The sleep loss insult of Spring Daylight Savings in the US is absorbed by Twitter users within 48 hours
Kelsey Linnell, Thayer Alshaabi, Thomas McAndrew, Jeanie Lim, Peter Sheridan Dodds, Christopher M. Danforth
TThe sleep loss insult of Spring Daylight Savings in the USis absorbed by Twitter users within 48 hours
Kelsey Linnell,
1, 2, ∗ Thayer Alshaabi, Thomas McAndrew, JeanieLim, Peter Sheridan Dodds,
1, 2 and Christopher M. Danforth
1, 2, † Computational Story Lab, Vermont Complex Systems Center, MassMutual Centerof Excellence for Complex Systems & Data Science, University of Vermont Department of Mathematics & Statistics, University of Vermont Department of Biostatistics & Epidemiology, School of PublicHealth & Health Sciences, University of Massachusetts at Amherst MassMutual Data Science
Abstract:
Sleep loss has been linked to heart disease, diabetes, cancer, and an increase in accidents,all of which are among the leading causes of death in the United States. Population-scale sleepstudies have the potential to advance public health by helping to identify at-risk populations, changesin collective sleep patterns, and to inform policy change. Prior research suggests other kinds of healthindicators such as depression and obesity can be estimated using social media activity. However,the inability to effectively measure collective sleep with publicly available data has limited large-scale academic studies. Here, we investigate the passive estimation of sleep loss through a proxyanalysis of Twitter activity profiles. We use “Spring Forward” events, which occur at the beginningof Daylight Savings Time in the United States, as a natural experimental condition to estimatespatial differences in sleep loss across the United States. On average, peak Twitter activity occursroughly 45 minutes later on the Sunday following Spring Forward. By Monday morning however,activity curves are realigned with the week before, suggesting that at least on Twitter, the lost hourof early Sunday morning has been quickly absorbed.
I. INTRODUCTION
The American Academy of Sleep Medicine recom-mends adults sleep 7 or more hours per night [1]. How-ever, studies show only 2/3 of adults sleep for this lengthof time consistently. In 2014, the Centers for DiseaseControl and Prevention’s (CDC’s) Behavioral Risk Fac-tor Surveillance System suggested that between 28% and44% of the adult population of each state received lessthan the recommended 7 hours of sleep [2]. Despite thescientific consensus that adequate sleep is essential tohealth, many adults are sleeping less than hours a nighton average—a state referred to as short sleep . Resultsfrom the most recent National Health Interview Sur-vey determined that since 1985, the age-adjusted aver-age sleep duration has decreased, and the percentage ofadults who experience short sleep, on average, rose by31% [3].Because adequate sleep is necessary for optimal cogni-tion, short sleep is adverse to productivity and learning,and reduces the human capacity to make effort- relat-ed choices such as whether to take precautionary safetymeasures [4–6]. Short sleep’s impact on human cognitionis harmful in the workplace, and poses a pronounced anddistinct threat to public safety when operating a vehi-cle [7–10]. Short sleep is linked to increased risk of seri-ous health conditions, including heart disease, obesity,diabetes, arthritis, depression, strokes, hypertension, and ∗ [email protected] † [email protected] cancer [2, 11–13], and a recent study found that disruptedsleep is also associated with DNA damage [14]. The linkbetween sleep loss and cancer is so strong that the WorldHealth Organization has classified night shift work as“probably carcinogenic to humans” [15]. Socio-economicstatus is positively correlated with quality of sleep [16–19]. Due to such detrimental effects, and high preva-lence among the population, insufficient sleep accountsfor between $280 and over $400 billion lost in the UnitedStates every year [20].Accurately measuring short sleep in a large populationis difficult, and there is often a trade-off between accuracyand the size of the study. Polysomnography—consideredthe most accurate way to measure sleep—can only mea-sure an individual’s sleep patterns in a controlled labora-tory setting [21, 22]. Large studies have relied on partici-pants recording their own sleep, but suffer from reportingbias [2, 23, 24].Wearable technology can measure short sleep at thepopulation scale, and has the potential to measure shortsleep accurately enough to study its association withadverse health risks [4, 21, 25]. One recent large sleepstudy enrolled 31,000 participants and used sleep datafrom wearable devices along with participant’s interac-tions with a web based search engine to compare sleeploss and performance [4]. The authors [4] showedthat measurements of cognitive performance (includingkeystroke and click latency) vary over time, follow a cir-cadian rhythm, and are related to the duration of par-ticipant’s sleep, results that closely mirrored those fromlaboratory settings and validated their methodology.While promising in the long run, present studies thatuse wearable devices have limitations. To infer fromTypeset by REVTEX a r X i v : . [ q - b i o . Q M ] A p r wearables that individuals are sleeping, data must firstgo through a pipeline of preprocessing, feature extrac-tion and classificiation. The pipeline for processing sleepdata is typically proprietary and dependent on the spe-cific wearable used, and changes to how data is processedcan impact results [26]. Moreover, validation studies haveyet to explore the effectiveness of these devices acrossgenders, ages, culture, and health [26].Social media may be an alternative way to measuresleep disturbances in a large population, for example bystudying the link between screen time and sleep [27, 28].Past work has found a correlation between sustainedlow activity on Twitter and sleep time as measured byconventional surveys, and these results were validatedagainst data collected from the CDC on sleep depriva-tion [27]. Other work has shown evidence of an increasein a user’s smart phone screen time as being associat-ed with an increase in short sleep [28]. Other mentaland physical characteristics have been measured fromsociotechnical systems. Several instruments developed bymembers of our research group including the Hedonome-ter [29], which measures population sentiment throughtweets, and the Lexicocalorimeter [30], which measurescaloric balance at the state level, have demonstrated anability to infer population-scale health metrics from Twit-ter data. Twitter data has also been used to identify userswho experience sleep deprivation and study the waystheir social media interactions differ from others [31].In urban, industrialized societies where social timingis synced to clock time, Daylight Savings- a biannualsudden upset to clock time- creates behavioral stabilityacross seasons [32, 33]. Past work has used Daylight Sav-ings as a natural experiment to show that a one hour col-lective sleep loss event has large and quantifiable effectson health, safety, and the economy [34–37], with twostriking findings being a one day increase in heart attacksby 24% and a loss of $31 billion on the NYSE, AMEX,and NASDAQ exchanges in the United States [34, 38].We hypothesize here that sleep loss is measurablein behavioral patterns on Twitter, and changes inpopulation-scale sleep patterns due to Spring Forwardcan be observed through changes in these behavioral pat-terns. In what follows, we first outline our methodologyfor estimating sleep loss from tweets, describing the dataand study design. We then visualize and describe theresults before concluding with a discussion of limitationand implications. II. METHODSData
We collected a 10% random sample of all publictweets—offered by Twitter’s Decahose API—for Sundaysand Mondays in the four weeks leading up to, the weekof, and the four weeks following Spring Forward eventsduring the years 2011-2014. Spring Forward is defined as the instantaneous clock adjustment from 2 a.m. to3 a.m. on the second Sunday of March each year. Weincluded tweets in the study if the user who created thetweet reported living in the U.S. in their bio, or if thetweet was geo-tagged to a GPS coordinate within theU.S. [39]. With these conditions, we ended up select-ing approximately 7% of the messages in the Decahoserandom sample for analysis [40].Twitter provided the time-zone from which each mes-sage was posted during the period from 2011 to 2014 (forprivacy purposes, Twitter discontinued publication oftime zone information in 2015). We used the time-zoneto determine the local time of posting for each tweet. Webinned tweets by 15 minute increments according to thelocal time of day they were posted.
