thank you very much for your support! In fact the data is a Time series. IIncreasing the number of iterations had the expected outcome and now the algorithm converges. Thank you very much again for your support I hope you have a great week.
Original Message:
Sent: Wed June 07, 2023 11:34 AM
From: David Dwyer
Subject: Mixed Regression not working
Hi @Marcel Hinss
I asked a couple of our statisticians to take a look at your query and provide some observations. Here is the result:
We had a discussion regarding the query on
Mixed Regression not working and we have come up with the following observations :
- Why the client choose ARMA(1,1)? Is it a time series data? Did he check whether data set satisfies the requirements of ARMA(1,1)?
- If the ARMA conditions are not satisfied with the dataset, then we suggest to proceed with other types of covariance such as Compound symmetry, etc.
- We also suggest increasing the number of iterations.
Another suggestion is to apply other methods of outlier detection.
You have indicated that you are a PhD student. Perhaps the best place for you to begin will be with the Methodologist on your Dissertation committee. That person will best be able to help you with a research design and correct analytic techniques to match.
------------------------------
David Dwyer
SPSS Technical Support
IBM Software
Original Message:
Sent: Wed May 31, 2023 10:01 AM
From: Marcel Hinss
Subject: Mixed Regression not working
Hello and thank you for taking the time to help me!
I want to run a mixed regression on some data that I collected, but whenever I run my syntax the SPSS either shuts down immediately or just tells me that it is running the procedure but it never produces an output. ( I will provide some more details on what I have tried further below).
About the data:
I collected data from 29 participants, each participant completed 5 blocks of a behavioral task, and the dependent variables that I want to investigate are reaction time (continuous) and correct (binary). Around 50.000 trials in total.
Now within each block there were 336 trials. Each trial could be allocated to a condition based on several factors:
- SWITCH: 3 levels
- Task: 4 levels
- Congruency: 7 levels
- Block: 5 levels (as mentioned above)
The design was balanced and after removing some outliers remains almost perfectly balanced ( I also tried all analyses without outlier removal without any difference).
Now what was my plan:
I wanted to run a mixed regression, starting with ARMA(1,1), the slowly reduce the complexity of the covariance matrix, and then look at the fixed effects.
The Syntax I was using:
MIXED log_RT BY SWITCH Congruency Task Block
/CRITERIA=DFMETHOD(SATTERTHWAITE) CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0.00000001, RELATIVE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0, ABSOLUTE)
/FIXED=SWITCH*Congruency SWITCH*Task SWITCH*Block Congruency*Task Congruency*Block Task*Block SWITCH Congruency Task Block | SSTYPE(3)
/METHOD=ML
/PRINT=G SOLUTION TESTCOV
/RANDOM=INTERCEPT | SUBJECT(sbj) COVTYPE(VC)
/REPEATED=trialnb_overall | SUBJECT(sbj) COVTYPE(ARMA11).
As mentioned above when I run the analysis above SPSS stops running ( however it never takes up more than 15% of the computational power of my PC: specs (16Go RAM, Intel Core i7, 500Go SSD, Nvidia Quadro P620).
When I run the same analysis on a 10% subset of the data I get a warning that the Covariance matrix did not converge, or that the Hessian is not positive definite.
So, first of all ... sorry for the long text ... I hope this makes the issue clear.
I am not quite sure whether the issue is that my data is too big or if I am not doing the analysis that I was wanting to do.
I would be very grateful for any kind of advice.
Kind Regards,
your fellow lost PhD student.
------------------------------
Marcel Hinss
------------------------------