About Dale

Dale · ‎11-19-2009

I responded to this thread on the SAS-L listserv. Rather than repeating that here, I'll just post a link to my response on SAS-L. http://listserv.uga.edu/cgi-bin/wa?A2=ind0911c&L=sas-l&F=&S=&P=37753 Dale

Dale · ‎11-16-2009

Ah, I like a problem with a simple solution. Sort the data by TREE prior to invoking the NLMIXED procedure. You will then get the results reported by Serroyen et al. Below is a note from the NLMIXED RANDOM statement documentation which discusses this requirement: NOTE: The input data set must be clustered according to the SUBJECT= variable. One easy way to accomplish this is to sort your data by the SUBJECT= variable prior to calling PROC NLMIXED. PROC NLMIXED does not sort the input data set for you; rather, it processes the data sequentially and considers an observation to be from a new subject whenever the value of its SUBJECT= changes from the previous observation.

Dale · ‎11-04-2009

Instead of testing a set of parameters using the CONTRAST statement as indicated by AA1973, you could construct a likelihood ratio test. The likelihood ratio test requires that you fit two models, one with the three psychological variables plus the anthropometric variables and a second model which has the anthropometric variables only. You must specify METHOD=ML when fitting each of these models. You can then take the difference in -2LL values returned by the two models and compare that difference to a table of the chi-square distribution with 3 df. If the difference is larger than the tabled value then the set psychological variables is significant at the specified alpha level.

Dale · ‎10-29-2009

KAZ, Try Googling "+pedigree +covariance +matrix +proc +mixed" (without the quotes). The first hit should be "Genetic analysis of complex traits using SAS - Google Books Result". If you click on that, you can see a couple of pages of text in which the INBREED procedure is used to construct a kinship matrix. The kinship matrix is then passed to the MIXED procedure. I would suggest that you purchase the book so that you can get the most from it. It looks like it could be quite useful Dale

Dale · ‎10-15-2009

Add the CORRCOEF option to the PROC ALLELE statement. The statistic produced is R, not R^2. See Example 3.2: Computing Linkage Disequilibrium Measures for SNP Data of the ALLELE procedure documentation. Dale

Dale · ‎10-14-2009

Andres, The p-values for male reported in the Parameter Estimates table represents a test of the effect of gender in the reference wave only. But the p-value for male reported in the Type III tests of fixed effects table represents a test of the gender effect across all waves. To better understand this, try adding the following ESTIMATE statements to your GLIMMIX run: estimate "Male effect in reference group" male 1 male*wave 0 0 1; estimate "Average male effect" male 3 male*wave 1 1 1 / divisor=3; The first ESTIMATE statement should give you the effect estimate and p-value returned in the Parameter Estimates table. The second ESTIMATE should give you the p-value returned in the Type III tests of fixed effects table. Your interest is probably not to test the effect of gender in wave 3 only. Rather, you probably would like to test the effect of gender in the entire data set. That is the test given by the second ESTIMATE statement representing the "Average male effect" across all waves and also the test presented by the Type III tests of fixed effects table. HTH, Dale

Dale · ‎10-05-2009

Can I ask why you are using GENMOD for estimating the parameters of your model instead of fitting your model using PROC LOGISTIC. It is not clear from your post why you cannot use PROC LOGISTIC all the way. But if you really need to fit your model with GENMOD and then construct an ROC curve for that fitted model, you can pass the linear predictor (X*beta_hat) from PROC GENMOD as the predictor variable in the LOGISTIC procedure. If you specify the NOINT option in PROC LOGISTIC and use the linear predictor from the GENMOD procedure (and take care to model the same level of the response as the event of interest), then you will fit exactly the same "model" in PROC LOGISTIC as you had fit using the GENMOD procedure. Dale

