An analysis of air mass effects on rail ridership in three US cities

Journal of Transport Geography 17 (2009) 198–207

Contents lists available at ScienceDirect

Journal of Transport Geography

journal homepage: www.elsevier .com/locate / j t rangeo

An analysis of air mass effects on rail ridership in three US cities

Adam J Kalkstein a,*, Michael Kuby a, Daniel Gerrity b, James J Clancy a

a School of Geographical Sciences, Arizona State University, Tempe, AZ 85287-0104, United Statesb Department of Civil and Environmental Engineering, Arizona State University, Tempe, AZ 85287-5306, United States

a r t i c l e i n f o a b s t r a c t

Keywords:Rail transitAir massUrban transportationWeatherClimateRidership

0966-6923/$ - see front matter � 2008 Elsevier Ltd. Adoi:10.1016/j.jtrangeo.2008.07.003

* Corresponding author. Tel.: +1 480 965 7533; faxE-mail address: [email protected] (A.J. Kalk

This paper examines whether daily weather affects ridership in urban transportation systems. Whenexamining human–weather relationships, it is often advantageous to examine air masses, which take intoaccount the entire parcel of air over a region. Spatial synoptic classification characterizes air massesbased upon numerous meteorological variables at a given location. Thus, rather than examining temper-ature or precipitation individually, here we compare daily ridership to synoptic air mass classificationsfor three urban rail systems: Chicago Transit Authority (CTA), Bay Area Rapid Transit (BART), and theHudson–Bergen light-rail line in northern New Jersey. Air masses are found to have a significant impacton daily rail ridership, with usage typically increasing on dry, comfortable days and decreasing on moist,cool ones, particularly on weekends. Although the comfort of a particular air mass changes throughoutthe year, seasonality is not a significant factor with respect to the air mass–ridership relationship. Theresults of this study can benefit rail system managers who must predict daily ridership or in the devel-opment of cost-benefit analyses for station improvements.

� 2008 Elsevier Ltd. All rights reserved.

1. Introduction

Cities in the United States are increasingly turning to rail transit,particularly light rail, to provide an attractive alternative to drivingand revitalize urban areas. Since 1980, two cities have built heavy-rail systems, and 17 have built light-rail systems, while one heavy-rail and 40 light-rail systems are currently proposed (AmericanPublic Transit Association, 2007). Controversy swirls around manyprojects, with some claiming that ridership forecasts are systemat-ically inflated (Pickrell, 1992; Flyvbjerg et al., 2005) while othersquestion these assertions (Demery, 2002). A host of factors are be-lieved to affect ridership, including accessibility, congestion, den-sity, size of central business district, car ownership, mixed-usetransit-oriented development, parking costs, fares, gas prices,employment levels, travel times, park-and-ride spaces, intermodalconnections, sporting and special events, and shopping opportuni-ties (Pushkarev and Zupan, 1982; Cervero, 1993; Parsons Brincker-hoff Quade Douglas and Inc, 1996; Filion, 2001; Kuby et al., 2004;Pucher, 2004). Some of these factors affect spatial variations in rid-ership at the neighborhood or metropolitan scale, while others af-fect temporal variations on a daily, weekly, annual, or long-termscale. It is imperative to understand all the factors that influencesystem ridership, but one factor in particular requiring increasedattention and research is weather.

ll rights reserved.

: +1 480 965 8313.stein).

Transit agencies have long recognized that weather conditionsaffect human comfort, which in turn affects transit ridership.Investments in bus shelters, sheltered pedestrian walkways inMinneapolis (Fielding, 1995), cooling towers and shade structuresin Phoenix (Levine, 1990), and other investments in weather pro-tection provide evidence of planners’ concerns about the effect ofweather on ridership. Guo et al. (2007) described a scenario inwhich a greater understanding of weather’s impact on transit rid-ership could assist in developing a cost-benefit analysis for transitstation investments. Despite the generally accepted relationshipbetween weather and ridership, studies attempting to qualify orquantify this relationship are limited and somewhat contradictory,as described below.