Experimental setup
To estimate behavioral change associated with Day-light Savings, we partitioned tweets into various groups,primarily a “Before Spring Forward” (BSF) group and a“Spring Forward” (SF) group. To establish a convenient‘control’ pattern of behavior, all tweets posted on any ofthe four Sundays before the Spring Forward event wereclassified as “Before Spring Forward” tweets. We classi-fied the ‘experimental’ set of tweets posted on the Sundaycoincident with the Spring Forward event as “Spring For-ward”. The above classification created, for every year, a4:1 matching of before to week of Spring Forward activi-ty. We analyzed tweets posted 1-4 weeks following SpringForward separately to quantify relaxation to the originalbehavior.
Analysis
We binned tweets by time in 15 minute intervals start-ing at the top of the hour, and normalized their frequen-cies by dividing by the total number of tweets posted onthe corresponding day. In this way, we establish a dis-crete description of the posting volume over the courseof a typical 24-hour period.We averaged the Before Spring Forward tweets overthe four Sundays, and the four years as follows: T BSF ( k ) = (4 × − (cid:88) Y =2011 4 (cid:88) S =1 C Y S ( k ) C Y S , where C Y S ( k ) is the number of tweets in the k th S th Sunday of year Y , C Y S is thetotal number of tweets posted on that Sunday and year,and T BSF ( k ) is the average fraction of tweets posted inthe k th
15 minute interval of a Sunday prior to SpringForward,We also noramlized the Spring Forward tweets against
Local Time
Sunday Mondaybreakfastlunch dinner N o r m a li z e d A c t i v i t y E a s t e r n P a c i f i c FIG. 1 . Diurnal collective attention to meals quantified, by normalized usage of the words ‘breakfast’, ‘lunch’,and ‘dinner’ for states observing Eastern Time (top) and Pacific Time (bottom), for the weeks before (solid),and of (dashed) Spring Forward.
The x-axis represents the interval between 3 a.m. Sunday and 9 p.m. Monday localtime. Counts for tweets containing each individual word were tallied in 15 minute increments, normalized by the total numberof tweets mentioning that word, and smoothed using Gaussian Process Regression. Each day has a clear pattern for frequencyof meal name appearance in tweets, with the peak for breakfast, lunch, and dinner occurring in the respective order of themeals themselves. For each of the meals, we observe a slight forward shift in the peak following Spring Forward, suggestingthat meals are taking place later than usual on the corresponding Sunday. By Monday, the peak for each meal name appearsto be aligned with the week before, with the exception of ’dinner’ on the west coast, which is still a bit later. daily activity: T SF ( k ) = (4) − (cid:88) Y =2011 C Y ( k ) C Y . To reduce noise that could depend on our choiceof bin size and spatial scale, we smoothed normal-ized tweet activity using Gaussian Process Regression(GPR) [41, 42]. We fit a GPR with a squared expo-nential kernel and characteristic length scale of 150 min-utes (a total of 10 bins of size 15-minutes) to normalizedtweets. We chose a characteristic length of 150 minutesfor consistency with previous work [27]. Tikhonov reg-ularization with an α penalty of 0.1 was included whenfinding weights ω k to prevent overfitting [42]. GPR yield-ed a smooth behavioral curve , B ( t ) , of the functionalform: B ( t ) = (cid:88) k =1 ω k exp (cid:34) − k (cid:18) t , t k (cid:19) (cid:35) , where ω k is a weight determined by the regression pro-cess, k is the squared-exponential kernel (commonlycalled a radial basis), t is the time in minutes since mid-night (00:00), and t k is the k th
15 minute interval of theday, i.e. t corresponds to 75 minutes past midnight,or 1:15 a.m. The sum to 96 refers to the number of 15minute intervals in a single 24 hour period. We generated behavioral curves B ( t ) for the BSF andSF groups by state, and for the U.S. in aggregate. Toestimate behavioral change induced by a Spring Forwardevent, we calculate two quantities from the behavioralcurves: (i) the time of peak activity and (ii) the time ofthe inflection point between the peak and trough. Theinflection point is referred to as a ‘twinflection’ point,and represents a point of diminishing losses in Twitteractivity for the night. Peak shift is defined as: arg max t { B SF ( t ) } − arg max t { B BSF ( t ) } and twinflection shift is defined as: arg min t ∈ N { B (cid:48) SF ( t ) } − arg min t ∈ N { B (cid:48) BSF ( t ) } , where N = { t : arg max t B ( t ) < t < arg min t B ( t ) } . Wewere able to reliably measure peak activity and twin-flection because behavioral curves exhibited a consistentdiurnal wave structure: a rise in the evening correspond-ing to peak Twitter posting activity, followed by a troughduring typical sleeping hours, and a plateau throughoutthe day.We measured the loss of sleep opportunity by calcu-lating the peak and twinflection times for the four weeksBefore Spring Forward and the week of Spring Forwarditself. We then characterize differences between the BSFand SF measures for each state, and for the total U.S.,as a proxy for sleep loss.
12 18 0 6 120.0050.0100.0150.0200.025 N o r m a li z e d A c t i v i t y a United States, 2013
BSFSF
12 18 0 6 12 b California, 2013 c California, 2013
BSFSFTwinflection
Local Time
FIG. 2 . Twitter activity behavioral curves B ( t ) . (a) Normalized count of tweets posted from a location within theUnited States between 12 p.m. Sunday and 12 p.m. Monday before (red) and the week of (blue) the 2013 Spring ForwardEvent. The time recorded for the tweet is that local to the author. Though the pattern of behavior is preserved followingDaylight Savings, peak activity is translated forward in time. (b)
The same plot, with location of tweet origin restricted tothe state of California. California is the state for which we have the most data, and therefore the most representative behaviorprofile after smoothing with Gaussian Process Regression (lines). We note that figure 5 shows behavioral curves for all states. (c)
The smoothed behavioral pattern for California during the hours of 9 p.m. to 3 a.m. Pacific Time. Activity peaks aredenoted by vertical dashed lines, and twinflection points are marked by squares. To estimate the behavioral shift in time, wecompute the distance along the temporal axis between these pairs of lines/points. California’s BSF peak is 30 minutes earlierthan the SF peak.
