These are now proportions that you could model with a logistic model and then similarly compute the percent change and confidence interval using NLMeans or directly with NLMIXED. If you just need the difference and its confidence interval, then this can easily be done in PROC FREQ without a fitting a model. data a; input year ncars ntot; y=1; count=ncars; output; y=0; count=ntot-ncars; output; datalines; 2012 89 3000 2022 111 5000 ; proc freq; weight count; table year*y / riskdiff; run; proc genmod data=a; class year; model ncars/ntot=year/dist=bin; lsmeans year / ilink e plots=none; ods output coef=c; store mod; run; %nlmeans(instore=mod, coef=c, link=logit, f=100*(mu2-mu1)/mu1, flabel=Pct Change, title=Percent change) ... View more

Actually, this can be done with just a single run of PROC NLMIXED rather than using the macros. But I find that people find it easier to work with a familiar, more specialized procedure, like GENMOD than with NLMIXED. The following provides essentially the same result as from GENMOD and NLMeans. proc nlmixed data=a df=1e8; mu=exp(b1*(year=1) + b2*(year=2)); model count ~ poisson(mu); estimate 'pct chg' ( 100*(exp(b2)-exp(b1))/ exp(b1) ); run; ... View more

I do not know that the tree-based method implemented in PROC HPSPLIT has any provision to specifically handle under- or oversampling of the response. However, this issue often comes up when using logistic modeling of a binary response and there are methods discussed in this note for dealing with that. While not designed for the tree-based method, you could consider using the weighting method presented in the note since HPSPLIT does support a WEIGHT statement. ... View more

I used the factorial example since it is the more complex scenario. The case of 4 distinct treatments is a simpler subcase where the model has only the treatment effect, just like the second model with AB. So, yes, that is still the model you could use for your ordinal response in the case of having 4 distinct, nonfactorial treatments. But since you want to estimate the proportions in response levels, it is simplest to treat the response as nominal rather than ordinal by using the generalized logit model (LINK=GLOGIT option). But the GLOGIT model (like the full non-proportional odds model) requires the estimation of many additional model parameters and this can cause estimation problems when the data are sparse, like these data (note the zero counts in the FREQ table), causing "separation" errors due to some model parameters trying to be infinite as happens in this case.   Nonetheless, suppose that the AB combinations are nonfactorial treatments. The following refits the AB model from before but using the GLOGIT model. The LSMEANS statement now produces the estimates of the individual response levels for each treatment. Despite the separation issues, you'll notice that the estimated proportions are mostly quite close to the observed proportions. So, for purposes of prediction, this works well but I would caution against trusting any tests since many parameters are very unstable as evidenced by the quite large standard errors.   proc logistic data=x; class ab / param=glm; model score(desc)=ab rep / link=glogit; lsmeans ab / ilink plots=none; run; proc freq data=x; table ab*score / nocol nopercent; run;     ... View more

As I showed, the model using TRT (or my AB variable which is the same thing), which is sometimes called the "mean model," is equivalent to the model with A, B, and A*B. The former simply combines the degrees of freedom (DF) into a single effect - there is 1 DF each for A and B and also 1 for the interaction, AB. TRT (or AB) just combine those 3 DF into a single effect. The proof that the models are equivalent, just reparameterized, is the identical likelihood values. The mean model is just more convenient for making comparisons among the 4 combinations (the same could be done with the A B AB model - it's just less obvious how and really messes people up).    Yes, for these data, the proportional odds test is significant. This test is known to be liberal, meaning that it shows significance more often than it should. But you can always investigate that assumption using the graphical and analytical methods described in this note . If you decide that the assumption really is not supported, then you can fit a fully or partially nonproportional odds model as shown in the note. Alternatively, you could ignore the ordinality of the response and fit a nominal multinomial model by adding the LINK=GLOGIT option in the MODEL statement. The nonproportional odds models or the nominal models just have more parameters to estimate, but the same general approach to your problem can be done.  ... View more

