yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-10-26 21:46
(5631 d 17:04 ago)

Posting: # 2584
Views: 30,718
 

 bear v2.0.1 (upwards) cont'd [🇷 for BE/BA]

Dear Elmaestro (EM) & Helmut,

I almost forget how to initiate a new thread in this Forum. have got used to click "Post reply" to send a replied message. :-D In previous message, I said that "...bear v2.0.1 is not designed to analyze the imbalanced data, but the balanced data, obtained from BE studies..." because of this replied message by EM. In that replied message, EM has suggested an lme method should be used first. Currently, we have not implemented lme method yet with bear. Right now, we use lm to do anova only. We still work on lme. Helmut also had a comment about analyzing unbalanced (or imbalanced) sequence BE data. So we got the book and studied MVUE. Please let me know if I am wrong about this. In this case, can bear still be used to analyze the imbalanced data of a BE study only with lm? Thanks.

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
ElMaestro
★★★

Denmark,
2008-10-27 21:36
(5630 d 17:14 ago)

@ yjlee168
Posting: # 2585
Views: 28,114
 

 bear v2.0.1 (upwards) cont'd

❝ In this case, can bear still be used to analyze the imbalanced data of a BE study only with lm?


Dear Dr Lee,

yes, you can easily do it with lm or glm as far as a 2-sequence, 2-period, 2-treatment study goes.
Do it this way:

MyFunkyFit=lm(LnParam~Subj+Seq+Pe+Treatm)
(You can use glm is stead of lm here, no problem)
The for the anova...
Anova with type I SS goes something like this:
A_typeI=anova(MyFunkyFit)
Anova with type III goes SS something like this:
A_typeIII=drop1(MyFunkyFit, test="F")

It is correct that I earlier suggested the use of lme in stead. The result wil be exactly the same, it needs to be said. I proposed that because I saw a thread in another forum which indicated that lm might be unstable or incorrect for imbalanced data. This seems not to be the case in practice; I have really tried hard to find a dataset that shows this phenomenon butI can't. Thus lm or lme or glm, they will all work, the choice is yours.

Good luck.
EM.
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-10-28 09:59
(5630 d 04:51 ago)

(edited by Jaime_R on 2008-10-28 10:38)
@ ElMaestro
Posting: # 2587
Views: 28,060
 

 bear v2.0.1 (upwards) cont'd

dear EM,

Thank you for your message. No wonder that I cannot find anything about the method to analyze imbalanced sequence, BE data from US-FDA guideline. Yes, we've already been able to calculated Type I SS and Type III SS with bear right now. Thus, bear can analyze imbalanced or balanced sequence, BE data using lm method. In your message of previous thread, you said it might be "unsafe" if we used lm to analyze imbalanced BE data. And in the following replied message, you said it might be "unstable or incorrect" (from another forum) if we do so. Could you please give us more details about why you said so? What have you seen if one uses lm to analyze an imbalanced sequence BE? It sounds like lme or SAS PROC MIXED is preferred when analyzing imbalanced sequence BE. Am I correct?

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
ElMaestro
★★★

Denmark,
2008-10-28 10:49
(5630 d 04:01 ago)

(edited by ElMaestro on 2008-10-28 13:04)
@ yjlee168
Posting: # 2588
Views: 28,282
 

 bear v2.0.1 (upwards) cont'd

Dear Dr Lee,

I am sorry if my signals have been wobbly.
I think at some point I read a thread on the mailing list for R http://tolstoy.newcastle.edu.au/R/ about the simple linear model, that it -in R's implementation- might be problematic or some other negative thing (will give the specific link if I find it again, promise) in case of imbalance. I could easily have misunderstood it, no doubt.
Due to this I suggested lme, which has some potential advantages:
  1. It will give the correct result.
  2. It should work for the more complex designs, too, not just the standard 2,2,2-BE design.
Since then I tried to challenge lm and glm in R with various stupid datasets and could not find that it is problematic (that is, I generally obtain the exact same results with lm/glm as with lme regardless of the imbalance in the dataset). On this basis I am inclined to think now that for a 2,2,2-BE design lm is as good as glm which is as good as lme. And of course, some might say lme is overkill for such simple datasets where in the strictest terms we do not need to work explicitly with mixed models.

Finally, in some *AHEM COUGH*iers I have seen for 2,2,2-BE studies, 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. I am not a statistician so I cannot qualify this in any way, I only relay it as an observation.

