Yes, sorry, I had just been using 1 initially. ... View more

See this note on one-sided testing and using a non-zero null value. For example, you can PROC ORTHOREG which is the modern regression procedure that has the most up-to-date capabilities available, including the LSMESTIMATE statement that allows you to estimate and compare means and to specify a null value other than zero (TESTVALUE=) and one-sided tests (UPPER and LOWER). See the ORTHOREG documentation for details. For example, the following tests whether the difference between the F and A drugs exceeds 4. data DrugTest; input Drug $ PreTreatment PostTreatment @@; datalines; A 11 6 A 8 0 A 5 2 A 14 8 A 19 11 A 6 4 A 10 13 A 6 1 A 11 8 A 3 0 D 6 0 D 6 2 D 7 3 D 8 1 D 18 18 D 8 4 D 19 14 D 8 9 D 5 1 D 15 9 F 16 13 F 13 10 F 11 18 F 9 5 F 21 23 F 16 12 F 12 5 F 12 16 F 7 1 F 12 20 ; proc orthoreg data=DrugTest; class Drug; model PostTreatment = Drug PreTreatment; lsmestimate Drug 'F-A>4' -1 0 1 / testvalue=1 upper; run; ... View more

If observations are matched then they are not independent. Study data can consist of matched sets of any size, not just pairs, with any number of events observed within each set if the response is binary. There are various analytical methods that can be used for binary response data of this type including conditional logistic regression which is available with the STRATA statement in PROC LOGISTIC. Other possibilities include the Generalized Estimating Equations and Alternating Logistic Regressions using the REPEATED statement in PROC GEE or PROC GENMOD and random effect models using the RANDOM statement in PROC GLIMMIX. A non-model-based approach using stratification is available with the CMH option in PROC FREQ when a multi-way table is specified. See the examples in the documentation for each of these procedures. ... View more

For the case of the lognormal, if you use FMM, the predicted values that it generates are on the original scale, but in GLIMMIX they are on the log transformed (normal) scale even though both use the identity link as has been noted. This distinction is pointed out in the description of the lognormal log likelihood in the FMM documentation.   Here is an example.  Again as has been pointed out, if you can use an alternative distribution like gamma or inverse gaussian with a suitable link function, that will avoid the problem by using ILINK in LSMEANS or when getting predicted values. data L; do x=1 to 10; do n=1 to 1000; y=rand('lognormal',x); logy=log(y); output; end; end; run; proc means data=l; class x; var y logy; run; proc fmm data=l; class x; model y=x/dist=lognormal cl; output out=l2 pred=p; run; proc means data=l2; class x; var p ; run; proc glimmix data=l; class x; model y=x/dist=lognormal cl s; lsmeans x/ilink; output out=l3 pred(ilink)=p; run; proc means data=l3; class x; var p ; run; proc glimmix data=l; class x; model y=x/dist=gamma cl s; lsmeans x/ilink; output out=l4 pred(ilink)=p; run; proc means data=l4; class x; var p ; run;   ... View more

If you find evidence of overdispersion, either from a too large Value/DF or from the test using SCALE=0 and NOSCALE with DIST=NEGBIN, then that suggests that the distribution is not Poisson and something else should be used like in the note I referred to. The warnings don't play into it. There is no hard and fast cutoff criterion on the Value/DF.  ... View more

You didn't provide any information about the nature of your response variable, but it sounds like it might be positively- or at least nonnegatively-valued. Such responses could be modeled using a generalized linear model using various distributions. If it is a discrete count response, then you could use the Poisson distribution (or negative binomial if it exhibits extra variability or excessive zeros). If it is continuous and positive, then the gamma, inverse Gaussian, or Tweedie distributions are commonly used. With these distributions, it is common to use the log link function to model the log of the distribution mean.   But yes, if your subjects or items are repeatedly measured, then the implied correlation should be taken into account by using an appropriate model such as a Generalized Estimating Equations model (PROC GEE) or a random effects model (PROC GLIMMIX) depending on the purpose of the model. Both procedures can use distributions like Poisson and gamma. See the examples in the documentation of those procedures. Also see this note on modeling positive, continuous responses and this note on modeling count responses using various distributions (though this is primarily about handling overdispersion). ... View more

