I'm looking for studies presenting in tables results of Generalized Regression Models (GRM's). I ran many models with two dependent variables each and I want to summarize the results from all models in a single table, but I'm a bit confused about the presentation of the statistics. Is it similar to Anova tables? Which statistical information can or can't be omitted?
I'd prefer confidence intervals over standard erors. And giving the AIC and/or logLikelihood for the complete model would not be bad.
Present the coeficient estimates and the standard errors.
From that anyone cam calculate intervals and tests
I'd prefer confidence intervals over standard erors. And giving the AIC and/or logLikelihood for the complete model would not be bad.
In the software I used (Statistica) the AIC and log Likelihood are not given as results. For the whole model(s) I have the R² and a typical Anova table of results (for each dependent variable). Besides, I was thinking in informing the summary of stepwise regressions (last step), in order to let the reader know the variables that did not affect both of dependent variables. In this latter case, I have the F-ratio and P values for each variable excluded (and included) from the models. Do you think this information would be redundant? Thanks for your help.
I don't know. I think that is is not very informative to have any of these values (F and P) for the variables excluded in the last step. Automated selection of a "best" regression model is always problematic. One should always better argue with expertise why a variable is used and why not. But if you have no expertise and simply want to follow "some rule" to define the "best" model, then I think it is sufficient to show the final model. After any model selection procedure you need to take care that statistics about this model are not valid anymore (the model will be biased, overestimating the effects of the remaining variables). I don't sa that this is bad or a problem. It only should be taken care of.
I have also ran manual model selection in R too, in order to understand a bit of the automated selection of many software packages. Similar final models were reached in both procedures. I have basically the same five/four variables tested for diferent temporal scales and dependent variables, so automated procedures reach a final model in few steps. It seems that models with too much variables are more problematic because the selection need more steps, and variables are prone to multicollinearity.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA