Multilevel modeling using SPSS (July, 2019) 

Thank you Leslie for visiting! I'm glad you found the video useful!

You're very welcome! best wishes

I'm so glad the video was helpful. Thank you for visiting!

I think you should not proceed with HLM further because the ICC value is below the 0.05 threshold.

Hi Adonis, the Generalized linear mixed model allows you to perform multilevel modeling where your level 1 outcome variable can be continuous or categorical (e.g., ordered categorical or nominal, count, etc., which generalizes to multilevel logistic, Poison regression etc). For example, here is something I put together awhile back on multilevel logistic regression using the generalized linear mixed model approach: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-roWTULimNPk.html Cheers!

Gender is technically not a continuous variable. However, since it is a binary variable it can be treated 'as if' it is continuous in the regression model (whether it is OLS regression or even multilevel regression, a binary predictor can be included in the model). The coding of 0 and 1 is based on a dummy coding system. The codings of 0 and 1 facilitate interpretation of the intercept in regression models. You might also check out a more recent video too: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-x5Z5KYODwNk.html

@Mike Crowson thanks for the helpful work you always share. what's more, is multilevelling model same as hierarchical regression? thanks in advance!

Hi there, Beatriz. SPSS can be used for multilevel analysis - and actually this video is an example of that. I also have another video demonstrating multilevel analysis in SPSS: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-x5Z5KYODwNk.html Here is the link to an entire book devoted to multilevel modeling in SPSS (which you should check out): www.routledge.com/Multilevel-and-Longitudinal-Modeling-with-IBM-SPSS/Heck-Thomas-Tabata/p/book/9780415817110 In your scenario, classroom serves as a level 2 unit whereas your students (level 1 units) are nested within those classroom. It is possible to model classroom-related predictors (those associated with the Level 2 units/classroom level) as predictors of a level 1 outcome; and that is demonstrated in this video (and the other referenced above). R and MPLUS are programs that also allow you to perform multilevel modeling (here's an example I made using R: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-8r9bUKUVecc.html). But they are not the only ones that allow you to do the analysis. Keep in mind that if you have only a small number of classrooms you are working with, that could pose a problem - as multilevel analysis generally requires a fairly large number of level 2 units (even though I have some literature indicating that perhaps a smaller number is ok; I'm not well-versed on that literature, however). I believe that you should have somewhere in the neighborhood of 30 level 2 units (e.g., classrooms) to use maximum likelihood estimation (the standard approach) to estimate model parameters; however, 15 or greater may be ok when estimating parameters with restricted maximum likelihood (which is actually the initial default in SPSS). I hope this is helpful to you. Best wishes!

It is because, after the standardisation, my grand mean values of predictors will be "0", and the predictors with the centre within the cluster will remain the same. By extension, my level 2 predictor will be "0" as well. This weird situation perplexes me......

Hi Chris, yes you can certainly do that using SPSS. I cover slopes as outcomes in the video: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-x5Z5KYODwNk.html Although I don't use the term 'slopes-as-outcomes', the cross-level interaction tested in the video is addressing whether the level 1 slope is a function of a level 2 predictor. I hope this helps!

Hi there. I have a video that demonstrates how these can be computed here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-NSFCibkocRA.html Also, I have a much more 'up-to-date' video that provides a lot more details on multilevel modeling with SPSS here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-x5Z5KYODwNk.html . It also includes a powerpoint link as well. And there are more resources here: sites.google.com/view/statistics-for-the-real-world/contents/multilevel-modeling-and-panel-regression Hope this helps!

Hi there. There is no global goodness of test of per se (as you might be used to in the context of SEM). However, you can evaluate the fit of a model by comparing it against others. Multilevel analysis tends to be carried out in a series of steps - often proceeding from a null model to increasingly complex models. When you have nested models, you can test whether the fit of the more complex model represents a significant improvement in fit relative to the reduced model using a chi-square difference test. I have a video (and calculator) for doing this here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-bzqMFnl3IV4.html You could also theoretically compare AIC values with competing models, with the preferred model among two being the one with the lowest AIC. [Although AIC was originally developed to compare non-nested models, it can still be used with nested models.] By the way, I have a more recent video on multilevel modeling in SPSS here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-x5Z5KYODwNk.html I hope this is helpful to you. Best wishes.

Hi Szeshing, I just saw your question. You can can compute pseudo-r-squares for within and between components using the Level 1 and Level 2 variance components [see table containing 'Estimates of covariance parameters' at the bottom of your output; look under Estimate]. For the between-groups pseudo- R-square, you can divided the variance of the Level 2 intercepts by [variance of level 1 residuals + variance of Level 2 intercepts]. You could compute a proportion of unexplained variance in y in a similar manner: variance at level 1 divided by variance of level 1 residuals+variance of Level 2 intercepts. I hope this helps! Another option might be to generate a pseudo R-square reflecting the proportionate reduction in a variance component associated with a full model against a reduced model (such as an intercept only model). For example, the proportionate reduction in variance in intercepts (i.e., between variance) can be computed as: 1 - [Level 2 variance full model / Level 2 variance reduced model ] I hope this helps!

@@mikecrowson2462 It does help, thanks.

Hi Qaiser, The links to the data and powerpoint should now be functional. Thanks for letting me know of the problem. Cheers!