About SteveDenham

SteveDenham · ‎05-24-2024

We all tend to see different things, but it appears that the absolute value of the residuals increase with increasing predicted values. You might consider that to be evidence of heteroskedasticity, but it might also be the result of model misspecification, improper scaling of design variables, collinearity (if this is a multiple regression) or not recognizing that whatever process generates the data is one where scale is a function of location, such as a gamma distribution. SteveDenham

SteveDenham · ‎05-24-2024

you might want to search Mr. Google for repeated AB testing, and see how others have analyzed their results. SteveDenham

SteveDenham · ‎05-24-2024

I was able to run your model without errors, but there are some results that make me wonder if your model could be over specified. Here is the code I used: PROC GLIMMIX DATA = test1/*(where=(stratclus<999))*/ METHOD=QUAD(FASTQUAD QPOINTS=20) plots=all; CLASS RAHHIDPN STRATCLUS GENDER(REF='1') RACE_ETH(ref='1') MARITAL(ref='1'); PARMS (0.01) (1.72) (0.5); MODEL MOBILITY = BSAGE_C BSAGE_C*BSAGE_C GENDER GENDER*BSAGE_C RACE_ETH RACE_ETH*BSAGE_C RACE_ETH*GENDER TIME TIME*TIME TIME*BSAGE_C TIME*GENDER TIME*RACE_ETH/solution obsweight=HCNSWGTR ;/**LEVEL 1 VARIABLES**/ RANDOM INTERCEPT /subject=STRATCLUS type=VC ; RANDOM INTERCEPT /subject=RAHHIDPN(STRATCLUS) type=VC ; NLOPTIONS GCONV=0.00001 tech=congra; COVTEST/WALD; ODS OUTPUT parameterestimates=FIXED1 covparms=RAND1; RUN; The two changes were to reduce the number of quadrature points, and to use the PARMS statement to provide some better starting values for the random effects. Regarding my thought of over-specification, the F values for five of the terms in the model went to infinity. These were GENDER, RACE_ETH*GENDER, TIME*TIME, TIME*RSAGE_C and TIME*GENDER. You may want to look to see if there is something unusual with the strata and the number of observations available in the various cells. SteveDenham

SteveDenham · ‎05-08-2024

Sometimes Mr. Google and Ms. Wikipedia can lead you to an interesting result: In SAS, PROC ARIMA can perform ADF tests So, do you have access to SAS/ETS? If so, this is not a difficult request. However, if your data are not from a time series (or something resembling a time series), I wonder if the effort to come up with an approximate ADF test is worth it, considering that many of the assumptions for the test are not being met. If you are looking for a test for autocorrelation of the residuals in PROC REG, then you can get the Durbin-Watson statistic, but that really only tests for a lag of 1, and requires that the response variable be ordered in some sense (time, location, etc.). SteveDenham

SteveDenham · ‎05-08-2024

Further, how can a duration be binomially distributed? Unless you are defining an incidence as a duration, such as behavior not seen --> duration=0, behavior seen -> 1. But your issue is with the first analysis, and the problem is that the log of 0 is negative infinity, so folks say you can't analyze the results when you have a zero and a log link. There are alternatives, and you are part way there by picking a multiplicative overdispersion. Consider adding a small (say 1e-6) to every observation, and fitting a gamma distribution (although you can still fit a Poisson, let's just move to the continuous version rather than staying with the discrete version). That should address your model. It is harder to compare means except on the log scale. The ILINK option gives a ratio for the exponentiated difference of two means on the log scale, not the difference on the measured scale of two exponentiated estimates. I don't know if the %NLest macro would be appropriate here, but it certainly is appropriate for the Poisson distribution example. Perhaps @StatDave could chime in here on that approach. SteveDenham

