Plotting predictions for each period along with confidence intervals?

[Peter Meier]

Dear colleagues,

given data on 12 months, I regressed my outcome on dummies for periods 2-12, plus controls c1-c5, to see where the outcome changes significantly relative to the baseline month:

reg outcome t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 c1 c2 c3 c4 c5

.

Now I can plot the 11 coefficients together with confidence intervals:

coefplot, keep(t*) vertical levels(95)

But I'm wondering how I can best plot the corresponding predictions? I know I could manually add to the coeffcient the constant, plus each of coefficients on c1-5 times the respective mean value. Or I could use predict, then average for each period, then plot those averages. But in neither case would be sure how to get the correct confidence intervals.

Is there a pre-programmed command for this, as option of -coefplot- or otherwise?

Thanks! PM

[Ken Chui]

First advice would be: "stop using manually computed dummies." It's nearly never necessary. Try refit the model with the original time variable (let's call that mos):

reg outcome i.mos c1 c2 c3 c4 c5

Then, you can tap into -margins- and -marginsplot-:

margins mos
marginsplot

It'd by default set all the other variables not mentioned here at their means. There are options to fine tune them, check -help margins- for use cases. Here is a quick demo using auto:

sysuse auto, clear

reg mpg i.rep78 weight headroom
margins rep78
marginsplot

[Peter Meier]

Hi Ken,

thanks so much, that seems to work so far!!

Just one follow-up: Ideally I'd also like to include in the regression interactions of the time dummies with a continous treatment variable T to see whether exposure to treatment T has a different effect in different periods. So I tried

reg outcome c.T##i.mos c1 c2 c3 c4 c5
margins c.T##i.mos
marginsplot

That did not work, presumably because this notation does not allow for continuous variables. Is there another way of getting there?

Thanks! PM

[Ken Chui]

Hi Ken,

thanks so much, that seems to work so far!!

Just one follow-up: Ideally I'd also like to include in the regression interactions of the time dummies with a continous treatment variable T to see whether exposure to treatment T has a different effect in different periods. So I tried

reg outcome c.T##i.mos c1 c2 c3 c4 c5
margins c.T##i.mos
marginsplot

That did not work, presumably because this notation does not allow for continuous variables. Is there another way of getting there?

Thanks! PM

Continuous variable cannot be a main argument in margins, which is for factors (categorical variables) only. If you need to see how predicted y changes with mos with T, you'll need to give some concrete T examples:

reg outcome c.T##i.mos c1 c2 c3 c4 c5
margins mos, at(T=(30 35 40 45 50))
marginsplot

[Peter Meier]

Thanks so much Ken, this works great now! PM