Traditionally, research relating weather and climate to urbantransportation has focused on safety and accidents (Changnon,1996; Andrey et al., 2003) or on maintenance issues (Arlinghausand Nystuen, 1985). Concerns about weather’s effect on safetyhave led some researchers to investigate weather’s effects on traf-fic volumes. Keay and Simmonds (2005) found a roughly 1–3%reduction in road traffic volume in Australia, depending on theamount of rainfall and the season. Knapp and Smithson (2000)found that wet or snowy weather in Iowa discourages driving, withwinter storms reducing traffic by 16–47%. Maze et al. (2006) esti-mated the effect of rain, snow, fog, cold, and wind on trafficvolumes, safety, and speed-flow relationships and noted thatinclement weather greatly increases the chances of accidents.

If weather causes a reduction in traffic volumes during adverseweather conditions, as indicated in the aforementioned studies, it

mailto:[email protected]

http://www.sciencedirect.com/science/journal/09666923

http://www.elsevier.com/locate/jtrangeo

A.J. Kalkstein et al. / Journal of Transport Geography 17 (2009) 198–207 199

is possible that people turn to alternative forms of transportationduring these times. A survey-based study in Geneva found thatover 50% of respondents indicated that weather was an importantfactor in determining what mode of transportation to use, while73% said it affects their departure time (de Palma and Rochat,1999). Khattak (1991) and Khattak et al. (1995) found that modeswitching – from automobile to public transit – occurred in Chi-cago and the Bay Area during adverse weather conditions. Aaheimand Hauge (2005) found small positive correlations between publictransportation and precipitation in Bergen, Norway, but only forcertain groups of people and areas of the city. High winds and hightemperatures also made people slightly more likely to prefer publictransportation over driving, biking, or walking, although the corre-lations for wind and temperature were even weaker and moreinconsistent than for precipitation. Finally, Khattak and de Palma(1997) found that transit ridership increases during periods ofbad weather in Brussels, Belgium.

Adverse weather conditions, however, do not always lead to in-creased use of public transit. In contrast to the preceding studies,Changnon (1996) found reduced ridership of public-transportationsystems on rainy days, especially during mid-day periods. Consis-tent with Changnon’s (1996) results, de Palma and Rochat (1999)found that adverse weather prompts some commuters who usuallytake transit to drive their cars instead. Guo et al. (2007) also foundthat bad weather, including decreased temperature, increased pre-cipitation, and windy conditions, caused significant decreases intransit ridership in Chicago. The degree of change was strongerfor bus than for rail, and on weekends than weekdays.

Researchers are also beginning to look at long-term climate ef-fects rather than short-term weather effects. For example, Ruthand Kirshen (2001) addressed the effects of climate change on ur-ban transportation. Kuby et al. (2004) found a strong relationshipbetween temperature extremes and light-rail ridership. Theydeveloped a multiple regression model to explain average weekdayridership at 268 stations in nine cities across the United States andfound average monthly heating-plus-cooling degree-days to behighly significant, accounting for differences of plus-or-minus sev-eral hundred riders per station per day. The implication is that cit-

Table 1Information on rail systems across the United States

Rail system Railtype

Annual ridership 2004(million unlinked trips)a

Directionaltrack milesa

Stationsa

Chicago (CTA) Heavy 64,328 206.3 144SF-Oakland

(BART)Heavy 62,373 209.0 43

New JerseyTransit(NJT)d

Light 7801 27.2 23

Other SystemsAtlanta

(MARTA)Heavy 69,089 96.1 38

New York City(MTA)

Heavy 339,819 493.8 468

Washington DC(Metro)

Heavy 58,205 206.6 83

Dallas (DART) Light 5153 87.7 34Salt Lake City

(UTA)Light 2969 37.3 23

Cleveland(GCRTA)

Light 1012 30.4 34

San DiegoTrolley

Light 6983 96.6 49

a Source: American Public Transportation Association: http://www.apta.com/researchtracked lines twice.

b Source: US Census Bureau: factfinder.census.gov (last accessed 11 June 2007).c Source: Texas Transportation Institute: mobility.tamu.edu/ums/congestion_data/tabld Ridership data for the NJT Hudson–Bergen line are from the NJT website for 2005. C

ies with extreme hot or cold climates experience lower ridershipthan peer cities with more moderate climatic conditions.