Dale · ‎09-24-2009

Laura, You must be using SAS 9.1 rather than SAS 9.2. I replicated your error for my 9.1 SAS install. It is a bug in version 9.1. The bug is fixed in SAS version 9.2. I looked for a hotfix for 9.1 but could not find anything having to do with the CLASS statement of PROC GENMOD. Can you update to SAS version 9.2? If not, then you will have to mimic the results that would be returned for the reference specification that you want by using ESTIMATE statements. Pursuing the use of ESTIMATE statements, suppose that Var2 has m=4 levels, "A" through "D", and you want to use "B", the second level of Var2, as the reference level. The following set of ESTIMATE statements would return the parameters that would have been estimated if you could use "B" for the reference for Var2. We are able to specify that the reference level for var1 is 'A'. proc genmod data=one; class var1(ref="A") var2 / param=ref; model y = var1 var2; estimate "Intercept" intercept 1 var2 0 1 0; estimate "Var1='B'" var1 1 0; estimate "Var1='C'" var1 0 1; estimate "Var2='A'" var2 1 -1 0; estimate "Var2='C'" var2 0 -1 1; estimate "Var2="D"' var2 0 -1 0; run; You will observe that the intercept under the model with the second level of Var2 as reference is the estimated intercept plus the value of the parameter which you wish to be the reference value. The effect estimates for Var2 are the observed parameter estimates minus the value of the parameter that is now fixed to be the reference level. This might not seem to be the case for the level which was the reference level actually employed by the GENMOD procedure since we don't see any contrast, just a subtraction of the value of the parameter which was estimated for Var2='B'. However, since the last level had a value of 0 and the value of the parameter for the new reference level subtracted from zero can be expressed simply by taking the negative of the intended reference level value, the specification for what was employed as the reference level is consistent with all other levels of Var2. The parameter estimates for Var1 are not affected by the reparameterization. HTH, Dale

Dale · ‎09-18-2009

It is incorrect to believe that the parameter estimates are the same for ML as they are for REML. The parameter estimates may be asymptotically the same, but in a finite sample you would expect the parameter estimates for ML and REML to differ. The approach whereby you find the "best" model for your fixed effects using IC to rank models fitted employing ML and subsequently re-estimating the parameters of the "best" model using REML estimation is acceptable. However, different IC can lead to different "best" models. Also, this general approach is very much like stepwise regression which is known to be problematic. It is usually better to employ a model which theory specifies a priori.

Dale · ‎09-16-2009

When you specify ML, you still get information criteria (AIC, etc.) which can be employed to rank models. All the usual caveats about using information criteria to select a model would still apply. Dale

Dale · ‎09-16-2009

In order to have some difference between lsmeans across levels of A at B=1 which are different from the difference between lsmeans across levels of A at B=2, the A*B interaction must be present in the model. Moreover, in order to get the lsmeans which capture these differences, you must then specify the LSMEANS statement: lsmeans A*B / diff ...; Why don't you want to assign a value for the covariate mean_bg3? Note that the LSMEANS statement does assign all continuous covariates their mean value by default. The ESTIMATE statement has no such default behavior. If you do not specify a value for the covariate in the ESTIMATE statement, then the ESTIMATE statement will zero fill for the covariate. Thus, you would be estimating mean values for factors A and B given mean_bg3=0. Is that what you want? I doubt it. Dale

Dale · ‎09-15-2009

My understanding is that you want something which operates for categorical covariates like the AT= option operates for continuous covariates. Is that correct? I believe that you will have to employ ESTIMATE statements in order to generate the results which you seek. You don't say how many levels there are for factors A and B, but I will assume that there are three for A (genotype) and two for B (period). With these assumptions, we could write the following code: proc glimmix data=Iastress.Iab3datenew; class genotype farm animal period datenum; model tf3(event='1' ref=first)=genotype period mean_bg3 / dist=negbin solution; random farm; random datenum(period) / subject=animal(farm) type=cs residual; lsmeans genotype period / diff cl ilink; estimate "LSM at genotype=1, period=1" intercept 1 genotype 1 0 0 period 1 0 mean_bg3 / ilink; estimate "LSM at genotype=2, period=1" intercept 1 genotype 0 1 0 period 1 0 mean_bg3 / ilink; estimate "LSM at genotype=3, period=1" intercept 1 genotype 0 0 1 period 1 0 mean_bg3 / ilink; estimate "LSM at genotype=1, period=2" intercept 1 genotype 1 0 0 period 0 1 mean_bg3 / ilink; estimate "LSM at genotype=2, period=2" intercept 1 genotype 0 1 0 period 0 1 mean_bg3 / ilink; estimate "LSM at genotype=3, period=2" intercept 1 genotype 0 0 1 period 0 1 mean_bg3 / ilink; run; Differences between LSMeans across levels of A at B=1 are identical to differences between LSMeans across levels of A at B=2, so your LSMEANS statement reproduced above can be employed to produce those differences. You will have to determine the appropriate value for the covariate mean_bg3. You can determine this value by running the code: proc means data=Iastress.Iab3datenew; where missing(genotype)=0 & missing(farm)=0 & missing(animal)=0 & missing(period)=0 & missing(datenum)=0 & missing(tf3); var mean_bg3; run; HTH, Dale

