Time varying covariates and logistic regression

[Mike Lyell]

I am trying to predict the outcome of a binomial dependent variable based on multiple measurents over time of a single variable measured prior to the outcome variable. Is it possible to do logistic regression in Stata with a time varying covariate nested within the individual? If not, any other suggestions on modeling this?

[Stephen Jenkins]

Have you looked at xtlogit, or am I missing something regarding your question?

[Steve Samuels]

I think Mike is referring to just a single outcome. If so, and if the measurements were taken at a small number of fixed intervals (after enrollment in a study, then all four can be entered as predictors. Otherwise, the predictors can be one or more per-person summaries of the measurements. The following come to mind, organized into categories of location, variation, and trend.

1. Mean
2. Median
3. A mean that puts more weight on recent measurements

4. Standard deviation.
5. Inter-quartile distance

6. Linear trend (slope from a simple linear regression on time).
7. Linear trend in more recent measurements

[Mike Lyell]

There are more than 15 measures, with equal spacing, for each individual. As correctly pointed out by Steven, the outcome is only measured once - hence, I am uncertain if xtlogit is appropriate. I've tried the different summary measurements also incl area under the curve-however, I thought the collapsing of the data like that would decrease the power of the test in detecting a difference by decreasing the amount of information and variability.

[Steve Samuels]

You don't have to restrict the analysis to just one summary measure at a time. A hypothesis test of of no association with >15 measurements is a simultaneous test of >14 comparisons. This test is likely to have lower power than one based on a few well-chosen summaries.

[Nazzarena]

I have the same problem. The issue is that on one hand, repeated measurements aren't independent, on the other, the shape of the relationship between time-varying covariates and outcome and the effect of time are lost using summary measures. Consider for example 3 timepoints, and individual i from group A has 5, 4, 6. Individual j from group B has 1, 4 ,8. Further, consider for example that the timepoint for i are 0, 60 , 90 days, and for j are 0, 100 , 400. Obviously, a single summary measure cannot contain all that info
Maybe I'm having a brain freeze and missing something obvious, any suggestions 9 years later?
Many thanks!

[Nazzarena]

Here is a portion of actual data where R is the response variable and v are the time-varying covariates. Beside the need to accommodate panel data I have a bit of zero inflation too .

	idn	days	R	v1	v2	v3	v4	v5	v6	v7	v8	v9
13	5	0	1	0	60	0	0	40	60	0	0	40
14	5	92	1	0	0	0	0	0	0	0	0	0
15	5	167	1	0	0	0	0	0	0	0	0	40
16	6	0	0	67	0	160	0	100	0	53	73	0
17	6	92	0	0	0	73	0	0	0	0	0	0
18	6	246	0	0	0	120	0	127	40	0	0	0
19	7	0	1	0	0	0	0	0	0	0	0	0
20	7	154	1	0	0	0	0	0	0	0	53	0
21	7	378	1	0	40	0	0	0	0	0	0	0
28	10	0	0	0	0	0	0	0	0	0	0	287
29	10	85	0	0	0	0	0	0	0	0	0	147
30	10	169	0	0	0	0	0	0	0	0	0	227

[Nazzarena]

well the formatting doesn't really help... Please see the attached if needed