Improving time_length for 29 Feb (when dealing with date time)

[larmarange]

With previous version, we had:

> naiss <- ymd_hms("1992-02-29 12:00:00")
> evt <- ymd_hms("2014-03-01 01:00:00")
> time_length(interval(naiss, evt), "years")
[1] 21.99874

We should consider that the anniversary occurs between 28 Feb at 23:59:59 and 1st March at 00:00:00. Therefore, %m++% should rollback at 00:00:00. In that case:

> time_length(interval(naiss, evt), "years")
[1] 22.00011

[larmarange]

A quick comment about period computation from intervals:

> naiss <- ymd_hms("1992-02-29 12:00:00")
> evt <- ymd_hms("2014-03-01 01:00:00")
> as.period(interval(naiss, evt))
[1] "22y 0m -1d 13H 0M 0S"

Should we avoid negative days?

[vspinu]

This should not happen. It's a bug in as.period.
On Jan 1, 2015 6:04 AM, "Joseph" notifications@github.com wrote:

A quick comment about period computation from intervals:

naiss <- ymd_hms("1992-02-29 12:00:00")> evt <- ymd_hms("2014-03-01 01:00:00")> as.period(interval(naiss, evt))
[1] "22y 0m -1d 13H 0M 0S"

Should we avoid negative days?

—
Reply to this email directly or view it on GitHub
#295 (comment).

[vspinu]

naiss <- ymd_hms("1992-02-29 12:00:00")
evt <- ymd_hms("2014-03-01 01:00:00")
time_length(interval(naiss, evt), "years")
[1] 21.99874

Well. There is a mirror problem with negative intervals that use %m+%:

> time_length(interval(ymd_hms('2000-02-29 12:00:00'), ymd_hms('1999-02-28 20:00:00')), "years")
[1] -0.9990868

So we will have to change %m+% as well, and I am not sure that's a good idea. %m+% will basically make all intraday timestamps for Feb 29 equal, and that's clearly not suitable for intraday analysis.

One easy solution is to add an optional argument to rollback and .month_plus and .month_plus_plus. Something like preserve_hms = TRUE.

BTW. your .rollback_day_one can also be integrated into rallback with an optional argument roll_to_first = FALSE or something similar.

[larmarange]

You are right.

We can also add roll_to_first to .month_plus. In that scenario, we don't need %m++ and .month_plus_plus (as it's not exported) and we can use directly .month_plus in time_length.

I will update the pull request.

[larmarange]

When we think with time, even for negative intervals we should not use %m-%. In fact, the anniversary should be the 28 Feb at 24:00:00, i.e. the 1st March at 00:00:00.

Therefore, the code is simpler.

Now, we obtain:

> time_length(interval(ymd_hms('2000-02-29 12:00:00'), ymd_hms('1999-02-28 20:00:00')), "years")
[1] -1.000457

[larmarange]

Tests and examples added for rollback. Code of time_length updated with ifelse(x@.Data < 0, -1, 1)

[vspinu]

Great! Thanks.