Mixed Effects Model for group-randomised trial
Locked
Mixed Effects Model for group-randomised trial
 
|
Hi,
I am trying to evaluate an intervention. Unfortunately I have some issues with my data and don't know how best to deal with this. A brief background to my design: 1) Schools were randomised into intervention and waitlist-control conditions (I have 27 intervention, 26 control schools) 2) Parent-child dyads were invited to take part in a questionnaire study (control schools) or a parenting intervention (intervention schools). They were all offered $30 as an incentive. 3) I have N = 225 (104 of these are control-dyads). Data was collected at baseline and 1 yr follow-up. I want to know if there is change on child and parent outcome variables. Problems: 1) I have a significant difference in student age (m difference is 3 months) due to control data being collected mostly in the second half of baseline year (goups do not differ in age at follow-up as all data was collected at the same time). Age is mostly uncorrelated with my outcome variables both baseline or follow-up. 2) I also have significantly more male children in my intervention group. Again, gender mostly is uncorrelated with my outcome variables. 3) My parent report data shows significant baseline differences on almost all outcome variables with the intervention group reporting being more distressed prior to the intervention. (I don't have any difference on the same child reported variables). Question: I want to use Mixed Effects Model (SPSS) analyses and the syntax I have come up with is this. Is this correct?: MIXED DV BY Condition Time WITH Age /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Condition Time Age Condition*Time | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(VC) /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(VC) /EMMEANS=TABLES(Condition) /EMMEANS=TABLES(Time) /EMMEANS=TABLES(Condition*Time) COMPARE(Condition) /EMMEANS=TABLES(Condition*Time) COMPARE(Time). |
Re: Mixed Effects Model for group-randomised trial
Administrator
|
Christiane wrote, "I want to know if there is change on child and parent outcome variables."
Are the child and parent outcome variables the same? If so, then child v parent is just another explanatory variable in the model, right? If not, I think you'll need two models, one for children and one for parents. Also, I assume your primary interest is in Intervention vs Control, specifically in whether the amount (or pattern) of change varies by Condition. Finally, if the Child and Parent outcome variables are the same (and Child-Parent is an explanatory variable), then I think you have variables at 5 levels, like this: Level 5: Intervention v Control Level 4: School -- clustered in Intervention/Control Level 3: Family/Dyad -- clustered in School Level 2: Role/person (child v parent) -- clustered in Dyad Level 1: Occasion (baseline v follow-up) -- clustered in Person If there are different outcomes for children and parents, and two models, then there would be 4 levels: Level 4: Intervention v Control Level 3: School -- clustered in Intervention/Control Level 2: Person -- clustered in School Level 1: Occasion (baseline v follow-up) -- clustered in Person Is this correct?
--
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: Mixed Effects Model for group-randomised trial
|
|
Thank you for replying. I am quite stuck with this one!
"Are the child and parent outcome variables the same? If so, then child v parent is just another explanatory variable in the model, right? If not, I think you'll need two models, one for children and one for parents." I think I will need two models (one for outcomes rated by the parent, and one for outcomes rated by the child). I have measured outcomes with a parent self report (e.g, Child Depression Inventory parent version) and an adolescent self-report version(e.g, Child Depression Inventory child self-report version). Yes,my primary interest is in Intervention vs Control, specifically in whether the amount (or pattern) of change varies by Condition on my outcome variables (e.g. child depression). It's a group randomised design, and unfortunately I do have a slight selection bias (more distressed parents, but not children in the intervention group). I hope this answers your clarification. Does my syntax get at all the levels? Kind regards from Melbourne, Christiane |
Re: Mixed Effects Model for group-randomised trial
|
Administrator
|
I have a couple comments/questions.