SteveDenham · ‎05-08-2024

The syntax would be the same as for the above design, so I will repeat that code, and then insert your factors. /* previous poster */ PROC GLIMMIX data=data; class site origin block cut tree; model recovery = site|origin|cut/ddfm=kr; random block block*site block*origin block*cut block*site*origin block*site*cut block*origin*cut block*origin*site*cut;; /* change variable names to organic, system, nitrogen */ PROC GLIMMIX data=data; class block organic system nitrogen; model recovery = organic|system|nitrogen/ddfm=kr2; random block block*organic block*system block*nitrogen block*organic*system block*organic*nitrogen block*organic*system*nitrogen; Again, I want to point out that you probably do not have enough data to estimate the three-way and four-way effects, and they probably should be considered residual error in any case. To improve the ability to converge, you may then want this for the RANDOM statement: random intercept organic system nitrogen/subject=block; SteveDenham

SteveDenham · ‎05-07-2024

Use an ODS OUTPUT statement to get the ESTIMATE value into a dataset, and then postprocess that dataset, comparing the result of the ESTIMATE statement to -1. This might be easier if you use LSMESTIMATE as you don"t need all of the terms that refer to anything except the lsmeans for visit*baseline. SteveDenham

SteveDenham · ‎05-07-2024

You may need to simplify your model, as I see 3 variables that are functions of one another: Chg = observed value - base Base Strata = median function determined by Base, so the induced collinearity between base and strata may well be driving the similarity of the estimates between the strata groups. What happens if you remove strata from the model, and (perhaps) post-process the predicted values obtained from an OUTP= option in your MODEL statement into quintiles, terciles or two groups? Does that address your research question? SteveDenham

SteveDenham · ‎05-07-2024

If you arrange your data in long format, with each record as bonus and then area, the following code is a start on the Bayesian t test in PROC MCMC: proc mcmc data=your_data outpost=postout seed=123 nmc=40000 monitor=(_parms_ mudif) statistics(alpha=0.01); ods select PostSumInt; /* This may need to be commented out if you want all the diagnostics, etc. */ parm mu0 0 mu1 0; parm sig20 1; parm sig21 1; prior mu: ~ general(0); prior sig20 ~ general(-log(sig20), lower=0); prior sig21 ~ general(-log(sig21), lower=0); mudif = mu0 - mu1; if bonus = 1 then do; mu = mu1; s2 = sig21; end; else do; mu = mu0; s2 = sig20; end; model area ~ normal(mu, var=s2); run; In this example, mu0 and sig20 are the mean and variance for bonus=0 and mu1 and sig21 are the mean and variance for bonus=1. Now as far as how MCMC handles missing data: There is a section in the Details: MCMC Procedure folder for PROC MCMC documentation that covers everything about missing data. Key point 1 - there was a big change in how MCMC handles missing data implemented for all versions after SAS/STAT 9.3. In the earlier versions, MCMC could only model complete cases. Now there are at least four different methods, with each appropriate for various types of missing data models. For MNAR data, the suggested approach is to fit more than one MODEL statement depending on whether you fit a conditional or marginal model to the response and the opposite type of model to the missingness vector. I hope this is more in line with the question you posed. SteveDenham

SteveDenham · ‎05-06-2024

I think this may be as simple as defining an additional indicator variable (say type as one of the following RCTPre_Rating, RCBPre_Rating, RCSPre_Rating, and putting the value of each into a single response variable in a long format dataset). You could then fit: model response= type OB|AT Teachertype|YOE INLevel INDisSchool INLOE PerDis FamDis EthnicitySel Race Gender Ableismselfassess ADCTO Vtrialorder / solution ddfm=bw; If you have sufficient data, you could even include some interactions of type with the other covariates. SteveDenham

SteveDenham · ‎05-06-2024

Would you care to share your MCMC code, and the dataset in sas7bdat format? Additionally, there are at least two other PROCs that are capable of providing a Bayesian analysis of this sort - GENMOD and BGLIMM. Also, the documentation for the Behrens-Fisher problem in the MCMC Getting Started may be of assistance. For me, trying MCMC as a first step was a bit like trying to outrun Usain Bolt when I was just learning to walk. SteveDemja,

