margins at specific ages or over(age groups)

[pierre martin]

Hi all, I have done the following model:

mixed cognitive_score c.real_age_10_c70##c.real_age_10_c70##c.bloodpress ure i.sexe || id: real_age_10_c70, residuals(ar 1, t(real_age))

I am studying the association of baseline blood pressure on cognition over time (age is the time scale)(incremented by 1 year at each annual visit where cognition is measured).
ps : chronological age is divided by 10 and then centered around 70 years but it does not change anything to my question

I had reported the coefficients of the interaction blood pressure * age and blood pressure * age2 but a reviewer is asking me this : consider reporting estimated effects across 10-year age groups rather than just reporting slopes for age

I am not sure I should do this 1)

margins, dydx(bloodpressure) at(real_age_10_c70=(-3 -2 -1 0 1 2))

OR this 2)

create age groups and then use over(age_group) to provide an average estimated effects for each decade instead of an estimated effect at 30 years, 40 years, 50 years etc

margins, dydx(bloodpressure) over(age_group)

I need an advice for major revision in a great journal IF > 10 and it is basically the only comment ... I am struggling with.

Thank you so much, I really appreciate your help.

Best, Pierre

[Clyde Schechter]

Good question. Here's how I would think about it.

Critically, the original regression model includes a variable, sex, which is probably confounded with age (unless your sampling design specifically did something to balance sex across ages.) And cognitive function, even at the same age, may well be different in men and women. Ditto for blood pressure. If you create an age-group variable and use the -over()- method, the difference in sex distribution across the age groups will not be adjusted for. Each age-group's results will be conditional on the observed sex distribution in the age group. By contrast, if you use -at()- the calculations adjust the sex distribution to be the observed sex distribution in the entire sample, so that comparisons across age-groups will be fully sex-adjusted.

So it depends on whether this potential sex-confounding is important for how these results will be used.

All of that said, I'm a little surprised that the original mixed model did not include an age#sex interaction. I would normally expect that the rate of change of cognitive function with age might differ between men and women. But perhaps you had good reason for your modeling choice: prior information that the interaction is negligible, or perhaps just a sample too small to further subdivide.

[pierre martin]

Hi Clyde,

Thank you so much for your feedback. This is extremely helpful to me. I will indeed use the -at()- instead of -over- method. Yes long story short but we finally did not include the age sex interaction.

Thank you again!

Best, Pierre