Dates and Times Made Easy with lubridate

This paper presents the lubridate package for R , which facilitates working with dates and times. Date-times create various technical problems for the data analyst. The paper highlights these problems and oﬀers practical advice on how to solve them using lubridate . The paper also introduces a conceptual framework for arithmetic with date-times in R .


Introduction
Date-time data can be frustrating to work with. Dates come in many different formats, which makes recognizing and parsing them a challenge. Will our program recognize the format that we have? If it does, we still face problems specific to date-times. How can we easily extract components of the date-times, such as years, months, or seconds? How can we switch between time zones, or compare times from places that use daylight savings time (DST) with times from places that do not? Date-times create even more complications when we try to do arithmetic with them. Conventions such as leap years and DST make it unclear what we mean by "one day from now" or "exactly two years away." Even leap seconds can disrupt a seemingly simple calculation. This complexity affects other tasks too, such as constructing sensible tick marks for plotting date-time data.
While base R (R Development Core Team 2011) handles some of these problems, the syntax it uses can be confusing and difficult to remember. Moreover, the correct R code often changes depending on the class of date-time object being used. lubridate acknowledges these problems and makes it easier to work with date-time data in R. It also provides tools for manipulating date-times in novel but useful ways. lubridate will enhance a user's experience for any analysis that includes date-time data. Specifically, lubridate helps users:

Motivation
To see how lubridate simplifies things, consider a common scenario. Given a character string, we would like to read it in as a date-time, extract the month, and change it to February (i.e, 2). Table 1 shows two ways we could do this. On the left are the base R methods we would use for these three tasks. On the right are the lubridate methods. Now we will go a step further. In Table 2, we move our date back in time by one day and display our new date in the Greenwich Meridian time zone (GMT). Again, base R methods are shown on the left, lubridate methods on the right.
lubridate makes basic date-time manipulations much more straightforward. Plus, the same lubridate methods work for most of the popular date-time object classes (Date, POSIXt, chron, etc.), which is not always true for base R methods. Table 3 provides a more complete comparison between lubridate methods and base R methods. It shows how lubridate can simplify each of the common date-time tasks presented in the article "Date and Time Classes in R" (Grothendieck and Petzoldt 2004). It also provides a useful summary of lubridate methods.

Parsing date-times
We can read dates into R using the ymd() series of functions provided by lubridate. These functions parse character strings into dates. The letters y, m, and d correspond to the year, month, and day elements of a date-time. To read in a date, choose the function name that matches the order of elements in your date-time object. For example, in the following date the month element comes first, followed by the day and then the year. So we would use the mdy() function: R> mdy("12-01-2010") [1] "2010-12-01 UTC" The same character string can be parsed as January 12, 2001 by reversing the month and day element with dmy().
[1] "2010-12-31 UTC" "2011-01-01 UTC" These functions create a POSIXct date-time object that matches the date described by the character string. The functions automatically recognize the separators commonly used to record dates. These include: "-", "/", ".", and "" (i.e., no separator). When a ymd() function is applied to a vector of dates, lubridate will assume that all of the dates have the same order and the same separators. ymd() type functions also exist for times recorded with hours, minutes, and seconds. Hour, minute, and second measurements that are not accompanied by a date will be parsed as period objects, which are a type of timespan object, see Section 5.4. These functions make it simple to parse any date-time object that can be converted to a character string. See Table 4 for a complete list of ymd() type parsing functions.

Manipulating date-times
Every date-time is a combination of different elements, each with its own value. For example, most date-times include a year value, a month value, a day value and so on. Together these elements specify the exact moment that the date-time refers to. We can easily extract each element of a date-time with the accessor function that has its name, as shown in Table 5. For example, if we save the current system time R> date <-now() we can extract each of its elements. Note that this was the system time when this example was written. now() will return a different date-time each time it is used.

Date component Accessor
Year week ( Table 5: Each date-time element can be extracted with its own accessor function.
[1] 51 For the month and weekday elements (wday), we can also specify whether we want to extract the numerical value of the element, an abbreviation of the name of the month or weekday, or the full name. For example,
R> day(date) <-1 R> month(date) <-month(date) + 1 R> day(date) <-day(date) -1 [1] "2010-03-31 09:51:48 CDT" lubridate also provides an update method for date-times. This is useful if you want to change multiple attributes at once or would like to create a modified copy instead of transforming in place.

