Estat trendplots takes hours

[Leon Broos]

Help needed!

I am trying to graph my pre and post-treatment period trends but it has been loading for hours and I am not sure if it will ever give me any output. This is about the fourth time I have tried it with me trying it for 30 minutes up to 2.5 hours without STATA providing me with the graphs.

My did command is as follows: didregress (hospitalcosts i.age i.year i.gender i.postalcode) (treatpost), group (age) time (year)

If I run estat ptrends it also takes a little while but it gives me some output. My data spans over 10 years and the dataset is rather large, I have 1.5 million observations.

Does anyone have any ideas on how to reduce the loading time? Will STATA ever provide me the parallel trends graphs? Thank you so much in advance!

[Andrew Musau]

Is your dataset a panel dataset at the post-code level? If so, see

help xtdidregress

instead of using indicators and didregress.

[Leon Broos]

Andrew Musau if I were to set it as panel data it would be grouped at the age, gender & postal code level. My observations are for all people in a country but grouped for age, gender & area (= postal code). Do you think changing it to a panel and running it as an xtdidregress would help? I have tried doing so with a panelid variable but now even the xtdidregress command is already taking 20 minutes. I am a little lost.

[Andrew Musau]

How many postcodes do you have and how many age groups are present? To reduce runtime, it would be sensible to absorb the larger of these.

[Leon Broos]

Andrew Musau I have 800 postcodes and 91 age groups. If I were to absorb these are you still talking about running an xtdid or a regular did model?

[Andrew Musau]

I mean:

xtset postalcode
xtdidregress (hospitalcosts i.age i.year i.gender) (treatpost), group (postalcode) time (year)

Check that the ATE is the same across both specifications.

[Leon Broos]

Andrew Musau Thank you so much! The ATE does indeed appear to be the same. However, still it takes a very long time to estat trendplots. Is there any way around this? And is it normal for STATA to take +2.5 hours? Should I just let it run all night?

[Andrew Musau]

This is the best you can do with the did set of commands at present. You could use xtreg in StataNow as it has an -absorb()- option or reghdfe from https://github.com/sergiocorreia/reghdfe to estimate the model. Then twoway to graph the parallel trends. For the tests, the formulas are given in the documentation of the commands.