Existing packages for working with dates and times in R expect them to
be tidy. That is, they should be in or coercible to the standard
yyyy-mm-dd and/or hh:mm:ss format.
But dates (and times) are often messy. Sometimes we only know the year when something happened, leaving other components of the date, such as the month or day, unspecified. This is often the case with historical dates, for instance. Sometimes we can only say approximately when an event occurred, that it occurred before or after a certain date, or we recognise that our best estimate comes from a dubious source. Other times there exists a set or range of possible dates for an event.
Although researchers generally recognise this messiness, many feel
expected to force artificial precision or unfortunate imprecision on
temporal data to proceed with analysis. For example, if we only know
that, for one observation, it happened sometime in 2021, then we might
revert to a panel data design even if greater precision is available
throughout the rest of the dataset, or opt to replace this date with the
start of that year (2021-01-01), assuming that erring on the earlier
(or later) side is more justifiable than a random date within that month
or year.
However, this can create inferential issues when timing or sequence is
important. {messydates} assists with this problem by retaining,
representing, and reasoning about messy dates and times, rather than
discarding them. It implements ISO 8601-2:2019 in R, introducing a new
mdate class that can represent a wide range of messy dates and times,
and provides methods for coercing into and from this class, as well as
tools for working with messy dates and times in a way that is compatible
with existing packages.
Base R and most date packages are built to convert input into a single,
precise instant, discarding anything that does not fit. {messydates}
instead treats imprecision as information to be retained, represented,
and reasoned about — and only resolved to a precise date when you
explicitly ask.
| Capability | base R | {lubridate} |
{anytime} |
{clock} |
{messydates} |
|---|---|---|---|---|---|
| Precise dates | ✓ | ✓ | ✓ | ✓ | ✓ |
| Precise times of day | ✓ | ✓ | ✓ | ✓ | ✓ |
| Parse flexible/ambiguous formats | ✓ | ✓ | ~ | ✓ | |
| Historical & era dates (BCE, far future) | ~ | ~ | ✓ | ||
| Retain unspecified components (year- or month-only) | ✓ | ||||
Approximate (~) & uncertain (?) annotations |
✓ | ||||
| Ranges, sets, and on-or-before/after dates | ✓ | ||||
| Times with uncertainty/approximation | ✓ | ||||
| Written/prose dates (e.g. “First of February…”) | ✓ | ||||
| Expand to (and contract from) all compatible dates | ✓ | ||||
| Transparent resolution (min/max/mean/median/random) | ✓ | ||||
| ISO 8601-2 (EDTF) support | ~ |
(✓ = supported, ~ = partially supported. {clock} excels at
high-precision arithmetic on valid, complete dates, which complements
rather than overlaps {messydates}.)
{messydates} covers the parts of ISO 8601-2 that carry imprecision:
unspecified components, approximate and uncertain annotations, ranges,
open ranges, sets, and times of day. Durations are covered too, by the
mduration class, which expresses them as a range of possible dates so
that uncertainty at either end is preserved. Not implemented are week
dates (2019-W12), ordinal dates (2019-123), season codes
(2019-21), significant digits (1234S3), extended years (Y17E7),
the P-style duration notation (P1Y2M), and repeating intervals;
these are rejected on coercion rather than silently passed through.
{messydates} implements the Extended Date/Time Format (EDTF)
annotations set by the International Organization for Standardization
(ISO) outlined in ISO
8601-2_2019(E) for R.
{messydates} introduces a new mdate class that embeds these
annotations, and offers a set of methods for constructing and coercing
into and from the mdate class, as well as tools for working with such
‘messy’ dates.