1. Why are you not including Sex as a factor in the model, especially since you are concerned about some imbalance? 2. Some authors (e.g. Twisk) recommend always starting with an unspecified covariance structure. After inspecting the unconstrained covariance matrix this produces, you may be able to impose some other structure (which will save you some degrees of freedom). 3. Have you centered Age on some convenient in-range value (e.g., a value near the minimum, or the mean)? This will make interpretation easier, particularly for the intercept. So I think your model would be something like: MIXED DV BY Condition Time Sex WITH Age /FIXED=Condition Time Sex Age Condition*Time Condition*Sex Time*Sex Condition*Time*Sex | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION TESTCOV /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(UN) /EMMEANS=TABLES(Condition) /EMMEANS=TABLES(Time) /EMMEANS=TABLES(Sex) /EMMEANS=TABLES(Condition*Time) /EMMEANS=TABLES(Condition*Sex) /EMMEANS=TABLES(Time*Sex) /EMMEANS=TABLES(Condition*Time*Sex) . 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/). |
|
|
Bruce,
Generally speaking, there is no difference between the following two RANDOM statements: /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(VC) The UNstructured specification is not necessary if all that is being estimated is a RANDOM intercept. Put another way, the intercept has nothing with which to covary, so there is no need to specify a covariance structure which allows for such covariance. Now, if we were estimating both a RANDOM intercept and slope, such as: /RANDOM=INTERCEPT x | SUBJECT(Schoolnumb) COVTYPE(UN) then the UNstructured covariance matrix may be appropriate. Ryan On Tue, Jun 7, 2011 at 8:44 AM, Bruce Weaver <[hidden email]> wrote: > I have a couple comments/questions. > > 1. Why are you not including Sex as a factor in the model, especially since > you are concerned about some imbalance? > > 2. Some authors (e.g. Twisk) recommend always starting with an unspecified > covariance structure. After inspecting the unconstrained covariance matrix > this produces, you may be able to impose some other structure (which will > save you some degrees of freedom). > > 3. Have you centered Age on some convenient in-range value (e.g., a value > near the minimum, or the mean)? This will make interpretation easier, > particularly for the intercept. > > > So I think your model would be something like: > > MIXED DV BY Condition Time Sex WITH Age > /FIXED=Condition Time Sex Age > Condition*Time Condition*Sex Time*Sex Condition*Time*Sex | > SSTYPE(3) > /METHOD=REML > /PRINT=SOLUTION TESTCOV > /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) > /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(UN) > /EMMEANS=TABLES(Condition) > /EMMEANS=TABLES(Time) > /EMMEANS=TABLES(Sex) > /EMMEANS=TABLES(Condition*Time) > /EMMEANS=TABLES(Condition*Sex) > /EMMEANS=TABLES(Time*Sex) > /EMMEANS=TABLES(Condition*Time*Sex) > . > > > HTH. > > > > christiane wrote: >> >> Thank you for replying. I am quite stuck with this one! >> >> >> "Are the child and parent outcome variables the same? If so, then child v >> parent is just another explanatory variable in the model, right? If not, >> I think you'll need two models, one for children and one for parents." >> I think I will need two models (one for outcomes rated by the parent, and >> one for outcomes rated by the child). I have measured outcomes with a >> parent self report (e.g, Child Depression Inventory parent version) and an >> adolescent self-report version(e.g, Child Depression Inventory child >> self-report version). >> >> >> Yes,my primary interest is in Intervention vs Control, specifically in >> whether the amount (or pattern) of change varies by Condition on my >> outcome variables (e.g. child depression). >> >> It's a group randomised design, and unfortunately I do have a slight >> selection bias (more distressed parents, but not children in the >> intervention group). >> >> I hope this answers your clarification. Does my syntax get at all the >> levels? >> >> Kind regards from Melbourne, >> >> Christiane >> > > > ----- > -- > Bruce Weaver > [hidden email] > http://sites.google.com/a/lakeheadu.ca/bweaver/ > > "When all else fails, RTFM." > > NOTE: My Hotmail account is not monitored regularly. > To send me an e-mail, please use the address shown above. > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4461642.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: Mixed Effects Model for group-randomised trial
|
Administrator
|
Absolutely right, Ryan. Good catch. If I'd looked at some of my own MIXED examples, I would have been reminded that I usually don't even bother including COVTYPE() when there is a random intercept only.