Arithmetic with date-times
Arithmetic with date-times is more complicated than arithmetic with numbers, but it can be done accurately and easily with lubridate. What complicates arithmetic with date-times? Clock times are periodically re-calibrated to reflect astronomical conditions, such as the hour of daylight or the Earth's tilt on its axis relative to the sun. We know these re-calibrations as daylight savings time, leap years, and leap seconds. Consider how one of these conventions might complicate a simple addition task. If today were January 1st, 2010 and we wished to know what day it would be one year from now, we could simply add 1 to the years element of our date.
January 1st, 2010 + 1 year = January 1st, 2011 Alternatively, we could add 365 to the days element of our date because a year is equivalent to 365 days.
January 1st, 2010 + 365 days = January 1st, 2011 Troubles arise if we try the same for January 1st, 2012. 2012 is a leap year, which means it has an extra day. Our two approaches above now give us different answers because the length of a year has changed. January 1st, 2012 + 1 year = January 1st, 2013 January 1st, 2012 + 365 days = December 31st, 2012 At different moments in time, the lengths of months, weeks, days, hours, and even minutes will also vary. We can consider these to be relative units of time; their length is relative to when they occur. In contrast, seconds always have a consistent length. Hence, seconds are exact units of time.
Researchers may be interested in exact lengths, relative lengths, or both. For example, the speed of a physical object is most precisely measured in exact lengths. The opening bell of the stock market is more easily modeled with relative lengths.
lubridate allows arithmetic with both relative and exact units by introducing four new time related objects. These are instants, intervals, durations, and periods. These concepts are borrowed from the Joda Time project (Colebourne and O'Neill 2010). Similar concepts for instants, periods, and durations also appear in the C++ library Boost.Date Time (Garland 2011). lubridate provides helper functions, object classes and methods for using all four concepts in the R language.

Instants
An instant is a specific moment in time, such as January 1st, 2012. We create an instant each time we parse a date into R.
R> start_2012 <-ymd_hms("2012-01-01 12:00:00") lubridate does not create a new class for instant objects. Instead, it recognizes any date-time object that refers to a moment of time as an instant. We can test if an object is an instant by using is.instant(). For example, We can also capture the current time as an instant with now(), and the current day with today().

Intervals
Intervals, durations, and periods are all ways of recording timespans. Of these, intervals are the most simple. An interval is a span of time that occurs between two specific instants. The length of an interval is never ambiguous, because we know when it occurs. Moreover, we can calculate the exact length of any unit of time that occurs during it.lubridate introduces the interval object class for modelling intervals.
We can create interval objects by subtracting two instants or by using the command new_interval().
Intervals always begin at the date-time that occurs first and end at the date-time that occurs last. Hence, intervals always have a positive length.

R> int_end(span)
[1] "2011-01-01 12:00:00 UTC" Unfortunately, since intervals are anchored to their start and end dates, they are not very useful for date-time calculations. It only makes sense to add an interval to its start date or to subtract it from its end date.

R> start_2010 + span
[1] "2011-01-01 12:00:00 UTC" Adding intervals to other date-times will not produce an error message. Instead lubridate will coerce the interval to a duration object, which is like an interval but without the reference dates. See Section 5.3.

R> start_2011 + span
coercing interval to duration [1] "2012-01-01 12:00:00 UTC" In most cases this will return the intended result, but accuracy can be ensured by first explicitly converting the interval to either a duration or a period, as described in the next two sections.

Durations
If we remove the start and end dates from an interval, we will have a generic time span that we can add to any date. But how should we measure this length of time? If we record the time span in seconds, it will have an exact length since seconds always have the same length. We call such time spans durations. Alternatively, we can record the time span in larger units, such as minutes or years. Since the length of these units varies over time, the exact length of the time span will depend on when it begins. These non-exact time spans are called periods and will be discussed in the next section.
The length of a duration is invariant to leap years, leap seconds, and daylight savings time because durations are measured in seconds. Hence, durations have consistent lengths and can be easily compared to other durations. Durations are the appropriate object to use when comparing time based attributes, such as speeds, rates, and lifetimes. difftime objects from base R are one type of duration object. lubridate provides a second type: duration class objects. These objects can be used with other date-time objects without worrying about what units they are displayed in. A duration object can be created with the function new_duration(): R> new_duration (60) [1] 60s For large durations, it becomes inconvenient to describe the length in seconds. For example, not many people would recognize 31557600 seconds as the length of a standard year. For this reason, large duration objects are followed in parentheses by an estimated length. Estimated units are created using the following relationships. A minute is 60 seconds, an hour 3600 seconds, a day 86400, a week 604800, and a year 31557600 (365.25 days). Month units are not used because they are so variable. The estimates are only provided for convenience; the underlying object is always recorded as a fixed number of seconds.
duration objects can be easily created with the helper functions dyears(), dweeks(), ddays(), dhours(), dminutes(), and dseconds(). The d in the title stands for duration and distinguishes these objects from period objects, discussed in Section 5.4. Each object creates a duration in seconds using the estimated relationships given above. The argument of each function is the number of estimated units we wish to include in the duration. For example, (1) [1] 60s

Periods
Periods record a time span in units larger than seconds, such as years, months, weeks, days, hours, and minutes. For convenience, we can also create a period that only uses seconds, but such a period would have the same properties as a duration. lubridate introduces the period class to model periods. We construct period objects with the helper functions years(), months(), weeks(), days(), hours(), minutes(), and seconds(). (3) [1] 3 months R> months(3) + days (2) [1] 3 months and 2 days

R> months
These functions do not contain a "d" in their name, because they do not create durations; they no longer have consistent lengths (as measured in seconds). For example, months(2) always has the length of two months even though the length of two months will change depending on when the period begins. For this reason, we can not compute exactly how long a period will be in seconds until we know when it occurs. However, we can still perform date-time calculations with periods. When we add or subtract a period to an instant, the period becomes anchored to the instant. The instant tells us when the period occurs, which allows us to calculate its exact length in seconds.
As a result, we can use periods to accurately model clock times without knowing when events such as leap seconds, leap days, and DST changes occur.

R> as.period(span)
[1] 1 year Periods can be added to instants, intervals, and other periods, but not to durations.

Division with timespans
We often wish to answer questions that involve dividing one timespan by another. For example, "how many weeks are there between Halloween and Christmas?" or "how old is a person born on September 1, 1976?" Objects of each timespan class-interval, duration, and period-can be divided by objects of the others. The results of these divisions varies depending on the nature of the timespans involved. Modulo operators (i.e, %% and %/%) also work between timespan classes.
To illustrate this, we make an interval that lasts from Halloween until Christmas.

R> interval / dweeks
Intervals are also exact measures of timespans. Although it is more work, we could divide an interval by another interval to get an exact answer. This gives the same answer as above because lubridate automatically coerces interval objects in the denominator to duration objects.

R> interval / new_interval(halloween, halloween + weeks(1))
interval denominator coerced to duration [1] 7.857143 Division is not possible with periods. Since periods have inconsistent lengths, we can not express the remainder as a decimal. For example, if we were to divide our interval by months, the remainder would be 24 days.

R> interval \%\% months(1)
[ If we attempt to divide by a period object, lubridate will return a warning message and automatically perform integer division: it returns the integer value equal to the number of whole periods that occur in the timespan. This is equivalent to the %/% operator. As shown above, we can retrieve the remainder using the %% operator.   (1) [1] 1

R> interval / months(1)
In summary, arithmetic with date-times involves four types of objects: instants, intervals, durations, and periods. lubridate creates new object classes for intervals, durations, and periods. It recognizes that most common date-time classes, such as POSIXt and Date, refer to instants. Table 6 describes which objects can be added to each other and what type of object will result.

Rounding dates
Like numbers, date-times occur in order. This allows date-times to be rounded. lubridate provides three methods that help perform this rounding: round_date(), floor_date(), and ceiling_date(). The first argument of each function is the date-time or date-times to be rounded. The second argument is the unit to round to. For example, we could round April 20, 2010 to the nearest day: R> april20 <-ymd_hms("2010-04-20 11:33:29") R> round_date(april20, "day") [1] "2010-04-20 UTC" or the nearest month.
ceiling_date() provides a second simple way to find the last day of a month. Ceiling the date to the next month and then subtract a day.