pkg_comparison <- data.frame(
Example = c("Normal date", "Future date", "Historical date", "Era date",
"Written date", "DMY date", "MDY date", "Wrongly specified date",
"Approximate date", "Uncertain date", "Unspecified date",
"Censored date", "Range of dates", "Set of dates",
"Date-time", "Approximate time"),
OriginalDate = c("2012-01-01", "2599-12-31", "476", "33 BC",
"First of February, two thousand and twelve", "10-31-2012",
"31-10-2012", "2012-31-10", "2012-01-12~", "2012-01-01?",
"2012-01", "..2012-01-12", "2012-11-01:2012-12-01",
"2012-5-26, 2012-11-19, 2012-12-4",
"2012-01-01 14:30", "2012-01-01 ~14:30"),
stringsAsFactors = FALSE)
pkg_comparison$base <- as.Date(pkg_comparison$OriginalDate)
pkg_comparison$lubridate <-
suppressWarnings(lubridate::as_date(pkg_comparison$OriginalDate))
pkg_comparison$messydates <-
messydates::as_messydate(pkg_comparison$OriginalDate)| Example | OriginalDate | base | lubridate | messydates |
|---|---|---|---|---|
| Normal date | 2012-01-01 | 2012-01-01 | 2012-01-01 | 2012-01-01 |
| Future date | 2599-12-31 | 2599-12-31 | 2599-12-31 | 2599-12-31 |
| Historical date | 476 | NA | NA | 0476 |
| Era date | 33 BC | NA | NA | -0032 |
| Written date | First of February, two thousand and twelve | NA | NA | 2012-02-01 |
| DMY date | 10-31-2012 | NA | NA | 2012-10-31 |
| MDY date | 31-10-2012 | 0031-10-20 | NA | 2012-10-31 |
| Wrongly specified date | 2012-31-10 | NA | NA | 2012-10-31 |
| Approximate date | 2012-01-12~ | 2012-01-12 | 2012-01-12 | 2012-01-12~ |
| Uncertain date | 2012-01-01? | 2012-01-01 | 2012-01-01 | 2012-01-01? |
| Unspecified date | 2012-01 | NA | 2020-12-01 | 2012-01 |
| Censored date | ..2012-01-12 | NA | 2012-01-12 | ..2012-01-12 |
| Range of dates | 2012-11-01:2012-12-01 | 2012-11-01 | 2012-11-01 | 2012-11-01..2012-12-01 |
| Set of dates | 2012-5-26, 2012-11-19, 2012-12-4 | 2012-05-26 | NA | {2012-05-26,2012-11-19,2012-12-04} |
| Date-time | 2012-01-01 14:30 | 2012-01-01 | 2020-12-01 | 2012-01-01 14:30 |
| Approximate time | 2012-01-01 ~14:30 | 2012-01-01 | 2020-12-01 | 2012-01-01 ~14:30 |
As can be seen in the table above, other date/time packages in R do not handle ‘messy’ dates well. Normal “yyyy-mm-dd” structures or other date formats that can easily be coerced into this structure are usually not a problem.
However, some syntaxes are entirely ignored, such as historical dates and dates from other eras (e.g. BCE), as well as written dates, frequently used in historical texts or treaties.
Other times, existing packages return a date, but strip away any annotations that express uncertainty or approximateness, introducing artificial precision.
And sometimes returning only a single date means ignoring other
information included. We see this here in how only the end of the
censored date, only the start of the date range, or the first in the set
of dates is returned. Sometimes date components even seem guessed, such
as how 2021-01 (January 2021) is assumed to be 1 December 2021 by
{lubridate}.
So only {messydates} enables researchers to retain all this
information. But most analysis does still expect some precision in dates
to work.
The first way that {messydates} assists researchers that use dates in
mdate class is to provide methods for converting back into common date
classes such as Date, POSIXct, and POSIXlt. It is thus fully
compatible with packages such as {lubridate} and {anydate}.
As messy date annotations can indicate multiple possible dates,
{messydates} allows e.g. ranges or sets of dates to be unpacked or
expanded into all compatible dates.
Since most methods of analysis or modelling expect single date
observations, we offer ways to resolve this multiplicity when coercing
mdate-class objects into other date formats. For example, researcher
might explicitly choose to favour the min(), max(), mean(),
median(), or even a random() date. This greatly facilitates research
transparency by demanding a conscious choice from researchers, as well
as supporting robustness checks by enabling description or inference
across dates compatible with the messy annotated date.