You didn't show the results of the model fit, particularly the goodness of fit table and the parameter estimates table showing the model parameter estimates including the negative binomial dispersion parameter. One possible reason for the warning, though certainly not the only one, is if the dispersion parameter is essentially zero which means that the distribution is really Poisson. You can test if there is significant overdispersion as can be caused by extra variability by rerunning your GENMOD step and adding the SCALE=0 NOSCALE options in the MODEL statement as discussed in this note. If the test is not significant, then consider using the Poisson distribution. Another possible problem is that the complexity of the model makes estimating all the parameters difficult. You could try fitting simpler models such as by eliminating some of the interaction terms. In fact you can use a model selection method, such as the Lasso method, by using the HPGENSELECT to find an adequate but simpler model. ... View more

With some work, there might be a way to simplify this down into a single call to PROC SGPANEL. But since you already have code that creates the two plots you want, you can just put them together in a side by side panel using ODS LAYAOUT GRIDDED and ODS REGION statements as illustrated in this note. ... View more

I think the easiest way to do this is to create separate data sets of the significant correlations for each of your two primary variables and then merge them together, keeping the correlations that are on the same WITH variable.  However, you are at risk of the common problem of multiple testing and its effect on your overall error rate. Because you are testing many correlations, presumably each at 0.05 level, the error rate over all of these tests is much higher than 0.05. This can be handled by using a p-value adjustment method. You can use PROC MULTTEST to do this. There are many adjustment methods available, but one that is generally acceptable is Holm's stepdown Bonferroni method which is much less conservative than the simple well-known Bonferroni method. The entire process is illustrated in the example below using the wine cultivar data that can be found in the PROC HPSPLIT documentation in the SAS/STAT User's Guide - see the Getting Started example. In the code below, Alcohol and Malic are treated like your two primary variables. Correlations with adjusted p-values less than 0.05 are kept.    proc corr data=wine fisher spearman; var alcohol malic; with _numeric_; ods output fisherspearmancorr = corr; run; proc multtest inpvalues(pvalue)=corr holm; ods output pvalues=adjp; run; data alc(where=(var='Alcohol' and stepdownbonferroni<.05)) mal(where=(var='Malic' and stepdownbonferroni<.05)); merge corr adjp; run; proc sort data=alc; by withvar; run; proc sort data=mal; by withvar; run; data done; merge alc(rename=(var=var1 corr=corralc raw=pvalalc stepdownbonferroni=adjpalc)) mal(rename=(var=var2 corr=corrmal raw=pvalmal stepdownbonferroni=adjpmal)); by withvar; drop zval--ucl; if var1 ne ' ' and var2 ne ' ' then output; run; proc print; id withvar; var var: corr: adj:; run; ... View more

You might want to take a look at PROC CAUSALGRAPH which can help to create an adjustment set of variables that can allow you to estimate the direct effect from X to Y. CAUSALGRAPH takes a DAG as input. Also look over the CAUSALTRT, CAUSALMED, or PSMATCH procedures to deal with confounding or mediating variables. ... View more

My previous suggestion using SUBJECT=MD, LOGOR=NEST1, and SUBCLUSTER=PATIENT will give you two log odds ratios - one for any pair of responses within a patient and one for any pair of responses from different patients. That uses all the responses from all patients in a physician as a cluster. Another possible analysis is using SUBJECT=PATIENT(MD) and LOGOR=LOGORVAR(MD). This treats all responses from each distinct patient as a cluster and estimates a log odds ratio for each physician essentially pooling across all his/her patients. ... View more

Questions like this that are about statistical methods or statistical procedures will be addressed faster and get more attention if you post them in the Analytics>Statistical Procedures community.   First, note that the GEE method is robust to not specifying exactly the right clustering structure, so it is not unreasonable the use the GEE model from the code you showed in your first post.   However, if you particularly want to use an ALR model to estimate log odds ratios, and if your data consists of patients clustered within physicians and with multiple observations from patients, then you need to change your REPEATED statement options. Instead of TYPE=EXCH, which requests a GEE model, specify SUBJECT=MD (I assume that MD is your physician indicator) then specify LOGOR=NEST1 and SUBCLUSTER=PATIENT: repeated subject=md / logor=nest1 subcluster=patient; While LOGOR=NESTK (with SUBCLUSTER=PATIENT) or FULLCLUST are other possible structures, they are probably not feasible since you indicate that, on average, there are 100 patients per physician. These structures would probably require the estimation far too many log odds ratios, and in any case would probably not be useful. ... View more