Timeline for Data from 2 eyes and repeated measures across time
Current License: CC BY-SA 4.0
24 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| S Nov 23, 2023 at 23:04 | history | bounty ended | CommunityBot | ||
| S Nov 23, 2023 at 23:04 | history | notice removed | CommunityBot | ||
| Nov 23, 2023 at 21:24 | history | edited | s.stats | CC BY-SA 4.0 |
added 560 characters in body
|
| Nov 23, 2023 at 19:33 | vote | accept | s.stats | ||
| Nov 20, 2023 at 13:02 | answer | added | Björn | timeline score: 3 | |
| Nov 20, 2023 at 12:21 | comment | added | Björn | Correlation structures like autoregressive, especially AR(1) over time tend to be terrible in terms of how they behave. They usually overestimate correlation of near-by observation and then - usually wrongly decide that observations that are sufficiently far apart are totally independent (when the data usually clearly contradicts that). | |
| Nov 16, 2023 at 17:56 | answer | added | Robert Long | timeline score: 4 | |
| S Nov 15, 2023 at 21:06 | history | bounty started | s.stats | ||
| S Nov 15, 2023 at 21:06 | history | notice added | s.stats | Authoritative reference needed | |
| Nov 7, 2023 at 11:59 | comment | added | s.stats | @PBulls Thanks very much! Since there is just single random effect for ID, does this mean that the model should be specified as follows: lmer(IOP ~ ....+ time ....+ (1|ID), data=data) ? Not all patients have 2 eyes eligible, would this be a major cause for concern, or does the model allow for this? | |
| Oct 26, 2023 at 14:28 | comment | added | PBulls | The nesting would matter if you also want to specify an effect across subjects (not eyes) or eyes (not subjects), but you only have the single random effect within both subjects and eyes, and since each subject has only one eye of each it doesn't matter what order you put them in. | |
| Oct 26, 2023 at 10:15 | comment | added | s.stats | @PBulls Thanks very much for the response. May I ask why the nesting hierarchy doesn't matter in this case? In that case would neither of the models I proposed specify the structure in my data | |
| Oct 13, 2023 at 19:08 | comment | added | Frank Harrell | Yes, in several mixed model packages you can specify a correlation structure that is added on top of the structure induced by random intercepts or slopes. | |
| Oct 13, 2023 at 17:42 | comment | added | s.stats | Is it possible to model two levels of correlation, 1) autoregressive for measurments between time points and 2) exchangeable or unstructured for measurements between eyes? | |
| Oct 13, 2023 at 17:40 | comment | added | s.stats | Would the autoregressive correlation structure be better in this case, if modelled using a linear mixed effects model. I'm not too familiar with generalised least squares or markov models so it was not my go to model, but thanks very much for the references which I shall go over. I was really just wondering how this case may be modelled using a linear mixed effects model, when you have time nested within eye nested within person. | |
| Oct 13, 2023 at 17:38 | comment | added | s.stats | @FrankHarrell , Sure, I will need to check the correct functional form for time , and then specify it as continuous in the model with the correct transformation if non-linear. | |
| Oct 13, 2023 at 16:55 | comment | added | Frank Harrell | I don’t understand what that has to do with it being discrete. Continuous-time models allow estimation at all possible times. | |
| Oct 13, 2023 at 16:14 | comment | added | s.stats | @EdM, it was specified as discrete to obtain the effect of corneal thickness at each time-point | |
| Oct 13, 2023 at 16:10 | comment | added | s.stats | @FrankHarrell , Some participants may have 2 eyes studied, depending on eligibility of the eye. So we have at most 2 eyes measured per person over 5 time points. | |
| Oct 13, 2023 at 15:23 | comment | added | Frank Harrell | I missed that there is only one eye studied per person. Then I would go with generalized least squares or a Markov model and not use random effects. If time is discrete (unlikely) and there are < 4 distinct measurement times then an unstructured covariance matrix can work well. | |
| Oct 13, 2023 at 15:15 | comment | added | PBulls | @FrankHarrell, doesn't the second model also allow for (unstructured) correlation over time? At least, given the choice between these two, I would certainly prefer the latter. The nesting hierarchy should not matter because each subject should only have one of each eye (per timepoint), so the choice is between 2 N*N blocks or N 2*2 blocks. | |
| Oct 13, 2023 at 14:23 | comment | added | Frank Harrell | Longitudinal data are often best handled using serial correlation models and not random effects so here you have an ideal case for having random effects for patients (assuming exchangeability of eyes) and a serial correlation structure for the longitudinal part, e.g., generalized least squares or a Markov model. A Bayesian random effects Markov model would be especially of interested. See here. | |
| Oct 13, 2023 at 14:15 | comment | added | EdM |
Is there some reason why you are treating Time as a factor instead of continuous?
|
|
| Oct 13, 2023 at 14:07 | history | asked | s.stats | CC BY-SA 4.0 |