III. RESULTS
Our overall finding is that peak Twitter activity occursroughly 45 minutes later on the Sunday evening imme-diately following Spring Forward, with this shift varyingamong states. By Monday morning, activity is back tonormal, suggesting that the hour of sleep lost is overcome,at least on Twitter, within 48 hours.In Fig 1, we plot B ( t ) for the subset of posts contain-ing the words ‘breakfast’, ‘lunch’, and ‘dinner’ for theperiod beginning 3 a.m. on Sunday and ending 9 p.m.on Monday, both before (solid) and after (dashed) SpringForward events. These curves were constructed for statesobserving Eastern Time (top row) and Pacific Time (bot-tom row).Meal-related language reveals a daily pattern of behav-ior in which peak volume occurs around the time thatmeal typically takes place. On an average Sunday, break-fast is most mentioned at 11 a.m., lunch at 1:45 p.m.,and dinner at 7 p.m. in Eastern Time Zone states (seeFig 1). On the average Monday, breakfast mentions peakat 10:15 a.m., lunch peaks at 1 p.m., and dinner at 8p.m. Breakfast is mentioned nearly twice as often on Sun-day than on Monday. Lunch shows the opposite trend,doubling on Monday in comparison to Sunday. There is essentially no discussion of meals during the period from2 a.m.-4 a.m. These plots also exhibit a small forwardshift in time following Spring Forward, suggesting thateach meal was tweeted about, and probably eaten, laterin the day on Sunday. The effect disappears by Monday.Broadening from messages mentioning specific mealsto all messages, daily activity plots of B BSF and B SF reveal a regular diurnal pattern of behavior that is con-sistently shifted forward in time the evening followingSpring Forward events. Figure 2 shows this shift for theyear 2013, but the results were similar for other years.Panel (a) suggests overall activity across the U.S. peaksaround 10 p.m. on Sundays before Spring Forward (redcircles), and experiences a minimum around 5am. Thepeak shifts approximately 45 minutes later on the Sundayof Spring Forward (blue squares) before synchronizingagain by early morning Monday. In panel (b) Califor-nia is used as an illustrative example of these patternsexisting at the state level, and the smooth behavioralpattern constructed using Gaussian Process Regression.The pattern is similar to that observed for the entirecountry, with the exception of a slightly reduced ampli-tude. Twinflection points are illustrated by black squaresin panels (b) and (c).Figure 2 demonstrates evidence that there is a shift FIG. 3 . Time of peak Twitter activity on Sunday night for each state before (top) and after (bottom) SpringForward for the four events observed between 2011 and 2014.
Before Spring Forward, the time of peak activity occursaround 10 p.m. in the Eastern Time Zone, and around 9:30 p.m. for the rest of the country. After Spring Forward, peakTwitter activity occurs between 0 and 90 minutes later for each state. Texas has the latest peak at 11 p.m. local time, a shiftof 90 minutes forward compared with prior Sundays. Pennsylvania, Hawaii, and Washington D.C. are the only states with noobserved change in peak time. We note again that the BSF estimates are based on the aggregation of four Sundays prior toSpring Forward, while the ASF estimates are based on the Sunday coincident with Spring Forward, and are therefore estimatedusing roughly 1/4 the data. in the peak time spent interacting with Twitter on Sun-day evening following Spring Forward, relative to priorSundays. Given the absence of a corresponding delay ininteraction Monday morning, we infer an increase in sleeploss experienced on Sunday night.To explore the spatial distribution of the behavioralchanges induced by Spring Forward, in Fig. 3 we mapthe time of peak Twitter activity on Sunday night for each state before (top) and the week of (bottom) SpringForward, averaged across the years 2011-2014. On theSundays leading up to Spring Forward (top), peak twitteractivity occurs near either 10 p.m. for states on the EastCoast, or 9:30 p.m., for the rest of the country. AfterSpring Forward, nearly all states exhibit peak activitylater in the night.Looking at Texas as an individual example, before
WY WVWIWA VTVAUTTX TNSD SC RIPAOR OK OH NYNV NM NJ NHNEND NCMT MSMOMN MI MEMD MALA KYKS INILID IAHI GA FL DE DCCTCOCA AZ AR ALAK a Peak Shift (mins)
WY WVWIWA VTVAUTTX TNSD SC RIPAOR OK OH NYNV NM NJ NHNEND NCMT MSMOMN MI MEMD MALA KYKS INILID IAHI GA FL DE DCCTCOCA AZ AR ALAK b -30 135 Twinflection Shift (mins)
WY WVWIWA VTVAUTTX TNSD SC RIPAOR OK OH NYNV NM NJ NHNEND NCMT MSMOMN MI MEMD MALA KYKS INILID IAHI GA FL DE DCCTCOCA AZ AR ALAK c
245 49,190
Raw Tweets
WY WVWIWA VTVAUTTX TNSD SC RIPAOR OK OH NYNV NM NJ NHNEND NCMT MSMOMN MI MEMD MALA KYKS INILID IAHI GA FL DE DCCTCOCA AZ AR ALAK d tweets per capita
30 60 30 1004/10,000 1/1000 3/1000250 1,000 10,000 50,000
Tweet count Tweets per capitaPeak shift (mins) Twinflection shift
FIG. 4 . The magnitude of Twitter behavioral shift following a Spring Forward event, averaged for the fouryears from 2011 to 2014. (a) Shift measured using behavioral curve peaks, the difference between the pair of maps inFigure 3 (bottom minus top). Texas is estimated to have experienced the greatest time shift. The effect of Spring Forwardis more pronounced in the South, and center of the country. No effect is measured for Hawaii. (b)
The same map, but withmeasurements calculated using twinflection shift instead. The states most affected are Texas and Mississippi, where the shiftwas 135 and 105 minutes respectively. Hawaii is the only state estimated to have a negative shift (30 minutes). Twinflectionshift produces similar spatial results to peak shift, with more exaggerated shift estimates. (c)
The number of tweets posted fromeach state in the period after Spring Forward. California and Texas both contributed over 40,000 tweets, while Alaska, Hawaii,Idaho, Wyoming, Montana, North Dakota, South Dakota, and Vermont each produced less than 1,000 tweets. (d)
The densityof data used to establish the experimental pattern of behavior, as measured by tweets per capita. This measurement reflectsthe ability of the data to capture the behavior of the tweeting population of each state. While Idaho, Wyoming, Montana andSouth Dakota have relatively little data compared to their populations, the remaining states have similar data density, withsomewhere between one and three tweets per thousand residents. Note: both panels (c) and (d) use logarithmically spacedcolorbars.