Best regards
EM.
d_labes
★★★

Berlin, Germany,
2008-10-28 18:14
(5629 d 20:36 ago)

@ ElMaestro
Posting: # 2592
Views: 28,441
 

 Imbalance or incomplete data, SASophylistic view

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
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-10-29 08:36
(5629 d 06:14 ago)

@ d_labes
Posting: # 2594
Views: 28,074
 

 Imbalance or incomplete data, SASophylistic view

dear DLabes,

We will follow your suggestions and examples given here to validate bear and will report the test result here. One more question: do we have to include all cases wuth incomplete data (only R or T), or we can treat these cases as dropouts? Which one is more acceptable based on regulatory guidelines (if any)? Thank you so much.

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
d_labes
★★★

Berlin, Germany,
2008-10-29 10:24
(5629 d 04:26 ago)

@ yjlee168
Posting: # 2595
Views: 28,084
 

 Incomplete cases

Dear Yung-yin,

❝ [...] One more question: do we have to include all cases wuth incomplete data (only R or T), or we can treat these cases as dropouts? Which one is more acceptable based on regulatory guidelines (if any)?[...]



By the way: wuth was a brewery in Wiesbaden with good German beer :-D . Nowadays an institute for learning advertising (buy things you do not need), no beer.

But let us be unsmiling scientific:
Generally spoken: Include all data you have.
But in the context of a classical 2x2 design and the use of Proc GLM or equivalent software it does not make a difference with respect to the bioequivalence test (90% confidence interval) as we have seen from our example analysis (3rd analysis) discussed just.

Therefore my attitude up to now was to exclude subjects with incomplete data from a 2x2 cross-over (reporting the data, if any, but not including in the statistical analysis).
But see Helmut's excerpt of SENN here.

In the new EMEA Draft it is stated under 4.1.8 Evaluation/Subject accountability:
"All treated subjects should be included in the statistical analysis, with the exception of subjects in a crossover trial who do not complete at least one period receiving each of the test and reference products (or who fail to complete the single period in a parallel group trial)."

A little bit confusing (at least to me :confused:) but according to Ohlbe it can be interpreted as not to include subjects with data for only one period (... receiving each of the test and reference products ...) from a 2x2 x-over.

Other regulatory recommendations I am not aware in the moment.
Any other out there with a comment?

Regards,

Detlew
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-10-30 08:12
(5628 d 06:38 ago)

@ d_labes
Posting: # 2610
Views: 28,151
 

 Incomplete cases

dear DLabes,

Thank you for your message.

❝ By the way: wuth was a brewery in Wiesbaden with good German beer :-D .

❝ Nowadays an institute for learning advertising (buy things you do not need), no beer.


German beer? I've already tasted some during August when attending useR!2008 Annual Conference at Dortmund, Germany. Here was one that I had in Kron. Don't know it's exact name. probably home-made. But it tasted great.
[image]

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
ElMaestro
★★★

Denmark,
2008-10-29 10:27
(5629 d 04:23 ago)

@ d_labes
Posting: # 2596
Views: 28,113
 

 Imbalance or incomplete data, SASophylistic view

Dear Dlabes,

excellent input. I learn a lot from these discussions.
Allow me a little digression here....
I have never seen a *AHEM COUGH*ier for a 2,2,2-BE study in which subjects with missing period data are included in the study. In fact, it is sort of standard to simply specify that the analysis will be based on data from subjects completing both periods (in some cases it is even written as 'the first 28 completers' etc etc).
I think this stems from the fact that the EMEA has been challenged on missing data (missing periods can be considered such a thing, I guess). For this reason they produced some documents with guidance. That guidance implicitly deals mainly with progression-types of data (i.e. intended measurements of the degree of syphilis/parkinsonism/crohn/hemorrhoids/tuberculosis/survival/whatever at week 4, 8, 12, 16 after treatment initiation but data from week 12 goes missing, and how do you then do the Kaplan-Meier etc), but equivalence/non-inf. studies are not excluded per se. Among realistic alternatives, EMEA says, is imputation. Who can imagine imputation for missing 2,2,2-BE period data*? I think it is wise to simply rule those subjects out.
In R this would correspond to "na.omit" if my memory serves me correct.

Please, please more discussions of this type.
EM.


