covariance structure in repeated measure using mixed model
Locked
covariance structure in repeated measure using mixed model
 
|
Thank you, Eugene, for your reply.
I read some materials and some suggested to use Likelihood ratio test to compare different covariance types. in my case, if I use UN (unstructured), the fit statistics: -2 Res Log Likelihood 3459.9 AIC (smaller is better) 3479.9 AICC (smaller is better) 3480.7 BIC (smaller is better) 3509.6 If I use AR(1), I got: -2 Res Log Likelihood 4137.2 AIC (smaller is better) 4141.2 AICC (smaller is better) 4141.2 BIC (smaller is better) 4147.1 And CS got slightly higher values for all the ICs. the chi2 is 677.3, df=10-2=8, so using UN is much better than using AR(1). I have only one group of subject (N=140), who were measured at 4 different times. In this case, unstructured covariance structure is reasonable? Thanks Rongjin Guan Rutgers Univ ________________________________________ From: SPSSX(r) Discussion [[hidden email]] On Behalf Of Maguin, Eugene [[hidden email]] Sent: Thursday, September 24, 2015 9:36 AM To: [hidden email] Subject: Re: SPSS repeated measure: how to do paired comparison? I'd like comment only on #3. Each choice implies a set of assumptions about the residual covariance matrix in terms of the number of non-zero values and a possible algebraic relationship between/among values. Together, those two elements define the number of parameters for the repeated effects shown in the model dimensions table. As someone somewhat recently pointed out, models can be compared by means of the difference in the log likelihoods against the difference in the degrees of freedom, i.e., number of parameters, for each model. (In a structural equation model, an added requirement is that the two models be nested. I don't know if that applies here.) The difficulty--and this may be the problem that you encounter--is that some residual covariance assumptions may imply the same number of parameters for a specific within structure. Others may be more knowledgeable but all you have for decision criteria are the information criteria numbers. Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Rongjin Guan Sent: Wednesday, September 23, 2015 2:38 PM To: [hidden email] Subject: Re: SPSS repeated measure: how to do paired comparison? Hi Ryan I have several more questions: 1. when I used the following syntax, I got each comparison in individual tables, but in SAS, these comparisons are grouped in the same table. Is SPSS, is there a way to group those test output into a single table? MIXED lstgmb BY Time /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Time | SSTYPE(3) /METHOD=REML /PRINT= SOLUTION /REPEATED=Time | SUBJECT(ClientID) COVTYPE(AR1) /TEST 'Time 2 versus Time 1' Time -1 1 0 0 /TEST 'Time 3 versus Time 1' time -1 0 1 0 TEST 'Time 4 versus Time 1' time -1 0 0 1 /TEST 'Time 3 versus Time 2' time 0 -1 1 0 /TEST 'Time 4 versus Time 2' time 0 -1 0 1 /TEST 'Time 4 versus Time 3' time 0 0 -1 1 2. Although the T-values, p values are all the same between SAS and SPSS output, the DFs are different for example, in SAS, the DF for all the comparison is 178, and in SPSS, each comparison has a different df value, not an integer, like 207.854. I do not know which DF to report. Do you have any idea about this? 3. How to choose the covariance structure type? I used AR(1). There are many different choices. Thank you Rongjin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: covariance structure in repeated measure using mixed model
|
|
Yes, unstructured is much better.
Gene Maguin -----Original Message----- From: Rongjin Guan [mailto:[hidden email]] Sent: Tuesday, September 29, 2015 2:50 PM To: Maguin, Eugene <[hidden email]>; [hidden email] Subject: covariance structure in repeated measure using mixed model Thank you, Eugene, for your reply. I read some materials and some suggested to use Likelihood ratio test to compare different covariance types. in my case, if I use UN (unstructured), the fit statistics: -2 Res Log Likelihood 3459.9 AIC (smaller is better) 3479.9 AICC (smaller is better) 3480.7 BIC (smaller is better) 3509.6 If I use AR(1), I got: -2 Res Log Likelihood 4137.2 AIC (smaller is better) 4141.2 AICC (smaller is better) 4141.2 BIC (smaller is better) 4147.1 And CS got slightly higher values for all the ICs. the chi2 is 677.3, df=10-2=8, so using UN is much better than using AR(1). I have only one group of subject (N=140), who were measured at 4 different times. In this case, unstructured covariance structure is reasonable? Thanks Rongjin Guan Rutgers Univ ________________________________________ From: SPSSX(r) Discussion [[hidden email]] On Behalf Of Maguin, Eugene [[hidden email]] Sent: Thursday, September 24, 2015 9:36 AM To: [hidden email] Subject: Re: SPSS repeated measure: how to do paired comparison? I'd like comment only on #3. Each choice implies a set of assumptions about the residual covariance matrix in terms of the number of non-zero values and a possible algebraic relationship between/among values. Together, those two elements define the number of parameters for the repeated effects shown in the model dimensions table. As someone somewhat recently pointed out, models can be compared by means of the difference in the log likelihoods against the difference in the degrees of freedom, i.e., number of parameters, for each model. (In a structural equation model, an added requirement is that the two models be nested. I don't know if that applies here.) The difficulty--and this may be the problem that you encounter--is that some residual covariance assumptions may imply the same number of parameters for a specific within structure. Others may be more knowledgeable but all you have for decision criteria are the information criteria numbers. Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Rongjin Guan Sent: Wednesday, September 23, 2015 2:38 PM To: [hidden email] Subject: Re: SPSS repeated measure: how to do paired comparison? Hi Ryan I have several more questions: 1. when I used the following syntax, I got each comparison in individual tables, but in SAS, these comparisons are grouped in the same table. Is SPSS, is there a way to group those test output into a single table? MIXED lstgmb BY Time /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Time | SSTYPE(3) /METHOD=REML /PRINT= SOLUTION /REPEATED=Time | SUBJECT(ClientID) COVTYPE(AR1) /TEST 'Time 2 versus Time 1' Time -1 1 0 0 /TEST 'Time 3 versus Time 1' time -1 0 1 0 TEST 'Time 4 versus Time 1' time -1 0 0 1 /TEST 'Time 3 versus Time 2' time 0 -1 1 0 /TEST 'Time 4 versus Time 2' time 0 -1 0 1 /TEST 'Time 4 versus Time 3' time 0 0 -1 1 2. Although the T-values, p values are all the same between SAS and SPSS output, the DFs are different for example, in SAS, the DF for all the comparison is 178, and in SPSS, each comparison has a different df value, not an integer, like 207.854. I do not know which DF to report. Do you have any idea about this? 3. How to choose the covariance structure type? I used AR(1). There are many different choices. Thank you Rongjin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |