Timeline for Estimates radically change when including Random Slopes in Multiple Logistic Regression
Current License: CC BY-SA 4.0
31 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 1, 2019 at 13:59 | comment | added | dtoribio | @BenBolker, I'd really appreciate if you could have a look at the data and send me your impressions on the issue. | |
| Feb 27, 2019 at 0:00 | history | tweeted | twitter.com/StackStats/status/1100546240285167616 | ||
| Feb 26, 2019 at 9:21 | comment | added | dtoribio | @BenBolker, I just shared the data. Feel free to run any analyses you consider appropriate. | |
| Feb 26, 2019 at 9:20 | history | edited | dtoribio | CC BY-SA 4.0 |
added 83 characters in body
|
| Feb 25, 2019 at 15:09 | comment | added | dtoribio |
@BenBolker, I run the analysis with the GLMMadaptive package without success. The model detects a large coefficient value and warns about that an "overly complex model is fitted to the data". In addition, this does not seem to solve the problem with the estimation of the log odds. Even the model without random effects estimates an intercept log odd of 30.93.
|
|
| Feb 25, 2019 at 10:12 | history | edited | dtoribio | CC BY-SA 4.0 |
added 271 characters in body
|
| Feb 25, 2019 at 10:06 | vote | accept | dtoribio | ||
| Feb 25, 2019 at 9:44 | history | edited | dtoribio | CC BY-SA 4.0 |
added 2980 characters in body
|
| Feb 21, 2019 at 23:48 | comment | added | Ben Bolker |
I would also strongly recommend that you try Gauss-Hermite quadrature instead of the default Laplace approximation; you might need the GLMMadaptive package to do this for your more complex random-effects model ...
|
|
| Feb 21, 2019 at 23:43 | comment | added | Ben Bolker | any chance that you can post your data somewhere ... ??? | |
| Feb 21, 2019 at 23:34 | history | edited | kjetil b halvorsen♦ | CC BY-SA 4.0 |
deleted 27 characters in body
|
| Feb 21, 2019 at 19:15 | answer | added | EdM | timeline score: 2 | |
| Feb 21, 2019 at 9:16 | history | edited | dtoribio | CC BY-SA 4.0 |
deleted 4 characters in body
|
| Feb 21, 2019 at 9:08 | comment | added | dtoribio | @EdM, I forgot to mention, my interest in modelling with random effects is to see if the fixed effects of my manipulations remain after accounting for the intra-individual sensitivity to these manipulations. If my approach is wrong, please, do not hesitate to point it out! | |
| Feb 21, 2019 at 9:00 | comment | added | dtoribio | @EdM, as you can see in my update, there is no coding problem with any of the variables. I added the frequencies for each within-subject condition, and as you can see Punishment is substantially more likely than No Punishment (could this be key?). | |
| Feb 21, 2019 at 8:54 | history | edited | dtoribio | CC BY-SA 4.0 |
added 2623 characters in body
|
| Feb 21, 2019 at 8:46 | history | edited | dtoribio | CC BY-SA 4.0 |
added 2623 characters in body
|
| Feb 20, 2019 at 23:48 | comment | added | EdM |
As you have an almost completely balanced design, it might be helpful to show a table of the percentage of choices of Punishment for each of the 4 combinations of Ambiguity and Uncertainty, putting aside for now any random effects. A small sample of your dichm data frame (maybe the results of head(dichm)) might also help troubleshooting. Note that you still need to think about just what you want to model with the random effects; that's the important statistical issue here.
|
|
| Feb 20, 2019 at 22:00 | comment | added | dtoribio | @GerardSanroma thank you for pointing out the intercept issue! Could you explain more in detail what you mean with the estimations produced by the model? Perhaps I can share them after discarding the coding problem. Cheers! | |
| Feb 20, 2019 at 21:57 | comment | added | dtoribio | @EdM, thank you very much for your comments and the references! But first, I´ll definitely check if there is any coding mistake and get back to you guys if the problem persists. Regarding the number of observations and ID values, that is exactly right, I have 4 repeated measures of every participant (163 subjects and a total of 651 observations) with one missing value. | |
| Feb 20, 2019 at 20:58 | comment | added | EdM | @GerardSanroma has a good point: the magnitude of the intercept (which I had been ignoring) is enormous. The intercept in the first model is the log-odds of PunishmentYes/No when there is no Ambiguity and no Uncertainty, giving an odds ratio of over 22000/1. Even in the case of AmbiguityYes an Uncertainly/No, the odds ratio only drops to about 2000/1. Yet you only have 657 observations. Please check that you have coded the problem correctly. | |
| Feb 20, 2019 at 20:16 | comment | added | gsanroma | Could you check the estimations produced by the model? | |
| Feb 20, 2019 at 20:15 | comment | added | gsanroma | @EdM but still with a maximum slope of ~2 in the 1st model and the data in 0, 1, isn't the logit highly biased to one side of the probability (up to the effect of the random intercept) ? | |
| Feb 20, 2019 at 19:57 | comment | added | EdM | @GerardSanroma That is a question about logistic regression per se: what you are modeling is the log of the odds ratio for PunishmentYes/No. That's what the logit link function is doing: it maps probabilities (from 0 to 1) onto the range of all real numbers. Thus intercepts and slopes can take on any real values. To convert from logit scale to probabilitiy, exponentiate to get the odds ratio, then convert the odds ratio to the corresponding probability. | |
| Feb 20, 2019 at 18:10 | comment | added | gsanroma | For a start, I do not understand how can the intercept be ~10 when the data are only 0s and 1s. Am I missing something? | |
| Feb 20, 2019 at 17:55 | comment | added | EdM | Also, from the number of observations (651) and the number of ID values (163), it seems that each ID only faced one set of the 4 combinations of Ambiguity and Uncertainty (with one missing data point). Your second model, with its large number of estimated random effects, might be overfitting. This difficulty here isn't with logistic regression per se, it's with what you want to and can adjust for with the mixed model. That choice depends on your understanding of the subject matter. | |
| Feb 20, 2019 at 17:27 | comment | added | EdM | Specifying random effects can be tricky, as different formulas allow for different potential correlations among estimates. Think carefully about exactly what you want to be modeling. See this answer on the lmer cheat sheet and pay attention to the difference between model M2 and model M3 in terms of how they handle correlations between intercept deviations and fixed-effect deviations for random effects. See this answer for enforcing independence of random-effect relations for 2 fixed effects. | |
| Feb 20, 2019 at 17:18 | history | edited | dtoribio | CC BY-SA 4.0 |
edited tags
|
| Feb 20, 2019 at 17:10 | history | edited | dtoribio | CC BY-SA 4.0 |
added 11 characters in body
|
| Feb 20, 2019 at 16:20 | review | First posts | |||
| Feb 20, 2019 at 16:43 | |||||
| Feb 20, 2019 at 16:18 | history | asked | dtoribio | CC BY-SA 4.0 |