I just answered your post in the Statistical Procedures forum. SAS is giving correct results (you just have to be careful how you use the parameters for the EFFECT type of parameterization). Interestingly, there was another post about the parameter differences between LOGISTIC and GENMOD. The answer is basically the same (has to do with the parameterization), and Dale gave a thorough answer. ... View more

There is no bug. The results are correct. LOGISTIC uses so-called effect parameterization for class variables; this means the last level of each factor has all -1 values. This is different from the GLM-type parameterization (used in MIXED and some other procedures). If you look at the rest of the output, you will find a table called Class Level Information. For x2, you will see that level a gets a 1 for the design variable and level b gets a -1. For getting estimated values, the parameters are multiplied by the design values and then summed. For the third observation, because it corresponds to level b of x2, the estimate it is: 0.1612 + (0.0806*1) + (0.3852*(-1)) = -0.14333. This can be tricky. There are ways of choosing other effect parameterizations. ... View more

You would be better off using LSMEANS rather than MEANS statement (under most circumstances). Then your request is also easy: lsmeans group / cl pdiff; will give you confidence limits and all possible pair-wise differences. ... View more

Dale makes a good point: it is best to avoid using parameter bounds, when possible. For your model, it should be possible to avoid explicit bounds, using the recommendations from Dale. With some nonlinear models, it can be challenging to do the optimization without using bounds, but always a good goal. ... View more

Presumably success in your example is a 0/1 variable (not a count out of n). The binomial is really the Bernoulli in this case. As can be shown (see book by Collett on binary data), the deviance (and many related statistics) "can tell us nothing about the agreement between the observations and their corresponding fitted probabilities" for binary data (page 65 in Modelling Binary Data by Collett, 1st edition [1991]). This is because the deviance in this special case depends only on the fitted values and not the observations. Presumably, this is why GENMOD does not display the deviance and related statistics. Now, if your binomial data was number of successes (s) out of n individuals per sampling unit (n), you could use: model s/n = / dist=bin link=logit type3; and you would get the deviance (etc.). ... View more

More than likely, your program is trying to take a log of 0 (including lgamma of 0), or creating infinity. For instance, with alpha=0 (starting value), m=1/alpha will cause big problems (infinity). Also, there are log or lgamma functions that involve alpha (which will also blow up with 0 for alpha). I took your data and added a few more dummy data points. The program ran when I used 1 for the starting value of alpha. But this is still dangerous, because parameter estimates can migrate to 'impossible' regions during the optimization. You should look into putting bounds on the parameters. Check out the syntax of the bounds statement. (I did not really look at the rest of the code to see if the customized ll is correct). ... View more

There is no explicit restrict command. Sometimes an explicit restriction can be re-formulated as an implicit restriction. For instance, with x1 and x2, and the same parameter for each, one could write: y = beta0 + beta*x1 + beta*x2 as y = beta0 + beta*(x1+x2). So, if you define xsum = x1+x2 in a data step, you could then use xsum as a single covariable in mixed. Of course, other restrictions are not so easy to formulate. Edwards and colleagues have done some good work on writing macros to allow for more arbitrary linear constraints on the parameters. Basically, "new" variables are constructed outside of mixed based on the constraints chosen, and then mixed is used to fit the model (with follow-up re-formulation). If you are statistically savvy, you might enjoy the article in Biometrics, vol. 57, pp. 1185-1190 (2001). I have not tried to use the the code made available by the authors. There may be other contributions in the last few years. ... View more

Usually, you should report floating point errors directly to Technical Support. Even if you have the wrong syntax, you should not get this kind of error. Are you using SAS 9.2 or 9.22? The procedure was experimental in 9.1, and experimental procedures are more likely to give this type of problem. COUNTREG is part of the ETS package of SAS. Thus, you might also get more advice in the Forecasting Forum. ... View more

