Linear Models - Mixed Model - SEs differ vastly from SPSS

[pblagov]

Hello,
I hope that someone with specialized knowledge of the Linear Models module, Mixed Models option, of jamovi 2.3.28.0 can take a look at this. After specifying a two-level multilevel model, making sure that the estimation method is the same, adding the random effect for the grouping variable, and testing a single fixed effect for the Level 2 predictor (covariate) with fixed intercepts, and testing several models like that with different sets of variables, jamovi consistently produced vastly different SEs for the estimate compared to SPSS. The estimates and p values for the estimates appear to be the same, but the SEs for the estimates are vastly different. I tried all the non-default options for estimating confidence intervals and centering, but that made the results depart from SPSS even more. Can someone help explain what is going on? I am concerned that this may be a flaw in the module, and I am weary about trying to publish results using it. Thank you ~ PB

[mcfanda@gmail.com]

Could you post an example ( a omv file)?

[mcfanda@gmail.com]

And a spss output to compare

[pblagov]

Thank you for your interest in time in responding to this question. Linked below are an SPSS data file and a jamovi data file with identical contents. The variables have been labelled for easy identification. Also linked are PDF bundles; each of them contains the same two analyses - one PDF is with SPSS, and the other PDF is with jamovi. Linked is also the SPSS output. The fixed-effects estimate ("Est.") in each of the two analyses does not different between SPSS and jamovi. But the degrees of freedom for the denominator ("Den df") differ dramatically, as do the SEs, CIs, and p values. It appears that SPSS and jamovi are treating the hierarchical structure of the dataset differently, but that should not be the case. In SPSS, the Participant ID was entered as a Participant variable, and in jamovi, the same Participant ID was entered as a clustering variable. This particular dataset has 10 observations per participant at Level 1 and participants are the clustering variable at Level 2. The dependent variable in each analysis is at Level 1, and there should be 1198 dfs of the denominator, unless I am mistaken (which is quite feasible). Here are the files:
https://drive.google.com/drive/folders/ ... sp=sharing

[mcfanda@gmail.com]

It seems that your spss model is not a mixed model. There are no random coefficients. In particular, in SPSS you need a /RANDOM command, which I couldn't find. With your SPSS code, you are running a simple regression. Indeed, if you run it in jamovi GLM, you get exactly the same results B=.4755, SE=.1512.

glm

f1.png (19.6 KiB) Viewed 7590 times

If you run a mixed model in SPSS, with at least a random intercept, you get this

f3.png (47.62 KiB) Viewed 7590 times

which is exactly what you get in jamovi.

I am also glad that SPSS latest version added the R-squared, that are in line with the R-squared jamovi has always produced:-)

[pblagov]

Thank you!!