Multiple Comparisons for a Random Factor

[ivanchuk]

Unfortunately Minitab does not currently perform multiple comparisons for a random factor or a term that includes a random factor.

Hi members:

I want to perform a Tukey test for factor A in this design Nested Anova but Minitab show this warning. What I can to do?
* WARNING * No multiple comparisons were calculated for the following terms which contain or interact with random factors.
​

[FONT=&quot]
[/FONT]
[FONT=&quot]Analysis of Variance for MPa, using Adjusted SS for Tests[/FONT]

[FONT=&quot]Source DF Seq SS Adj SS Adj MS F P[/FONT]
[FONT=&quot]A 2 467.53 467.53 233.76 5.31 0.012[/FONT]
[FONT=&quot]B 1 1188.97 1188.97 1188.97 27.03 0.000[/FONT]
[FONT=&quot]A*B 2 82.72 82.72 41.36 0.94 0.404[/FONT]
[FONT=&quot]Tooth(A B) 24 1055.63 1055.63 43.98 1.55 0.064[/FONT]
[FONT=&quot]Error 120 3396.74 3396.74 28.31[/FONT]
[FONT=&quot]Total 149 6191.59[/FONT]

.................

Comments about Mixed Model in Minitab​

For a lot of you who are using GLM instead of the balanced command because you wanted to get the standard error term easily, note that you can get the standard error term easily only if it's a fixed effects model. There's actually a history to this. When they wrote GLM originally, they didn't allow random effects. What this meant in practice is that if you had random effects, you had to use the balanced command and if you had covariates, you had to use the GLM command, so you couldn't have a model like this example that had both random effects and covariates in it. Minitab just wouldn't analyze them. So, Minitab tacked a program on that analyzed mixed models, but they never finished it. This means that a lot of the output that we're used to

expecting in the fixed effect models we just can't get in the mixed models.

However, as long as the random effect is messed up with the fixed effect, it just won't complete the analysis. For instance, it won't do Tukey in this case. We've seen it in general in GLM where it will do Tukey, but in many mixed models it just won't. In my mind, Minitab should be able to do these things, but it won't.
...................................

 

[Miner]

Minitab is correct in their handling of this. It is not an error nor an oversight on their part, and the routine is quite finished beyond future enhancements.

You will first need to clearly understand the differences between a fixed term and a random term.

• FIXED term: A fixed term is one where the levels are 100% of all possible levels, or one where the levels are the specific levels that are of interest to you.
• RANDOM term: A random term is one where the levels were chosen at random, and the levels are intended to represent the effect of all levels.

Now consider the following:

1. A Fixed term where the levels are 100% of all possible levels. In this scenario, you are interested in the actual means of each level. You may want to select the best mean level. An example might be the highest material strength from supplier A, B or C. These suppliers are the only supplier available. Therefore, a means test such as Tukey's is appropriate.
2. A Fixed term where the levels do not represent all levels, but they are the SPECIFIC levels in which you are interested. Again, you want to select the level with the best mean. An example of this might be 3 versions of a material. They are not all versions, but they are the only versions in which you are interested. Again, a means test is appropriate.
3. A Random term where the levels do not represent all levels, and you are not interested in the specific mean of any particular level. You are only interested in the overal EFFECT of that variable. An example is the Part in a Gage R&R. You are not interested in the exact mean of a particular part. You are only interested in the overall EFFECT of the variable, Part. A means test is NOT appropriate in this scenario. You will never see these parts again. The only represent the effect of the variable.

Since a Mixed model includes both types of terms, it will perform a means test for the FIXED terms, but not for the RANDOM terms. It will not perform the test for interactions between fixed and random because you have no interest in the random variable other than its effect.

 

[ivanchuk]

Thanks Marc. I was needing an explanation like that.
Ivan