Here is an example (with a different model). The output statement stores the residuals and other stuff, as seen when the created file is printed. I also show how to use GPLOT. data a; input x y; datalines; 0 0 1 2 2 5 3 10 4 10 5 12 6 12 7 15 8 14 9 15.5 ; proc nlin data=a; parameters a 20 b .2; model y = a*(1 - exp(-b*x)); output out=a_pred predicted=p student=s residual=r; *-file a_pred contains residuals; run; proc print data=a_pred;run; proc means data=a uss css mean var ; *-css is corrected sum of squares; var y; run; proc gplot data=a_pred; *-plots of observed and predicted y versus x, and residuals; symbol1 color=blue h=2 v=dot i=none; symbol2 color=red w=2 line=1 v=none i=join; symbol3 color=black w=2 v=dot i=none; plot (y p)*x / overlay; plot s*p=3; *-plot studentized residuals versus predicted y; plot r*p=3; *-plot regular residuals versus predicted y; run; Based on your questions, you probably need to learn more about sas, in general. There are many on-line resources, and many books where sas is used throughout. Note, from the above output, the pseudoR2 is 1 - (8.4491/266.225) = 0.968. ... View more

There are different views on how to assess goodness of fit for nonlinear models. Many investigators want a R^2 value, but some statisticians feel that R^2 values should not be calculated for nonlinear models. The book by Ratkowsky summarizes this perspective. However, I side with others that an "R^2-type" of statistic has value. This is not reported by NLIN, but can be determined manually. The statistic is sometimes called a pseudoR^2, and is defined using just one of the standard definitions for linear models: pseudoR2 = 1 - (SSerror/SStotal(corrected)) You get SSerror directly from the table in the NLIN output. However, the SStotal(corrected) is not given in this table. There are reasons for this, but I won't go into these here. The NLIN output gives the uncorrected total SS (sum of squares around 0). You can get the total corrected sum of squares (sum of squares around the mean) for y using css with proc means: proc means data=a css mean var ; var y; run; Then do the calculation by hand.With a very bad fit of a model, the pseudo-R2 could actually be negative. Note: in linear models with an intercept, the mean y corresponds to a reduced model (with an intercept and no other parameters). Thus, the regular R2 is a nice comparison of the relative change in sums of squares between a full and reduced model. But with most nonlinear models, the mean y does not correspond to a reduced model compared with the full model (the nonlinear one being fitted). This is partly why some do not like the idea of R2 for nonlinear models. To me, however, it is still interesting to see the fit of the nonlinear model relative to a model with only the mean y (even though this is not a special case of the other). I realize that others may write in disagreement with my view. I do agree that one must be cautious in interpretation. There are other statistics, such as MSE, that can/should be used. You probably want to look at residuals. Just add an output statement in NLIN like: output out=preds predicted=p residual=r student=s; There are other keywords that could be added. You can then plot the studentized residuals versus predicted values using GPLOT or SGSCATTER. Another caution: it can be shown that there can still be a trend in the residual plot even when the appropriate nonlinear model is being fitted. The textbook by Schabenberger and Pierce discuss this at length. There are other types of residuals to consider, but these are tedious to calculate. Finally: the NLIN procedure will have a nice upgrade in 9.3 of SAS. In particular, there will be some nice ods graphics that will enable you to do a wide range of model assessments. ... View more

I gave an answer previously for this problem. Repeated here. Here is one way. You can use ods output to create a sas file for each relevant part of the output from a procedure. For instance, you could use: proc glimmix ; ods output tests3=tests3; class trt b ; model y = trt|b; run; to store the Type 3 test results in a file called tests3 (these are the F tests, degrees of freedom, P values). The p value is called ProbF in this file (just print it to check this out). Then in a print procedure, you can choose a different format for ProbF. Below is one way, where I request 12 decimal places for P. proc print data=tests3; format ProbF pvalue15.12; *<--no space in pvalue15.12 ; run; The relevant ods file is different for each procedure. Check the documentation for your desired procedure. Usually, one does not need to know p to more than 5 or 6 decimal places. But it does come up with multiple testing problems, where one wants to adjust the p values for multiple testing. Then, it can be useful to see more than 5 or 6 decimal places. ... View more

Virtually any procedure in SAS that can calculate a mean can also calculate a weighted mean. Why not start with the obvious one, PROC MEANS. Just add a weight statement. You probably should weight by the INVERSE of the variance, not the standard deviation. Get the weight variable in a data step first. proc means data=a; weight wt; var y; run; ... View more