Spring Forward we see peak activity around 9:30 p.m.local time, and after Spring Forward it occurs at 11p.m. local time. While Texas is one of the latest peaksobserved on the evening following Spring Forward, sever-al other states are up late including Oklahoma, Georgia,and Mississippi each peaking around 10:45 p.m.In the appendix, we show maps estimating the time ofpeak activity for each of the individual 9 weeks centered on Spring Forward (Figure A1). There is some week-to-week variation, most notably in the second week priorto Spring Forward, which was the night of the AcademyAwards for three of the four years. By four weeks afterSpring Forward, the peak activity map has relaxed toroughly the same pattern as BSF.The magnitude of the forward shift in behavior illus-trated in Figure 3 is considered a proxy for the loss ofsleep opportunity on the Sunday night following SpringForward. We used two distinct methods to estimate thismagnitude, namely the peak shift and the twinflectionshift. A comparison of the spatial estimates made usingeach method are shown in Figure 4.Panel (a) illustrates the average shift in peak activi-ty observed for 2011-2014 by computing the differencebetween the pair of maps in Figure 3 (bottom minustop). While all states exhibit a shift forward in time onthe night of Spring Forward, there is clear spatial varia-tion. The peak in Twitter behavior for the east and westcoasts occurred 15-30 minutes later Sunday night, whileit occurred 45-90 minutes later for the central U.S. (Fig4 panel a).Figure 4 panel (b) estimates the change using twin-flection, namely the change in concavity of the behavioractivity curve from down to up. Every state but Hawaiiexhibits a shift forward in time, and with similar spatialregularity. When measured with twinflection shift, Texasand Mississippi are seen to have the greatest temporalshift following Spring Forward. Texans were tweeting135 minutes later than usual following a Spring Forwardevent. Most of the east and west coast states were mea-sured as tweeting 30 to 45 minutes later (Fig 4 panel b).Both measures agreed on a positive shift for the countryas a whole, and for all states exclusive of Hawaii. How-ever, the two measures yielded different results for themagnitude of these shifts, with twinflection shift gener-ally estimating a greater effect size.Figure 4 panels (c) and (d) illustrate the amountof data contributing to calculations for the behavioralcurves, and the density of this data with respect to eachstate’s population. Idaho, Alaska, Hawaii, Montana,Wyoming, North Dakota, South Dakota, and Vermontwere the states offering the smallest amount of data, andsubsequently have the highest potential for a poor behav-ioral curve model fit.Though the amount of data available for California andTexas is much greater than the other states, when con-sidering their large population size we find their twitteractivity per capita to be similar to most other states.Based on our estimate of tweets per capita, we expectbehavioral curves for most states to be more or less equal-ly representative of their tweeting populations.Looking at the diurnal cycle of Twitter activity for eachindividual state, we see remarkable consistency. Fig. 5shows the 24 hour period spanning noon Sunday to noonMonday local time for the year 2014. Plots for the other3 years exhibit similar behavior. Before Spring Forward(red), most states show a peak between 9:30 and 10:15p.m., local time. After Spring Forward (blue), nearly allstates have a peak after 10:15 p.m. By Monday morn-ing, nearly all curves have re-aligned. We also consis-tently observe higher peaks for the BSF curves whichwe believe to be driven by televised events such as theOscars. The Sunday of Spring Forward does not havea regularly scheduled popular television event, and as aresult the SF curves have lower amplitude. Both the peak and twinflection demonstrate that it ispossible to observe a measurable decrease in the amountof sleep opportunity people in the United States receiveon average due to Spring Forward. They also bothdemonstrate uneven geographic distribution of the effectof Spring Forward, and therefore the ability to determinegeographic disparity in sleep loss.We also discovered that the Super Bowl occurredexactly 5 weeks prior to Spring Forward in each of theyears studied. This annual event watched by over 100million individuals in the U.S. caused peak Twitter activ-ity to synchronize at roughly the same time nationally,around 9 p.m. Eastern, during the second half of thefootball game. The map in Figure 6 shows the time ofpeak activity for each state on Super Bowl Sunday, aver-aged over the years 2011 to 2014. The colormap is thesame as the scale used for 3, with the additional coolerrange brought in to reflect the earlier peaks in Mountainand Western time zones. The map bears a remarkableresemblance to the timezone map, demonstrating a syn-chronization of collective attention across the country.Data from Super Bowl Sunday was not included in theBefore Spring Forward data, as it does not accuratelyreflect the spatial distribution of typical posting behav-ior on a Sunday evening.
IV. DISCUSSION
Technically speaking, Spring Forward occurs very earlySunday morning, and the instantaneous clock adjustmentfrom 2 a.m. to 3 a.m. is witnessed by very few wakingindividuals. In addition, we speculate that the majorityof individuals do not set an alarm clock for Sunday morn-ing. As a result, we expect that the hour lost to SpringForward will be felt by our bodies most meaningfully onMonday morning. Indeed, we are likely to experiencethe Monday morning alarm as occurring an hour early,as Spring Forward shortens the time typically reservedfor sleep opportunity Sunday night by one hour.Considering the correlation between screen time andlack of sleep, the Sunday evening shift, and the cor-responding Monday morning re-synchronization, weobserve strong evidence that sleep opportunity is lost theevening of Spring Forward. By estimating the magnitudeand spatial distribution of the shift in Twitter behavioralcurves, we have approximated a lower bound on sleep lossat the state level.Our pair of measurement methodologies have a Pear-son correlation coefficient of 0.715, and a Spearman cor-relation coefficient of 0.645 (See Figure A2). While theyproduced slightly different estimates of the magnitudeof temporal shift in behavior, the resulting geographicprofiles of sleep loss were similar. Both suggest thatstates along the coast are least affected by Spring For-ward, while Texas and the states surrounding it to theNorth and East are the most affected.Peak shift suggests the temporal shift in behavior due F I G . . N o r m a li z e d T w i tt e r a c t i v i t y b e t w ee n p . m . Sund a y a nd p . m . M o nd a y p r i o r t o a nd f o ll o w i n g t h e Sp r i n g F o r w a r d e v e n t f o r e a c h s t a t e . R e d i nd i c a t e s a n agg r e ga t i o n o f d a t a f r o m t h e s p e c i fi e dp e r i o d o v e r f o u r w ee k s b e f o r e t h e Sp r i n g F o r w a r d E v e n t . B l u e i nd i c a t e s d a t a f r o m t h e s i n g l e h o u r p e r i o d a f t e r Sp r i n g F o r w a r dh a s o cc u rr e d . D o t s a r e i nd i c a t i v e o f ‘ r a w ’ d a t a , w h il e t h e c o rr e s p o nd i n g c u r v e s d e m o n s t r a t e s G a u ss i a n s m oo t h i n g . T e x a s e x h i b i t s t h e l a r g e s t c h a n g e f o ll o w i n g Sp r i n g F o r w a r d . C u r v e s f o r n e a r l y a ll s t a t e s h a v e a li g n e db y M o nd a y m o r n i n g . T h e B S F p e a k s a r e c o n s i s t e n t l y h i g h e r t h a n t h e S F p e a k s , l a r g