Dale · ‎09-11-2009

Perhaps you want to use the LIFEREG procedure instead of PROC GENMOD. The LIFEREG procedure allows specification on the MODEL statement that the distribution is exponential. Dale

Dale · ‎09-11-2009

I don't know how you performed your search, but there are a number of SAS procedures which employ generalized least squares estimation methods. Since you don't tell us anything about the problem other than that you want some way to deal with heteroskedasticity, it is a bit difficult to advise on what methods to employ. However, assuming that you are fitting a regression model in which the residual error is porportional to a power of the mean, then I would suggest looking at the MIXED procedure. Such estimation is discussed in the documentation of the REPEATED statement. Actually, you might find it easier to use the NLMIXED procedure to estimate a model in which the residual variance is proportional to a power of the mean. NLMIXED code that fits a model in which the variance is proportional to the mean can be fit with code like the following: proc nlmixed data=mydata; mu = b0 + b1*x1 + b2*x2 + ...; model y ~ normal(mu, mu*exp(2*log_sig)); estimate "Var(eps) for mu=1" exp(2*log_sig); run; There will be problems with the above model if your parameters produce a value of mu which is non-positive (zero or negative). A generalization of the above model which is consistent with the discussion of the POM model in the PROC MIXED documentation would be to fit the model: proc nlmixed data=mydata; mu = b0 + b1*x1 + b2*x2 + ...; model y ~ normal(mu, (mu**theta)*exp(2*log_sig)); estimate "Var(eps) for mu=1" exp(2*log_sig); run; I made the assumption above that you were interested in a model in which the heteroskedasticity produced a residual variance which was proportional to a power of the mean. However, you might be interested in a model in which the residual variance differs according to the level of some categorical predictor variable. Such models can be handled quite easily with the MIXED procedure. Again, look at the documentation of the REPEATED statement and pay particular attention to the GROUP= option. HTH, Dale

Dale · ‎06-26-2009

Your problem specification is incomplete. Let me respond by first asking a few questions. Why do you think you should use powers of X and Z? Are those powers prespecified? Are X and Z nonnegative? If the powers are not prespecified integer values, then you could run into problems if the algorithm which is used to fit the model tries to raise a negative value to a noninteger power. Do you want to consider polynomial functions of X and Z? The model below is representative of a polynomial model: Y = a1(X) + a2(X**2) + a3(X**3) + b1(Z) + b2(Z**2) + c Finally (for now, anyway), what motivates the model that you want to fit?

Online Status	Offline
Date Last Visited	‎12-20-2017 01:32 PM

Three-way tables in Proc Freq not producing "Controlling for ..." text

Re: PROC LOGISTIC vs PROC GENMOD (different results)

PROC GENMOD: no automatic selection method?

Re: NLMIXED: Alternative covarinace structures?

Re: NLMIXED: Alternative covarinace structures?

Re: A data set with over 5million records and 30 variables

Re: A data set with over 5million records and 30 variables

Re: How to adjust for spatial autocorrelation in the logistic model

Re: Calculating Mallow's CP correctly

Re: out of memory error using proc GENMOD poisson

Re: Can you get standard deviations for lsmeans, instead of standard e...

Re: Multinomial Logistic models

Re: Different Somer's D from Freq and Logistic procedures

Re: Nested random effects in proc mixed

Re: NLMIXED: Alternative covarinace structures?

Re: R square in proc mixed?

Re: NLMIXED Convergence

Re: PROC MIXED - Testing sset of predictors

Re: Specifying a covariance matrix. PROC GLIMMIX?

Re: R^2 Command for finding LD HELP!

Re: p-values in Proc Glimmix

Re: ROC curves

Re: GENMOD CLASS REF=

Re: Model selection in "proc mixed"

Re: Model selection in "proc mixed"

Re: Proc glimmix - lsmeans output

Re: Proc glimmix - lsmeans output

Re: Proc genmod - Response variable exponentially distributed

Re: Generalized Least Squares

Re: Curve fitting on Data with three variable X,Y,Z