--
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/). |
|
|
Still, Bruce, I think you raise an important point regarding an
UNstructured covariance type, particularly with regard to handling repeated measures. I can't remember if there are more than two time points and/or if the OP thinks the residual variance may be different depending on the condition and/or time point. If yes to either of these points, then he/she should consider replacing the second RANDOM statement with a REPEATED statement and make the appropriate specifications. Typically, the RANDOM intercept statement is equivalent to a REPEATED statement with a compound symmetric (CS) structure. If a CS structure is too restrictive, then a REPEATED statement should be considered in place of the second RANDOM statement. Ryan On Tue, Jun 7, 2011 at 6:10 PM, Bruce Weaver <[hidden email]> wrote: > Absolutely right, Ryan. Good catch. If I'd looked at some of my own MIXED > examples, I would have been reminded that I usually don't even bother > including COVTYPE() when there is a random intercept only. > > > > R B wrote: >> >> Bruce, >> >> Generally speaking, there is no difference between the following two >> RANDOM statements: >> >> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) >> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(VC) >> >> The UNstructured specification is not necessary if all that is being >> estimated is a RANDOM intercept. Put another way, the intercept has >> nothing with which to covary, so there is no need to specify a >> covariance structure which allows for such covariance. Now, if we were >> estimating both a RANDOM intercept and slope, such as: >> >> /RANDOM=INTERCEPT x | SUBJECT(Schoolnumb) COVTYPE(UN) >> >> then the UNstructured covariance matrix may be appropriate. >> >> Ryan >> >> On Tue, Jun 7, 2011 at 8:44 AM, Bruce Weaver >> <[hidden email]> wrote: >>> I have a couple comments/questions. >>> >>> 1. Why are you not including Sex as a factor in the model, especially >>> since >>> you are concerned about some imbalance? >>> >>> 2. Some authors (e.g. Twisk) recommend always starting with an >>> unspecified >>> covariance structure. After inspecting the unconstrained covariance >>> matrix >>> this produces, you may be able to impose some other structure (which will >>> save you some degrees of freedom). >>> >>> 3. Have you centered Age on some convenient in-range value (e.g., a value >>> near the minimum, or the mean)? This will make interpretation easier, >>> particularly for the intercept. >>> >>> >>> So I think your model would be something like: >>> >>> MIXED DV BY Condition Time Sex WITH Age >>> /FIXED=Condition Time Sex Age >>> Condition*Time Condition*Sex Time*Sex Condition*Time*Sex | >>> SSTYPE(3) >>> /METHOD=REML >>> /PRINT=SOLUTION TESTCOV >>> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) >>> /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(UN) >>> /EMMEANS=TABLES(Condition) >>> /EMMEANS=TABLES(Time) >>> /EMMEANS=TABLES(Sex) >>> /EMMEANS=TABLES(Condition*Time) >>> /EMMEANS=TABLES(Condition*Sex) >>> /EMMEANS=TABLES(Time*Sex) >>> /EMMEANS=TABLES(Condition*Time*Sex) >>> . >>> >>> >>> HTH. >>> >>> >>> >>> christiane wrote: >>>> >>>> Thank you for replying. I am quite stuck with this one! >>>> >>>> >>>> "Are the child and parent outcome variables the same? If so, then child >>>> v >>>> parent is just another explanatory variable in the model, right? If >>>> not, >>>> I think you'll need two models, one for children and one for parents." >>>> I think I will need two models (one for outcomes rated by the parent, >>>> and >>>> one for outcomes rated by the child). I have measured outcomes with a >>>> parent self report (e.g, Child Depression Inventory parent version) and >>>> an >>>> adolescent self-report version(e.g, Child Depression Inventory child >>>> self-report version). >>>> >>>> >>>> Yes,my primary interest is in Intervention vs Control, specifically in >>>> whether the amount (or pattern) of change varies by Condition on my >>>> outcome variables (e.g. child depression). >>>> >>>> It's a group randomised design, and unfortunately I do have a slight >>>> selection bias (more distressed parents, but not children in the >>>> intervention group). >>>> >>>> I hope this answers your clarification. Does my syntax get at all the >>>> levels? >>>> >>>> Kind regards from Melbourne, >>>> >>>> Christiane >>>> >>> >>> >>> ----- >>> -- >>> Bruce Weaver >>> [hidden email] >>> http://sites.google.com/a/lakeheadu.ca/bweaver/ >>> >>> "When all else fails, RTFM." >>> >>> NOTE: My Hotmail account is not monitored regularly. >>> To send me an e-mail, please use the address shown above. >>> >>> -- >>> View this message in context: >>> http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4461642.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 >> > > > ----- > -- > Bruce Weaver > [hidden email] > http://sites.google.com/a/lakeheadu.ca/bweaver/ > > "When all else fails, RTFM." > > NOTE: My Hotmail account is not monitored regularly. > To send me an e-mail, please use the address shown above. > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4463258.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: Mixed Effects Model for group-randomised trial
|
Administrator
|
From the original post: "Data was collected at baseline and 1 yr follow-up."
