Question about graphing observed means in parallel trends diagnostics after didregress

[Sangno Lee]

Hello Statalist community,

I’m having trouble understanding the graphical diagnostics for the parallel trends assumption after running the didregress command, as shown in the Stata documentation here: https://www.stata.com/manuals/causaldidregress.pdf.

According to the documentation, a visual diagnostic of this assumption can be obtained by plotting the means of the outcome over time for both groups, or by visualizing the results of the linear-trends model. The estat trendplots command displays two graphs: one for the observed means and another for the linear-trends model.

Initially, I assumed that the graph of observed means could be generated using the following commands:

collapse (mean) satis, by(procedure month)
twoway connected satis month if procedure == 0 || connected satis month if procedure == 1

But the graph is different in that the treatment group doesn't have observations from January to March and the values of control group are higher than trendplots.

I then tried plotting the means using a model-based approach:

regress satis i.procedure i.month
margins, at(procedure=0 month=(1(1)7) )
margins, at(procedure=1 month=(1(1)7) )

But the predicted values from this approach differ from those shown in the estat trendplots graph.

Is there a way to reproduce the observed means graph shown in estat trendplots using separate commands?

Thank you in advance for your help.

[George Ford]

In the first one, I suspect that procedure does not measure ever treated observations, which is why there's no graph pre for the treated. create a variable == 1 if the unit is ever treated. Might make sense to center as well, so the lines are of the same scale.

[Sangno Lee]

Thank you, Your response was very helpful. I appreciate your support!

bys hospital: egen evertreated = max(procedure)
collapse (mean) satis, by(evertreated month)
twoway connected satis month if evertreated == 0 || connected satis month if evertreated == 1

[George Ford]

you could use lgraph, which collapses the data behind the scenes but does not disturb your data.

ssc install lgraph
lgraph satis month, by(evertreated)