As the prior research shows, the relationship between transpor-tation and weather appears to be complex. Several studies cite de-creases in traffic volumes during bad weather, but there are alsoseveral studies citing increased traffic volumes during similarweather conditions. Similarly, some studies report increases inpublic transit ridership during adverse weather conditions,whereas others claim bad weather leads to decreased ridership.These inconsistencies may be related to the general focus on sin-gle-variable weather conditions, such as precipitation, tempera-ture, wind speed, etc. It is possible that a more encompassingweather variable may provide a better explanation for the weath-er–transportation relationship.

In light of the mixed prior results, this paper attempts to answerthe question: is there a relationship between weather, as repre-sented by air mass type, and rail-transit ridership? When examin-ing human–weather relationships, it is often advantageous to usean air mass approach. Air masses are large, relatively homogeneousparcels of air that are classified based upon numerous meteorolog-ical variables such as temperature, humidity, cloud cover, andwind speed. Air masses account for many meteorological variablessimultaneously and present a more accurate picture of the actualweather conditions that were present. Also, by examining airmasses, researchers can focus on a single category that takes intoaccount many weather variables (Kalkstein, 1991; Sheridan,2002). Guo et al. (2007) thought it was important to analyze the ef-fect of weather that was substantially cooler or warmer than thehistorical average (by >12 �F). When their cool and warm dummyvariables turned out to be generally insignificant or inconsistent,they stated that a ‘‘better definition which combines both temper-ature changes and human perceptions is necessary to explore thepossible effects of extreme temperature on transit ridership”(p. 17). Air masses may satisfy this need because they simulta-neously encompass many variables that affect public perceptionof weather conditions.

Air masses have been shown to have a large influence on humanhealth, and certain air masses have been correlated with sharp

Households with novehicles, 2005 (%)b

Metropolitan transitcommuting share, 2005 (%)b

Annual delay pertraveler, 2003 (h)c

10.6 10.8 5811.3 13.7 72

29.9 37.7 49

5.3 3.3 67

28.4 29.8 49

9.5 13.2 69

4.6 1.5 604.9 3.6 31

9.4 4.1 10

5.5 2.9 52

/stats/rail/index.cfm (last accessed 11 June 2007). Directional miles counts double-

es/national/table_1.pdf (last accessed 11 June 2007).ensus data are for Hudson County, NJ. Delay data are for New York metro area.

http://www.apta.com/research/stats/rail/index.cfm

200 A.J. Kalkstein et al. / Journal of Transport Geography 17 (2009) 198–207

increases in human mortality (Smoyer et al., 2000). Surprisingly,while there is a clear relationship between air mass type and hu-man health, we know of no studies that have investigated theirinfluence on human behavior. This study uses an air mass approachbecause it provides a more complete picture of weather conditionsthan any single weather variable and may therefore have a signif-icant effect on rail ridership.

Table 2Spatial synoptic classification system

Air mass type Description

Dry moderate (DM) Mild and dry, although slightly more humid than either DP orDT

Dry polar (DP) The coldest air mass at a location for the time of year,generally advected from Canada

Dry tropical (DT) The hottest and driest air mass at a given location for thetime of year; usually desert air advected from Mexico or thesouthwestern United States; can also be produced by rapidlysinking air such as the Chinook or Santa Ana winds

Moist moderate(MM)

Similar to MP but warmer; often associated withprecipitation

Moist polar (MP) Cloudy and cool; often located north of a warm front or neara cool ocean body

Moist tropical (MT) Warm and humid; usually in the warm sector of a frontalsystem

Transition (TR) When two or more air masses are present at a location in agiven day; often associated with frontal passages

Moist tropical plus(MT+)

A particularly warm, humid subset of MT

Table 3Average afternoon temperature (in �C) for each air mass in both summer and winter

Chicago San Francisco New York

Winter Summer Winter Summer Winter Summer

DM 6 28 13 21 7 28DP �7 23 9 18 �1 24DT Insufficient

data34 18 30 16 34

MM 4 24 13 19 6 23MP �1 18 9 17 0 17MT 13 30 16 28 13 29TR �3 25 13 24 4 27MT+ 17 32 17 Insufficient

data16 33

2. Study area and data

Daily urban rail ridership data were obtained for Chicago, IL(CTA – Chicago Transit Authority), the San Francisco Bay Area, CA(BART – Bay Area Rapid Transit), and Northern New Jersey (NJT –New Jersey Transit’s Hudson–Bergen light rail). These data consistof the daily sum of all passengers in the entire system for each city.The Chicago data span from April 1, 1998 through May 31, 2004;the BART data span from January 1, 1998 through February 28,2005; and NJT data span from April 16, 2000 through December31, 2003. We contacted other US transit agencies with rail transitsystems, but these three were the only ones that provided a dailytime series of system ridership for a multi-year period. It was for-tuitous that the three cities represent distinctly different climatetypes in the West Coast, Midwest, and Northeast United States.The three rail systems differ on several dimensions, and Table 1illustrates how these systems, and their corresponding cities, com-pare to others across the United States. The New York, Chicago, andSan Francisco Bay areas are ranked 1st, 3rd, and 12th in metropol-itan area population, and transit captures a higher share for com-muting than in most US cities. The Chicago and BART systemsare quite large by US standards, while the NJT Hudson–Bergen lineis small even for a new light-rail system.

Chicago built its first street railway in 1892, and the city grewradially around its various routes. Today, the CTA operates almost1,200 heavy-rail cars over eight separate lines and 103 miles (206directional miles) of track that cover much of the City of Chicago.The system takes its informal name – the ‘‘EL” – from the elevatedsection of track circling the downtown ‘‘Loop.” In total, 144 sta-tions serve downtown Chicago, Midway and O’Hare Airports, the-atres and museums, universities and sports stadiums, inner-cityand outlying neighborhoods, and suburbs. Fares are based on a flatrate (Chicago Transit Authority, 2006). The CTA averages about500,000 trips on weekdays, 250,000 trips on Saturdays, and170,000 trips on Sundays. Frequency of service varies between 7and 10 min on weekdays and 10–15 min on weekends.

BART, which opened in 1972, is a heavy-rail system with trainstraveling at up to 80 miles per hour. Its five different lines stretchover 209 directional miles, tunneling under San Francisco Bay andthrough topographic barriers and connecting numerous suburbsand urban neighborhoods with downtown San Francisco, down-town Oakland, Berkeley, the San Francisco and Oakland airports(the latter via shuttle), and the Oakland Coliseum. The system aver-ages over 300,000 trips on weekdays, 135,000 on Saturdays, and93,000 on Sundays. Approximately 730 trains operate on the week-days with service intervals of about 15 min. The number of trainsdecreases to approximately 480 and 400 on Saturdays and Sun-days, respectively, with a service frequency of approximately20 min. The 43 BART stations are roughly evenly divided amongsubway, surface, and elevated locations, and offer 46,000 park-and-ride spaces. Fares are distance-based (San Francisco Bay AreaRapid Transit District, 2006). The system was extended to theSan Francisco International Airport and three other stations onJune 22, 2003, during our study period.

The Hudson–Bergen light-rail line is one of three rail systemsoperated by NJT. Opened in 2000 and expanded four times since,the system consists of 23 stations on three surface lines totaling

27.2 directional miles. The predominantly north–south systemconnects residential areas in North Bergen, Union City, Weehaw-ken, Hoboken, Jersey City, and Bayonne with the business andshopping opportunities at Jersey City’s Exchange Place and New-port Center and Hoboken Terminal. These centers, located directlyacross the Hudson River from Lower Manhattan, provide easy con-nections to New York City via PATH and NY Waterway. The Hud-son–Bergen line serves about 21,000 trips each weekday,dropping to 4100 on Saturdays and 3100 on Sundays, and chargesa flat rate (New Jersey Transit, 2006). Service frequencies vary fromabout 10–20 min on weekdays to 15–20 min on weekends.

The ridership data were compared to a daily air mass calendarbased on spatial synoptic classification (SSC) (Sheridan, 2008).The SSC places each day at each location into one of the eight airmass types shown in Table 2 based upon daily meteorological con-ditions (Kalkstein et al., 1996; Sheridan 2002). As a result, thisstudy was conducted on a daily rather than hourly time scale be-cause the SSC examines meteorological variables taken throughoutthe entire day. It is worth noting that days with dramatic changesin meteorological conditions, in which one type of air mass re-places another, are classified into a separate ‘‘transitional”category.

