I don't know about "instead of", but I do know how to get "in addition to." proc sort data=test;   by sortnames name; run; proc print data=test ;   by sortnames name; run; This should at least put both in the byvar header block. Steve Denham ... View more

It doesn't look like it.  The only workaround I see is to specify each term in a separate ODDSRATIO statement, and get what you need from the ODS table OddsRatiosPL. Good luck, Steve Denham ... View more

One other idea, the LSMESTIMATE went out the window, as it looks like the inverse link is applied to the difference, so you would get something like an odds ratio. So, with an alpha=.05, it looks like tinv=1.96 within rounding error, which suggests df for each mean of infinity.  That won't work, so I am going to pick a large n for each group, say n=100. Thus upper minus lower should be about 3.92 "standard errors". Infected: 33721 - 19944 = 13777 13777 /  3.92 = 3514.54 (as a standard error) Stddev1 = sqrt(100) * 3514.54 = 35145.4 Non-infected: 17117 - 10139 = 6978 6978 / 3.92 = 1780.10 (as a standard error) Stddev2 = sqrt(100) * 1780.10 = 17801.0 Pooled standard error of the difference = sqrt ( (35145.4)*(35145.4)/100 + (17801)*(17801) / 100) = 3939.64 (formula from page 81 of Steel and Torrie, Principles and Procedures of Statistics, 1960) Thus an approximate 95% confidence interval on the difference of 25933 - 13174 = 12759 would be      Lower 12759 - 3939.64 * 1.972 = 12759-7769.04 = 4989.96      Upper: 12759 + 3939.64 * 1.972 = 12759 + 7769.04 = 20528.04, where 1.972 is the t value for a 2 tailed distribution with 198 degrees of freedom. I note that the standard errors I came up with from the asymmetric confidence bounds are not much different from what is reported, so you could probably use an ODS statement to capture all of the needed values in a single dataset, never bother with the upper minus lower business, and get what you need fairly easily. I would recommend moving from GENMOD to GLIMMIX, in case you want to regard any of your factors as random in a hierarchical sense, and to get the degrees of freedom for the estimates in the dataset (as it looks as though GENMOD doesn't print them). Good luck, Steve Denham ... View more

Ron, Point well taken.  OK, I'm still battling some brain congestion, so let's get some more detail.  First, what distribution and link are you using in GENMOD?  Second, could you post some of the output from the LSMeans part.  I think if I can do one specific calculation, it should be a lot easier to generalize. For instance, for a lognormal distribution, I use the following code (dataset is output from PROC GLIMMIX): data lsmeans_logn_bt; set lsmeans_logn; omega=exp(stderr*stderr*df); btmean=exp(estimate)*sqrt(omega)-1; btvar=exp(2*estimate)*omega*(omega-1); btsem=sqrt(btvar/df); run; Note that here I don't calculate from the confidence bounds.  If I had to, say for some other distribution, it would look something like: /* UNTESTED CODE */ stderr = (upper - lower) / (2 * tinv (alpha, df)); stddev = sqrt (df + 1) * stderr; and then plug the stddev for each mean into: sddiff = sqrt ( (stddev1 * stddev1 / (df1 + 1) ) + (stddev2 * stddev2 / (df2 + 1))); in all this upper and lower are confidence bounds with alpha selected to make them one standard error above and below, and then I indexed the individual calculations.  If this isn't working, then I think I really need to see more of what you are getting in your output. Steve Denham ... View more

If you have means and standard errors, you can calculate differences and standard errors of the difference.  Things work out more easily if you have a balanced design, so that degrees of freedom are equal, but if not: SE diff = sqrt( 1/ n1 + 1/n2) * s where n1 is df for mean 1+1, n2 is df for mean 2 +1, and s is the common standard error, which you can get from an elementary stat text (too many square roots, divisions and exponents to do in plain text). I would change the lsmeans statement to include the ILINK option, so that the bounds you get are on the original scale. Steve Denham ... View more

I doubt very much that you can get the differences or standard errors of differences on the original scale as direct output.  Using the ILINK option will give LSmeans on the original scale, and by post-processing, you could calculate differences and approximate asymptotic standard errors of the difference, using the confidence bounds on the back-transformed means to get values to plug through various approximations.  For instance, setting ALPHA=0.308 should give approximately 2 SEM as the bounds.  Granted these are asymmetric, but an approximate value can be obtained as StdErr/2, which then can be used for each mean to calculate an approximate standard error of the difference.  Not pretty, and if you have more than three or four comparisons, not necessarily easy to program. Steve Denham ... View more