Binary logistic regression using Stata 17 drop-down menus (May 2021)

[@Camilika]

Hi Mike, why don't you separate the sample to train and test the model? Does it only apply to ML algorithms? Thank you. Best regards from Colombia.

[@rubaazfrali]

Hi I did not undestand the relationship between the OR and coeff. lets sya I want to calculate the OR from the coeff, how Iam gonna do that?

[@mikecrowson2462]

Hi Ruba, Take the natural log of the OR , to convert to the regression slope: ln(OR) To calculate the OR from slope: exp(b) , where you raise e to the power of b I hope this helps! Best wishes

[@fanaaraf8663]

Hello, I have noticed that your log likelihood is negative. Shouldn't the log likelihood be a positive number to establish that the model is a good fit? From what I know, the log likelihood should be a positive number. Is this true? I'm dealing with a similar issue, such that I have over 300k observations. Although my variables are significant and prob>chi2 = 0.000 , my log likelihood is - 207079 . Why is the log likelihood so large and negative? Is the log likelihood interpreted in absolute value? I would appreciate your assistance.

[@mikecrowson2462]

The log likelihood you are referencing is computed as the sum of the log likelihood of the cases, where the log likelihood for a case is the natural log of the probability of a case being in it's observed group. The natural log of a number of a probability is a negative number since it ranges from 0-1 (excluding 0 of course). There's a nice illustration in Tabachnick and Fidell (2013). The sum this is negative, where a value closer to 0 is better fit and more negative numbers indicating worse fit. If you multiply this by -2, you get the -2LL, which theoretically ranges from 0 to positive infinity. Some programs explicitly report this value where closer to 0 indicates better fit and higher values indicate worse fit. This value is approximately chi square distributed and used when calculating the chi square test that is reported. Hope this helps. Cheers