topic Re: Multiple comparison in the unequal variance case (Games Howell) in New SAS User

Multiple comparison in the unequal variance case (Games Howell)

AndreaSch — Thu, 03 Jan 2019 20:42:28 GMT

Dear Sir or Madame,

I want to do a Multiple Comparison. I have data that is normally distributed, but has unequal variance. So I want to perform the Games-Howell-Test. I tryed to write a code but I'm not sure if it's right. Could anyone have a look at it and tell me if it's correct or how to correct it?

DATA Peak_715;
INPUT Gruppe Intensitaet @@;
CARDS;
0.1	1599.265639	0.1	1154.661803	0.1	1399.502189
0.2	1041.65137	0.2	812.660909	0.2	841.5435114
0.5	2043.370755	0.5	1082.422559	0.5	1167.995499
0.7	1780.851446	0.7	1938.736796	0.7	1636.642438
1	1275.664136	1	1392.723226	1	1234.064438
1.4	1271.28798	1.4	864.408437	1.4	1148.410745
1.5	1312.517333	1.5	1230.61054	1.5	1068.233446
2	1132.318401	2	1530.504886	2	1068.480662
2.1	1301.379045	2.1	1462.262261	2.1	1267.615
2.8	1109.274158	2.8	1375.193526	2.8	1287.310669

RUN;
PROC GLIMMIX DATA= Peak_715;
CLASS Gruppe;
MODEL Intensitaet = Gruppe; 
RANDOM _residual_/GROUP= Gruppe;
LSMEANS Gruppe/ ADJUST= tukey ADJDFE= row;
run;

The results viewer shows me "Adjustment of Multiple Comparisons: Tukey-Kramer". Why is it showing Tukey-Kramer? My data are balanced.

My other question is: How do I have to interpret the results? Do I have to compare Pr>t with alpha? I am running SAS version 9.4. on Windows.

I appreciate any help. Thanks in advance.

Regards

Re: Multiple comparison in the unequal variance case (Games Howell)

PaigeMiller — Thu, 03 Jan 2019 21:21:54 GMT

@AndreaSch wrote:

The results viewer shows me "Adjustment of Multiple Comparisons: Tukey-Kramer". Why is it showing Tukey-Kramer? My data are balanced.

Because, according to this documentation, Tukey-Kramer is the equivalent to Games-Howell.

If you have unequal variances in the different groups, you can do a WELCH test of the MEANS statement in PROC GLM (but not PROC GLIMMIX). You probably also want to use the HOVTEST option to actually test if your variances are unequal or not.

My other question is: How do I have to interpret the results? Do I have to compare Pr>t with alpha?

If the value of Pr>|t| is less than 0.05, then the two groups on that row have statistically different values of Intensitaet (for a test where alpha = 0.05). If you are testing all possible combinations of Gruppe, then the Adj P value is adjusted due to the multiplicity of tests, and again, the ones where Adj P < 0.05 are the statistically significant ones.

Re: Multiple comparison in the unequal variance case (Games Howell)

AndreaSch — Fri, 04 Jan 2019 11:12:56 GMT

Thank you for your answer.

In advance i have already used the option HOVTEST.

According to my code. what exactly do I have to write. Is GLIMMIX wrong? I don't want to use the Welch-Test because it only says if there are differences and not where they are.