e l y du e t o t e l e v i s e d e v e n t s B e f o r e Sp r i n g F o r w a r d s u c h a s t h e O s c a r s . T h e Sund a y o f Sp r i n g F o r w a r dd o e s n o t h a v e a r e g u l a r l y s c h e du l e dp o pu l a r t e l e v i s i o n e v e n t , a nd a s a r e s u l tt h e S F c u r v e s h a v e l o w e r a m p li t ud e . to Spring Forward is of a similar magnitude to the actualclock shift (1 hour). California, the state for which wehave the most data and therefore the most representa-tive behavior profile after smoothing, was found to havea peak shift of 30 minutes. Considering the clock adjust-ment of exactly one hour, the peak shift measurementseems likely to be directly representative of the sleep lost.Twinflection measured similar shifts for most states, butfor a few estimated much larger effects. While Californiawas measured as having the same 30 minute shift, Texas,the state for which we have the second most data, wasestimated by twinflection to be delayed by an additional45 minutes. While the relationship between magnitude oftwinflection shift and magnitude of sleep loss is uncertain,this measure made spatial disparities more apparent.Hawaii presents interesting and extreme results. Inboth cases, Hawaii is the state with the least measuredsleep loss by both accounts; for twinflection shift, thereis even a demonstrated gain in sleep. Considering thatHawaii does not observe DST, these results are plausi-ble. However, they should be considered tentative atbest, given the sparsity of data available. Caution shouldlikewise be extended to measurements ascribed to SouthDakota, North Dakota, Wyoming, Idaho, Montana, Ver-mont, New Hampshire, and Maine. These states havesmaller populations, less population density, and lowervolume of tweets. As a result, the behavioral curves asso-ciated with these states are less reliable.Discrepancies in available data were determined to belargely accounted for by differences in population. Thus,we expect results for each state (exclusive of those men-tioned earlier) to be comparably reliable in their repre-sentation of sleep loss for the state as a whole.Incremental future work in this area could include look-ing at the end of Daylight Savings in November, where weare ostensibly given an additional hour of sleep opportu-nity. Our findings suggest that the sleep behavior associ-ated with other annual events including New Year’s Eveand Thanksgiving ought to be visible through tweets.More ambitiously, proxy data such as this could be ver-ified by matching wearable measurements of sleep (e.g.Fitbit) with social media accounts. Limitations
Our study suffers from several limitations associatedwith our data source, we describe a few such exampleshere. The geographic location users provide in their Twit-ter bio is static and unlikely to be updated when travel-ing. As a result, user locations (time zone, state) inferredfrom this field will not always reflect their precise loca-tion. The GPS tagged messages included in our analysiswill not suffer from this same uncertainty. Furthermore,the tweeting population of each state is likely to havecomplicated biases with respect to their representationof the general population [43].Our dataset likely contains automated activity. Indeed, an entire ecology of algorithmic tweets evolvedduring the period in which we collected data for thisstudy. However, we expect the majority of this activityto be scheduled using software that updates local timeautomatically in response to Daylight Savings. As such,this ‘bot’ type activity should largely serve to reduce ourestimate of the time shift exhibited by humans.As we showed for the Super Bowl, live televised events(e.g. sports, awards shows) have the potential to be aforcing mechanism to synchronize our collective attentionthroughout the week, and especially on Sunday evenings.Indeed, many individuals take to Twitter as a secondscreen during such events to interact with other viewers.In addition, streaming services such as Netflix and HBOoften release new episodes of popular shows on Sundaynight to align with peak consumption opportunity. Thesecultural attractions exert a temporal organizing influenceon our leisure behavior, and the Spring Forward distur-bance translates this synchronization forward in time.It is worth noting that early March is a rather dulltime of year for popular professional sports in the UnitedStates. While the National Basketball Association andNational Hockey League are finishing up their regularseasons, the National Football League is in its off-seasonand Major League Baseball beginning pre-season exercis-es. Arguably the most engaging live-televised sportingcontests taking place in early March are the NCAA Col-lege Basketball Conference Championship games, withMarch Madness happening weeks after Spring Forward.In 2014, the Academy Awards were hosted by EllenDeGeneres on Sunday March 2. Her famous selfie tweetcontaining many famous actors was posted that evening,a message which held the record for most retweeted sta-tus update for several years [44]. The event happenedthe week before Spring Forward, and led to anomalousbehavior compared with all other Sundays we looked at.Finally, Twitter (and other social media companies)have access to much higher fidelity information regardinguser activity than we have analyzed here. We are not ableto analyze consumption activity on the site, e.g. whenindividual messages are interacted with via views, likes,or clicks. These forms of interaction with the Twitterecosystem are likely to occur chronologically followingthe final posting of a message in the evening, and priorto the initial posting of a message in the morning. As aresult, we expect our estimate of the sleep opportunitylost due to Spring Forward to be a lower bound.
Conclusion
Privacy preserving passive measurement of dai-ly behavior has tremendous potential to transformpopulation-scale human activity into public healthinsight. The present study demonstrates a proof-of-concept along the path to a far more ambitious goal: con-struction of an ‘Insomniometer’ capable of real-time esti-mation of large-scale sleep duration and quality. Which0
Super Bowl Sunday
FIG. 6 . Peak activity time (local) for Super Bowl Sunday, 5 weeks prior to Spring Forward, averaged overthe years 2011 to 2014.
Activity exhibits a clear resemblance to the U.S. timezone map, with a peak near 9 p.m. EasternTime just following the halftime performance. The data suggests a national collective synchronization in attention. Green BayPackers d. Pittsburgh Steelers (2011), New York Giants d. New England Patriots (2012), Baltimore Ravens d. San Francisco49ers (2013), and Seattle Seahawks d. Denver Broncos (2014). Performers included The Black Eyed Peas, Usher, and Slash(2011), Madonna, LMFAO, Cirque du Soleil, Nicki Minaj, M.I.A., and Cee Lo Green (2012), Beyoncé, Destiny’s Child (2013),and Bruno Mars, Red Hot Chili Peppers (2014). We note that the colormap here the same as the scale used for 3, with bluecolors included to reflect the earlier peaks seen in Mountain and Western time zones. cities in the U.S. slept well last night? Which states areincreasingly suffering from insomnia? Answers to ques-tions like these are not available today, but could leadto better public health surveillance in the near future.For example, communities exhibiting disrupted sleep ina collective pattern may be in the early stages of the out-break of the flu or some other virus. Current methodolo-gies for answering these questions are not scalable, butsocial media, mobile devices, and wearable fitness track- ers offer a new opportunity for improved monitoring ofpublic health.