The characteristics of each air mass can differ substantiallyacross locations and seasons (Table 3). For example, a moist trop-ical (MT) air mass in Northern New Jersey would differ from MTair in the Bay Area. Although the meteorological conditions withineach air mass vary among locations, the relative impacts are simi-lar; for example, MT air in all locations would consist of tempera-tures and dew points above average for that location and time ofyear. With respect to human perception, relative conditions are of-


ten more important than actual weather variables. For example, anexceptionally hot summer day in Chicago might have an identicaltemperature to an unseasonably cool and comfortable day in Phoe-nix, Arizona. This further highlights the advantages of using an airmass approach when comparing human responses between cities,since the relative conditions, as opposed to actual conditions, aremost influential. Similarly, dry polar (DP) air in Chicago duringthe winter would be far colder than DP air during the spring orsummer. We accounted for seasonal differences by examiningthe air mass–ridership relationship separately for each season, inaddition to year-round analyses, to ensure that ridership was com-pared only among days with similar meteorological conditions. Weconducted these seasonal analyses because warm days in the sum-mer, which would likely be uncomfortable, might elicit differenthuman responses than warm days in the winter.

3. Methods

The dependent variable for this analysis is the percentage dif-ference between the actual daily ridership, rt, and an average rider-ship variable. We will refer to this dependent variable as the dailyridership residual. Our methodology is designed to control for fluc-tuations in ridership associated with the day of the week, the timeof the year, and other non-weather-related trends over time.

Fig. 1, using BART data as an example, shows the complex tem-poral relationships that the standardization procedure needs tohandle. Three distinct horizontal linear groupings correspond, fromtop to bottom, to weekdays, Saturdays, and Sundays. Within each

Fig. 1. Raw, smoothed, and residual ridership time series for BART, 1998–2005. The fouSaturdays, Sundays, and the ridership residual-shown here in absolute rather than perc

of these groups, a general upward trend is visible through theend of 2000, followed by a decline until 2003 most likely associ-ated with the dot-com economic bust, and then a modest recovery.Within most years, ridership dips during winter and peaks duringlate summer and fall, but not in any purely symmetrical cycle.

To create a valid ‘‘base measure” to which to compare the dailyridership, we smoothed the ridership data by applying a 9-termmoving average. The nine terms in each moving average are theridership totals for the day in question (t) and the same day ofthe week in the prior four and following four weeks. For example,the moving average for a given Monday is calculated from the aver-age of the Monday in question, the four Mondays before, and thefour Mondays after. The 9-term moving average is defined as

�rMA�4t ¼

P4s¼�4rtþ7s

9ð1Þ

The index t is continuous from 1 to T. The index s representsweeks, so as s goes from �4 to 4, ðt þ 7sÞ goes from 28 days beforeto 28 days after, by sevens. We chose nine terms in order to com-pare a day’s ridership only to other days during a similar time ofyear, which we defined as the four weeks on either side of theday in question. On the one hand, averaging over as long a periodas possible creates a more representative average against which tocompare the ridership on that day of week. On the other hand,averaging over too long a period could lead to including days thatare substantially different than the day in question for non-weath-er reasons, such as seasonal changes in ridership, student ridership,and so on. Comparing a day’s ridership to the average for the same

r distinct horizontal linear groupings correspond, from top to bottom, to weekdays,entage terms for visibility-which centers on zero. Day 1 is January 1, 1998.


day of the week within one month on either side represents a com-promise between these two concerns. We included the ‘‘test day”in the comparison average because it was, in fact, part of the timeperiod in question. An additional analysis was conducted using 7and 11-term moving averages to evaluate the sensitivity of the re-sults with respect to the number of terms used.

The 9-term moving average is shown in Fig. 1 with black dots.Visually, it appears to track the uneven ups and downs of theraw data. In addition, the moving average for Saturdays and Sun-days closely tracks the raw daily data for those weekend days. Insome years, the moving averages for each different weekday arevisibly distinct from the other weekdays, as in 2002 (Mondaysand Fridays are lowest).

