Fitting mixed models in R (with lme4)
ฝัง
- เผยแพร่เมื่อ 16 ต.ค. 2024
- Learning Objectives:
Understand lmer syntax (fixed, random, cluster)
Understand how to interpret fixed effect parameters
Here's the dataset I'm using: quantpsych.net...
Here's my video that shows you how to identify your cluster variable: • How to identify your c...
Here's a video about how to determine whether an effect is fixed or random: • How to decide whether ...
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/r...
Undergraduate curriculum playlist (GLM-based approach): www.youtube.co....
Graduate curriculum playlist (also GLM-based approach): www.youtube.co....
Exonerating EDA paper: psyarxiv.com/5...
Download JASP (and visual modeling module): www.jasp-stat.org
Yes, I know my head is in the way of the output. Sorry! But you should still get an idea of how to do these things in R.
When will you upload the next video? I'm desperately waiting for it!
You're funny. I feel glad to came across your content!! I'm using lme4 for my dissertation project
07:40 unfortunately your head is blocking the summary table you are talking about :d
5 referred to fixed effect intercept estimate
3.261 referred to random effects residual Variance
1.8 referred to random effects residual SD
cba to do it for the next table atm
Your videos are excellent!Thanks for helping me understand this subject.
i know this is an annoying comment but pretty sure lme4 took so long to install not because it's a "big library" so much as because it was also compiled for your system. anyway, great content, thank youuu
In the CSV file, the variable is named dojo_id not jedi_id. I always tell my students, "KNOW THY DATA" and that applies to me too. Bad zim for blindly typing what was in the video instead of verifying with the data first :)
Good catch!
Interested in your view about the following question. Your third model allows slope(darkness,anger) to depend on jedi_id. You could also just add an interaction term darkness*factor(jedi_id) right? Assuming this model is identifiable (and I think it is because anger is varying over time within jedi_id) what is the value of the random effects model which makes the arbitrary assumption that the jedi-specific slopes follow a normal distribution?
Not a fistbump...😳You punched my face!🤣😂
Thank goodness for your videos you're so frrreakin FUNNIEEEEE!
As always, love ur videos, specially if they involve R. However, while it did not prevent me from following you, your face camera occluded the results window..just being picky :)
Different topic...do u consider uploading some machine learning tutorial for dummies?
Hello! Thank you for your videos, they help a lot. Would it be possible to get more information about the warnings one gets when fitting a random effect model, called "full" in your videos (e.g. singularity, did not converge, etc.) and how to solve them? I used the model comparison to verify whether I should fit a variable as a fixed effect only or as a fixed/random effect and some of my "full" models do not converge or "are singular". I also get warnings when trying to fit my final models (linear mixed model and generalized mixed models (poisson and negative binomial). Thank you for your help!
This was extremely helpful... Thank you so much for this!
Thanks, I found your video super helpful. Could you explain what it means when the error "singular model" apprears?
i have a repeated measures dataset. there are 90 patients with 3 observations each(at baseline 6 months and 12 months), so when made long i will have a dataset of 270 rows. The issue is i have a situation where 57 out of the 270 observations have missing data at 6months and 12months. How do i approach modelling this data using lmer? do i need to handle missing values or just model it like that? I would appreciate if you can help me or possibly we can even hop on a call, really need hep right now
No need to handle missing data, assuming the data are MAR or MCAR. lmer will drop them, but it will still estimate a regression line for those missing data at 6/12 months. You might have convergence issues, but the estimation should be unbiased (again, provided it's mar or mcar).
Hi! Thank you for this clear explanation. I followed the same for my data (data is in the same format that you have shown). My models worked fine up to fixed part. The moment I added the random part (1 + *** | ID), I got this error : number of observations (=100) < = number of random effect for term (1 + *** | ID).
What am I missing here or doing wrong?
It is not 1+*, I guess the * would mean everything but you can't do everything
Thank you so much for your wonderful video, I have a question, I have a model in which my fixed effect has 3 levels but when I ask for the summary one of the levels is not reported in the fixed effect, do you have any idea why is that?
This video is quite old but maybe you see it. Can I have multiple cluster variables? my dataset is over multiple years and in different areas. So as I understand it I would cluster it by the year but also to the areas?
You can, but it's more complicated and I don't have any videos on that. But you can google something like "3 level mixed models"
@@QuantPsych Wow thanks for the quick answer! I don't even know if its necessary 🐒
awesome video. I have a question : I have two clustering variables for a finance dataset where data is nested in banks and scores are time series reported quarterly. How to include the variable date in the script for the fixed model :
fixed_y = lmer(y ~ 1 + x1 + x2 + x4 + x4 + (1 | Bank))
Good question. I think the proper notation is "...(1|Bank:Time)"
Hi prof Fife, I have a question for you: let's assume you want to add age_started as a predictor of darkness. Of course age_started doesn't vary within each level of the jedi_id cluster variable (i.e., each jedi has just a unique value of age_started). Could you add it with no problems? Is it possible to use age_started as fixed effect + jedi_id as random effect together? Or would you encounter some separation / convergence issues? I'm struggling to understand this point but I feel kind of lost. Thank you for your enlightening videos!
Hello, thanks for uploading these tutorial videos. I'm a uni student trying to understand statistics in R and I found your channel is extremely helpful, thanks!
Besides, I'd like to know is it reasonable to consider a factor as fixed and random at the same time?
Looking forward to seeing you how explain the mixed model in following videos!
Yes. All random effects also have fixed effects. You cannot have a random effect without a fixed effect.
Head's in the way bro! 😜
Still luvs ya videos though, as always thanks for sharing 😊
Thank you..
If you can explain the difference between the model with ~1 and the last one, I 'd really appreciate that..
~1 just means to fit an intercept. If it's omitted, R will fit an intercept anyway. The first model (baseline) has to have it because there are no predictors (i.e., fitting this will throw an error: lmer(y~ ( | id), data=d)). For the remaining models, it's redundant, but I put it there so you can easily track what I've added to the model.
In there a scenario where makes sense to have random slope and intercept for jedi_id? If so, how would you code it? thx
I'm not sure I understand your question...cluster variables don't have random slopes and intercepts. Variables do.
Cool, but if we could actually see the output while you're interpreting it and not your face, that would be tremendously helpful. Informative nonetheless 👍.
Hey, i really like your video and your way of explaining things. Way clearer and simpler than other channels out here :D it is a great example on how mixed models work. However, i may have an understanding problem when it comes to the data format. At 3:40 you say that in their first year, when they are 5, they already killed someone. Does the dataset offers any time-specific variable? because the 5 stands for the point in time they started training, which should be always the same when looking at one jedi. (here jedi_id = jedi_1). Did I miss something? I'm only a poliscience student who stumbled upon that one :)
Ah, good point.
I was getting really worried that five year olds were murdering people, but you're right. They might have waited until they were seven :)
Dear Dustin I really enjoy your videos and love flexplot, but you always fit linear models. I'm a neonatologist, and almost 90% of my outcomes are binary o dichotomous. Is it possible to give usa video about non linear mixed models, aka logistic, ordinal etc.. ? I guess you'll use the nlme function but I'd really love to have your explanations. Thanks!!!
I'm pretty sure I made a generalized mixed model video. I think it was poisson mixed models, but it might have been logistic. Just take my logistic regression videos, combine those with my mixed models videos, and you'll get it.
Repeated measures....to make sense of it, I'd say it could be that every measure per jedi_id, could be the therapist measuring their anger level for every session?
That works :)
Definitely not watching this as a SAS to R transplant for my dissertation 😂
Your head is in front of the output 😕
I can't help having a fat head :)
Ignacio Wells
Lera Ferry
9292 Jordi Place
Why is it a 1? You reallj don't know? C'mon man! (Just call me Joe Biden). It obviously refers to a column of 1's in the design matrix.
Could you please stop with the clowning and get on with the explanations. Your attempt at deadpan humor sucks. You're way better at teaching, do that please !
No, it's so deeply ingrained in my personality. I do me, you do you.
And it seems you're outvoted. Everyone else seems to like it :)