Mixed Model Analysis: Real Example
In this video, I'm analyzing the data from this paper: journals.plos.org/plosone/art...
I briefly mention the idea of visual partitions. It wasn't until after I recorded the video, I realized I'd never explained those on my channel. Here's a paper that explains the idea: psyarxiv.com/avu2n/
Link about EDA versus CDA: • Ethics in Statistics P...
My Multivariate playlist: • Multivariate Statistics
And here's a paper I wrote about my eight step approach to data analysis: psyarxiv.com/r8g7c/
Undergraduate curriculum playlist (GLM-based approach): kzread.info?list...
Graduate curriculum playlist (also GLM-based approach): kzread.info?list...
Exonerating EDA paper: psyarxiv.com/5vfq6/
Download JASP (and visual modeling module): www.jasp-stat.org
Пікірлер: 24
I cannot thank you enough! I realize all KZread how-to vids use narration, but I relate SO MUCH to the way you think out loud at each point in the model building process 🔥
Oh my god this video is amazing... This really gives me the insight into how powerful the R is...
I love your energy and the way you explain things! Tysm :)))
Love this video…and I think I also love you 😂 you are so cool and helpful.
Please do that convergence video issues :)
Hey! Cool video, thanks! but... have you ever considered doing a video to explain how to deal with some warnings? For example: "boundary (singular) fit: see ?isSingular". It would be actually very useful.
@QuantPsych
Жыл бұрын
I have. It's on my to-do list! Here's what I'd do: (1) scale the variables (if you can), (2) simplify the model (if you can), (3) use a Bayesian mixed model. There may have been one or two other things to do, I just can't remember off the top of my head.
Do you have videos like this one that show your process of adjusting for confounding within a linear mixed model???
Thanks for your videos, they are helpful for some modelling I'm attempting! QUICK QUESTION - Usually I see one of the random effects of model described as repeated measures/same people across multiple time points. I have data where I have multiple time points but slight differences in the groups of people (i.e. most are different people but some people show up across multiple time points). How would this type of data be handled in a mixed model? Would I ignore this as a random effects factor as most people are not present in the data across multiple time points?
I have a question on reporting LME models. When doing them in r, spss, or jamovi, you get the summary of the anova table and f values. What is the recommended way to report these models? F values like an anova framework or estimates like in regression. I see them reported both ways, so I wanted to get your thoughts on it. Should the anova table be paid any attention to, and instead, we jump right into the estimates? Thank you.
Heyo, Dustin, thank you for your enthusiasm and effort in teaching! I've started using "flexplot" and am trying to utilize it for analysing GLMM (log-link, Gamma distr), and it doesn't compile any graphs... when calling the 'visualize(model, plot="model")' R gives out this message Error in draw_axis(break_positions = guide$key[[aesthetic]], break_labels = guide$key$.label, : lazy-load database '/Library/Frameworks/R.framework/Versions/4.2/Resources/library/gtable/R/gtable.rdb' is corrupt my question is - is flexplot suitable for plotting, analysing (also for R-sq statistic) when used with GLMM (glmer) not just linear MM? many thanks and keep it up! :)
Nice videos! Do you have one where you explain why we should use mixed-effect modelling instead of doing separate regressions on data subsets?
@QuantPsych
23 күн бұрын
Kinda of. You can check out this video: kzread.info/dash/bejne/Z6iDy8iGZZTAf84.html
I need your suggestions for my data analysis using mixed models. Please let me know if you are available on zoom or teams at the earliest. It would be of great help.
How can I use multilevel modeling with a binary outcome variable? Any chance you can do a video example of that? I have an ordinal independent variable (3), ordinal moderator (3), and annual time waves (10).
@QuantPsych
Жыл бұрын
something like this: require(lme4) mod = glmer(outcome~var1 + var2 + Time + (Time | cluster), family="binomial", data=d) Or something similar.
Hi, thanks for the video! I did not really yet understand these two things: 1. why do we compare two possible models, does that have to do with controlling the interaction of the effects (trying to make sure that no things are interpreted as output effects that in reality are interactions that are only occuring due to the combination in the model)? And: How can I interpret a PB Test run by using the Modcomp function? This would be extremely helpful. Thanks for the video
@nosaosawe3158
7 ай бұрын
For your first uestion: You want to be sure you are not making the model to complex (with many independent variables) so you compare with a simpler model. In this context, it helps to decide which should remain as random and fixed effects. Simplifications are just the best. It helps with interpretation too.
I couldn't install flexplot, it says it's not available for the latest version of R. Any chances there's a workaround for this? thank you!
@Elrorosa
Жыл бұрын
did you try on Github?
@QuantPsych
Жыл бұрын
Hard to tell. You could post an issue on github.
I sympathize with your general approach to data analysis, but I also think that you can not disentangle the study design from the analysis and the question asked, which might have been the case here. The last models use only fixed effects without interaction for group and/or time, so group effects are estimated including baseline and follow-up measurements, where the first has nothing to do with the intervention but the group effects were interpreted as such. These models completely neglect potential treatment effects (e.g. modelled as interaction term as was done in the previous models). Actually, the self-care had the lowest baseline values and the lowest post-pre difference of about 1.3 where other groups were higher than 2, which might explain the lower group effect for the self-care group in your analysis. However, these effects are overall very small... I think it would rather be interesting to see, how large the treatment effect in disability is, while controlling for pain intensity (which was an interesting idea). While this is a randomized trial, an ANCOVA approach could be used, which has a nice mixed model translation: disability ~ 1 + time + group:time - see here doi.org/10.1080/00273171.2013.831743 and here: solomonkurz.netlify.app/post/2022-06-13-just-use-multilevel-models-for-your-pre-post-rct-data I am enjoying your videos and would love to see a follow-up video on that topic someday!
@QuantPsych
Жыл бұрын
I think I know what you're saying, but let me say it in my own words: I can't just ignore the interaction between group and time because the groups have different baseline levels. If I just evaluate main effects, I'm really (kinda sorta) asking whether the groups have different baselines levels (which they might by chance). Is that what you mean?
@robinschafer597
Жыл бұрын
@@QuantPsych yes, that is the main argument. At least in this case, when treatment effects are overall small this seems to apply. The main effect for group would be kind of a weighted average over all time points including baseline. Useful questions you can answer with that term in such a model in RCTs are very rare I'd say...