Acknowledgements
KL, TA, PSD, and CMD thank MassMutual for con-tributing funding in support of this research. Theauthors thank Adam Fox, Marc Maier, Jane Adams,David Dewhurst, Lewis Mitchell, and Henry Mitchell forhelpful conversations. [1] C. C. Panel, N. F. Watson, M. S. Badr, G. Belenky, D. L.Bliwise, O. M. Buxton, D. Buysse, D. F. Dinges, J. Gang-wisch, M. A. Grandner, et al. Joint consensus statement of the american academy of sleep medicine and sleepresearch society on the recommended amount of sleepfor a healthy adult: Methodology and discussion.
Sleep , Short Sleep Duration AmongUS Adults , 5 2017.[3] E. S. Ford, T. J. Cunningham, and J. B. Croft. Trends inself-reported sleep duration among us adults from 1985to 2012.
Sleep , 38(5):829–832, 2015.[4] T. Althoff, E. Horvitz, R. W. White, and J. Zeitzer. Har-nessing the web for population-scale physiological sens-ing: A case study of sleep and performance.
WWW’17 Proceedings of the 26th International Conference onWorld Wide Web , pages 113–122, 2017.[5] G. Curcio, M. Ferrara, and L. De Gennaro. Sleeploss, learning capacity and academic performance.
SleepMedicine Reviews , 10(5):323–337, 2006.[6] M. Engle-Friedman. The effects of sleep loss on capacityand effort.
Sleep Science , 7(4):213–224, 2014.[7] M. R. Rosekind, K. B. Gregory, M. M. Mallis, S. L.Brandt, B. Seal, and D. Lerner. The cost of poorsleep: Workplace productivity loss and associated costs.
Journal of Occupational and Environmental Medicine ,52(1):91–98, 2010.[8] B. Dean, D. Aguilar, C. Shapiro, W. C. Orr, J. A. Isser-man, B. Calimlim, and G. A. Rippon. Impaired healthstatus, daily functioning, and work productivity in adultswith excessive sleepiness.
Journal of Occupational andEnvironmental Medicine , 52(2):144–149, 2010.[9] J. Owens, T. Dingus, F. Guo, Y. Fang, M. Perez,J. McClafferty, and B. Tefft. Prevalence of drowsy-driving crashes: Estimates from a large-scale naturalisticdriving study.
AAA Foundation for Traffic Safety , 2018.[10] C. Anderson, S. Ftouni, J. M. Ronda, S. M. Rajarat-nam, C. A. Czeisler, and S. W. Lockley. Self-reporteddrowsiness and safety outcomes while driving after anextended duration work shift in trainee physicians.
Sleep ,41(2):zsx195, 2018.[11] G. Medic, M. Wille, and M. E. Hemels. Short-and long-term health consequences of sleep disruption.
Nature andScience of Sleep , 9:151, 2017.[12] M. Walker.
Why we sleep: Unlocking the power of sleepand dreams . Simon and Schuster, 2017.[13] M. Nagai, S. Hoshide, and K. Kario. Sleep duration asa risk factor for cardiovascular disease-a review of therecent literature.
Current cardiology reviews , 6(1):54–61,2010.[14] V. Cheung, V. Yuen, G. Wong, and S. Choi. The effectof sleep deprivation and disruption on dna damage andhealth of doctors.
Anaesthesia , 74(4):434–440, 2019.[15] M. Fox. Shift work may cause cancer, world agency says.
Reuters , 2007.[16] N. P. Patel, M. A. Grandner, D. Xie, C. C. Branas, andN. Gooneratne. “sleep disparity” in the population: Poorsleep quality is strongly associated with poverty and eth-nicity.
BMC Public Health , 10(1):475, 2010.[17] V. K. Chattu, S. K. Chattu, D. W. Spence, M. D. Man-zar, D. Burman, and S. R. Pandi-Perumal. Do dispari-ties in sleep duration among racial and ethnic minoritiescontribute to differences in disease prevalence?
Journalof Racial and Ethnic Health Disparities , 6(6):1053–1061,2019.[18] M. E. Ruiter, J. DeCoster, L. Jacobs, and K. L. Lich-stein. Normal sleep in African-Americans and Caucasian-Americans: A meta-analysis.
Sleep Medicine , 12(3):209–214, 2011. [19] D. S. Curtis, T. E. Fuller-Rowell, M. El-Sheikh, M. R.Carnethon, and C. D. Ryff. Habitual sleep as a contribu-tor to racial differences in cardiometabolic risk.
Proceed-ings of the National Academy of Sciences , 114(33):8889–8894, 2017.[20] M. Hafner, M. Stepanek, J. Taylor, W. M. Troxel, andC. Van Stolk. Why sleep matters—the economic costs ofinsufficient sleep: A cross-country comparative analysis.
Rand Health Quarterly , 6(4), 2017.[21] M. T. Bianchi. Sleep devices: Wearables and nearables,informational and interventional, consumer and clinical.
Metabolism , 84:99–108, 2018.[22] N. J. Douglas, S. Thomas, and M. A. Jan. Clinical val-ue of polysomnography.
The Lancet , 339(8789):347–350,1992.[23] A. G. Harvey and N. K. Tang. (Mis) perception of sleepin insomnia: A puzzle and a resolution.
PsychologicalBulletin , 138(1):77, 2012.[24] D. S. Lauderdale, K. L. Knutson, L. L. Yan, K. Liu,and P. J. Rathouz. Self-reported and measured sleepduration: How similar are they?
Epidemiology , pages838–845, 2008.[25] M. Marino, Y. Li, M. N. Rueschman, J. W. Winkelman,J. Ellenbogen, J. M. Solet, H. Dulin, L. F. Berkman,and O. M. Buxton. Measuring sleep: Accuracy, sensi-tivity, and specificity of wrist actigraphy compared topolysomnography.
Sleep , 36(11):1747–1755, 2013.[26] S. Roomkham, D. Lovell, J. Cheung, and D. Perrin.Promises and challenges in the use of consumer-gradedevices for sleep monitoring.
IEEE Reviews in Biomedi-cal Engineering , 11:53–67, 2018.[27] E. Leypunskiy, E. Kıcıman, M. Shah, O. J. Walch,A. Rzhetsky, A. R. Dinner, and M. J. Rust. Geo-graphically resolved rhythms in Twitter use reveal socialpressures on daily activity patterns.
Current Biology ,28(23):3763–3775, 2018.[28] M. A. Christensen, L. Bettencourt, L. Kaye, S. T. Motu-ru, K. T. Nguyen, J. E. Olgin, M. J. Pletcher, and G. M.Marcus. Direct measurements of smartphone screen-time: Relationships with demographics and sleep.
PLOSONE , 11(11), 2016.[29] P. S. Dodds, K. D. Harris, I. M. Kloumann, C. A. Bliss,and C. M. Danforth. Temporal patterns of happiness andinformation in a global social network: Hedonometricsand Twitter.
