No, if the test of the interaction is significant (less than 0.5, or whatever criterion you want to use) then that indicates that the slopes are different and are not parallel. ... View more

If you are concerned about normality, you can use a nonparametric test like the Kruskal-Wallis test in PROC NPAR1WAY. And since you'll be doing multiple tests - one test for each combination of continuous response variable and categorical variable -  you might want to consider using a p-value adjustment method to control the overall, family error rate. This can be done by using an adjustment method, like Holm's stepdown Bonferroni method, on the multiple p-values in PROC MULTTEST.   An efficient way to do all of this is to rearrange your data from 200 observations into 200*k observations, where k is the number of your categorical variables. Each block of 200 will have your three continuous variables and a variable containing one of your categorical variables. You can then use BY processing in a single call of PROC NPAR1WAY to do all of the analyses. Use the OUTPUT statement to save the Kruskal-Wallis test results in a data set and then use PROC MULTTEST on that data set to do the p-value adjustment. All of that is illustrated below. The ODS statements are used to avoid displaying the results from the multiple NPAR1WAY runs.  data a; set <your-data>; array x (*) <list-of-your-categorical-variables>; do byvar=1 to dim(x); c=x(byvar); output; end; run; proc sort; by byvar; run; ods exclude all; proc npar1way wilcoxon; by byvar; class c; var <your-continuous-variables>; output out=out wilcoxon; run; ods select all; proc multtest inpvalues(P_KW)=out holm; run; ... View more

If you use a spline effect, you can avoid choosing knot points by using a natural cubic spline. See the example in this note. Note that if the curve on the log odds scale is as you expect, it might look somewhat different on the event probability scale as in the example. When you say you want to find the point after which "X is no longer beneficial," I assume you mean on the probability scale but are thinking of the log odds curve. So, it would be best to plot both curves to help you be certain of what you want.   Consider the probability curve in the example in the note and suppose you wanted to pick the value where the curve begins to drop. You could use the method shown in the Risk Difference section of the note to estimate the sequential one-unit probability differences. Using the tests or confidence intervals of those differences you could select the point at which the difference becomes significantly negative. In that example, you can see that the change from StartVert=7 to 8 is -0.035 and is the first difference to become significant (p=.038). ... View more

It's not clear what you want. What "adjustments" do you mean?  If you mean that you want to adjust for performing multiple tests, then you can do that by saving the table of the comparisons and using an p-value adjustment method in PROC MULTTEST. You can save the comparisons table by adding this ODS OUTPUT statement in the PROC LOGISTIC step:   ods output roccontrastestimate=roccon;   and then do the comparison in MULTTEST using the INPVALUES= option like below.  proc multtest inpvalues(probchisq)=roccon holm; See the example titled "Inputting Raw p-Values" in the MULTTEST documentation. ... View more

The Holm step-down method available in PROC MULTTEST is generally applicable for adjusting a set of p-values. It is powerful, much less conservative than Bonferroni, controls the family-wise error rate, and is not sensitive to dependence that might exist among the tests. ... View more

The point is that you don't need to run PROC GENMOD twice. The TYPE3 test for GROUP is the likelihood ratio test you want. So, you just need the results from your first PROC GENMOD step. ... View more

The main issue is that the linear probability model using the identity link often cannot fit binary response data well. And might not even be successful at all. That is suggested in the Note at the very end of the usage note you refer to. With the identity link, the predicted values can fall outside the valid [0,1] range and fail to converge. This is particularly likely if the data are such that some responses are near 0 or 1. The linear probability model might be okay if all of the responses are in the middle range, such as between 0.25 and 0.75 where the curve of a logistic model is roughly straight. But generally, the logit link is the best way to go to properly deal with binomially distributed data.  ... View more