--
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: Mixed Effects Model for group-randomised trial
|
|
In reply to this post by Bruce Weaver
Thank you all for taking the time and answering my question! I have found all suggestions though provoking and helpful!
Qu: 1. Why are you not including Sex as a factor in the model, especially since you are concerned about some imbalance? A: Due to my between groups imbalance, I have considered 'child age' and 'gender' as covariates for all my analyses, however, child age is uncorrelated with all my outcome variables at baseline and follow-up. It is my understanding that there is no point including a covariate if it is unrelated to the DV (but correct me if I'm wrong!). Gender correlates with two of my DV's. I have run your suggested syntax, and for one DV gender/interactions with gender do not reach significance, although the Model is improved by including gender (as indicated by lower AIC). For the other variable it turns out that gender is the only significant predictor. Qu: 2. Some authors (e.g. Twisk) recommend always starting with an unspecified covariance structure. After inspecting the unconstrained covariance matrix this produces, you may be able to impose some other structure (which will save you some degrees of freedom). A: yes, I did start off using unspecified covariance structure, but as per the other comments, SPSS just changes UN to identity. Qu: 3. Have you centered Age on some convenient in-range value (e.g., a value near the minimum, or the mean)? This will make interpretation easier, particularly for the intercept. A: Yes, I used grand mean centering. Thank you again, Christiane So I think your model would be something like: MIXED DV BY Condition Time Sex WITH Age /FIXED=Condition Time Sex Age Condition*Time Condition*Sex Time*Sex Condition*Time*Sex | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION TESTCOV /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(UN) /EMMEANS=TABLES(Condition) /EMMEANS=TABLES(Time) /EMMEANS=TABLES(Sex) /EMMEANS=TABLES(Condition*Time) /EMMEANS=TABLES(Condition*Sex) /EMMEANS=TABLES(Time*Sex) /EMMEANS=TABLES(Condition*Time*Sex) . HTH.
|
Re: Mixed Effects Model for group-randomised trial
|
|
In reply to this post by Bruce Weaver
Thanks again for your help. I have just one last query regarding the estimated means output:
For two of my outcomes I get a non-significant interaction between time and condition (suggesting the intervention did not help), but when looking at the table from the same analyses produced by requesting '/EMMEANS=TABLES(Condition*Time) COMPARE(Time)' I get a significant p value suggesting that the estimated change for the intervention group reached significance p = .016, but the control group did not change p = .250. I have checked the assumptions and all appears ok. Kinds regards, Christiane |
|
|
Christiane,
The group difference in mean change from time 1 to time 2 is not significant (p>.05), even though the mean change from time 1 to time 2 for one group is significantly different than zero (p=.016) while the mean change from time 1 to time 2 for the other group is not significantly different than zero (p=.250). This is all plausible. Ryan On Fri, Jun 10, 2011 at 3:05 AM, christiane <[hidden email]> wrote: > Thanks again for your help. I have just one last query regarding the > estimated means output: > > For two of my outcomes I get a non-significant interaction between time and > condition (suggesting the intervention did not help), but when looking at > the table from the same analyses produced by requesting > '/EMMEANS=TABLES(Condition*Time) COMPARE(Time)' I get a significant p value > suggesting that the estimated change for the intervention group reached > significance p = .016, but the control group did not change p = .250. > > I have checked the assumptions and all appears ok. > > Kinds regards, > > Christiane > > > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4475397.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: Mixed Effects Model for group-randomised trial
