds are…
Likelihood of something happening to not happening.
Probability / (1 - probability)
Odds ratio meaning
Comparing odds, what happens to the odds when X changes….
Odds ratio; formula and meaning
odds ratio = exp(B1* delta Xi)
value of 1: no relation
value of < 1: negative relation
value of > 1: positive relation
Marginal effects interpretation
How much does the probability of observing a 1 increases or decreases with a increase in the variable, assuming all else stays equal.
Validation
1) Log likelihood ratio test
2) AIC / BIC
3) Hitrate
4) P-values & Wald (for coef.)
5) Psuedo R2
Two extra measures of validity
1) Top decile lift
2) Gini Coefficient
Top decile lift -> formula
TDL = (actual response rate group 1 / actual overall response rate) * 100%
