Multilevel modeling equivalent to random effects panel regression (SPSS demo)

[@lucal.1487]

Thanks a lot, it really helped. Is it also possible to carry out the Hausman test in SPSS?

[@MahamudTawel]

Thanks Mike, I have benefited a lot from your videos on fixed and random effects regressions , I wish you could show us how can we choose which is the appropriate model using SPSS results. Thanks again.

[@DB-in2mr]

Very informative! thanks Mike. This cross view between stata and spss modelling is really cool. Thanks again for this nice piece of tutorial/presentation

[@mikecrowson2462]

Hi Daniele, I'm glad you found this useful! best wishes

[@christinedanitha4871]

Thank you Mike, how to detecting outliers ?

[@felo13579]

Dear Professor Mike Crowson, Thank you very much for the lessons! I wanted to clarify some doubts: how exactly to model time units in mixed models? For example, in SPSS's Generalized Linear Mixed Models should I put the time units in Repeated Measures and as a fixed effect, and can I treat individuals as a random effect? Or just individuals as a random effect? I've seen videos where they only include individuals as a random effect and don't put time into Repeated Measures. I have this doubt as I need to model a neutropenia effect over time for patients receiving treatment that also varies over time. Could you give me some suggestions? Thanks in advance for your help.

[@farnazmarianoshokrifard8446]

Hello, can you suggest, how to enter the data in SPSS if we want to make a within-country mixed-effects logistic regression model for the two factors?

[@juiliballal361]

Thank you for this informative video. I have a few things to ask. When I inserted the control variables in Covariates box, the model was ready in quick time. However, when I insert independent variables in Covariates box, it is taking a lot of time to load results. Also, could you please let me know where should we insert moderator term? Thank you. Appreciate your help.

[@929bleh]

Thank you for the video. Could I use this to look at a laboratory marker to predict a binary outcome measure (sepsis vs no sepsis), but where some of the markers are taken from the same patient at different times during their stay in hospital (ie they were tested on separate unrelated occasions for infection)? So can this work with a dichotomous outcome?

[@mikecrowson2462]

Hi there, if you have a binary (repeated measures) outcome (e.g., sepsis vs no at time 1, time 2...) with that outcome nested within individual/patient then it would seem you may have two options. Option 1 might be multilevel binary logistic regression. Option 2 could be using generalized estimating equations with a binary outcome. Both of these can be accomplished via SPSS. I don't have any videos on Option 2, but here's a video demo using cross-sectional data for Option 1. Both approaches (including option 1 with repeated measures data) are described in www.routledge.com/Multilevel-Modeling-of-Categorical-Outcomes-Using-IBM-SPSS/Heck-Thomas-Tabata/p/book/9781848729568. Cheers!

[@929bleh]

@@mikecrowson2462 Thanks for your reply! I was actually looking at doing GEE with binary outcome measure ie, splitting into episode numbers, working out the odds ratio of each episode and then pooling the data, as in this video (kzbin.info/www/bejne/iKXWlYevmbqcfdk - 9 mins 18 seconds), is this what you meant by Option 2? I saw your video on Option 1 multiple binary logistic regression, but it seemed to have completely different patients in each 'group'. In my case, in my 'groups' (constituting episode number of infection, 1-10), there will be the same patients in each group. Would that still work? Which would be the most accurate method, if any? Thanks so much!

[@nicholaslentz4689]

First: Thanks for your videos. These helped me a lot! One question about this: Why did you only add the constant term to the random effects and not the correlation coefficients as well? The reason for my question is that I carry out a panel data analysis myself (using the random effects model) with 19 covariates and I am not sure which random effects are to be calculated? In most literature, only the intercept and 1 regression coefficient are actually considered. Is it also possible to calculate the RE for all slopes at once or does this make any sense at all?