|
|
Thanks for all your help Ryan, |
Automatic reply: Mixed Effects Model for group-randomised trial
|
|
I am currently out of the office, but will return Monday, June 20. I will have intermittant e-mail access through 6/15, but will not have
access after that time. I will follow up with you 6/20 or so. If your question is urgent, please contact Kathy Cook, 503-883-2561. |
|
|
In reply to this post by Bruce Weaver
I just saw this message. This topic is interesting enough for me to
respond, despite my response being a couple days late. :-) With only two time points, the following RANDOM statement probably suffices: /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) which, by the way, is typically equivalent to /REPEATED=TIME | SUBJECT(Student*Schoolnumb) TYPE=CS If we suspected, however, that the variances differed across both measurement periods, then one could capture this invariance by employing a REPEATED statement with an UNstructured covariance structure. HTH, Ryan On Wed, Jun 8, 2011 at 5:57 PM, Bruce Weaver <[hidden email]> wrote: > From the original post: "Data was collected at baseline and 1 yr follow-up." > > > R B wrote: >> >> Still, Bruce, I think you raise an important point regarding an >> UNstructured covariance type, particularly with regard to handling >> repeated measures. I can't remember if there are more than two time >> points and/or if the OP thinks the residual variance may be different >> depending on the condition and/or time point. If yes to either of >> these points, then he/she should consider replacing the second RANDOM >> statement with a REPEATED statement and make the appropriate >> specifications. Typically, the RANDOM intercept statement is >> equivalent to a REPEATED statement with a compound symmetric (CS) >> structure. If a CS structure is too restrictive, then a REPEATED >> statement should be considered in place of the second RANDOM >> statement. >> >> Ryan >> >> On Tue, Jun 7, 2011 at 6:10 PM, Bruce Weaver >> <[hidden email]> wrote: >>> Absolutely right, Ryan. Good catch. If I'd looked at some of my own >>> MIXED >>> examples, I would have been reminded that I usually don't even bother >>> including COVTYPE() when there is a random intercept only. >>> >>> >>> >>> R B wrote: >>>> >>>> Bruce, >>>> >>>> Generally speaking, there is no difference between the following two >>>> RANDOM statements: >>>> >>>> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) >>>> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(VC) >>>> >>>> The UNstructured specification is not necessary if all that is being >>>> estimated is a RANDOM intercept. Put another way, the intercept has >>>> nothing with which to covary, so there is no need to specify a >>>> covariance structure which allows for such covariance. Now, if we were >>>> estimating both a RANDOM intercept and slope, such as: >>>> >>>> /RANDOM=INTERCEPT x | SUBJECT(Schoolnumb) COVTYPE(UN) >>>> >>>> then the UNstructured covariance matrix may be appropriate. >>>> >>>> Ryan >>>> >>>> On Tue, Jun 7, 2011 at 8:44 AM, Bruce Weaver >>>> <[hidden email]> wrote: >>>>> I have a couple comments/questions. >>>>> >>>>> 1. Why are you not including Sex as a factor in the model, especially >>>>> since >>>>> you are concerned about some imbalance? >>>>> >>>>> 2. Some authors (e.g. Twisk) recommend always starting with an >>>>> unspecified >>>>> covariance structure. After inspecting the unconstrained covariance >>>>> matrix >>>>> this produces, you may be able to impose some other structure (which >>>>> will >>>>> save you some degrees of freedom). >>>>> >>>>> 3. Have you centered Age on some convenient in-range value (e.g., a >>>>> value >>>>> near the minimum, or the mean)? This will make interpretation easier, >>>>> particularly for the intercept. >>>>> >>>>> >>>>> So I think your model would be something like: >>>>> >>>>> MIXED DV BY Condition Time Sex WITH Age >>>>> /FIXED=Condition Time Sex Age >>>>> Condition*Time Condition*Sex Time*Sex Condition*Time*Sex | >>>>> SSTYPE(3) >>>>> /METHOD=REML >>>>> /PRINT=SOLUTION TESTCOV >>>>> /RANDOM=INTERCEPT | SUBJECT(Schoolnumb) COVTYPE(UN) >>>>> /RANDOM=INTERCEPT | SUBJECT(Student*Schoolnumb) COVTYPE(UN) >>>>> /EMMEANS=TABLES(Condition) >>>>> /EMMEANS=TABLES(Time) >>>>> /EMMEANS=TABLES(Sex) >>>>> /EMMEANS=TABLES(Condition*Time) >>>>> /EMMEANS=TABLES(Condition*Sex) >>>>> /EMMEANS=TABLES(Time*Sex) >>>>> /EMMEANS=TABLES(Condition*Time*Sex) >>>>> . >>>>> >>>>> >>>>> HTH. >>>>> >>>>> >>>>> >>>>> christiane wrote: >>>>>> >>>>>> Thank you for replying. I am quite stuck with this one! >>>>>> >>>>>> >>>>>> "Are the child and parent outcome variables the same? If so, then >>>>>> child >>>>>> v >>>>>> parent is just another explanatory variable in the model, right? If >>>>>> not, >>>>>> I think you'll need two models, one for children and one for parents." >>>>>> I think I will need two models (one for outcomes rated by the parent, >>>>>> and >>>>>> one for outcomes rated by the child). I have measured outcomes with a >>>>>> parent self report (e.g, Child Depression Inventory parent version) >>>>>> and >>>>>> an >>>>>> adolescent self-report version(e.g, Child Depression Inventory child >>>>>> self-report version). >>>>>> >>>>>> >>>>>> Yes,my primary interest is in Intervention vs Control, specifically in >>>>>> whether the amount (or pattern) of change varies by Condition on my >>>>>> outcome variables (e.g. child depression). >>>>>> >>>>>> It's a group randomised design, and unfortunately I do have a slight >>>>>> selection bias (more distressed parents, but not children in the >>>>>> intervention group). >>>>>> >>>>>> I hope this answers your clarification. Does my syntax get at all the >>>>>> levels? >>>>>> >>>>>> Kind regards from Melbourne, >>>>>> >>>>>> Christiane >>>>>> >>>>> >>>>> >>>>> ----- >>>>> -- >>>>> Bruce Weaver >>>>> [hidden email] >>>>> http://sites.google.com/a/lakeheadu.ca/bweaver/ >>>>> >>>>> "When all else fails, RTFM." >>>>> >>>>> NOTE: My Hotmail account is not monitored regularly. >>>>> To send me an e-mail, please use the address shown above. >>>>> >>>>> -- >>>>> View this message in context: >>>>> http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4461642.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 >>>> >>> >>> >>> ----- >>> -- >>> Bruce Weaver >>> [hidden email] >>> http://sites.google.com/a/lakeheadu.ca/bweaver/ >>> >>> "When all else fails, RTFM." >>> >>> NOTE: My Hotmail account is not monitored regularly. >>> To send me an e-mail, please use the address shown above. >>> >>> -- >>> View this message in context: >>> http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4463258.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 >> > > > ----- > -- > Bruce Weaver > [hidden email] > http://sites.google.com/a/lakeheadu.ca/bweaver/ > > "When all else fails, RTFM." > > NOTE: My Hotmail account is not monitored regularly. > To send me an e-mail, please use the address shown above. > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Mixed-Effects-Model-for-group-randomised-trial-tp4447503p4471060.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: Mixed Effects Model for group-randomised trial
|
|
Thanks Ryan, |
|
|
In reply to this post by Bruce Weaver
Hi there,
I am wondering if anyone would like to help me with an analysis- I have a dataset from 500 individuals and 26 clusters- clusters were randomized to receiving either the intervention or control. The individuals in both groups filled out some Quality of life questionnaires and I am trying to analyse the data to see whether the intervention had any effect on the quality of life although the clusters were randomized- individual is the unit of analysis- Cuold you please let me know how you would go about doing this? Many thanks, |
| Free forum by Nabble | Edit this page |