Split-plot with missing data points

[ACFrasca]

Hi,

I am analyzing the 2nd year of data from my experiment. I have a split-plot design with 4 replications in which my main plot is plant spacing (2 levels) and my subplots are varieties (8 levels).

For my 1st year of data I had 6 levels for varieties and no missing data points. Therefore, I analyzed using proc glm with mean separation by Duncan. The code used follows.

Proc GLM data = Tomato;

class rep spac var;

Model FXL FL FM FC SXL SL SM SC FH SH XLT LT MT CT TTH = rep spac rep*spac var var*spac;

test h=spac e=rep*spac;

means var/duncan;

means spac/duncan e=rep*spac;

run;

This year I have 8 levels of varieties and 4 missing data points (out of 64). Two of these missing data points are 2 varieties on rep 1, while the other two are the same variety on rep 4.

I would like to know what's the most appropriate way to analyze the data for this year differing the least possible from last year (since it has already been published).

Thanks,

Aline

[SteveDenham]

It is well-known, and has been for at least 20 years, that the standard errors produced by PROC GLM for split plot models are incorrect, as are many of the standard errors of differences (see SAS System for Mixed Models (1996) by Littell et al., see Section 2.4.2).  The best thing you could possibly do is analyze this in PROC MIXED, where the problem of missing values is adequately handled.

If you absolutely must do this in GLM, then consider the use of LSMEANS rather than MEANS.  For the balanced case (previous year) the values should be exactly the same.

However, I would reanalyze the results of the previous year in PROC MIXED, in order to maintain consistency, and if the biological conclusions are different, I would prepare a retraction of the previous analysis.  Sounds severe, but I am very surprised that this got through peer review.

Steve Denham

[ACFrasca]

Hi Steve.

Thanks for your reply. I've been studying a lot since I posted the question and I did learn about the mixed model being more appropriate for split-plot than the glm. I published the first year of data in the proceedings of a conference and therefore it wasn't peer reviewed. I wish it had been so I'd learn the correct way to do it.

I am glad I ran into missing data problems this year in order to study again and learn the correct way. I ran my data for this year using proc glimmix this time and re-ran last year's data using glimmix as well and results are fairly similar but mean separation is a little bit different since I've used Tukey-Kramer instead of Duncan this time.

By the way, I never quite understood how you pick the most appropriate mean separation test for a specific set of data. Do you have any suggestions of a good reading about it?

Also, using glimmix, and I believe using proc mixed also, I need to choose a method for the calculation of the denominator degrees of freedom. I used kr as suggested by a stats person but I would like to know more about the different methods that can be used. I'd appreciate if you know any reading about that too.

Thanks,

Aline

[SteveDenham]

Well, the second edition of SAS for Mixed Models does address the denominator degrees of freedom.  I would recommend ddfm=satterthwaite here; if there were repeated measures, I would recommend Kenward-Rogers.

For mean separation tests, you should go through Multiple Comparisons and Multiple Tests Using SAS, 2nd. edition, by Westfall, Tobias and Wolfinger.  I would recommend Edward and Berry's simulation method.

Steve Denham