# logistic regression+test of hypothesis

#### Lisa92

##### New Member
Hello guys!

I need your advice. I have done a logistic regression to estimate a probability of impaiment occurs or not. My indep.variables were different factors that can inflduence dep.variable. The data is taken from the stock exchange so it is not normally distributed.

I got results from the whole model and then i made 5 different versions of that where each time i took out variables with a least signif.level and did analysis again.

I used stata

I want to do an extra test to check whether it is right or not. And I want to test a hypotheisis that some of this variables do influence dep.variable.I am not that good at statistics but if you advise me a test I would be very thankfull!
Maybe a lrtest in stata?

#### noetsi

##### No cake for spunky
If your dependent variable is a probability which is usually 0-100 percent, why did you use logistic regression? The fact that the data is not normally distributed is not a good reason to run logistic regression if what you are predicting ranges from 0 -100.

What do you want to test if it is right or not? The log likelihood and other model tests show you if the overall model has predictive value. The individual wald test for variables show if they have predictive value.

A practical problem of doing multiple test the way you are is you increase the chance of type 1 error but that tends to get ignored in regression models. Dropping variables and rerunning the results, even if commonly done, is really not the right way to run regression. If your theory says the variables are supposed to be in the model you should leave them there and report they are not significant.

#### Lisa92

##### New Member
If your dependent variable is a probability which is usually 0-100 percent, why did you use logistic regression? The fact that the data is not normally distributed is not a good reason to run logistic regression if what you are predicting ranges from 0 -100.

What do you want to test if it is right or not? The log likelihood and other model tests show you if the overall model has predictive value. The individual wald test for variables show if they have predictive value.

A practical problem of doing multiple test the way you are is you increase the chance of type 1 error but that tends to get ignored in regression models. Dropping variables and rerunning the results, even if commonly done, is really not the right way to run regression. If your theory says the variables are supposed to be in the model you should leave them there and report they are not significant.

No, no, my dep.variable is binomal, it can be 0 or 1!