Interpretation of subject-specific effect in a cross-over study

I am not entirely sure whether I am interpreting a within-subject treatment effect produced by a Poisson mixed effects regression model correctly, so I am hoping to get some help on this forum.

In the study that generated the data for this model, half the patients were assigned to Treatment A in Period 1 and Treatment B in Period 2. (There was a washout period between treatments.) The other half of the patients were assigned to Treatment B in Period 1 and Treatment A in Period 2.

The outcome in each period was the number of days (out of the number of completed treatment days) patients were able to achieve a pre-defined clinical goal.

The Poisson mixed effects regression model used to analyse the data included fixed effects for Treatment and Period, as well as random intercepts and random slopes for Treatment across patients. The model used a log link and an offset term accounting for the fact that different patients completed a different number of treatment days in each period.

Since the model controls for Period and both Treatment and Period are dynamic (or time-varying predictors), I want to make sure I interpret the fixed effect of Treatment in this model correctly.

I know this effect is a subject-specific effect of Treatment in a particular period. After exponentiating the reported effect, can I interpret it as quantifying the multiplicative factor by which the expected number of days (per number of completed treatment days) differs for the typical patient on Treatment B relative to Treatment A in a particular treatment period?

I guess what confuses me is that we’re typically interested in how the same subject responds to switching treatments from one period to another, but here we are controlling for period. Does that mean that we would hypothetically give the typical subject both treatments in the same period (in either order?) and then monitor how well they respond on Treatment B relative to Treatment A?

2 Likes

Isabella, I’m not entirely familiar with the Poisson mixed effects model; could you offer a citation? Does this model treat the count of days when patient was clinically at goal as a Poisson variate? I would have to think that most clinical outcomes would be highly autocorrelated, and that this would violate the independence assumption. (Or does this model somehow happen to exhibit robustness to violations of this assumption?)

1 Like

David, here’s one citation I was able to find: Count data and treatment heterogeneity in
crossover trials, by N. T. Longford, Appl. Statist. (1998) 47, Part 2, pp. 217-229.

Yes, this model treats the count of days when the patient as clinically at goal as a Poisson variable. (It’s possible to relax this assumption, but I would like to understand the interpretation of the treatment effect in the current setting, before considering other - more complex - settings.)

Note that, in my model, I used dummy coding for the Treatment and Period variables, whereas the citation above used different coding.

1 Like

TY. By way of curation, here’s link to article: https://www.jstor.org/stable/2988351.

1 Like

Hi all!

I have a related question if you don’t mind: it seems that the study @Isabella_Ghement described is a single crossover trial with repeated measurements within each treatment period, as opposed to a repeated crossover trial (or a series of n-of-1 trials). Stephen Senn mentions here that single crossovers cannot estimate subject-specific treatment effects, but I failed to find specific mention of single crossover with repeated measurements. He also mentions this necessity in this comment (faster read).

I understand I can fit a random intercept/slope model with data from a single crossover trial with repeated measurements within each treatment period. Still, I wonder if that gives me valid estimates of the individual treatment effect. If so, why run repeated crossover trials at all?

Regardless of the model, it seems to me that, in the single crossover trial with many measurements, each individual treatment effect would be analogous to “contrasting average individual outcomes”; in the repeated crossover trial, on the other hand, each individual effect would come from “averaging individual contrasts”. Of course, you could do the latter in the single crossover trial too if you map individual measurements by, e.g., measurement day, but is this subject to bias or some sort of confounding?

Is my understanding correct?

Any insight is much appreciated!

Thank you very much.

1 Like

Follow-up after reading section 5.8.1 of [1]: it seems that estimating subject-specific effects from a single crossover trial with repeated measurements assumes the exchangeability of measurements within each period for each subject. If there are time effects (e.g. measuring blood sugar after a few days of exposure to a strict diet), then this assumption would be broken.

[1] Jones & Kenward (2015). Design and analysis of cross-over trials. CRC Press. 3ed. ISBN: 978-1-4398-6143-1.