multivariate multilevel model
Locked
 
|
I am running a multivariate multilevel model and testing if the effect of Gender (coded -0.5 and 0.5) varies across schools. Index1 is the indicator variable for the responses, treatment is coded -0.5 and 0.5, pretest is group mean centered (pretest_cen) and grand mean centerd (meanpretest_cen), and there is an interaction between treatment and gender (TxG).
When I run this code . . . MIXED Response BY Treat Index1 Gender WITH Pretest_cen MeanPretest_cen TXG /FIXED=Index1 Treat*Index1 Gender*Index1 Pretest_cen*Index1 MeanPretest_cen*Index1 TXG*Index1 | NOINT SSTYPE(3) /METHOD=ML /PRINT=G R SOLUTION /RANDOM=Index1 Gender | SUBJECT(School) COVTYPE(UN) /REPEATED=Index1 | SUBJECT(School*Student) covtype (UN). . . . I get this error message: Warnings Iteration was terminated but convergence has not been achieved. The MIXED procedure continues despite this warning. Subsequent results produced are based on the last iteration. Validity of the model fit is uncertain. If I move Gender from BY to WITH, the model converges but the estimates are wrong. Please help! Thank you! |
Re: multivariate multilevel model
|
|
Three off the cuff ideas:
1. In your syntax you put GENDER under BY but at the same time you have an interaction involving GENDER under the WITH keyword (TxG). This may cause problems with maximum likelihood estimates. 2. Besides, perhaps the number of cases in your study is not enough to support the number of variables and interactions in your model. Too many predictors and interactions certainly reduce the chances of a meaningful solution, even if not incurring in the above-mentioned problem of using GENDER under both BY and WITH. Notice that your specification ' BY Treat Index1 Gender' would create a group for each combination of values of these three variables, and they could be too many. 3. Last but not least, even models that are formally correct may fail to converge because they are not a good representation of the data. Changing the model specifications may get it to converge to a meaningful solution. By the way, you report that after moving GENDER to WITH the model converges but 'the results are wrong'. How do you know they are wrong? And what do you mean by 'wrong'? Are they just different from what you expected? In case of conflict between theory and data, it is theory (and the model expressing it) that should be questioned, not the data. HTH Hector -----Mensaje original----- De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de wchang Enviado el: Tuesday, July 02, 2013 11:56 Para: [hidden email] Asunto: multivariate multilevel model I am running a multivariate multilevel model and testing if the effect of Gender (coded -0.5 and 0.5) varies across schools. Index1 is the indicator variable for the responses, treatment is coded -0.5 and 0.5, pretest is group mean centered (pretest_cen) and grand mean centerd (meanpretest_cen), and there is an interaction between treatment and gender (TxG). When I run this code . . . MIXED Response BY Treat Index1 Gender WITH Pretest_cen MeanPretest_cen TXG /FIXED=Index1 Treat*Index1 Gender*Index1 Pretest_cen*Index1 MeanPretest_cen*Index1 TXG*Index1 | NOINT SSTYPE(3) /METHOD=ML /PRINT=G R SOLUTION /RANDOM=Index1 Gender | SUBJECT(School) COVTYPE(UN) /REPEATED=Index1 | SUBJECT(School*Student) covtype (UN). . . . I get this error message: Warnings Iteration was terminated but convergence has not been achieved. The MIXED procedure continues despite this warning. Subsequent results produced are based on the last iteration. Validity of the model fit is uncertain. If I move Gender from BY to WITH, the model converges but the estimates are wrong. Please help! Thank you! -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/multivariate-multilevel-model- tp5720966.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== 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 ===================== 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: multivariate multilevel model
|
|
In reply to this post by wchang
Regarding my first remark, note that in your syntax you attempt to estimate
the effect of treatment (Treat) and the interaction of treatment and gender (TxG) BY GENDER, i.e. within each gender. Since within males the gender is homogeneous, TxG = Treat for all those cases, and the same for the females. In other words, you appear to have two predictor variables (Treat and TXG) that are exact copies of each other insofar as you examine them BY GENDER. Taking TXG from the WITH variable list would eliminate this problem. In fact, using BY is a way of achieving the same effect, i.e. ascertaining whether the effect of treatment varies between genders. Hector -----Mensaje original----- De: Hector Maletta [mailto:[hidden email]] Enviado el: Tuesday, July 02, 2013 13:10 Para: 'wchang'; '[hidden email]' Asunto: RE: multivariate multilevel model Three off the cuff ideas: 1. In your syntax you put GENDER under BY but at the same time you have an interaction involving GENDER under the WITH keyword (TxG). This may cause problems with maximum likelihood estimates. 