PLOS ONE , 6(12), 2011.[30] S. E. Alajajian, J. R. Williams, A. J. Reagan, S. C. Alaja-jian, M. R. Frank, L. Mitchell, J. Lahne, C. M. Danforth,and P. S. Dodds. The Lexicocalorimeter: Gauging publichealth through caloric input and output on social media.
PLOS ONE , 12(2), 2017.[31] D. J. McIver, J. B. Hawkins, R. Chunara, A. K. Chat-terjee, A. Bhandari, T. P. Fitzgerald, S. H. Jain, andJ. S. Brownstein. Characterizing sleep issues using twit-ter.
Journal of Medical Internet Research , 17(6):e140,2015.[32] J. M. Martín-Olalla. the long term impact of daylightsaving time regulations in daily life at several circles oflatitude.
Scientific Reports , 9(1):1–13, 2019.[33] J. M. Martín-Olalla. Scandinavian bed and rise times inthe age of enlightenment and in the 21st century showsimilarity, helped by daylight saving time.
Journal ofSleep Research , Sep 2019.[34] A. Sandhu, M. Seth, and H. S. Gurm. Daylight Sav-ings Time and myocardial infarction.
Open Heart , Sleep Medicine , 27:20–24, 2016.[36] J. Varughese and R. P. Allen. Fatal accidents followingchanges in Daylight Savings Time: The American expe-rience.
Sleep Medicine , 2(1):31–36, 2001.[37] J. M. Martín-Olalla. Traffic accident increase attributedto daylight saving time doubled after energy policy act.
Current Biology , 30(7):R298–R300, Apr 2020.[38] M. J. Kamstra, L. A. Kramer, and M. D. Levi. Los-ing sleep at the market: The Daylight Saving anomaly.
American Economic Review , 90(4):1005–1011, 2000.[39] T. J. Gray, A. J. Reagan, P. S. Dodds, and C. M. Dan-forth. English verb regularization in books and tweets.
PLOS ONE , 13(12), 2018.[40] Twitter. Developer application program interface (API).https://developer.twitter.com/en/docs/tweets/sample-realtime/overview/decahose, 2020.[41] C. E. Rasmussen. Gaussian processes in machine learn-ing. In
Summer School on Machine Learning , pages 63–71. Springer, 2003.[42] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learn-ing in python.
Journal of Machine Learning Research ,12(Oct):2825–2830, 2011.[43] S. Wojcik and A. Hughes. Sizing up twitter users, 2019.[44] E. DeGeneres. https://twitter.com/theellenshow/status/440322224407314432, March 2, 2014. FIG. A1 . Peak activity time (local) for the Sunday of the four weeks prior to, the week of, and the four weeksfollowing Spring Forward, aggregated from 2011 to 2014.
We have used the same colormap as for Fig. 3 in the mainmanuscript. States shown in white had a peak time that was 9 pm or earlier. From 2011 to 2013, the Academy Awards tookplace two weeks prior to Spring Forward, while in 2014 they took place one week prior. A clear discontinuity is visible betweenthe “One Week Before” and “Week Of” maps. State count State count State log(TPC) State log(TPC)AK 747 CA 49190 AK 1.02e-03 DC 2.93e-03AL 8072 TX 45406 AL 1.67e-03 LA 2.35e-03AR 3560 FL 27339 AR 1.21e-03 DE 2.24e-03AZ 6567 NY 25833 AZ 1.00e-03 MD 1.87e-03CA 49190 OH 21061 CA 1.29e-03 NJ 1.83e-03CO 4310 PA 18790 CO 8.31e-04 OH 1.82e-03CT 5263 MI 17259 CT 1.47e-03 TX 1.81e-03DC 1854 IL 16620 DC 2.93e-03 RI 1.76e-03DE 2051 NJ 16216 DE 2.24e-03 MI 1.75e-03FL 27339 GA 15952 FL 1.42e-03 NV 1.74e-03GA 15952 NC 13600 GA 1.61e-03 AL 1.67e-03HI 1309 VA 11761 HI 9.40e-04 SC 1.65e-03IA 4233 MD 11030 IA 1.38e-03 GA 1.61e-03ID 934 LA 10822 ID 5.85e-04 MA 1.50e-03IL 16620 MA 9995 IL 1.29e-03 PA 1.47e-03IN 8138 TN 8173 IN 1.24e-03 CT 1.47e-03KS 4063 IN 8138 KS 1.41e-03 KY 1.45e-03KY 6373 AL 8072 KY 1.45e-03 WV 1.45e-03LA 10822 SC 7817 LA 2.35e-03 OK 1.44e-03MA 9995 WA 7469 MA 1.50e-03 VA 1.44e-03MD 11030 AZ 6567 MD 1.87e-03 FL 1.42e-03ME 965 KY 6373 ME 7.26e-04 KS 1.41e-03MI 17259 MO 6099 MI 1.75e-03 MS 1.40e-03MN 5258 WI 5705 MN 9.77e-04 NC 1.39e-03MO 6099 OK 5495 MO 1.01e-03 IA 1.38e-03MS 4182 CT 5263 MS 1.40e-03 NY 1.32e-03MT 369 MN 5258 MT 3.67e-04 CA 1.29e-03NC 13600 NV 4792 NC 1.39e-03 IL 1.29e-03ND 780 CO 4310 ND 1.11e-03 TN 1.25e-03NE 2262 IA 4233 NE 1.22e-03 IN 1.24e-03NH 1128 MS 4182 NH 8.54e-04 NE 1.22e-03NJ 16216 KS 4063 NJ 1.83e-03 AR 1.21e-03NM 1846 OR 3871 NM 8.85e-04 ND 1.11e-03NV 4792 AR 3560 NV 1.74e-03 WA 1.08e-03NY 25833 WV 2683 NY 1.32e-03 AK 1.02e-03OH 21061 UT 2495 OH 1.82e-03 MO 1.01e-03OK 5495 NE 2262 OK 1.44e-03 AZ 1.00e-03OR 3871 DE 2051 OR 9.93e-04 WI 9.96e-04PA 18790 DC 1854 PA 1.47e-03 OR 9.93e-04RI 1845 NM 1846 RI 1.76e-03 MN 9.77e-04SC 7817 RI 1845 SC 1.65e-03 HI 9.40e-04SD 540 HI 1309 SD 6.48e-04 NM 8.85e-04TN 8173 NH 1128 TN 1.25e-03 UT 8.74e-04TX 45406 ME 965 TX 1.81e-03 NH 8.54e-04UT 2495 ID 934 UT 8.74e-04 CO 8.31e-04VA 11761 ND 780 VA 1.44e-03 VT 7.59e-04VT 475 AK 747 VT 7.59e-04 ME 7.26e-04WA 7469 SD 540 WA 1.08e-03 SD 6.48e-04WI 5705 VT 475 WI 9.96e-04 ID 5.85e-04WV 2683 MT 369 WV 1.45e-03 WY 4.25e-04WY 245 WY 245 WY 4.25e-04 MT 3.67e-04
TABLE A1 . Tweet Counts.
