Penalized spline models with -mixed-

[Geoff Dougherty]

Greetings, all.

I recently ran across a StataCorp presentation by Roberto Gutierrez that discussed using -xtmixed- to obtain penalized spline estimates (https://www.stata.com/meeting/uk08/gutierrez_mixed.pdf, slides 27 et seq). I am interested in updating this approach to use -mixed-, and to incorporate on multiple subjects per time point (I have states and years in this case, with the outcome being a score for each state/year combination).

I haven't had much luck beyond fitting the basic crossed effects model : mixed outcome || _all: R.state || all: R.year, which of course is treating years as factors and not providing a smoothed slope.

Curious to hear if anyone else has tried this and might have some advice.

Thanks!

[Geoff Dougherty]

Hi, everyone. Pleased to report I have this running in mixed as a multi-level model that adapts the original data and code:

use "http://www.stata-press.com/data/r13/motorcycle.dta", clear
bys time: gen id=_n // set up an ID to simulate multilevel data. There are multiple measurements for some time points

local i 1
forvalues k = 1(1)60 {
gen time_`i' = cond(time - `k' > 0, time - `k', 0)
local ++i
}

mixed accel time || _all: R.id || _all: time_*, noconstant cov(identity)
predict p, fitted

However, I am not tracking how the time_* variables function in this model. If anyone is able to share some intuition on that point, I would appreciate it. Thanks!