I'm trying to do a pairwise test to calculate p-values between various products that are being evaluated by several assessors. I'm trying to reproduce what SAS is doing in R, and I'm running into an issue in one particular situation: when the items being compared (in my case SAMPLE) are complete for one item in the pairwise comparison but not the other. Here's an example:
data raw; INPUT JUDGE REP SAMPLE X1; datalines; 1 1 1 3.0 1 2 1 4.0 2 1 1 7.0 2 2 1 2.0 1 1 2 1.0 1 2 2 6.0 2 1 2 3.0 ; run; proc glm data = raw; class judge sample rep; model X1 = judge sample rep judge * sample / ss3; lsmeans sample / stderr tdiff alpha=0.10 e=judge*sample; run;
In this case we have a JUDGE (2) who is missing REP 2 for SAMPLE 2. SAMPLE 1 has a complete set of data. When I calculate the pairwise comparison between SAMPLE 1 and 2, I get a p-value of 0.48973, whereas in SAS it's an even 0.5. To get my value I'm doing the following:
mse = interaction_sum_of_squares / degrees of freedom = 0.66666666 / 1
constant = sum of (1/n1j + 1/n2j) = (1/2 + 1/2 + 1/2 + 1) = 2.5
sd = sqrt( (mse / num_judges^2) * constant)) = 0.6454972
t-value = lsmean1 - lsmean2 / sd = (4.0 - 3.3333333) / (0.6454972) = 1.032796
Meanwhile, the SAS t-value is 1.0.
I'm matching the lsmeans that SAS is getting, and the interaction_sum_of_squares and the degrees of freedom match what SAS is putting out as well. That really only leaves the constant, or something else SAS is doing which I can't figure out.
I match SAS perfectly when the data is balanced, it's only when it's unbalanced that it's not working.