I am currently doing some work on annual reports and am planning on using the Logit method. I was wondering if someone could point me in the right direction. When running this in Microfit do I include an intercept term (set as 1) in the regression? If I use an intercept term then non of my models are showing any variable that are statistically significant. Yet without the intercept term I seem to get results suggesting that the majority of my variables are statistically significant. So Would it be ok for me to run the models in logit without an intercept term? Thanks

HiI wish I could say intercept could be chucked away, but there is normally no reason for doing so. I believe if the intercept were not significant, including it should not have made much difference anyway.Normally the problem is down to the choice of explanatory variables. Check whether you could drop some (such as the MOST insignificant ones, of course it is best if you can try to explain why in theory), or whether some of them are correlated since this creats large biases. Sometimes even the choice of unit measures can result in insignificant results, scale them down.Universities normally offer good documents on Logit models (using Microfit or STATA).Hope this helps, good luck!

Thanks for the reply. I am currently using simple models with just one explanatory variable. None of these have been statistically significant due while including an intercept term which is why I asked. Also there are some econometric models that do not use an intercept term so I was hoping that I could get away without using one here.I have also taken logs of the data and used these in the model but I am still getting results that are not statistically significant.

Think about what your results would say with regular regression with one independent variable (i.e. with intercept, variable is not significant, but w/o it is).This could happen if you took some uncorrelated data and centered it on the point (10, 10). With intercept, you'd get an intercept of 10 and a slope of 0...or not significantly different than zero. Without intercept, you'd get a slope of 1 which could be statistically different than zero.Trying to choose a model that "gives significance" isn't really a good way to go about it. It reminds me of people in physics labs who would use the model (say for the time it takes for ball to roll down a slope) to adjust "poor" experimental data.Ideally, as suppiii mentions, it is best to have some economic theory to guide the form of the model. Also, as suppiii mentions, including the intercept is allowing a larger class of models...if true intercept is zero, then including it shouldn't hurt.Sometimes the insignificance is as much information as the significance. Variables that we think are important may not be.Good luck.

Thanks for your reply KO. What you have said makes alot of sense and my approach has been the wrong one.

IF it is possible you can run the logit in SAS. Where among the various tests one of the tests try tp tests the hypothesis that all estimates other than the intercept are zero. You can do a seperate hypothesis testing also.Besides in your existing exercise you can transform the input variables. A very simple transformation that may help your exercise is:From individual variables(Say Var A) substract the sample mean of the variable(Sample mean of Var A) and use it in the model instead of the original variable.

I;ve done the same research before, and your question made me go back and rethink what my results were.My results showed that the intercept played a very strong role. both in terms of the size of the coefficient and the T-stats significance level.If I recall upon deeper investigation, we found that most of the initial variables we used were quite useless (explaining why the had low power and T-stats)I suspect the intercept captured a significant source of unobservable bias.First, the variables you used may be what you think is important but proabably was useless. Consider interactive variables e.g Gearing x Cash Flow. Interepretation gearing may only be important for firms with weak cash flowsSecond, how is you data balanced - 90% good and only 10% bad. This usually leads to lousy prediction models.How many data points do you have.What is your objective - Prediction, Descriptive or Hypothesis testing (causality)Going based on theory - helps a bit. But I found that almost all the variables that analyst tell me is important, seldom are