Inclusion of endpoints when executing xteventtest after adjusting with a linear trend

[Constantino-Carreto-Romero]

After running xtevent with the trend option to adjust for a linear trend, when executing xteventtest the endpoints should not be considered in the tests.

For instance, when executing:

*Setup
webuse nlswork, clear
*year variable has many missing observations.
*Create a time variable that ignores these gaps.
by idcode (year): gen time=_n
xtset idcode time
*Generate a policy variable that follows staggered adoption.
by idcode (time): gen union2 = sum(union)
replace union2 = 1 if union2 > 1
*Adjust the pre-trend by estimating a linear trend by OLS
xtevent ln_w age c.age#c.age ttl_exp c.ttl_exp#c.ttl_exp tenure , pol(union2) w(-5 9) cluster(idcode) impute(stag) trend(-4, method(ols) saveoverlay) 

If we then run:

*leveling off
xteventtest, overidpost(2)

we get:

but the coefficients considered for the test should be 8 and 9.

[Constantino-Carreto-Romero]

@jorpppp

• in https://github.com/JMSLab/xtevent/commit/1153e3aa257ea0fc14023b12c60819a6edd2eec7 I implemented small changes, so the endpoints are not considered in xteventtest's tests if the trend option was used in the previous call of xtevent.
• in https://github.com/JMSLab/xtevent/commit/ed64fdeda3b2fade92dc9d22f5c70f3a7640b511 I uploaded an issue folder with a do-file and its log-file where I tested xteventtest after changes in https://github.com/JMSLab/xtevent/commit/1153e3aa257ea0fc14023b12c60819a6edd2eec7. The only option showing an error is overid. Nonetheless, I think it is an expected error because the option inherits overidpre's value from xtevent's call (which is stated in the helpfile). Then, it tries to test pre-trends with that number of coefficients, but after the linear trend adjustment (with ols method), the number of pre-event coefficients available for the test is smaller.

[jorpppp]

Thanks @Constantino-Carreto-Romero . We should be able to adjust the value of overidpre and overidpost that xtevent will use if the xtevent call used the trend option. Why don't you give it a shot?

@jmshapir @SimonFreyaldenhoven FYI

[Constantino-Carreto-Romero]

@jorpppp sure thing, I'll try changing xteventtest's overid option. For instance, if first calling xtevent with overidpre(5) (and pre(0)) and trend(-3, method(ols)), then the available pre-event coefficients for xteventtest's are only -5 and -4. xteventtest will try to test pre-trends using the first 5 pre-event coefficients, since now it will not find 5 pre-event coefficients available, but a smaller number of coefficients, it should try using all available pre-event coefficients. I think it should also show some message saying the change in the number of coefficients used for the test and what these coefficients are.

[jorpppp]

@Constantino-Carreto-Romero Yes, this sounds like a good plan. On some instances there won't be any coefficients available to test, in which case we may also want to have an informative error message.

[Constantino-Carreto-Romero]

update: I've not made progress, but I'm planning to retake today's afternoon.

[Constantino-Carreto-Romero]

@jorpppp

• in https://github.com/JMSLab/xtevent/commit/ea08baf5e7b211a3d0f0e01bd51cfa0e12e27730 I implemented https://github.com/JMSLab/xtevent/issues/166#issuecomment-1775458910 and https://github.com/JMSLab/xtevent/issues/166#issuecomment-1775465658
• in https://github.com/JMSLab/xtevent/commit/8e1a88bca3d995cccbb1210dea78024307e5757f I updated the do file and its log file with examples to test the implementation.

[Constantino-Carreto-Romero]

Summary: In this issue, we made changes to xteventtest, so it doesn't use the endpoints in any of its tests if previously xtevent used linear trend adjustment. We also modified the number of coefficients it uses to test pre-trends in some cases if previously xtevent used linear trend adjustment with the ols method.

Changes brought to main in https://github.com/JMSLab/xtevent/commit/4198174fab4a012e3ab69eca0b76f2ecc9199e40

Thread continues in https://github.com/JMSLab/xtevent/pull/167