2. Besides, perhaps the number of cases in your study is not enough to support the number of variables and interactions in your model. Too many predictors and interactions certainly reduce the chances of a meaningful solution, even if not incurring in the above-mentioned problem of using GENDER under both BY and WITH. Notice that your specification ' BY Treat Index1 Gender' would create a group for each combination of values of these three variables, and they could be too many. 3. Last but not least, even models that are formally correct may fail to converge because they are not a good representation of the data. Changing the model specifications may get it to converge to a meaningful solution. By the way, you report that after moving GENDER to WITH the model converges but 'the results are wrong'. How do you know they are wrong? And what do you mean by 'wrong'? Are they just different from what you expected? In case of conflict between theory and data, it is theory (and the model expressing it) that should be questioned, not the data. HTH Hector -----Mensaje original----- De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de wchang Enviado el: Tuesday, July 02, 2013 11:56 Para: [hidden email] Asunto: multivariate multilevel model I am running a multivariate multilevel model and testing if the effect of Gender (coded -0.5 and 0.5) varies across schools. Index1 is the indicator variable for the responses, treatment is coded -0.5 and 0.5, pretest is group mean centered (pretest_cen) and grand mean centerd (meanpretest_cen), and there is an interaction between treatment and gender (TxG). When I run this code . . . MIXED Response BY Treat Index1 Gender WITH Pretest_cen MeanPretest_cen TXG /FIXED=Index1 Treat*Index1 Gender*Index1 Pretest_cen*Index1 MeanPretest_cen*Index1 TXG*Index1 | NOINT SSTYPE(3) /METHOD=ML /PRINT=G R SOLUTION /RANDOM=Index1 Gender | SUBJECT(School) COVTYPE(UN) /REPEATED=Index1 | SUBJECT(School*Student) covtype (UN). . . . I get this error message: Warnings Iteration was terminated but convergence has not been achieved. The MIXED procedure continues despite this warning. Subsequent results produced are based on the last iteration. Validity of the model fit is uncertain. If I move Gender from BY to WITH, the model converges but the estimates are wrong. Please help! Thank you! -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/multivariate-multilevel-model- tp5720966.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== 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 ===================== 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: multivariate multilevel model
|
|
In reply to this post by wchang
I'd like to understand your datastructure and model. Is this an accurate summary?
Schools were assigned, maybe randomly, maybe not, to condition and schools in one condition received something, something that schools in the other condition didn't receive. Students in all schools were tested twice with multiple tests. However, you are treating the pretest as a covariate; thus, eliminating time as a level. So if that is all true, then doesn't that give you as three level model: tests within students with in schools? Or, does the repeated statement represent the tests level? Thanks, Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of wchang Sent: Tuesday, July 02, 2013 10:56 AM To: [hidden email] Subject: multivariate multilevel model I am running a multivariate multilevel model and testing if the effect of Gender (coded -0.5 and 0.5) varies across schools. Index1 is the indicator variable for the responses, treatment is coded -0.5 and 0.5, pretest is group mean centered (pretest_cen) and grand mean centerd (meanpretest_cen), and there is an interaction between treatment and gender (TxG). When I run this code . . . MIXED Response BY Treat Index1 Gender WITH Pretest_cen MeanPretest_cen TXG /FIXED=Index1 Treat*Index1 Gender*Index1 Pretest_cen*Index1 MeanPretest_cen*Index1 TXG*Index1 | NOINT SSTYPE(3) /METHOD=ML /PRINT=G R SOLUTION /RANDOM=Index1 Gender | SUBJECT(School) COVTYPE(UN) /REPEATED=Index1 | SUBJECT(School*Student) covtype (UN). . . . I get this error message: Warnings Iteration was terminated but convergence has not been achieved. The MIXED procedure continues despite this warning. Subsequent results produced are based on the last iteration. Validity of the model fit is uncertain. If I move Gender from BY to WITH, the model converges but the estimates are wrong. Please help! Thank you! -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/multivariate-multilevel-model-tp5720966.html Sent from the SPSSX Discussion mailing list archive at Nabble.com. ===================== 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 ===================== 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 |
Automatic reply: multivariate multilevel model
|
|
I will be out of office until July 8th, with limited access to email. I will respond to your email when I return. |
|
|
In reply to this post by Hector Maletta
That you for your responses.