*: For example "Analyses that investigate missing data imbalance in all relevant factors and whether patients with and without missing values have different characteristics at baseline might also be conducted" and that sort of thinking [http://www.emea.europa.eu/pdfs/human/ewp/177699EN.pdf, under revision]
d_labes
★★★

Berlin, Germany,
2008-10-29 10:56
(5629 d 03:53 ago)

@ ElMaestro
Posting: # 2597
Views: 28,013
 

 Gold in discussion

Dear ElMaestro,

full ACK with your points. At least for the classical 2x2 cross-over. See my two coins in the comment above.

A further justification: In Equivalence trials (and BE studies are a sort of, really ;-) ) the primary population usually is the per protocol population. We already had it here in this thread. And missing data are mostly the consequence of protocol deviations/violations.

❝ Please, please more discussions of this type.

❝ EM.

I also wish this. Really!

I know, time is spare with us. But YOU out there, do not follow "Talk is silver, silence is golden". Take silver. Gold is spare in the nowadays time of financial disasters over the world.

Regards,

Detlew
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2008-10-29 14:03
(5629 d 00:47 ago)

@ d_labes
Posting: # 2602
Views: 27,847
 

 Gold?

Dear all!

I haven’t performed analyses of incomplete data of a 2×2×2 study until now – and haven’t seen one in my work as a consultant performed by anybody else.

Another quote* (referring to their example 3.2 in 49 subjects):

Before we can continue to fit a linear model to the (log-transformed) data, we must decide what to do with the data from those subjects who did not provide a value in both periods. Such a comparison is not possible for those subjects with only a single value. If however, there are two such subjects and has a value only on T and the other has a value only on R, then a between-subject com­parison of T and R is possible by taking the difference of these two single values. However, the precision of such a comparison will be low because the between-subject variation […] is much higher than the within-subject vari­abi­lity. Because we have assumed the subject effects are random variables, these between-subject comparisons can be recovered in the analysis if we fit what is known as a mixed model. […] However, the recovery of between-subject infor­ma­tion on the comparison of T and R is unlikely to make much difference to the results, and so nothing of significance will be lost by ignoring the data on those subjects that provided only a single value. To justify this assertion, we will also report the results of fitting the mixed model to the complete data set […]

Personally I think it’s a little bit ambiguous to perform an analysis of the complete data set (simple model) as primary and ‘justify’ it by running a second one on the full (mixed) model. For our example the width of the confidence interval is 0.198015 (simple model) and 0.189859 (full model). As DLabes pointed out in this post the mixed model is scientifically valid, but I would be wary to present it as the only one – since without suitable software it’s not possible to reproduce the results. If done the other way ’round (from Patterson/Jones), I would expect a lot of discussions (imagine if model #1 shows BE, and #2 does not)…


  • Patterson S, Jones B. Bioequivalence and Statistics in Clinical Pharmacology. Boca Raton: Chapman & Hall/CRC; 2006. p. 61–7.

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2008-10-29 13:12
(5629 d 01:38 ago)

@ d_labes
Posting: # 2601
Views: 28,447
 

 Imbalance or incomplete data, WinNonlin5.2.1

Dear all,

let’s continue the example in WinNonlin 5.2.1 (which uses also linear mixed effects modeling). Identical results to SAS are obtained both for the complete data set, and for the data set with #24 missing in both periods.
If data of #24 are missing in period 2 only, I obtained:

Denominator df option : residual
Partial Tests of Model Effects
Hypothesis  Numer_DF  Denom_DF     F_stat   P_value
---------------------------------------------------
  Sequence         1        21  0.988858     0.3313
    Formul         1        21  0.383815     0.5422
    Period         1        21  0.0564614    0.8145
90% Confidence intervals
point est.  lower     upper
0.963755    0.869808  1.067849


Denominator df option : satterthwaite
Partial Tests of Model Effects
Hypothesis  Numer_DF  Denom_DF     F_stat   P_value
---------------------------------------------------
  Sequence         1      21.8  0.988858     0.3309
    Formul         1      21.0  0.383815     0.5422
    Period         1      21.1  0.0564614    0.8145
90% Confidence intervals
point est.  lower     upper
0.963755    0.869820  1.067835


The second table is in agreement with SAS’ Proc MIXED.

Cave: although Satterthwaite degrees of freedom are default in WinNonlin, it’s possible to change this behaviour in the program’s preferences. This doesn’t make a difference if complete subjects are missing, but should be avoided in this case.

