1
$\begingroup$

I'm analyzing customer satisfaction for a certain program. My response variable is ordinal- where 0 means dissatisfied and 1 means satisfied. I have 10 predictor variables. I used Binary logistic Regression on my data, and the results showed that 6 of my 10 variables are statistically significant. My question is, what can I safely say about my output? Can I say that there is significant evidence to suggest the variables have an affect on the response?

$\endgroup$
1
$\begingroup$

A common practice is to perform a likelihood ratio test for testing the assumptions that all coefficients are equal to 0. Most of the statistical softwares provide this test in the output of the logistic regression. Once this test has been rejected, you can analyse your results.

$\endgroup$
  • $\begingroup$ +1; Another way of saying the same thing is that it sounds like you are interested in a test of your model against a null model that estimates the intercept only (e.g. base rate of satisfaction). This approach will lump together all variables and see how much better the set of them do at predicting satisfaction than just using the base rate. I'm not sure I agree with the 'protection' suggestion (requiring a significant overall model before looking at individual coefficients) perse... but the gist of this answer is right. $\endgroup$ – russellpierce Aug 2 '13 at 13:26
  • $\begingroup$ @rpierce Thanks, so would the null model be the one with all my predictor variables? And is it okay for me to say that there is evidence to suggest my predictor variables have an affect on response? $\endgroup$ – Stat01 Aug 2 '13 at 14:03
  • $\begingroup$ The "null model" is the model under the null hypothesis. In this case, typically, it would be the model with only an intercept. The alternative model is the one that has all of your predictor variables. Getting a description of your result that is both easy to understand and accurate is difficult to do without getting bogged down in NHST language. Something like "a model with these predictors was a statistically significant improvement over a model without these predictors. This suggests that these predictors are related to this outcome". $\endgroup$ – russellpierce Aug 2 '13 at 15:08
  • $\begingroup$ Be careful about "affect" unless you experimentally manipulated your IVs. There is no special sauce in the math that lets you assess causality. That will depend on your reasoning and research design. $\endgroup$ – russellpierce Aug 2 '13 at 15:09
  • $\begingroup$ Global null hypothesis tests using the likelihood ratio or score test are standard output from most packages and you don't have to actually re-fit a null model. $\endgroup$ – Frank Harrell Aug 2 '13 at 16:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.