GLIMMIX for analysis of visual scoring data (ordinal)

[lrae86]

Hi all,

I have a very introductory understanding of SAS, but it is widely used in my department. I am working on a rapeseed pod shatter evaluation trial where pod shatter tolerance of 194 genotypes is being studied.

The design is RCBD, each genotype is replicated 3 times (rep=block) and there are 4 siteyears (environments). Fixed variables include genotype and siteyear, the random variable is block nested in siteyear. Shatter tolerance is my dependent variable, where a rating of 1-9 is approximated based on visual observation in the field.

1 = 10% or < pods intact

2 = 20% pods intact

3 = 30% pods intact

4 = 40% pods intact

5 = 50% pods intact

6 = 60% pods intact

7 = 70% pods intact

8 = 80% pods intact

9 = 90% or > pods intact

I have been instructed to analyse the data using PROC GLIMMIX with a multinomial distribution and cumulative logit link. An alternative method has been suggested using the beta distribution and converting my ratings into proportions – although I’m not sure that is correct since the ratings are an estimate of percentage not an actual measurement. I am interested in the statistical significance of genotype, siteyear and genotype*siteyear, as well as performing a multiple comparison test (which I believe can be done with LSMEANS?).

Is GLIMMIX the best option for analysing this type of ordinal data? If anyone could provide me with some insight, it would be much appreciated. I have attached the code I have been using and a sample of my data. Thank you!

proc Glimmix data=shatter method=laplace;
class siteyear genotype block;  
model sh_tol = genotype|siteyear / dist=mult link=cumlogit;
random block(siteyear);
store glimres;
covtest / wald; 
run;

proc plm restore=glimres;
lsmeans genotype siteyear/ ilink pdiff adjust=bon;   
ods output diffs=ppp lsmeans=mmm;
ods listing exclude diffs lsmeans;
run;

Sample data

data shatter;
input siteyear$ block genotype sh_tol; 
cards;
siteyear	block	genotype	sh_tol
22_SH1_CM	1	10	6
22_SH1_CM	2	10	5
22_SH1_CM	3	10	3
22_SH1_CM	1	11	5
22_SH1_CM	2	11	6
22_SH1_CM	3	11	6
22_SH1_CM	1	12	4
22_SH1_CM	2	12	3
22_SH1_CM	3	12	3
22_SH1_PT	1	10	1
22_SH1_PT	2	10	4
22_SH1_PT	3	10	1
22_SH1_PT	1	11	2
22_SH1_PT	2	11	4
22_SH1_PT	3	11	3
22_SH1_PT	1	12	2
22_SH1_PT	2	12	1
22_SH1_PT	3	12	1
23_SH1_CM	1	10	7
23_SH1_CM	2	10	7
23_SH1_CM	3	10	3
23_SH1_CM	1	11	4
23_SH1_CM	2	11	2
23_SH1_CM	3	11	3
23_SH1_CM	1	12	7
23_SH1_CM	2	12	6
23_SH1_CM	3	12	6
23_SH1_PT	1	10	8
23_SH1_PT	2	10	6
23_SH1_PT	3	10	7
23_SH1_PT	1	11	5
23_SH1_PT	2	11	2
23_SH1_PT	3	11	4
23_SH1_PT	1	12	6
23_SH1_PT	2	12	4
23_SH1_PT	3	12	5
;;;;

[sbxkoenk]

Some remarks :

• The combination of DIST=MULTINOMIAL and LINK=CUMLOGIT requests the proportional odds model. You should check that proportional odds assumption.
  Like here:
  Usage Note 22954: The PROC LOGISTIC proportional odds test and fitting a partial proportional odds model
  https://support.sas.com/kb/22/954.html
• You don't need to store your model in an item store for restoring in PROC PLM (for Post-fitting statistical analyses on Linear Models). You can put the LSMEANS statement directly in PROC GLIMMIX.
• For modelling continuous proportions (not sure this is appropriate for your use case), see here:
  Usage Note 57480: Modeling continuous proportions: Normal and Beta Regression Models
  https://support.sas.com/kb/57/480.html

Koen

[sbxkoenk]

By the way ... I have moved tis topic to "Statistical Procedures" - board under Analytics header.

Home > Analytics > Stat Procs

is a better place for this question than
Home > Programming > SAS Procs.

Koen