Imbalance or incomplete data, SASophylistic view [🇷 for BE/BA]

posted by d_labes  – Berlin, Germany, 2008-10-28 18:14 (6076 d 14:54 ago) – Posting: # 2592
Views: 34,157

Dear Yung-yin, dear ElMaestro

❝ [...] the protocol has explicitly stated that proc glm would be used if the dataset was balanced, and that proc mixed would be used in case of imbalance.


May be the confusion comes from the meaning of imbalance.
Usually imbalance in the context of 2x2 cross-over means unequal number of subjects in the sequence groups.

But sometimes incomplete data are meaned, i.e. drop-outs with data for the first period. This is of course imbalance to number of data in periods.

In the first case Proc GLM and Proc MIXED give identical results.

Only in the case of missings Proc MIXED can recover some information from the subject with incomplete data and will then give different results to GLM.

Let us take an example: Lets use Helmuts data here on the forum.
This is a balanced study.
The results (without subject effects, wich are dealt different in GLM or mixed effects analysis and wich are seldom from interest):

Proc GLM
ANOVA typeIII F-tests
effect    F-value df1 df2 p-value
Formul     0.38    1   22  0.5429   
Period     0.08    1   22  0.7856
sequence   1.02    1   22  0.3239

90% Confidence intervals
point est. lower   upper
 0.9653    0.8750  1.0648

Proc MIXED
Fixed effects F-tests (type III)
effect    F-value df1 df2 p-value
Formul     0.38    1   22  0.5429   
Period     0.08    1   22  0.7856
sequence   1.02    1   22  0.3239

90% Confidence intervals
point est. lower   upper
 0.9653    0.8750  1.0648

As you see, identical results.

Now let us deal with imbalance in sequence. Let's drop subject 24.
The results:

Proc GLM
ANOVA F-tests
effect    F-value df1 df2 p-value
Formul     0.26    1   21  0.6129   
Period     0.12    1   21  0.7349
sequence   0.54    1   21  0.4725

90% Confidence intervals
point est. lower   upper
 0.9698    0.8753  1.0746

Proc MIXED
Fixed effects F-tests
effect    F-value df1 df2 p-value
Formul     0.26    1   21  0.6129   
Period     0.12    1   21  0.7349
sequence   0.54    1   21  0.4725

90% Confidence intervals
point est. lower   upper
 0.9698    0.8753  1.0746

As you see, again complete identical results.

And now with incomplete data. Set subject 24, period 2 to missing.
The results:

Proc GLM
ANOVA F-tests
effect    F-value df1 df2   p-value
Formul     0.26    1   21    0.6129   
Period     0.12    1   21    0.7349
sequence   1.09    1   22.17 0.4725

90% Confidence intervals
point est. lower   upper
 0.9698    0.8753  1.0746

Proc MIXED
Fixed effects F-tests
effect    F-value df1 df2 p-value
Formul     0.38    1   21  0.5422   
Period     0.06    1   21  0.8145
sequence   0.99    1   22  0.3308

90% Confidence intervals
point est. lower   upper
 0.9638    0.8698  1.0678

Here we have differences in the results. Proc GLM gives the same confidence interval as with subject 24 dropped from the analysis, whereas Proc MIXED yields some result near the analysis of the complete data.

Now its up to you to figure out what the R-ophylistic analysis will give.
Hopefully the same :-D . Then this could be a validation example for BEAR.
—
Regards,

Detlew

Complete thread:

UA Flag
Activity
 Admin contact
23,424 posts in 4,927 threads, 1,672 registered users;
36 visitors (0 registered, 36 guests [including 22 identified bots]).
Forum time: 10:09 CEST (Europe/Vienna)

Truth and clarity are complementary.    Niels Bohr

The Bioequivalence and Bioavailability Forum is hosted by
BEBAC Ing. Helmut Schütz
HTML5