Tweet count and tweets per capita ( log ) sorted alphabetically and in order of volume for thefour ASF Sundays observed in 2011-2014. State BSF State BSF State SF State SFAK 09:45 PA 10:15 AK 10:45 TX 11:00AL 09:30 FL 10:15 AL 10:15 AK 10:45AR 09:45 NY 10:15 AR 10:30 GA 10:45AZ 09:30 KY 10:15 AZ 09:45 OK 10:45CA 09:30 OH 10:15 CA 10:00 OH 10:45CO 09:00 IN 10:15 CO 10:00 ND 10:45CT 10:00 MI 10:15 CT 10:30 MI 10:45DC 10:00 GA 10:15 DC 10:00 KY 10:45DE 10:00 SC 10:15 DE 10:15 MS 10:45FL 10:15 NJ 10:15 FL 10:30 FL 10:30GA 10:15 VA 10:15 GA 10:45 LA 10:30HI 06:00 WV 10:15 HI 06:00 NM 10:30IA 09:30 DE 10:00 IA 10:15 NY 10:30ID 09:30 DC 10:00 ID 10:15 RI 10:30IL 09:30 CT 10:00 IL 10:15 SC 10:30IN 10:15 RI 10:00 IN 10:30 CT 10:30KS 09:45 NC 10:00 KS 10:30 MD 10:30KY 10:15 NH 10:00 KY 10:45 AR 10:30LA 09:30 MA 10:00 LA 10:30 KS 10:30MA 10:00 MD 10:00 MA 10:15 IN 10:30MD 10:00 ME 10:00 MD 10:30 VA 10:30ME 10:00 NM 09:45 ME 10:15 VT 10:30MI 10:15 AK 09:45 MI 10:45 WV 10:30MN 09:30 OK 09:45 MN 10:15 NJ 10:30MO 09:30 TN 09:45 MO 10:15 UT 10:15MS 09:45 VT 09:45 MS 10:45 DE 10:15MT 09:00 NE 09:45 MT 10:00 TN 10:15NC 10:00 MS 09:45 NC 10:15 SD 10:15ND 09:30 AR 09:45 ND 10:45 PA 10:15NE 09:45 KS 09:45 NE 10:00 WI 10:15NH 10:00 ND 09:30 NH 10:15 NH 10:15NJ 10:15 SD 09:30 NJ 10:30 MN 10:15NM 09:45 WI 09:30 NM 10:30 ME 10:15NV 09:30 WA 09:30 NV 10:00 IA 10:15NY 10:15 AZ 09:30 NY 10:30 NC 10:15OH 10:15 CA 09:30 OH 10:45 ID 10:15OK 09:45 UT 09:30 OK 10:45 IL 10:15OR 09:30 TX 09:30 OR 10:00 AL 10:15PA 10:15 IA 09:30 PA 10:15 MO 10:15RI 10:00 ID 09:30 RI 10:30 MA 10:15SC 10:15 OR 09:30 SC 10:30 WA 10:00SD 09:30 IL 09:30 SD 10:15 DC 10:00TN 09:45 LA 09:30 TN 10:15 NE 10:00TX 09:30 NV 09:30 TX 11:00 CO 10:00UT 09:30 MN 09:30 UT 10:15 MT 10:00VA 10:15 MO 09:30 VA 10:30 OR 10:00VT 09:45 AL 09:30 VT 10:30 CA 10:00WA 09:30 MT 09:00 WA 10:00 NV 10:00WI 09:30 CO 09:00 WI 10:15 AZ 09:45WV 10:15 WY 09:00 WV 10:30 WY 09:45WY 09:00 HI 06:00 WY 09:45 HI 06:00
TABLE A2 . Time of Peak Twitter Activity by State.
Time of peak Twitter activity Before Spring Forward (BSF) andthe week of Spring Forward (SF) for each state, listed alphabetically and by time of peak. State Peak State Peak State Twin State TwinAK 60 TX 90 AK 60 TX 135AL 45 ND 75 AL 60 MS 105AR 45 AK 60 AR 60 LA 90AZ 15 LA 60 AZ 30 ID 75CA 30 OK 60 CA 30 TN 75CO 60 MT 60 CO 75 CO 75CT 30 MS 60 CT 30 ND 75DC 0 CO 60 DC 45 MN 75DE 15 WY 45 DE 15 IL 75FL 15 MO 45 FL 45 WI 60GA 30 MN 45 GA 45 OK 60HI 0 UT 45 HI -30 NM 60IA 45 AR 45 IA 45 AL 60ID 45 VT 45 ID 75 AK 60IL 45 SD 45 IL 75 AR 60IN 15 KS 45 IN 30 IA 45KS 45 NM 45 KS 45 MO 45KY 30 IL 45 KY 45 VT 45LA 60 ID 45 LA 90 UT 45MA 15 IA 45 MA 45 SC 45MD 30 WI 45 MD 45 OH 45ME 15 AL 45 ME 15 NJ 45MI 30 TN 30 MI 45 NE 45MN 45 CA 30 MN 75 FL 45MO 45 NV 30 MO 45 DC 45MS 60 OH 30 MS 105 GA 45MT 60 MI 30 MT 30 KS 45NC 15 OR 30 NC 30 MI 45ND 75 MD 30 ND 75 KY 45NE 15 CT 30 NE 45 MD 45NH 15 KY 30 NH 30 MA 45NJ 15 WA 30 NJ 45 MT 30NM 45 GA 30 NM 60 WA 30NV 30 RI 30 NV 30 VA 30NY 15 VA 15 NY 30 AZ 30OH 30 SC 15 OH 45 CA 30OK 60 WV 15 OK 60 SD 30OR 30 NE 15 OR 30 RI 30PA 0 AZ 15 PA 30 PA 30RI 30 NY 15 RI 30 OR 30SC 15 NJ 15 SC 45 CT 30SD 45 NH 15 SD 30 NY 30TN 30 DE 15 TN 75 NV 30TX 90 NC 15 TX 135 IN 30UT 45 ME 15 UT 45 NH 30VA 15 MA 15 VA 30 NC 30VT 45 IN 15 VT 45 ME 15WA 30 FL 15 WA 30 DE 15WI 45 PA 0 WI 60 WV 15WV 15 HI 0 WV 15 WY 15WY 45 DC 0 WY 15 HI -30
TABLE A3 . Spring Forward Time Shift (minutes) by State.
The temporal shift in (1) peak activity and (2)twinflection sorted alphabetically and by magnitude. Times reported are differences between columns in the preceding table,and reported in minutes.
20 0 20 40 60 80 100 120 140
Peak Shift (mins) T w i n f l e c t i o n S h i f t ( m i n s ) DCHIPA NYWVFL CAOHTN ALMOILSDWY OKCOLAMSMT ND TX