About SteveDenham

SteveDenham · ‎06-16-2022

Friedman's occurred to me too @PaigeMiller , but two of the assumptions for Friedman's test are that there is no interaction between the two factors, and that the levels of both factors are at least uncorrelated. However, I think this is a repeated measures design (whether with 3 timepoints or 2). It also looks like the range of values is bounded above by 100, which implies that the response variable might be some sort of percentage-like value. If that is the case, I would try the Stroup and Claassen approach for a GLMM: proc glimmix data=yourdata_IN_LONG_FORMAT; resp = response/100; class group time pid; model resp =group time group*time/dist=binom; random time/residual subject=pid type=chol; lsmeans group|time/diff ilink; run; By long format I mean that for each combination of group, time and pid there is a single record in the dataset. To compare the "percentages" on the original scale, you'll probably need to invoke the %NLmeans macro (which in turn calls the %NLest macro). See appropriate posts by @StatDave which point to the note where the most recent versions of these can be downloaded, and a lot of really good and helpful tips show up. SteveDenham

SteveDenham · ‎06-16-2022

I am glad the Z score method worked for you, as it makes my original guess as to the issue less likely. That guess is that you have overspecified the model, such that there is enough collinearity that the SWEEP operator ends up with no variability "remaining" for those terms. Is that at all likely? SteveDenham

SteveDenham · ‎06-16-2022

Buried down in the fine print of the documentation are some equations for putting the estimates in the lognormal space into the normal space. Note that @Rick_SAS 's method of exponentiating results gives the geometric mean, which likely the best estimate in the original space. However, it is not the expected value - to get that you need: omega = exp(sigma*sigma) expected value of Y = exp(mu)*sqrt(omega) Variance of Y = exp(2*mu)*omega8(omega - 1) Note that these values are the expected values, which in a right skewed distribution (like a lognormal) may be much larger than the exp(mu), which should be approximately the value that gives the peak of the distribution. SteveDenhm

SteveDenham · ‎06-16-2022

I am not quite sure I understand the question. By "post hoc power", do you mean that you have run an analysis and now wish to calculate the power of that analysis? The power in that case is either 0 or 1, depending on what cutoff is applied to your resultant p value - you either did not find a significant difference or you did find one. With a sample of 1 (=number of analyses), the only possible values are 0 or 1. However, a bootstrap approach may give you what you want. Generate 5000 (or some other fairly large number) potential outcomes based on the results you currently have by sampling with replacement. Then repeat the analysis for each of those potential outcomes, saving the p value of interest. The proportion of samples which yield greater p values is a fairly good estimate of beta, so the power would be 1 - beta. SteveDenham

SteveDenham · ‎06-16-2022

I suspect that the Hessian matrix is nearly singular - the zero count for that date isn't handled well by the log link. I would be very tempted to remove the 7/16/21 records, if in fact there are nothing but zeroes observed then. Also, since you have no RANDOM effects in your model, you may wish to do the analysis in PROC GENMOD, where you at least have the option of fitting a zero-inflated model. SteveDenham

SteveDenham · ‎06-02-2022

Here are some things to consider: It appears that your response variable (Distance? Job?) is multinomial, as is the predictor. That means that PROC MIXED is probably the wrong method for any analysis, as it assumes that the distribution of errors is normal. You may want to look at other mixed model procedures (GLIMMIX), or generalized linear model/estimating equation procedures (GENMOD, GEE). What is the role of ID1 and ID2? Is there ever a case where the values for these two variables are identical? Are there more than 4 levels? Would you consider these as predictors? If so, the association between the two could be measured by including an interaction term in the model. If not predictors, would you consider them random effects, such as blocks? Since ID levels B and C appear in both ID variables, this may lead to an inability to estimate the random effects unless you have a lot of levels and observations per level. If there is only a small number of levels, you might be better off considering them fixed effects, in which case you might not need a mixed model at all. SteveDenham

SteveDenham · ‎05-31-2022

There is a difference between obtaining a negative estimate for a parameter in the log space, and having a negative number as a result of calculating difference in difference. The former may arise if there is a decrease in the response variable considered as a whole within a parameter of interest (say age group), where you may get a smaller difference in difference for an older age cohort. My concern was what happens when a measurement for a subject is negative. In that case, trying to use a log link will result in a missing value for that subject as the log of a negative number cannot be calculated. As far as the other (exponentiating the estimates) question, you can do that through the use of an LSMESTIMATE statement and the ILINK option. SteveDenham

SteveDenham · ‎05-19-2022

You might be able to get what you want by doing a restricted mean analysis. Here is a link to an example: https://documentation.sas.com/doc/en/statug/15.2/statug_lifetest_examples05.htm#:~:text=The%20following%20statements%20use%20PROC%20LIFETEST%20to%20perform,time%20SurvTime%2ACensor%281%29%3B%20strata%20Cell%3B%20run%3B%20ods%20graphics%20off%3B Depending on which level of SAS/STAT you have, you could consider moving to PROC RMSTREG, which has an LSMEANS statement. If it supports the DIFF and CL option, and especially the AT= option, you are in good shape. See this example for a first step: https://documentation.sas.com/doc/en/statug/15.2/statug_rmstreg_examples02.htm and this for a more complete analysis https://documentation.sas.com/doc/en/statug/15.2/statug_rmstreg_examples03.htm SteveDenham.