md <- pkg_comparison$messydates
resolve_mdate <- data.frame(
messydates = as.character(md),
min = md |> as.Date(vmin),
median = md |> as.Date(vmedian),
max = md |> as.Date(vmax),
stringsAsFactors = FALSE)| messydates | min | median | max |
|---|---|---|---|
| 2012-01-01 | 2012-01-01 | 2012-01-01 | 2012-01-01 |
| 2599-12-31 | 2599-12-31 | 2599-12-31 | 2599-12-31 |
| 0476 | 0476-01-01 | 0476-07-01 | 0476-12-31 |
| -0032 | -032-01-01 | -032-07-02 | -032-12-31 |
| 2012-02-01 | 2012-02-01 | 2012-02-01 | 2012-02-01 |
| 2012-10-31 | 2012-10-31 | 2012-10-31 | 2012-10-31 |
| 2012-10-31 | 2012-10-31 | 2012-10-31 | 2012-10-31 |
| 2012-10-31 | 2012-10-31 | 2012-10-31 | 2012-10-31 |
| 2012-01-12~ | 2012-01-12 | 2012-01-12 | 2012-01-12 |
| 2012-01-01? | 2012-01-01 | 2012-01-01 | 2012-01-01 |
| 2012-01 | 2012-01-01 | 2012-01-16 | 2012-01-31 |
| ..2012-01-12 | 2012-01-12 | 2012-01-12 | 2012-01-12 |
| 2012-11-01..2012-12-01 | 2012-11-01 | 2012-11-16 | 2012-12-01 |
| {2012-05-26,2012-11-19,2012-12-04} | 2012-05-26 | 2012-11-19 | 2012-12-04 |
| 2012-01-01 14:30 | 2012-01-01 | 2012-01-01 | 2012-01-01 |
| 2012-01-01 ~14:30 | 2012-01-01 | 2012-01-01 | 2012-01-01 |
As can be seen in the table above, all ‘precise’ dates are respected as
such, and returned no matter what ‘resolution’ function is given. But
for messy dates, the choice of function can make a difference. Where
only a year is given, e.g. 0476 or -0032, we draw from all the days
in the year. The minimum is the first of January and the maximum the
31st of December. Dates are also drawn from a set or range of dates when
given.
When only an approximate or censored date is known, then depending on whether the whole date or just a component of the date is annotated, then a range of dates is imputed based on some window (by default 3 years, months, or days), and then a precise date is resolved from that.
This translation via an expanded list of compatible dates is fast, robust, and extensible, allowing researchers to use messy dates in an analytic strategy that uses any other package.
Following ISO 8601-2:2019, mdate objects can also carry a time of day,
appended to a date with a space (as ISO 8601-1 and RFC 3339 both permit
as an alternative to T, and messydates uses for readability; T
continues to be accepted on input). Hours, minutes, and seconds (with
fractional seconds), am/pm times, the UTC designator Z, and numeric
offsets such as +02:00 are all parsed and standardised. Time
components accept the same approximate (~), uncertain (?), and
unspecified (X) annotations as dates.
library(messydates)
as_messydate(c("2019-03-01 14:30:00Z", "2019-03-01 2:30pm", "2019-03-01 ~14:30"))
#> 'mdate' chr [1:3] "2019-03-01 14:30:00Z" "2019-03-01 14:30" ...
# extract and measure sub-day components
hour(as_messydate("2019-03-01 14:30:00"))
#> [1] 14
precision(as_messydate("2019-03-01 14:30")) # 1440 = 1/minute of a day
#> [1] 1440
# sub-day arithmetic and sequences
as_messydate("2019-03-01 14:30:00") + "2 hours"
#> 'mdate' chr "2019-03-01 16:30:00"
seq(as_messydate("2019-03-01 09:00"), as_messydate("2019-03-01 12:00"), by = "hour")
#> [1] "2019-03-01 09:00:00" "2019-03-01 10:00:00" "2019-03-01 11:00:00"
#> [4] "2019-03-01 12:00:00"Because : also serves as a range separator, times are detected first,
so 2009-01-01:2019-01-01 is still read as a range. Sub-day arithmetic,
sequences, and coercion to and from POSIXct/POSIXlt all preserve the
time of day.
Please see the cheat sheet and the messydates
website for more information
about how to use {messydates}.
The easiest way to install {messydates} is directly from CRAN:
install.packages("messydates")However, you may also install the development version from GitHub.
# install.packages("remotes")
remotes::install_github("globalgov/messydates")The package was developed as part of the PANARCHIC project, which studies the effects of network and power on how quickly states join, reform, or create international institutions by examining the historical dynamics of institutional networks from different domains.
The PANARCHIC project is funded by the Swiss National Science Foundation (SNSF). For more information on current projects of the Geneva Global Governance Observatory, please see our Github website.