From what you describe, your response variable is ordinal. Since it is integer valued, it is probably not reasonable to assume normality and use GLM. You can use PROC LOGISTIC to model an ordinal multinomial response like yours.  Using the data you provided, these statements fit the model allowing for interaction in the factors. It simultaneously  models the cumulative probabilities - Pr(score=3), Pr(score=2 or 3), Pr(score=1,2, or 3). This ordering and grouping of level is the result of the DESC response variable option. The ILINK option in the LSMEANS  statement provides the estimated cumulative probabilities for each A-B combination (in the Mean column). In the results you'll see that there is a significant interaction. As with any model, the interaction means that the effect of one factor changes with the level of the other factor. So, it is necessary to assess the effect of each factor at each level of the other factor. The effect of REP is not significant and could possibly be dropped from the model but is left in for the following. proc logistic data=x; class factora factorb / param=glm; model score(desc)=factora|factorb rep; lsmeans factora*factorb/ilink plots=none; store mod2; run; You can visually plot the interaction effect using the fitted and stored model (in MOD2) and then using the EFFECTPLOT statement in PROC PLM to produce the plot. The following statements show the interaction two different but equivalent ways. Note that there is a plot for each of the three cumulative probabilities. proc plm source=mod2; effectplot interaction(x=factorb sliceby=factora); effectplot interaction(x=factora sliceby=factorb); run; To assess the interaction, it is reasonable to use the approach you mentioned which is to include just the 4 level TRT variable in the model and then make suitable comparisons among the factor combinations. That is done by the following, but a better labeled variable, AB, is used instead of TRT. data x; set x; ab=cats(factora,factorb); run; proc logistic data=x; class ab / param=glm; model score(desc)=ab rep; lsmeans ab / ilink plots=none; store mod; run; Note that the model is equivalent to the first one - same likelihood value, same estimated cumulative probabilities. Using the order of the factor combinations shown in the LSMEANS table, you can use LSMESTIMATE statements to easily test the effects within the interaction as described above.  proc plm source=mod; effectplot interaction(x=ab sliceby=score); lsmestimate ab 'A1-A2 in Yes' 0 1 0 -1 / joint e; lsmestimate ab 'A1-A2 in No' 1 0 -1 0 / joint e; lsmestimate ab 'diff in diff' -1 1 1 -1 / joint e; *A1-A2 in Yes - A1-A2 in No; run; The EFFECTPLOT statement shows the cumulative probabilities for each of the 4 factor combinations. The first two LSMESTIMATE statements test the effect of A in each level of B. The last LSMESTIMATE statement tests the so-called "difference in difference" effect which is the difference of the two differences in the first two statements. This is, in fact, the interaction effect and you'll see that it has the same p-value as from the first model fit. The results form the first two statements show that the effect of A is significant in B=Yes but not in B=No. You can easily see that best in the plot, but also in the interaction plots earlier. ... View more

You seem to be describing the Peto mortality-prevalence test that I assume you have seen since you specifically mention PROC MULTTEST which offers this test. See the example of using this test in the Examples section of the MULTTEST documentation. ... View more

If you want smooth confidence lines as well as the fitted line on the hazard ratios, you can slightly modify the code in my last post to compute the confidence limits and plot them as shown below. Note that your model is fitting a spline curve to the log hazards, not to the hazard ratios, so a smoothed curve on the hazard ratios doesn't come from the model. Hence the use of the PBSPLINE statement. data ref; set Myeloma; protein=10; time=.; run; data both; set Myeloma ref(in=inref); ref=inref; run; proc phreg data=both; effect spl=spline(protein/naturalcubic); model Time*VStatus(0)=spl ; output out=out xbeta=xb stdxbeta=sexb; run; data xb; set out; where ref=0; lhz=xb; keep lhz protein lhz sexb; run; data xbref; set out; where ref=1; lhzref=xb; keep lhzref; run; data hr; merge xb xbref; hr=exp(lhz-lhzref); lower=exp(lhz-lhzref-1.96*sexb); upper=exp(lhz-lhzref+1.96*sexb); run; proc sort data=hr; by protein; run; proc sgplot data=hr noautolegend; pbspline y=upper x=protein /nomarkers; pbspline y=lower x=protein /nomarkers; pbspline y=hr x=protein; refline 1 / axis=y; yaxis label='HR'; run; ... View more

As I mentioned, the ILINK option applies to the individual means which the SLICE statement (unlike the LSMEANS statement) does not provide unless you also specify the MEANS option. It depends on what you want to do, but I would think that once you have a model that you feel fits the data reasonably, it is that model that provides all of the inferences about mean estimates and differences, so it is the model-based results you would show that correspond to those estimates and inferences. ... View more

Yes, you can use the LSMEANS, SLICE, or LSMESTIMATE statement to make comparisons. With the ILINK option these will provide estimates of the Poisson mean for the individual levels. With the EXP option, the table of differences provides estimates of the ratios of Poisson means comparing two levels at a time. ... View more

Did you also download the latest version of the NLEST macro? NLEST is called by NLMeans so you will need the latest version of that too. A link to NLEST is in the NLMeans documentation. If you still get an error after updating both, attach the section of your log showing your NLMeans call and any messages that follow. ... View more

It's always best to avoid CONTRAST or ESTIMATE statements when the LSMEANS, SLICE, or LSMESTIMATE statement can be used. This avoids having to come up with appropriate and estimable coefficients needed in the former statements. The following statement will estimate and compare the means among the Location*Treatment1 combinations: lsmeans location*treatment_1 / diff e; As with your CONTRAST statements, the E option shows the coefficients that LSMEANS uses to compute each mean. While it's possible that some will be nonestimable, depending on your data structure, you could try options like OM and BYLEVEL to make changes to the coefficients. With the list of LS-means that are produced, if you want to make some comparison among sets of them, you could either use a SLICE statement or the LSMESTIMATE statement. For example, if you want the effect of treatment1 in each location: slice location*treatment1 / sliceby=location; For comparisons other than pairwise comparisons as done by the DIFF option in LSMEANS, you can use the LSMESTIMATE statement to much more simply compare, say, groups of LS-means than can be done with a CONSTRAST or ESTIMATE statement.   ... View more