SteveDenham · ‎05-06-2024

The GLIMMIX documentation goes through several caveats regarding the ordering of the categories for a nominal multinomial analysis. In the end, I think their approach would look like this when translated to your situation: proc glimmix data=lab.merge INITGLM; class occ_reduced region(ref='SK') ; model occ_reduced(order= freq ref='LOW') = diffH diffM diffL region / solution DIST=MULTINOMIAL link=glogit; where labour=1; weight weight; run; Quoting the documentation: " In generalized logit models (for multinomial data with unordered categories), one response category is chosen as the reference category in the formulation of the generalized logits. By default, the linear predictor in the reference category is set to 0, and the reference category corresponds to the entry in the "Response Profile" table with the highest Ordered Value. You can affect the assignment of Ordered Values with the DESCENDING and ORDER= options in the MODEL statement. You can choose a different reference category with the REF= option. The choice of the reference category for generalized logit models affects the results. It is sometimes recommended that you choose the category with the highest frequency as the reference (see, for example, Brown and Prescott 1999, p. 160). You can achieve this with the GLIMMIX procedure by combining the ORDER= and REF= options..." I don't know if that will help or not, but my experience with generalized logit models is that critical clues to its function in GLIMMIX aren't stressed enough in the examples. SteveDenham

SteveDenham · ‎05-06-2024

I would caution you about any inferences regarding drug or treatment. The design I see completely confounds drug*treatment with pen, and for a lot of folks that would mean that pen is the experimental unit, and animals are observational units measured on each of the pens. You might be able to use the following, but I provide no guarantees. I moved to the spatial power covariance structure because I believe the residuals will be correlated with those closer in time more highly correlated. The autoregressive error structures (including ANTE and ARMA) treat the time points as being equally spaced in time. I also fit pen as an effect, rather than trt or drug. You can get LSMEAN-like estimates for the various levels of each by using LSMESTIMATE statements to manipulate the pen lsmeans. I added the /e option to the LSMEANS statement to facilitate construction of the LSMESTIMATES: proc mixed data= XX; class animal pen day; model BW = pen|day initial_BW ddfm=kr; random intercept / subject = animal; repeated days / subject = animal type = sp(pow)(t); /* t is a continuous version of 'day' */ lsmeans pen|day /e adjust = tukey; /* Add a section of LSMESTIMATE code here to calculate trt and drug means and differences */ run; SteveDenham

SteveDenham · ‎05-06-2024

If you have 20G observations, and have DEN F coming out as Inf, think about what that means - you are out at the point where F tests are indistinguishable from chi squared tests, and the t distribution doesn't differ from the normal until you get out to several decimal places. So, rather than trying DDFM= basically anything, try using the CHISQ option, and use those results rather than the F tests. If you ran the analysis in GLIMMIX, you could use ddfm=none. SteveDenham

SteveDenham · ‎04-04-2024

I am not sure what is going to happen with this code, but it might trigger something on your side. What happens if you change the RANDOM _residual_ to RANDOM _residual_/group=sex ? I would expect to see two estimates for residual error in that case. To actually see them and their confidence bounds you need to a COVTEST statement. In this case it would look like: covtest homogeneity /cl; I don't see a good method in GLIMMIX to get the SE for the residual variance, but that doesn't mean there isn't one. GENMOD does give an SE for the scale parameter in a Poisson regression, that may or may not have an SE>0, depending on the data. 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: Heteroskedasticity

Re: Nonparametric ANOVA F (MIVQUE0) for neurosciense

Re: PROC GLIMMIX ERROR: QUANEW Optimization cannot be completed.

Re: Augmented Dickey Fuller Test with AIC and BIC option

Re: Most suitable model: GLIMMIX or something else?

Re: How to analyze RCBD with 3 factors

Re: Specify non-inferiority margin using mixed model for repeated meas...

Re: How to test the interaction between strata and treatment groups in...

Re: Bayesian TTest by Proc MCMC

Re: Multivariate Multilevel model using proc mixed

Re: Bayesian TTest by Proc MCMC

Re: Unexpected estimates for ref category with proc GLIMMIX

Re: Help Random Effect and Multiple Comparison Analysis

Re: What parameter to use for DDFM= in proc mixed?

Re: No standard error for the overdispersion factor in a simple Poisso...

SAS Analytics Explorers