Replacing observations with imputed date

[Tsi Njim]

Hi, please I have a dataset which is in this format (not real dataset)

S/N	date last seen	year last seen	outcome
1	19/05/2015	2015	Alive
2	2015	2015	Alive
4		2018	Dead
5	21/06/2018	2018	Dead
6		2018	Dead

I am trying to replace the "date last seen" with the midpoint for the "year last seen" (for example replacing observation 2 with - 01/06/2015).

Please are there any suggestions for an easy method to do this?

Thanks!

[William Lisowski]

Welcome to Statalist.

Perhaps the following will point you in a useful direction. Note that July 1st is (approximately) the midpoint of a calendar year, not June 1.

generate midpoint = mdy(7,1,year_last_seen)
format midpoint %td

[Steve Samuels]

Mid-year midpoints can be disastrous if you intend to use the dates to calculate durations. Even monthly midpoints (if you knew month and year) could have undesirable consequences. Both distort the distribution of what would otherwise be smoothly random durations. Better would be to randomly impute one of the 365 or 366 days of the year. (The reference below imputes dates in a known month.)

If you have long enough follow-up (say five years) and many missing dates, as in your example, consider a grouped data analysis, with "year" as the time unit.

Reference:
Samuels, S.J. and Cox, N.J., 2012. Stata tip 105: Daily dates with missing days. Stata Journal, 12(1), pp.159-161.

[Tsi Njim]

Welcome to Statalist.

Perhaps the following will point you in a useful direction. Note that July 1st is (approximately) the midpoint of a calendar year, not June 1.

generate midpoint = mdy(7,1,year_last_seen)
format midpoint %td

Thanks so much for your help William.

[Tsi Njim]

Mid-year midpoints can be disastrous if you intend to use the dates to calculate durations. Even monthly midpoints (if you knew month and year) could have undesirable consequences. Both distort the distribution of what would otherwise be smoothly random durations. Better would be to randomly impute one of the 365 or 366 days of the year. (The reference below imputes dates in a known month.)

If you have long enough follow-up (say five years) and many missing dates, as in your example, consider a grouped data analysis, with "year" as the time unit.

Reference:
Samuels, S.J. and Cox, N.J., 2012. Stata tip 105: Daily dates with missing days. Stata Journal, 12(1), pp.159-161.

Thanks Steve. The link you referenced was really helpful.

[Steve Samuels]

You are very welcome, Tsi.

I realized since I wrote that imputation has a cost: standard errors must reflect the randomness of the process. To get proper standard errors, you will need to create multiple imputation (MI) datasets by hand; then add Stata's mi estimate prefix to your estimation commands.

To avoid this, I recommend the grouped-data survival models, as they require no imputation.