@DLabes: I gave the PE/CI to 6 decimal places in order to see a difference (if any); Satterthwaite dfs were 21.760741 (Sequence), 21.049560 (Formul), and 21.059500 (Period); can you give us your results also?

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
d_labes
★★★

Berlin, Germany,
2008-10-29 17:38
(5628 d 21:12 ago)

@ Helmut
Posting: # 2605
Views: 27,891
 

 Incomplete data, SAS9.2

Dear HS, dear all,

Here the results of SAS9.2 Proc MIXED

Denominator df option : residual
Type 3 Tests of fixed effects
Effect        Num DF    Den DF   F value    Pr > F
---------------------------------------------------
  Sequence         1        43  0.383808     0.3256
    Formul         1        43  0.383808     0.5388
    Period         1        43  0.056455     0.8133
90% Confidence intervals
 point est.  lower     upper
 0.963754    0.871885  1.065304


Denominator df option : satterthwaite
Type 3 Tests of fixed effects
Effect        Num DF   Den DF     F value   Pr > F
---------------------------------------------------
  Sequence        1   21.78633   0.988893    0.3309
    Formul        1   21.05951   0.383808    0.5422
    Period        1   21.05951   0.056455    0.8145
90% Confidence intervals
 point est.  lower     upper
 0.963754    0.869838  1.067811


Seems there is a very distinct definition, what residual df option is. :confused:

The SAS standard method is Containment (whatever this is) and this was what I reported earlier in this thread.
I think I must delve into my SAS/Stat users guide and struggle again with the dragon "The power to know".

Regards,

Detlew
ElMaestro
★★★

Denmark,
2008-10-30 09:57
(5628 d 04:53 ago)

@ d_labes
Posting: # 2611
Views: 28,179
 

 Missing values, imputation; separate study

Hey hey hey,

Having thought a little about this discussion, I almost feel for initiating a dry study investigating some of the issues in this thread. It would be interesting to find out which approach performs better (if an objective measure of that can be defined) when there are missing period data for one or more subjects. In particular it would be great FUN (and of potential interest to the PK oracles at EMEA) to play around with imputation methods and see if it is possible to find an imputation approach that outperforms linear mixed effect models.

Comments?

EM.
d_labes
★★★

Berlin, Germany,
2008-10-30 16:54
(5627 d 21:56 ago)

@ d_labes
Posting: # 2612
Views: 27,926
 

 Incomplete data, SAS9.2, DDF option residual

Dear all,

after RTFM and reading tons of I-net places dealing with SAS Proc MIXED I am more confused as before.
For German speakers: "Da steh ich nun, ich armer Tor und bin genauso klug als wie zuvor!" (Johann W. von Goethe - Faust).

The only thing I could figure out was that DDFM=residual creates degrees of freedom which neglect totally the covariance structure of the random effects in the model. Not a good idea in case of data from x-over trials, I think.
But I couldn't figure out yet what this option is good for.

Interesting enough this DDFM was the default in SAS before Version 8. Thus users not aware of this had gotten inappropriate F-test!

For our data the consequence is: Simply forget it. The degrees of freedom are not appropriate anyhow. Full stop.


@for HS: Could it be that the DDF for the sequence effect was 22 instead of 21? Then WinNonlin method could be compared to SAS DDFM=contain or DDFM=betweenwithin.

Regards,

Detlew
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2008-10-30 17:30
(5627 d 21:20 ago)

@ d_labes
Posting: # 2613
Views: 27,871
 

 Never belief in any software you haven't written yourself

Dear DLabes!

❝ after RTFM…


:lol3:

❝ Interesting enough this DDFM was the default in SAS before Version 8. Thus users not aware of this had gotten inappropriate F-test!


Wow!

❝ @for HS: Could it be that the DDF for the sequence effect was 22 instead of 21? Then WinNonlin method could be compared to SAS DDFM=contain or DDFM=betweenwithin.


