Binomial multivariate repeated GLM
Locked
Re: Binomial multivariate repeated GLM
 
|
Ok, I conflated measurement and structural invariance. My error. Measurement invariance implies equality of (for continuous indicator items) factor loadings, intercepts and item variances. I am less confident about the requirements for ordinal indicators but definitely factor loadings and thresholds. Mplus has the delta and theta parameterization and for delta, the scale factors can be constrained and for theta the item variances can be constrained but there is quite a bit of complexity. They describe invariance for categorical indicators in the manual and, I think, on the demo videos/handouts and on the discussion list. I don’t completely understand but it seems that Muthen’s recommendations differ from Roger Millsap’s in his 2004 MBR article. >>Strong longitudinal measurement invariance basically does not exist when factor means increase or decrease drastically over time (as is the case in my sample). Not even the number of factors is likely to be invariant across time, because lower factor means = lower variation in the items = lower intercorrelation of items = lower probability to detect multiple factors. Given variables measured as real numbers (range +/- infinity, decimal values) from a multivariate normal distribution, there’s no reason for the variance or covariances to differ as factor means decrease. However, the typical 1-5 or 1-7 likert scale treated as continuous could well be a different story because of floor or ceiling effects. The same seems like it ought to be true for categorical variables because changing thresholds can only mean, I think, changing amounts of skew and, therefore, different correlations. Given what I know about your analysis, I’d use mplus. You have a question about factor composition. Certainly, separate ESEMs would give you insight into factor composition stability and let you also look at residual covariances. Regardless of whether you declare items to be categorical or continuous, you’re also going to find out about factor variances, item intercepts/thresholds and residual variances. Those numbers have to be similar for measurement invariance to hold. Gene From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of torvon Gene , You grasped my question perfectly - these are indeed 9 items of a pre-existing questionnaire that is supposed to measure one underlying disorder. We want to show, however, that symptoms react very differently to what happens to the subjects between time 1 and time 2. Unfortunately, I would not know what analysis to perform exactly to test whether symptoms change differentially over time. I can perform measurement invariance tests in MPLUS with ordered variables, ad know much more about MPLUS that SPSS, actually, but don't think it would help. Strong longitudinal measurement invariance basically does not exist when factor means increase or decrease drastically over time (as is the case in my sample). Not even the number of factors is likely to be invariant across time, because lower factor means = lower variation in the items = lower intercorrelation of items = lower probability to detect multiple factors. However, differential change of symptoms over time is just one of many explanations for lack of measurement invariance across time in samples with drastically increasing factor means, so measurement invariance wouldn't really directly tackle my question, or am I missing something? Would you have a recommendation here as to what to test? Also, item intercepts wouldn't exist in ordered models, so I couldn't compare these. Thank you for all the helpful comments so far ta-ta Eiko View this message in context: Re: Binomial multivariate repeated GLM |
|
|
In reply to this post by Alex Reutter
Alex, Bruce, Gene.
Let's stick to the modeling of observed variables for now. I'll read up on ways to model this in MPLUS differently and might just report both analyses.
I had the data coded in the "long long" format originally ("ID * time * symptoms" rows), the way you suggested it now. However, in the recommended post above (http://spssx-discussion.1045642.n5.nabble.com/Mixed-models-Repeated-Measures-multiple-DVs-SPSS-td5720624.html#a5720639) the data are set up differently: the variable x is used as pre-test DV, and the variable y as post-test DV. That's why I recoded accordingly to this syntax, and that's where the problems come from.
(1) Doing it as MIXED now with the long long format looks like this:
MIXED symptoms BY time phq_index /FIXED=time phq_index time*phq_index | SSTYPE(3)
/METHOD=REML /PRINT= g r SOLUTION
/RANDOM=INTERCEPT | SUBJECT(UserID) COVTYPE(VC) I added UserID as random intercept here, hope that makes sense. It gives me significant fixed effects "time" (because all symptoms increase), "symptoms" (because symptoms differ from each other?) and "index*time" (because symptoms change differentially over time). Correct interpretation and syntax?
(2) Doing it as GENLIN looks like this:
GENLIN symptoms (ORDER=ASCENDING) BY time phq_index (ORDER=ASCENDING) /MODEL phq_index time time*phq_index DISTRIBUTION=MULTINOMIAL LINK=CUMLOGIT
/PRINT CPS MODELINFO FIT SUMMARY SOLUTION. The model finds the same significant effects. (3) Lastly, GENLINMIXED: GENLINMIXED /DATA_STRUCTURE SUBJECTS=UserID
/FIELDS TARGET=symptoms /TARGET_OPTIONS DISTRIBUTION=MULTINOMIAL LINK=LOGIT /FIXED EFFECTS=phq_index time phq_index*time USE_INTERCEPT=TRUE
/RANDOM EFFECTS=UserID USE_INTERCEPT=TRUE SUBJECTS=UserID COVARIANCE_TYPE=VARIANCE_COMPONENTS /BUILD_OPTIONS TARGET_CATEGORY_ORDER=ASCENDING INPUTS_CATEGORY_ORDER=ASCENDING MAX_ITERATIONS=100
CONFIDENCE_LEVEL=95 DF_METHOD=RESIDUAL COVB=MODEL. Which just runs for an hour and then crashes SPSS. Tried it 3 times now.
Thank you very much for your time & help! Eiko On 16 July 2013 17:09, Alex Reutter [via SPSSX Discussion] <[hidden email]> wrote:
|
|
Hi Eiko,
I know this post may be a bit old but I have the similar research question like yours (with multinomial variables). I tried using the GENLINMIXED but it either crashes or I get this error message Repeated measurement analysis is not supported for the multinomial probability distribution'. Did you manage to analyze your data using GENLINMIXED in the end? Regards, Wendy |
|
|
Hi Wendy, Both Mplus and R can very easily deal with this problem, which is what I used in the end. Best Eiko On 17 April 2015 at 07:25, Wendylim [via SPSSX Discussion] <[hidden email]> wrote: Hi Eiko, |
| Free forum by Nabble | Edit this page |