Use one of the methods discussed and illustrated in this note.  ... View more

To answer briefly: yes, each category is represented in the model by a dummy variable. Your output makes it clear that you used the default dummy variable coding (GLM coding), so the parameter for a given category is interpreted as the difference in effect of that category compared to the reference category, which is the last category. The last category has a zero parameter because its effect is the difference with itself. ... View more

Since you want the clinic odds ratio for each sex and for each race, I assume that you want to allow for the clinic odds ratio to change in each sex and each race.  If so, then you model needs to accommodate that. You could then just use the ODDSRATIO statement to get the clinic odds ratios for each combination of sex and race. But you seem to want them for sex alone and for race alone. To do that you have to be comfortable averaging over the other variable (that is, over race for the odds ratios for sex). You can do that using the SLICE statement. For example, using C, S, R for Clinic, Sex, and Race: proc logistic plots=none; class c s r / param=glm; model disease(event='Yes')=c c*s c*r; oddsratio c; slice c*s / sliceby=s diff oddsratio cl; slice c*r / sliceby=r diff oddsratio cl; run; ... View more

Just add the clinic*race interaction in the MODEL and LSMEANS statements. But you'll notice in the ODDSRATIO results that now the clinic effect depends on both sex and race. If you want clinic odds ratios for sex alone then you can ignore any variability over race. Similarly for the clinic odds ratio for race alone. That can be done using SLICE statements and the ODDSRATIO option. model SecondDose (event = "Yes") = CLINIC_TYPE|RECIP_SEX RECIP_AGE CLINIC_TYPE|RECIP_RACE_ETH; oddsratio CLINIC_TYPE; slice CLINIC_TYPE*RECIP_SEX / sliceby=RECIP_SEX diff cl oddsratio; slice CLINIC_TYPE*RECIP_RACE_ETH / sliceby=RECIP_RACE_ETH diff cl oddsratio; ... View more

As I said, your model does not allow for the odds ratio comparing clinics to change by sex (or race or age). To allow the model to estimate different clinic odds ratios for the two sexes you need to include the clinic*sex interaction as shown in the following statements. The LSMEANS statement provides the estimated event probability for each clinic-sex combination (in the Mean column). The ODDSRATIO statement provides the odds ratio estimates comparing clinics in each sex level. proc logistic data=vaccine; class CLINIC_TYPE (Ref = '1') RECIP_SEX RECIP_RACE_ETH / param=glm; model SecondDose (event = "Yes") = CLINIC_TYPE|RECIP_SEX RECIP_AGE RECIP_RACE_ETH; lsmeans CLINIC_TYPE*RECIP_SEX / ilink cl; oddsratio CLINIC_TYPE; where RECIP_SEX = "M" or RECIP_SEX = "F"; run; ... View more

Before, you had your race variable in the BY statement, not your sex variable. In any case, if you want to obtain one odds ratio estimate for the clinic effect that is adjusted for sex, race, and age, then you need a single model that includes all of those variables; not separate models by sex (or race). The code I showed before provides that. Otherwise, if you have your sex variable in a BY statement, then you'll get two analyses with two estimates of the clinic effect. If you want that, it is better to do it with a single model that includes the clinic*sex interaction, not with a BY statement. With the interaction in the model, you would include an ODDSRATIO statement: oddsratio clinic; ... View more

You definitely don't want to put a variable in the BY statement in the MODEL statement since then it will only have one value and therefore no variability in each BY group analysis. If you want your odds ratios adjusted for sex, race, and age, then put them all in the model and don't use the BY statement. You also don't need to specify the response variable in the CLASS statement. So, these statements should provide a table including an estimate of the odds ratio that compares the two clinics. Note that since you include no interactions in your model, you are assuming that the odds ratio does not change within different levels of the other variables. proc logistic data=vaccine; class CLINIC_TYPE (Ref = '1') RECIP_SEX RECIP_RACE_ETH / param=ref; model SecondDose (event = "Yes") = CLINIC_TYPE RECIP_AGE RECIP_SEX RECIP_RACE_ETH; where RECIP_SEX = "M" or RECIP_SEX = "F"; run; ... View more

See this note. When you are looking for a particular statistic in SAS, check the list of Frequently Asked-for Statistics (FASTats) in the Important Links section of the Statistical Procedures Community page. ... View more

It would help if you confirmed whether your response is a proportion of a binary variable and if you have the numerator and denominator counts., and then exactly what it is that you are trying to compute or estimate. It is not at all clear what "all 3 conditions" are and how that figures into the model. ... View more

I see that your response values are proportions. If these are proportions responding on some binary variable and you have the numerator and denominator counts making up the proportions, then you should be using a logistic model. If the data are not survey data (you aren't using the CLUSTER or STRATA statement), then you just need PROC LOGISTIC to fit the model. You can add an LSMEANS statement with the ILINK option to get the fitted proportions for the STATE*YEAR combinations. For example proc logistic data=work.project3; class state year / param=glm; model num_events/total = state*year / noint; lsmeans state*year / ilink; run; If the model fit is still a problem, as it appears to be in the results you showed, then you will probably need to merge together some levels of STATE and/or YEAR to remove any sparseness in the data.    But if your goal is to make difference in difference (DID) comparisons then you should review the "Generalized Linear Models with a Non-Identity Link" section of this note.  As shown there, you can use Margins macro with appropriate contrasts data set to make the various DID comparisons since you have many levels of both STATE and YEAR. ... View more

The line you quoted is correct - read through the Details section of the Margins macro documentation for more info. Predictive margins do not generally set the values of all of the predictors other than specified in margins=. If you want all of the other predictors fixed, then you want predicted means, not predictive margins. You can get predicted means by simply using the PRED= option in the OUTPUT statement as I showed in my earlier reply. That will give you a data set with predicted means for all of your observations. If you want predicted means for other predictor settings than you have in your input data set, then create a data set with the desired predictor settings and then use the SCORE statement. See the example titled "Scoring data sets" in the LOGISTIC documentation. ... View more

That is correct. You can put in any 2 or more values in the joint_years variable in the MDAT data set and the Margins macro will give all pairwise comparisons among those values. But as I noted earlier, predictive margins are computed differently from individual predicted values which change depending on the values of the other variables in the model. The predictive margin is the average of the predicted values from all of the observations with joint_years fixed at a given value. ... View more

If you want a plot and analysis for each species separately, then you probably just need a simple correspondence analysis as in the first part of the example titled "Simple and Multiple Correspondence Analysis of Automobiles and Their Owners" in the CORRESP documentation. For these data, the plot is just one dimensional. ... View more

See "Linearity in the logit (or link), testing" in the list of Frequently Asked-for Statistics (FASTats) in the Important Links section of the Statistical Procedures Community page. As suggested there, you can use the method shown in the "Nonparametric logistic regression" example in the GAMPL procedure documentation, or the ASSESS statement in GENMOD to check for linearity/adequacy of the specified form. ... View more

The reference in the abstract stating the GEEGOR models "the association of ordered categorical responses within a cluster using the global odds ratio as a measure of association" sounds to me like the Alternating Logistic Regressions (ALR) capability that is available in the GENMOD and GEE procedures. See the discussion of the ALR method in the GENMOD documentation. ... View more