Timeline for Regression spline for time to allow for slope changes
Current License: CC BY-SA 4.0
22 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Sep 21 at 14:22 | vote | accept | user167591 | ||
| Sep 8 at 17:05 | comment | added | EdM | @user167591 the choice between the last pre-treatment baseline value or an average over prior baseline values to calculate the differences is a matter of your understanding of the subject matter, not statistics. If you don't have reason (before looking at the data) to expect differences among pre-treatment values, then their average might be more precise. That probably isn't the case at the switch to the second phase of the study, however, as the prior values presumably change during the first phase. Using differences should be OK if you aren't also using baseline values as predictors. | |
| Sep 8 at 11:16 | comment | added | user167591 | Cont…my initial thought was to use Y at the single month immediately before start treatment as time0 for the contrasts. The previous baseline months (-1, -2, -3 etc) are disregarded in the contrasts; their inclusion in the model as a continuous spline is solely to get a better estimate of t0 ie the month immediately before starting treatment. The other idea I had was to use the average of all negative time points (all baseline months) as the baseline Y in the contrasts. This uses all baseline information available to us but I don’t think it works if there is a trend over baseline months? | |
| Sep 8 at 10:42 | comment | added | user167591 | sorry can I please clarify the contrasts I had in mind after fitting your suggested first model. Here time is scaled relative to baseline-active transition. I plan to present treatment effect as double differences ie y(time t) - y(time0) for treatment group minus the same for control where time t is each follow-up month and time0 is baseline. A log link means we could present such contrasts as ratio of rate ratios (RRR). Does that sound reasonable? If so, I’m still unsure about t0 for the contrasts since there are several months worth of baseline data (month 0, -1, -2, -3 etc). | |
| Jul 13 at 18:15 | comment | added | user167591 | since I plan to use splines for the time variable in each of your suggested models, I thought of presenting effects as difference in difference contrasts. Specifically, this would be (post-pre) - (post-pre) where brackets are treatment and control groups respectively. Post refers to predicted Y at several scaled times >0, and pre is predicted y = 0, the latter being constant. That way we could compute a smooth curve of double difference contrasts for the treatment effect over time. Does that sound reasonable? | |
| Jul 11 at 18:46 | comment | added | EdM | @user167591 I don't see a way to join up the fitted lines between the two models, given the different durations of active phase among pairs, but that might say more about me than about what's possible. ANCOVA is closely related to mixed modeling. Change scores per se aren't necessarily a problem; a paired t-test is on a change score. Problems arise if the change scores are necessarily associated with the baseline values in some way. This page has more extensive discussion. | |
| Jul 11 at 10:39 | comment | added | user167591 | The mixed models you suggest seem to be equivalent to change score models, see here: middleprofessor.com/files/applied-biostatistics_bookdown/_book/…, sect. 19.3. This "estimates the effect as the coefficient of the interaction...[and] will give the same estimate as Model M2" (M2 is change score model). However, the above Stata page implies the mixed model is equivalent to ANCOVA ...which one is right?! Worried as @Frank Harrell appears to warn against change scores fharrell.com/post/errmed/#change-from-baseline | |
| Jul 8 at 19:50 | comment | added | user167591 | I agree that the single comprehensive model would make most sense. I had in mind getting predicting expected values from such a model across the entire time period of the experiment. Am I right that with two models we can’t do this properly so that the fitted lines join up? | |
| Jul 8 at 18:11 | comment | added | EdM | @user167591 see my first comment in response to your first two comments on my answer, above. In general a single comprehensive model would make the most sense, but with different durations of active phases in your situation I couldn't figure out a way to make that work. Collinearity is often much less of a problem than it's purported to be. | |
| Jul 8 at 17:26 | comment | added | user167591 | Thank you @EdM. Could you also briefly explain why it’s not wise to model two time scales in one big model? Is it perhaps concerns about co-linearity? | |
| Jul 8 at 15:26 | comment | added | EdM | @user167591 yes, based on this quote from that page: "The pure mixed-effects model has both pre- and post-intervention observations for each patient and does not include the baseline outcome measurement as a covariate." That said, the choice between that modeling approach and using the average of baseline values as a covariate mostly depends on your understanding of the subject matter and how you want to explain your results to your audience. | |
| Jul 8 at 14:00 | comment | added | user167591 | thanks so much for your patience and explaining. Is your approach what Clyve refers to as the "pure mixed model" here: statalist.org/forums/forum/general-stata-discussion/general/… | |
| Jul 8 at 13:39 | comment | added | EdM | @user167591 my proposed baseline-to-active model wouldn't use an additional baseline-average covariate, as it would model changes in time directly. If the baseline values are sufficiently stable in time, you could just start analysis at the baseline-to-active transition, ignore the details of data prior to that time, and just use the prior average as a covariate. If there is substantial month-to-month variation and the transition happened at different months among pairs, however, that might lead to some problems; haven't thought that through completely. | |
| Jul 8 at 7:52 | comment | added | user167591 | can you please explain the data structure (columns on which to fit model) for your first model? The way I see it, there will be columns for dependant variable, time (negative values in baseline i.e. pre active phase), baseline, plus other covariates. The dependent variable will already have values covering the baseline phase (negative time). Then, what is the use of an additional baseline covariate? | |
| Jul 4 at 13:04 | comment | added | user167591 | yes I intend the dependant variable to be uptake rate adjusted for baseline, not change from baseline. Regarding your first model with time centred on baseline to active transition, I dont understand how one can use mean (several months) baseline as a covariate. | |
| Jul 4 at 11:38 | comment | added | EdM | @user167591 there might be negative values of time, but that's not a problem. If you use the baseline value as a covariate you should not be using changes from baseline as the outcome. See this page. You could include the duration of the active phase as a covariate in the model for the active-to-post transition. That would implicitly "define the time since start of active phase" in a model of the time since the start of the post phase. | |
| Jul 4 at 9:08 | comment | added | user167591 | Also, consider your second model, with time centered on active to post transition. At e.g. time= -5 (all sites will be in active phase, 5 months prior to start of post phase). But their time relative to start active phase may differ e.g. one pair might have had active treatment for longer than another pair. Would it not be better with 2 time scales in the same model to precisely define the time since start active phase and time since start post phase for each site? | |
| Jul 3 at 17:14 | comment | added | user167591 | if we center on transition times, wouldn’t we get negative values extending into baseline phase? I was thinking to use mean baseline as a covariate as per the ANCOVA approach. | |
| Jul 3 at 13:03 | comment | added | EdM | @user167591 I don't see a way to use a single time scale without losing information about changes at the transition from active to post phase. Say that some pairs had 12 months of active phase while others had 15. With a single time scale and an interaction with phase, at 16 months (all pairs now in post phase) you would be assuming that all pairs had the same outcomes. In fact some pairs would be 4 months post-active while others were only 1 month post-active, so you might expect differences on that account. That's why I suggest two models, centered on the transition times. | |
| Jul 3 at 10:45 | comment | added | user167591 | Thanks so much @EdM. Do you think a single time variable with time zero start of active phase, phase as a categorical predictor, and their interaction would work? Im concerned about the high correlation of phase and time though. | |
| Jul 3 at 10:24 | comment | added | user167591 | Thanks so much @EdM,, | |
| Jun 29 at 21:01 | history | answered | EdM | CC BY-SA 4.0 |