I am spending this weekend working on the results section of my study. I spent 2 days on the descriptive statistics (all 134 variables) and probably still need to add more discussion on some of them. Those sections included t-tests and chi-square based on my key outcome variable to see which items matter. I even limited the ones I would move into the logistic regression based upon effect sizes.

So I’m trying to run the logistic regressions today. First I put each of the most promising set of variables in individually. When you do a logistic regression in SPSS it’s base model is to look at the distribution of the outcome variable. If more than 50% have the outcome variable, guess all have it. If less, guess none have it. For me this makes a blind guess around 70% correct.

The ONLY version of the model that adds anything substantive (around 10% better prediction) uses the two variables that you get latest in the process – like at the same time you get the outcome variable. Specifically you can make a better guess as to whether someone will need placement into remedial math if you know whether they needed to be placed in remedial reading or writing. Unfortunately that’s a bit too late to do anything about the situation.

All of the other potentially predictive variables, including high school GPA, math test scores at two different points, math self-efficacy, or demographic stuff, fails to add more than 0.1% to the predictiveness of the model. Worse, some of these actually reduce it by up to 0.7%.

I don’t know what I’m doing wrong here….I mean, I didn’t expect a nobel prize for this thing, but I expected to find something. People do it all the time. I have to assume I’m doing something wrong…..I just don’t know what…

### Like this:

Like Loading...

*Related*

Joseph

said:Prediction is hard. One can make a very good model of tomorrow’s weather by guessing that it will be the same as today’s. In some places this approach can work really well. Yet all of the work of meterology adds about 10% to the model . . .

So I wouldn’t use 70% as the threshold. It is a non-deterministic system and it can be hard to separate people in the middle. In the financial industry we often had risk models with 65% concordance — but that was enough to do a lot with.