The final step is to calculate the difference between the actualridership and moving average of ridership for day t, as a percentageof the moving average

et ¼rt � �rMA�4

t

�rMA�4t

ð2Þ

This daily ridership residual, et, is standardized in percentageterms in order to compare data across all days of the week. Other-wise, because absolute ridership is much lower on weekends, theabsolute residuals would also be much lower, which would makeit appear (falsely, as it turns out) that weather has less effect onweekends than weekdays.

The daily ridership residual is the dependent variable in the sta-tistical analysis relating ridership to air mass types. Fig. 1 showsthe ridership residuals, in absolute rather than percentage terms,by the line fluctuating around zero. The absence of any pattern inthe absolute residual indicates that the technique for smoothingnon-weather effects, such as business cycle and annual patterns,was successful. As mentioned earlier, the standardized residualfurther smoothes the data for fluctuations due to day of the week.

Given that four weeks on either side of t are needed to calculateet, we begin our analysis with t = 29 and end it with t = T � 29. Incases where new stations or lines were opened – thus increasingridership ‘‘overnight” – the entire four weeks prior to and followingthe opening were removed because the moving average wouldotherwise be influenced by days on either side of this divide.

We also removed days when ridership was impacted by majorevents unrelated to weather. These included official holidays, festi-vals, and special sporting events. With the exception of specialevents that were particular to just one city (such as the ChineseNew Year festival or the Bay to Breakers race in the San FranciscoBay Area), the same days were deleted for all three cities. In addi-tion to official holidays, surrounding days were removed if theyshowed a consistent outlier pattern, such as July 3–5, Wednes-day–Friday of Thanksgiving week, and December 23–January 2inclusive of every year. September 11, 2001 was also removed, aswell as September 12, 2001 in the case of Hudson–Bergen. Theseoutlier days were removed from both the set of observations andthe computation of the moving averages, which would then beaveraged over fewer than nine terms.

Using these standardized daily ridership residuals, we were ableto examine what impact air mass type has on daily ridership ateach location. For each city, we first compared daily rail ridershipto air mass type across the entire period of record. We then sepa-rated the dataset into weekdays and weekends to see if air massesaffect ridership differently depending on the day of the week. Fi-nally, since the character of each air mass varies considerablythroughout the year, we examined how seasonality might impactthe relationship between air mass type and daily rail ridership.The winter season is defined as December through February, thespring extends from March through May, June through Augustmarks the summer, and September through November defines

the fall season. Examining air masses within a particular seasonfor each city ensures that ridership will only be compared on daysin which similar meteorological conditions were present.

It is important to note that not all air masses occur with thesame frequency. For example, dry polar air rarely makes it intothe Bay Area due to its proximity to water. As a result, we requireda minimum sample size of 10 days for all of our bar graphs show-ing average residuals. The intent of this restriction was to preventfalse impressions of the size of a ridership response. Thus, we ex-cluded winter bars for DT, MT, and MT+ in both Hudson–Bergenand Chicago. MT+ also had fewer than 10 occurrences in the fallin both cities and the spring along Hudson–Bergen, as did DT forthe summer in Chicago. Finally, nearly all bars for both DP andMT+ for the Bay Area do not meet the 10-day requirement. Thesesmall samples, however, were included in the statistical analyses,which take sample size into account.

Because air mass data are categorical, we used one-way analysisof variance (ANOVA) to compare the mean ridership residualsamong the different air mass types. ANOVA is the statistical tech-nique used for analyzing the relationship between a numericaldependent variable and a categorical independent variable withthree or more categories. The null hypothesis is that the mean rid-ership residuals of the air mass types are equal, and air mass typehas no effect on ridership. For each city, we performed seven ANO-VAs: one for each season; one each for weekdays and weekends;and one for all days combined. Given that the dependent variableis standardized as a percentage deviation from average for eachcity, we were able to pool the data for all three cities and conductthe same set of seven ANOVAs for the combined dataset. Finally,we ran a different set of 24 ANOVAs for each city and air mass type(3 � 8) with daily ridership residual as the dependent variable (DV)and season as the categorical independent variable (IV). Theseanalyses tested whether there was a significantly different rider-ship response to a given air mass (e.g., dry polar) in a given cityacross the different seasons (e.g., summer vs. winter).