Time zones
Time zones give multiple names to the same instant. For example, "2010-03-26 11:53:24 CDT" and "2010-03-26 12:53:24 EDT" both describe the same instant. The first shows how the instant is labeled in the United States' central time zone (CDT). The second shows how the same instant is labelled in the United States' eastern time zone (EDT). Time zones complicate date-time data but are useful for mapping clock time to local daylight conditions. When working with instants, it is standard to give the clock time as it appears in the Coordinated Universal time zone (UTC). This saves calculations but can be annoying if your computer insists on translating times to your current time zone. It may also be inconvenient to discuss clock times that occur in a place unrelated to the data.
lubridate eases the frustration caused by time zones in two ways. We can change the time zone in which an instant is displayed by using the function with_tz(). This changes how the clock time is displayed, but not the specific instant of time that is referred to. For example, R> date with_tz() and force_tz() only work with time zones recognized by your computer's operating system. This list of time zones will vary between computers. See the base R help page for Sys.timezone() for more information.

Daylight savings time
In many parts of the world, the official clock time springs forward by one hour in the spring and falls back one hour in the fall. For example, in Chicago, Illinois a change in daylight savings time occurred at 2:00 AM on March 14, 2010. The last instant to occur before this change was 2010-03-14 01:59:59 CST.

R> dst_time + hours(2)
[1] "2010-03-14 03:59:59 CDT" displays the clock time that usually occurs two hours after 1:59:59 AM. When using periods, we do not have to track DST changes because they will not affect our calculations. This will prevent errors when we are trying to model events that depend on clock times, such as the opening and closing bells of the New York Stock Exchange. Adding a duration would give us the actual clock time that appeared exactly two hours later on March 14, 2010. This prevents changes in clock time from interfering with exact measurements, such as the life times of a light bulb, but will create surprises if clock times are important to our analysis.
We can also avoid the complications created by daylight savings time by keeping our datetimes in a time zone such as "UTC", which does not adopt daylight savings hours.

Case study 1
The next two sections will work through some techniques using lubridate. First, we will use lubridate to calculate the dates of holidays. Then we will use lubridate to explore an example data set (lakers).

Thanksgiving
Some holidays, such as Thanksgiving (US) and Memorial Day (US) do not occur on fixed dates. Instead, they are celebrated according to a common rule. For example, Thanksgiving is celebrated on the fourth Thursday of November. To calculate when Thanksgiving will occur in 2010, we can start with the first day of 2010.

[1] Monday
This implies November 4th will be the first Thursday of November.

Memorial Day
Memorial day also occurs according to a rule; it falls on the last Monday of May. To calculate the date of Memorial day, we can again start with the first of the year.
R> month(date) <-5 [1] "2010-05-01 UTC" Now our holiday occurs in relation to the last day of the month instead of the first. We find the last day of the month by rounding up to the next month and then subtracting a day.
R> date <-ceiling_date(date, "month") -days (1) [1] "2010-05-31 UTC" We can then check which day of the week May 31st falls on. It happens to be a Monday, so we are done. If May 31st had been another day of the week, we could've subtracted an appropriate number of days to get to the last Monday of May.

Case study 2
The lakers data set contains play by play statistics of every major league basketball game played by the Los Angeles Lakers during the 2008-2009 season. This data is from http://www. basketballgeek.com/downloads/ (Parker 2010) and comes with the lubridate package. We will explore the distribution of Lakers' games throughout the year as well as the distribution of plays within Lakers' games. We choose to use the ggplot2 (Wickham 2009) package to create our graphs.
The lakers data set comes with a date variable which records the date of each game. Using the str() command, we see that R recognizes the dates as integers. Before we can work with them as dates, we must parse them into R as date-time objects. The dates appear to be arranged with their year element first, followed by the month element, and then the day element. Hence, we should use the ymd() parsing function.
R> lakers$date <-ymd(lakers$date) R> str(lakers$date) R now recognizes the dates as POSIXct date-time objects. It will now treat them as date-times in any functions that have POSIXct specific methods. For example, if we plot the occurrences of home and away games throughout the season, our x axis will display date-time information for the tick marks ( Figure 1).
R> qplot(date, 0, data = lakers, colour = game_type) Figure 1 shows that games are played continuously throughout the season with a few short breaks. The frequency of games seems lower at the start of the season and games appear to be grouped into clusters of home games and away games. The tick marks and breaks on the x axis are automatically generated by the lubridate method pretty.dates().
Next we will examine how Lakers games are distributed throughout the week. We use the wday() command to extract the day of the week of each date.
R> qplot(wday(date, label = TRUE, abbr = FALSE), data = lakers, geom = "histogram") The frequency of basketball games varies throughout the week (Figure 2). Surprisingly, the highest number of games are played on Tuesdays.
Now we look at the games themselves. In particular, we look at the distribution of plays throughout the game. The lakers data set lists the time that appeared on the game clock for each play. These times begin at 12:00 at the beginning of each period and then count down to 00:00, which marks the end of the period. The first two digits refer to the number of minutes left in the period. The second two digits refer to the number of seconds.
The times have not been parsed as date-time data to R. It would be difficult to record the time data as a date-time object because the data is incomplete: a minutes and seconds element are not sufficient to identify a unique instant of time. However, we can store the minutes and seconds information as a period object, as defined in Section 5.4, using the ms() parse function.