SteveDenham · ‎05-19-2022

Well, I think you have a handle on it using GLIMMIX. Check out @Rick_SAS 's blog post on ways to obtain the Hessian https://blogs.sas.com/content/iml/2019/02/13/hessian-maximum-likelihood-covariance.html#:~:text=Some%20regression%20procedures%20support%20the%20COVB%20option%20%28%22covariance,the%20PROC%20NLMIXED%20statement%20to%20display%20the%20Hessian. If you have a different endpoint (continuous with normal errors), I think you can use a STORE statement in PROC MIXED, and then use the SHOW statement in PROC PLM to get the Hessian. I don't see an equivalence for the gradient, though. SteveDenham

SteveDenham · ‎05-19-2022

A bit more information would help. Which survival analysis PROC are you using? All of them support a TEST statement, and most support the LSMEANS and LSMESTIMATE statement. SteveDenham

SteveDenham · ‎05-19-2022

I agree. If you do not have pre-planned comparisons, some form of adjustment for multiplicity should be done to control Type I error. One of the advantages of Bayesian approaches is that since the results do not have distributional or hypothesis assumptions and you can specify your utility (misclassification), you probably shouldn't be doing adjustments for multiplicity (per F. Harrell). SteveDenham

SteveDenham · ‎05-18-2022

Provided the a priori hypotheses are clearly stated, the only p values you need are those for the comparisons. The stagewise procedure of looking at the Type III p value as a gateway to post hoc comparisons isn't needed. SteveDenham PS. A Bayesian analysis would make this more apparent.

SteveDenham · ‎05-16-2022

One issue with DID designs is that the value may be negative, which would mean that you are going to drop all of those values due to the log link. You may be better off modeling the actual values, with the baseline value as a covariate. SteveDenham

SteveDenham · ‎05-16-2022

The output reflects the model. You have age pathogen_new diarrhea bloodydiarr fever as your right hand side variables, and that is what shows in your output. Until you add "specimen" to the model, it will not appear in the output. I suspect that fever is confounded completely with at least one other variable (this looks like a good question for @StatDave ). SteveDenham

SteveDenham · ‎05-10-2022

@Brendan42 - these are great replies, and really cover everything that I thought of. At this point, the only difference between the literature approach and your code is the PARMS vector. In stage 1, the lit approach uses .0.1, with #3 and #5 bounded below by zero, and in stage 2, fixing 17 parameters to 1. I assume your code is for stage 1, and until we get that to behave, there really isn't much sense in trying stage 2. The first thing I noticed is that you are applying a hold to 42 parameters, but the list has 20 parameters bounded below by 0, 20 parameters with some bounded below and others allowed to vary, and 1 other that I can't attribute bounded below by 0. The last of the 42 must be the residual variance and is not included in the code (yes/no ?). Is there anything in the Macholdt paper that explains the values that they plug into the stage 1 parms statement? Note also that since they are using PROC MIXED, they are using a REML approach, rather than the maximum likelihood method you specify. That could be saving them from some of the issues you are having (and maybe not, but it seems like something to try). SteveDenham

Online Status	Offline
Date Last Visited	‎03-19-2026 03:00 PM

Re: Assessing Variable Redundancy for Mixed Effects Modeling

Re: Assessing Variable Redundancy for Mixed Effects Modeling

Re: Randomized block design and meaning of LSMEAN/STDERR

Re: Help with Restricted Cubic Splines : Code Optimization and Graphic...

Re: PROC POWER for Cox regression

Re: Repeated measures model executes in MIXED but not in GLIMMIX

Re: Model heteroscedasticity directly or use log transformation

Re: Model heteroscedasticity directly or use log transformation

Re: Repeated measures model executes in MIXED but not in GLIMMIX

Re: Repeated measures model executes in MIXED but not in GLIMMIX

Re: Passing TESTVALUE option in LSMESTIMATE statement in glimmix

Re: What is the "estimate" in the SolutionR output of the proc mixed. ...

Re: Assessing Variable Redundancy for Mixed Effects Modeling

Re: SAS OnDemand Outpm option in Proc Mixed

Re: question about lsmean pdiff=;option in proc glm step, st102d03, SA...

Re: Proc mixed, defining data structure for desired comparison (Random...

Re: Assessing Variable Redundancy for Mixed Effects Modeling

Re: Assessing Variable Redundancy for Mixed Effects Modeling

Re: Help with Restricted Cubic Splines : Code Optimization and Graphic...

Re: Repeated measures model executes in MIXED but not in GLIMMIX

Re: Is there a non-parametric equivalent of a two-way ANOVA? how to c...

Re: PROC MIXED - Infinity t-values in moderated growth curve model

Re: Interpreting proc glimmix results on different dist

Re: How to do the Post hoc power analysis for cohort study in SAS?

Re: Why does zero variance in raw data have the greatest variance in e...

Re: How to create a mixed effects model with PROC MIXED with pairwise ...

Re: Am I using the RANDOM statement properly? (PROC GLIMMIX, event stu...

Re: how to get the confidence interval of difference in the survival r...

Re: Hessian and gradient in proc mixed

Re: how to get the confidence interval of difference in the survival r...

Re: Should I report type III fixed effects stats and a priori contrast...

Re: Should I report type III fixed effects stats and a priori contrast...

Re: Am I using the RANDOM statement properly? (PROC GLIMMIX, event stu...

Re: GENMOD dropping variable in parameter estimates table

Re: How do we include a linear g-side variance function in mixed/glimm...

SAS Analytics Explorers