I am trying to replicate results obtained from MLwiN, which is the reason I said my estimates were wrong even when the model did converge. (I am a grad student working on a project with a professor, and my task is to run models in SPSS and SAS that the professor ran in MLwinN.)
The multivariate multilevel model is: Level 1 with 2 posttest scores, Level 2 with Gender and group mean centered Pretest as predictors, and Level 3 with Treatment and grand mean centered Pretest.
I removed TXG from the WITH variable list as you suggested but was still unable to get the same estimates as from MLwiN. There are 800 cases total in the dataset, each with 2 dependent variables.
The issue seems to have to do with putting Gender in the Random statement. Without Gender in the Random statement, the model runs regardless of whether Gender is listed as a BY or WITH variable. (I am running into the same problem with getting SAS to match MLwiN . . . I have no problems except when it comes to testing if the effect of Gender varies across schools).
More insight is greatly appreciated! On Tue, Jul 2, 2013 at 11:25 AM, Hector Maletta <[hidden email]> wrote: Regarding my first remark, note that in your syntax you attempt to estimate |
Re: multivariate multilevel model
Administrator
|
Re the TXG term, why would you not specify that interaction as Treat*Gender, like this:
MIXED Response BY Treat Index1 Gender WITH Pretest_cen MeanPretest_cen /FIXED=Index1 Treat*Index1 Gender*Index1 Treat*Gender*Index1 Pretest_cen*Index1 MeanPretest_cen*Index1 TXG*Index1 | NOINT SSTYPE(3) /METHOD=ML /PRINT=G R SOLUTION /RANDOM=Index1 Gender | SUBJECT(School) COVTYPE(UN) /REPEATED=Index1 | SUBJECT(School*Student) covtype (UN). Another observation: It seems odd to me to treat Gender as a random effect. Was that really done in the model you're trying to replicate? I'm not familiar with MLwIN, but maybe some other list members are. Can you post the MLwIN code for the model? HTH.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
Re: multivariate multilevel model
|
|
In reply to this post by wchang
In addition to what Bruce said, you said that you are trying to match mlwin results with both spss and sas. How different are they? Actually I’d like to see the mlwin syntax and output along with the spss output. Just the model fit and coefficients tables. Nothing else. Can you copy them into an email posting? Gene Maguin From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Wanchen Chang That you for your responses. I am trying to replicate results obtained from MLwiN, which is the reason I said my estimates were wrong even when the model did converge. (I am a grad student working on a project with a professor, and my task is to run models in SPSS and SAS that the professor ran in MLwinN.) The multivariate multilevel model is: Level 1 with 2 posttest scores, Level 2 with Gender and group mean centered Pretest as predictors, and Level 3 with Treatment and grand mean centered Pretest. I removed TXG from the WITH variable list as you suggested but was still unable to get the same estimates as from MLwiN. There are 800 cases total in the dataset, each with 2 dependent variables. The issue seems to have to do with putting Gender in the Random statement. Without Gender in the Random statement, the model runs regardless of whether Gender is listed as a BY or WITH variable. (I am running into the same problem with getting SAS to match MLwiN . . . I have no problems except when it comes to testing if the effect of Gender varies across schools). More insight is greatly appreciated! On Tue, Jul 2, 2013 at 11:25 AM, Hector Maletta <[hidden email]> wrote: Regarding my first remark, note that in your syntax you attempt to estimate
|
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |