# Looking for help with interpreting my OLS models

I have been busy building an explainatory regression model for a companies' social behavior. In the table below you find the variables, ordered by theme. The English to French variables are dummies, indicating the legal system a company has to comply with.

From these models, conclusions about what influences a companies' social behavior will be drawn. Since I'm a law student (with no statistics course in the curriculum) I've been looking at other scientific papers in which regression results are presented. I have some questions:

1. I present the coefficients and the significance levels. Many authors present the coefficients and the t-values, and indicate significance level with * (for 1%) and ** (for 5%). What does the t-value indicate / add to the information this table presents?
2. Papers regularly test 6 to 8 models. Why? I guess this is to test the robustness of the coefficients but am unsure. Also, is there a rule of thumb to see which models need testing. In other words: do the six models presented here cover model testing or do the results call for more models to be tested? And if so, which models?
3. About the conclusions: take, for instance, (under Culture), uncertainty avoidance. It's effect on company behavior is significant in models (A) to (D), but not in (E) and (F). What should I report? That uncertainty avoidance does of does not have an effect on social behavior?

Many questions, most of which basic probably. But I'm trying my best to get this right and any help is much appreciated :-) Also other comments regarding these models that are not covered in my questions, are welcome.

• These are complicated questions & the issues involved are subtle. As fond as I am of CV, I am not confident that people can steer you through the maze of issues involved to address this situation appropriately via Q&A over the internet. I don't want to be discouraging, but I think you need to either take some statistics classes, read & work through the textbooks on your own, or work w/ a consultant. – gung - Reinstate Monica May 6 '13 at 18:54
• I certainly agree that my knowledge of statistics is under par. However, I now tought myself to the point that I can perform a regression analysis. (DV has normal division, IV's have relatively low correlation to DV, multicollinearity in all models is low, durbin-watson around 2 (as rule of thumb indicates)). I learned that, when looking at model (A) all effects are significant except for ROA and French dummy. I learned that dummy coefficients should not be interpreted against the DV, but rather the left-out dummy (English) in this case. I believe at least some subtleties can be hinted? – Pr0no May 6 '13 at 23:19

While agreeing with @gung 's comment above, it might be possible to point you to some general ideas, please bear in mind that these aren't complete answers. A good book will help.

1. The t-values are another effect size measure (like the coefficient) but they are on a standard scale, so that, according to some people you can compare them across variables, saying which is most important.

2. Model testing and selection is a huge area. Perhaps, in your field, it is typical to present 6 to 8 models, but it is by no means always done. My view is that model selection should be driven by substantive concerns and by the MAGIC criteria (see the link for what those are).

3. Don't confuse significance with importance and don't confuse either with what should be reported. My view is that you should report results that are interesting.

However, all three questions will elicit somewhat different views from different people.

• I take it from the way you accentuate some people, you are not one of those people? – Pr0no May 6 '13 at 23:20
• That is correct; but there are highly competent statisticians who do (and who don't). – Peter Flom - Reinstate Monica May 7 '13 at 10:19