Checked: it’s given with 21 indeed…
For other WinNonlin-Users I uploaded a workbook containing the full data set, the imbalanced one (#24 both periods missing, and the incomplete one (#24 2nd period missing) in separate sheets. Feel free to fiddle around with it…

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-11-07 06:12
(5620 d 08:38 ago)

@ d_labes
Posting: # 2640
Views: 27,793
 

 Imbalance or incomplete data, SASophylistic view

Dear DLabes,

In order to validate bear with your SAS outputs, could you please provide detailed anova output obtained from your SAS runs? Such as that we got anova output obtained from bear with Helmut's data (of balanced seq.) as follow:
Dependent Variable: Cmax

Type I SS
Analysis of Variance Table

Response: Cmax
          Df Sum Sq Mean Sq F value    Pr(>F)   
seq        1  239.9   239.9  5.1601   0.03324 *
prd        1   14.2    14.2  0.3053   0.58615   
drug       1    5.3     5.3  0.1133   0.73960   
seq:subj  22 5987.3   272.2  5.8547 5.389e-05 ***
Residuals 22 1022.6    46.5                     
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Type III SS
Single term deletions

Model:
Cmax ~ seq + subj:seq + prd + drug
         Df Sum of Sq    RSS    AIC F value     Pr(F)   
<none>                1022.6  198.8                     
prd       1      14.2 1036.8  197.5  0.3053    0.5861   
drug      1       5.3 1027.9  197.1  0.1133    0.7396   
seq:subj 22    5987.3 7010.0  247.2  5.8547 5.389e-05 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Tests of Hypothesis for SUBJECT(SEQUENCE) as an error term

Error: subj
          Df Sum Sq Mean Sq F value Pr(>F)
prd:drug   1  239.9   239.9  0.8813  0.358
Residuals 22 5987.3   272.2               

Error: Within
          Df  Sum Sq Mean Sq F value Pr(>F)
prd        1   14.19   14.19  0.3053 0.5861
drug       1    5.27    5.27  0.1133 0.7396
Residuals 22 1022.65   46.48


Thank you.

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
d_labes
★★★

Berlin, Germany,
2008-11-11 09:35
(5616 d 05:14 ago)

@ yjlee168
Posting: # 2648
Views: 27,690
 

 Imbalance or incomplete data, SASophylistic view

Dear Yung-jin,

sorry I'm a bit late because I missed your post.

Here are the complete ANOVA tables (the F-tests were already contained in the post above) from Proc GLM:

Complete dataset:
Dependent Variable: logCmax

                                Sum of
Source              DF         Squares     Mean Square    F Value    Pr > F
Model               25      5.57251321      0.22290053       5.68    <.0001
Error               22      0.86275554      0.03921616

Corrected Total     47      6.43526875


       Tests of Hypotheses for Mixed Model Analysis of Variance

Dependent Variable: logCmax

Source              DF     Type III SS     Mean Square    F Value    Pr > F

formul               1        0.014981        0.014981       0.38    0.5429
period               1        0.002975        0.002975       0.08    0.7856
Subject(Sequence)   22        5.308795        0.241309       6.15    <.0001

Error: MS(Error)    22        0.862756        0.039216

Source              DF     Type III SS     Mean Square    F Value    Pr > F

Sequence             1        0.245762        0.245762       1.02    0.3239

Error               22        5.308795        0.241309
Error: MS(Subject(Sequence))


Inbalanced dataset (Subject 24 excluded):
Dependent Variable: logCmax
                                Sum of
Source              DF         Squares     Mean Square    F Value    Pr > F

Model               24      5.05598321      0.21066597       5.16    0.0002
Error               21      0.85689625      0.04080458

Corrected Total     45      5.91287946


                           The GLM Procedure
       Tests of Hypotheses for Mixed Model Analysis of Variance

Dependent Variable: logCmax

Source              DF     Type III SS     Mean Square    F Value    Pr > F

formul               1        0.010764        0.010764       0.26    0.6129
period               1        0.004803        0.004803       0.12    0.7349
Subject(Sequence)   21        4.914505        0.234024       5.74    <.0001

Error: MS(Error)            21        0.856896        0.040805

Source              DF     Type III SS     Mean Square    F Value    Pr > F

Sequence             1        0.125256        0.125256       0.54    0.4725

Error               21        4.914505        0.234024
Error: MS(Subject(Sequence))


Type I sum of squares are not contained by default. If you need them, let me know. In case of balanced data they are same as type III, isn't it?

Regards,

Detlew
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-11-11 09:54
(5616 d 04:55 ago)

(edited by yjlee168 on 2008-11-11 11:39)
@ d_labes
Posting: # 2649
Views: 27,753
 

 Imbalance or incomplete data, SASophylistic view

Dear DLabes,

Thank you for your replied message. We've done the part of balanced and unbalanced data with SAS PROC GLM and PROC MIXED when we're waiting for your replied message. Fortunately, bear can produce the exactly same results as SAS. I will post both results at this Forum later. However, we need the part of incomplete data (obtained from both PROC GLM and PROC MIXED). Actually, we did that part a little with beat with lme package, too, and found there were some differences between SAS (with PROC MIXED and PROC MIXED) and bear.