[@mikecrowson2462]

Hi Nicholas, I made this video in response to inquiries by students about whether it is possible to perform panel regression in SPSS - as you can using other programs such as R or Stata. In Stata, for instance, you have menu options for directly performing fixed and random-effects panel regression - and neither of these approaches depend on the user adopting a multilevel framework when specifying the models. In another video (kzbin.info/www/bejne/a3Wtg2mAaK13gbc), I demonstrated how one might obtain fixed effects estimates using SPSS using standard OLS regression (based on Allison's 2009 presentation: us.sagepub.com/en-us/nam/fixed-effects-regression-models/book226025). Of course, this model treats between cluster information as something to be covaried out, so it does not allow you to model cluster-level predictors of the outcome variable (although you could still model interactions across levels). This video was simply to demonstrate how you can obtain the same output as a random effects panel regression using ML estimation (an option in Stata) in SPSS by using a multilevel framework and randomly varying the intercepts. [Side note: Although not shown here, it is also possible to run a panel regression in SPSS using Generalized Estimating Equations in the same way you could in Stata. Again, this would essentially be a 'single-level' approach to modeling]. I personally prefer the multilevel framework because it allows you to treat intercepts and slope parameters as randomly varying - and you can model the covariation in those parameters as well. So, I'm not advocating the approach in the video as much as I am demonstrating how you can obtain the same results for ML panel regression by recasting it into a multilevel framework. The question about the number of slopes to randomly vary is a good one. In my experience, randomly varying a lot of slopes at once is more likely to result in convergence failures. As you add in more random slopes (with or without covariances), you are more likely to end up with variance estimates at - or pretty darned near - zero. This will result in model convergence problems or estimates that cannot be trusted. You will notice in pedagogical books and articles on multilevel modeling, the authors generally start with simpler models and then add parameters in a series of steps (i.e., models of increasing complexity). At points along the way, if model convergence becomes a problem, you are able to better diagnose the source of the problem and address it. But if specify your model in a single step - adding everything at once - then you (a) are more likely to have problems with your model converging AND (b) are less able to diagnose the possible sources of the problem & remedy them. I should also mention that you probably want to be choosy about those slope parameters that you are randomly varying. Apart from the convergence problems that are likely to result if you randomly vary all 19 slopes for your predictors, you should ask yourself whether the additional parameters being estimated might yield additional useful information. And ordinarily if the a slope exhibits significant variation, you would attempt to explain that variation by including a cross-level interaction in your model specification. If you do not have any theoretical framework as a guide, then you could just simply be adding in additional parameters that do not contribute as much to your understanding of the phenomenon. It's kind of like adding in correlated errors in the context of SEM. We might be able to 'improve the fit' of a model by adding in correlated errors, but that improvement in fit comes at the cost of parsimony - and may offer little in the way of additional useful information. I hope that makes sense :) FYI, I have a number of other videos on multilevel modeling. Here's one using cross-sectional data in SPSS: kzbin.info/www/bejne/rma9Zn6PhKmqhM0 Here's another on growth curve modeling in SPSS: kzbin.info/www/bejne/ZmnIfayBna-Nqs0 Thanks for your thought-provoking questions. They were a good way to start my Sunday morning! :) Cheers, and good luck with your research!

[@nicholaslentz4689]

@@mikecrowson2462 Hi Mike, Thank you for the very detailed, quick and helpful answer! Overall, I find the SPSS and Stata comparison very helpful in understanding the topic. Do you possibly also know whether it is possible to carry out the Hausman test in SPSS? I did a lot of research and came to the conclusion that at least by default there is no way to do this in SPSS (which is relatively easy with Stata). Best regards from Germany

[@lenakatharina9369]

Thank you Mike, this video really helped me a lot. I do have one question though: I am currently working on a uni project where we have been given panel data. However, our variables do not have a linear trajectory course. Is it still possible to use a random effects panel regression with these variables or do we have to adjust them i.e. ln-transform them? Additionally, would it be possible to include moderating effects in a panel regression? Thank you so much in advance!

[@yulinliu850]

Thank-you Mike!

[@mikecrowson2462]

You are very welcome. Thanks for watching, Yulin!