Multilevel binary logistic regression using IBM SPSS (March 2020)

[@diarlisoncosta516]

Thank you! I hope you can make a video about Multilevel Multinomial Logistic regression using SPSS.

[@yulinliu850]

Many thanks, Mike! One question: in the ICC of Model 1 and Model 2, where does 3.29 come from? Should it be 1, the variance of leve-1 residual term?

[@mikecrowson2462]

Hi Yulin, Heck et al. (2014) state that the residual variance is scaled to 1 and is not tested for significance. They note: "It is merely a means of providing a metric for the underlying latent predictor" (which is eta; i.e., variance of predicted logits). They state that "The variance of a logistic distribution with scale factor of 1 is pi^2/3, or approximately 3.29. (p. 157). Hope this helps!

[@yulinliu850]

@@mikecrowson2462 Thank you for the helpful reply!

[@martatessema4875]

@@mikecrowson2462 Thank you dear it was also my Q

[@timpowers9320]

Hi Mike. Love your videos. Have a question on the ICC calculation in this video (around 14 min mark). I see where the level 2 variance (intercept) comes from, but was unsure where the "3.29" variance for level 1 came from. The residual variance was the scaled identity of 1.000 on the output ... so wondering how the 3.29 was devised or sourced.

[@erikthibaut7220]

Do you have any idea yet? I have the same question :-)

[@anaidlindemann4268]

@@erikthibaut7220 Me too!

[@anandsundaralingam7843]

Thanks Mike, great vid. If a random effects variance (intercept) is non-significant can you safely conclude there is no random effect and can revert to a fixed effects model?

[@mikecrowson2462]

Hi Anand, a significant random intercept variance is generally regarded as an indication of clustering in the data. However, I would not base my entire decision on it. I would also consult the ICC for evidence of clustering. See the Powerpoint (drive.google.com/file/d/16UJsWJodaVFdxJesu7OTQFgGWtrsITzv/view) that accompanies this video - slides 16 and 17 for more details. It is possible to have non-trivial clustering even though the significance test is non-significant. Cheers.

[@willdenomme1]

Great video! One question though: why does the output for the multilevel binary logistic regression use F and T significance tests, while the single-level fixed-effects binary logistic regression uses Chi-square and Wald statistics in the output (on SPSS)? I also noticed that R uses Z scores to compute significance tests instead of any of the aforementioned significance test.

[@xiaoyuanzhao4713]

Very helpful! Can you also make a video of the same model but with repeated measures?

[@Johnceau]

Thank you Mike! Your clear and easy to follow explanation helped me a lot!

[@shukryzawahir6075]

Thank you Mike, It's really useful. I have a question, If I have a continues level 2 predictor how should i treat in the model.

[@caijimin7806]

Thanks Mike. I wonder whether this method can be applied to multilevel multinomial logistic regression?

[@andreab2114]

Does the formula to calculate the ICC (variance/(variance + 3.29) apply to any model including multilevel multinomial?

[@Duressa1000]

It's different for multilevel linear regression where it's calculated from table of estimates of covariance parameters. It's calculated as 'intercept variance'/(intercept variance + residual).

[@이명화-l7v]

In SPSS multilevel binary analysis, how do you analyze categorical independent variables? The output displays the message "Duplicate parameters" ..Please make sure to respond, thank youㅜㅜ

[@Chipalo23]

Hi Mike, I have enjoyed learning about the multilevel regressions. Do you also have videos on multilevel linear rgressions?

[@johncapitanio7294]

Thanks Mike, very helpful, as with your other videos. I have one question that you didn't touch on, and that is cross-classified models. Pretty easy in MIXED, but there is very little out there for GLMM. I am interested in how x, y, and z impact my binary outcome, but I am working with animals and must control for relatedness. So I am using dam_id and sire_id as level 2 effects. I am not interested in them, per se, just want my fixed effects to be "net" of any contribution due to relatedness. The problem is I don't how to specify my subject variable in the Data Structure menu. Should *both* dam_id and sire_id be listed there? And then later identified as the random effects? Or should only 1 be listed (and does it matter which one)? Thanks for any advice.

[@bipinthapa3338]

Thank you very much! This is so clear and pretty much helpful.

[@edonuzi]

Thanks for the lovely video. I want to ask how can I combine clustering with neural network using spss? How to cluster using k-means to filter the data and use the filtered (training data) as input for the neural network?

[@bilalasghar9616]

Thank you so much!!! One more time...

[@mustafaali5222]

Do you have a video that shows how to account for repeated measures in the GLMM model if we have longitudinal/panel data?

[@vincentchirimwami6599]

Thank you, Mike.

[@carlosgarciarivero]

Thank you very much for your excellent work and making it public. One question if I may: How can I perform a 3-Level logistic binary analysis. My question is about how to incorporate the third level. Do you have any video or doc about it? Thanks a lot again!!

[@ermiasayalew1475]

Mike I really like your video. You are great instructor. I was following your tutorial. I was working using dhs data. The first problem I encounter is when I did the unconditional model. When I add intercept for the fixed effect I get error and the variance result will be to large. When I omit the fixed intercept it will run normally without error. What shall I do? What is the result trying to tell me?

[@anikogregor2046]

Very informative video, thanks! At the same time, I know it's not a troubleshooting forum but has anyone bumped into not having an output displayed of GENLINMIXED command in SPSS 28 on Mac? I found no solution anywhere. :/

[@Alejandrosluck]

Thank you very much Mike, it's really helpful. May I ask what are the consequences of not specifiying "subject combination" in random effects? Since the documentation says it's "optional"...

[@929bleh]

Thanks Mike! I am doing multilevel binomial logistic regression with repeat measures per subject (each repeat totally unrelated and at random time points), I have added each time point as my repeat measure and used Scaled Identity. But do I also need to add intercept for Random effect?

[@luamghebrehiwot5150]

I am planning the same. Did you figure it out?!

[@indrayadigunardi4695]

Thank you so much

[@northenights]

mr if i can know what version is your SPSS is?

[@mikecrowson2462]

Hi there. When I put this together it was either version 25 or 26. I'm now on 27. There's not any difference in terms of how to carry out the analysis between these versions. Best wishes

[@northenights]

@@mikecrowson2462 i use SPSS 25 but the output not same as yours. may i have your email for consultation with you?

[@abat5991]

Sorry to say that this is extremely not understandable. Try another practical example and talk much less than you did in this video. Focus on what multilevel means, how to do that on SPSS, and the interpretation of the findings.