Thanks.

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
d_labes
★★★

Berlin, Germany,
2008-11-11 14:13
(5616 d 00:37 ago)

@ yjlee168
Posting: # 2651
Views: 27,735
 

 Imbalance or incomplete data, SASophylistic view

Dear Yung-yin,

here comes the incomplete data
                          The GLM Procedure

Dependent Variable: logCmax

                                Sum of
Source              DF         Squares     Mean Square    F Value    Pr > F

Model               25      5.38508084      0.21540323       5.28    0.0001
Error               21      0.85689625      0.04080458

Corrected Total     46      6.24197709


                         The GLM Procedure
       Tests of Hypotheses for Mixed Model Analysis of Variance

Dependent Variable: logCmax

Source              DF     Type III SS     Mean Square    F Value    Pr > F

formul               1        0.010764        0.010764       0.26    0.6129
period               1        0.004803        0.004803       0.12    0.7349
Subject(Sequence)   22        5.162645        0.234666       5.75    <.0001

Error: MS(Error)    21        0.856896        0.040805


Source              DF     Type III SS     Mean Square    F Value    Pr > F

Sequence             1        0.250808        0.250808       1.09    0.3082

Error           22.166        5.110348        0.230549
Error: 0.9788*MS(Subject(Sequence)) + 0.0212*MS(Error)


                                  The Mixed Procedure

                                  Model Information

                Data Set                     WORK.BEBAC24LONG
                Dependent Variable           logCmax
                Covariance Structure         Variance Components
                Subject Effect               Subject
                Estimation Method            REML
                Residual Variance Method     Profile
                Fixed Effects SE Method      Model-Based
                Degrees of Freedom Method    Satterthwaite


                               Class Level Information

                 Class       Levels    Values

                 formul           2    R T
                 period           2    1 2
                 Subject         24    1 2 3 4 5 6 7 8 9 10 11 12 13
                                       14 15 16 17 18 19 20 21 22 23
                                       24
                 Sequence         2    RT TR


                                   Iteration History

              Iteration    Evaluations    -2 Res Log Like       Criterion

                      0              1        48.72149331
                      1              2        34.40082009      0.00001350
                      2              1        34.40051663      0.00000000


                                  The Mixed Procedure

                              Convergence criteria met.


                            Covariance Parameter Estimates

                       Cov Parm              Subject    Estimate

                       Subject(Sequence)     Subject      0.1002
                       Residual                          0.04102


                                    Fit Statistics

                         -2 Res Log Likelihood            34.4
                         AIC (smaller is better)          38.4
                         AICC (smaller is better)         38.7
                         BIC (smaller is better)          40.8


                            Type 3 Tests of Fixed Effects

                                  Num     Den
                    Effect         DF      DF    F Value    Pr > F

                    formul          1    21.1       0.38    0.5422
                    period          1    21.1       0.06    0.8145
                    Sequence        1    21.8       0.99    0.3309


                                  Least Squares Means

                          Standard
Effect  formul  Estimate     Error    DF  t Value  Pr > |t|   Alpha     Lower     Upper

formul  R         3.4844   0.07757  29.6    44.92    <.0001     0.1    3.3527    3.6161
formul  T         3.4475   0.07671  28.8    44.94    <.0001     0.1    3.3171    3.5779


                          Differences of Least Squares Means

                                        Standard
 Effect   formul   _formul   Estimate      Error     DF   t Value   Pr > |t|    Alpha

 formul   R        T          0.03692    0.05959   21.1      0.62     0.5422      0.1

                          Differences of Least Squares Means

                   Effect   formul   _formul      Lower       Upper

                   formul   R        T         -0.06561      0.1394

As you see, there is no such thing like the ANOVA table in Proc MIXED.

Regards,

Detlew
yjlee168
★★★
avatar
Homepage
Kaohsiung, Taiwan,
2008-11-11 14:35
(5616 d 00:15 ago)

@ d_labes
Posting: # 2653
Views: 27,717
 

 Imbalance or incomplete data, SASophylistic view

Dear DLabes,

Thanks for your quick response. Also, Helmut's good tip though I've used that way for a while.