R> lakers$time <-ms(lakers$time)
Since periods have relative lengths, it is dangerous to compare them to each other. So we should next convert our periods to durations, which have exact lengths.

R> lakers$time <-as.duration(lakers$time)
This allows us to directly compare different durations. It would also allow us to determine exactly when each play occurred by adding the duration to the instant the game began.
(Unfortunately, the starting time for each game is not available in the data set). However, we can still calculate when in each game each play occurred. Each period of play is 12 minutes long and overtime-the 5th period-is 5 minutes long. At the start of each period, the game clock begins counting down from 12:00. So to calculate how much play time elapses before each play, we subtract the time that appears on the game clock from a duration of 12, 24, 36, 48, or 53 minutes (depending on the period of play). We now have a new duration that shows exactly how far into the game each play occurred. -dminutes(c(12, 24, 36, 48, 53)[lakers$period]) -+ lakers$time

R> lakers$time <
We can now plot the number of events over time within each game ( Figure 3). We can plot the time of each event as a duration, which will display the number of seconds into the game each play occurred on the x axis, Figure 3: The graph on the left displays seconds on the x axis. The graph on the right uses a more intuitive division with minutes.
R> qplot(time, data = lakers, geom = "histogram", binwidth = 60) or we can take advantage of pretty.date() to make pretty tick marks by first transforming each duration into a date-time. This helper function recognizes the most intuitive binning and labeling of date-time data, which further enhances our graph. To change durations into datetimes we can just add them all to the same date-time. It does not matter which date we chose. Since the range of our data occurs entirely within an hour, only the minutes information will display in the graph.
R> lakers$minutes <-ymd("2008-01-01") + lakers$time R> qplot(minutes, data = lakers, geom = "histogram", binwidth = 60) We see that the number of plays peaks within each of the four periods and then plummets at the beginning of the next period, Figure 3. The most plays occur in the last minute of the game. Perhaps any shot is worth taking at this point or there's less of an incentive not to foul other players. Fewer plays occur in overtime, since not all games go to overtime. Now lets look more closely at just one basketball game: the game played against the Boston Celtics on Christmas of 2008. We can quickly model the amounts of time that occurred between each shot attempt.
Since we have recorded the time of each shot attempt as a duration (above), we can record the differences by subtracting the two durations. This automatically creates a new duration whose length is equal to the difference between the first two durations. Figure 4: Wait times between shot attempts rarely lasted more than 30 seconds.
We can also examine changes in the score throughout the game. This reveals that though the game was eventful, the Lakers maintained a lead for the most of the game, Figure 5. Note: the necessary calculations are made simpler by the ddply() function from the plyr package, which lubridate automatically loads.

Conclusion
Date-times create technical difficulties that other types of data do not. They must be specifically identified as date-time data, which can be difficult due to the overabundance of date-time classes. It can also be difficult to access and manipulate the individual pieces of data contained within a date-time. Arithmetic with date-times is often appropriate, but must follow different rules than arithmetic with ordinal numbers. Finally, date-related conventions such as daylight savings time and time zones make it difficult to compare and recognize different moments of time.
Base R handles many of these difficulties, but not all. Moreover, base R's date-time methods can be complicated and confusing. lubridate makes it to easier to work with date-time data in R. The package provides a set of standard methods for most common date-time classes. These methods make it simple to parse, manipulate, and perform calculations with date-time objects. By implementing the time concepts pioneered by projects such as Joda Time and Boost.Date Time, lubridate helps researchers perform precise calculations as well as model tricky time-related processes. lubridate also makes it simple to switch between time zones and to use or ignore daylight savings time.
Future work on the lubridate package will develop methods for handling partial dates and for modeling recurrent events, such as stock market opening times, business days, or street cleaning hours. In particular, we hope to create methods for R that work with reoccurring temporal date patterns, which were introduced by Fowler (1997) and have been implemented in Ruby by the runt project (Lipper 2008).