❝ (...)

                         The GLM Procedure

       Tests of Hypotheses for Mixed Model Analysis of Variance

❝ (...)

❝ As you see, there is no such thing like the ANOVA table in Proc MIXED.


still has ANOVA Table with PROC MIXED? ;-) BTW, in your previous post, you showed a 90%CI as well as a point estimate for each run with SAS. What 90%CI method did you use? We got the results from bear and there were some different from yours after the second decimal digits.

All the best,
-- Yung-jin Lee
bear v2.9.1:- created by Hsin-ya Lee & Yung-jin Lee
Kaohsiung, Taiwan https://www.pkpd168.com/bear
Download link (updated) -> here
d_labes
★★★

Berlin, Germany,
2008-11-11 15:02
(5615 d 23:47 ago)

@ yjlee168
Posting: # 2655
Views: 27,871
 

 Imbalance or incomplete data, SASophylistic view

Dear Yung-ying,

❝ still has ANOVA Table with PROC MIXED? ;-)


The heading you cited is from Proc GLM :-P.
Under Proc MIXED there are no SSq nor MSq.
This is because the REML estimation method does not rely on SSq's and their decomposition.

❝ What 90%CI method did you use?


The shown point estimates and 90% CI's are from the difference between Least square means.

Here is my code for Proc GLM / mixed (to play around with, it seems you have access to "The power to know"):
ODS output LSMeandiffCL=ratios; *<- this saves the 90% CIs;
Proc GLM data=BEBAC24long;
  class formul period subject sequence;
  model logCmax=formul period sequence subject(sequence);
  random subject(sequence) /test;
  lsmeans formul/cl pdiff alpha=0.1;
run; quit;
*log diff R-T is calc. because of lexicographic order, but we need T-R;
Data ratios;
  set ratios;
  point=Exp(-difference);
  lower=exp(-upperCL);
  upper=exp(-lowerCL);
run;


ODS output Diffs=ratiosMixed;*<- this saves the 90% CIs;
Proc mixed data=BEBAC24long;
  class formul period subject sequence;
  model logCmax=formul period sequence /ddfm=satterth;
  random subject(sequence) /subject=subject;
  lsmeans formul/cl diff alpha=0.1;
run;

Regards,

Detlew
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2008-11-11 13:24
(5616 d 01:26 ago)

@ d_labes
Posting: # 2650
Views: 27,768
 

 Notification about new posts

Dear DLabes,

❝ sorry I'm a bit late because I missed your post.


Just some hints: ;-)
You may bookmark the Latest Posts instead of the forum for a quick overview what's going on…
Or you may use the RSS-Feed
Or you may opt in for e-mail notification about new posts in your personal profile…

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
d_labes
★★★

Berlin, Germany,
2008-11-11 14:24
(5616 d 00:26 ago)

@ Helmut
Posting: # 2652
Views: 27,621
 

 Notification about new posts

Dear HS,

thanx for the hints. But all that has not helped me in that case. I think I need some more time or money, or both :-D .

Regards,

Detlew
Helmut
★★★
avatar
Homepage
Vienna, Austria,
2008-11-11 14:39
(5616 d 00:11 ago)

@ d_labes
Posting: # 2654
Views: 27,943
 

 3 Prerequisites for biostatistics

Dear DLabes,

❝ I think I need some more time or money, or both :-D .


I was once told by a very famous biostatistician that there are only three things which actually count:
  • Hardware
  • Software
  • Boss
He didn’t mention Time/Money:cool:

Dif-tor heh smusma 🖖🏼 Довге життя Україна! [image]
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. 🚮
Science Quotes
d_labes
★★★

Berlin, Germany,
2008-11-11 15:09
(5615 d 23:40 ago)

@ Helmut
Posting: # 2656
Views: 27,781
 

 3 Prerequisites for biostatistics

Dear HS,

in the old days:
  • Hardware -> Paper, pencil
  • Software -> Staff's brain or some sort of both cheap and doesn't count much.
  • Boss: always ever the only factor which counts :-D

Regards,

Detlew
UA Flag
Activity
 Admin contact
22,957 posts in 4,819 threads, 1,638 registered users;
89 visitors (0 registered, 89 guests [including 8 identified bots]).
Forum time: 14:50 CET (Europe/Vienna)

Nothing shows a lack of mathematical education more
than an overly precise calculation.    Carl Friedrich Gauß

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