We recognize that many factors not related to weather also af-fect ridership. Although non-weather effects are not modeledexplicitly in the one-way ANOVAs, we factor them out of the rider-ship data by manually removing holidays and special events andby computing the ridership residuals using the same-day-of-week,9-term moving average. By comparing a day’s ridership to the rid-ership on the same day of the week within four weeks of the day inquestion, we compare it only to similar days at similar times ofyear. This is a far more practical approach than collecting data onevery school and university class schedule, concert or performance,sporting event, and religious holiday, or trying to incorporate dataon migration, retail, jobs, and business activity in areas served bythe rail systems. Given that some unaccounted-for, non-weather-related noise will remain in the signal even after computing theridership residuals, the relationship between rail ridership andair mass type will have to be that much stronger to be statisticallysignificant.

4. Results

Air mass type has a clear impact on the daily rail ridershipresidual in all three locations, and these relationships were almostalways similar for each city, despite their differing climates (Fig. 2).For example, on the Hudson–Bergen line, average daily ridershipwas 2.3% higher on days with a dry moderate air mass in place.Similarly, Chicago and the Bay Area experienced ridership in-creases of 1.4% and 0.7%, respectively, on days with dry moderateair. Conversely, moist polar air was associated with decreases indaily ridership in all three locations, with the Hudson–Bergen linedecreasing by more than 3.7% on these days. Ridership decreased

Fig. 2. Percent change in daily ridership vs. air mass type for all three cities.

Table 6ANOVA results for NJT: daily ridership residual (DV) vs. air mass type (IV)

Sub-sample Total d.f. F-statistic Significance

All 1338 7.005 .000Weekdays 944 3.639 .000Weekends 393 4.895 .000Winter 285 3.930 .000Spring 380 2.036 .041Summer 422 2.875 .004Fall 248 1.569 .135

Table 7ANOVA results for all rail systems combined: daily ridership residual (DV) vs. air masstype (IV)




in all three cities on dry polar, moist polar, moist moderate, andtransitional days. Furthermore, ridership increased in all three cit-ies on dry moderate, dry tropical, and moist tropical days. In fact, ofthe eight different air mass types, seven had similar responsesacross all three cities, with the only exception being moist tropicalplus, which was associated with elevated ridership on Hudson–Bergen and near-normal ridership on CTA and BART.

Analysis of variance confirms that the differences among themeans of the ridership residuals for the eight air mass types are al-most always highly significant, and most relationships in all threecities are significant at the .001 level (Tables 4–6). The only non-significant results were found in the Bay Area during the summerand fall, likely due to little weather variability during these sea-sons, and in the fall along the Hudson–Bergen line, largely due toa reduced sample size. When data are combined for all three cities,air mass–ridership relationships are also significant at the .001 le-vel (Table 7). Air masses can feel more or less comfortable in differ-ent cities and in different seasons. However, when all data arecombined, the difference in means of the ridership residualsamong the eight air mass types is highly significant, illustratingthe relative nature of air mass–ridership responses.

Finally, sensitivity analysis tested whether these results dependon the use of a moving average with nine terms for calculating theridership residuals for each day. We repeated the ANOVA analysis

Table 4ANOVA results for CTA: daily ridership residual (DV) vs. air mass type (IV)



Note: SPSS rounds the significance levels to 3 decimal places.

Table 5ANOVA results for BART: daily ridership residual (DV) vs. air mass type (IV)



based on 7-term and 11-term moving averages, in addition to the9-term methodology used in this paper. Table 8 illustrates that thedifferences among the means of the ridership residuals for the dif-ferent air mass types remain highly significant regardless ofwhether the moving average term averages ridership for the sameday of the week within three, four, or five weeks of each day.

4.1. Day of week effects

The impacts of air mass type on daily ridership varied substan-tially between weekdays and weekends, and for all three locations,variations in ridership were much larger on weekends as a result ofan increased number of discretionary travelers (Fig. 3). For exam-ple, weekdays in Northern New Jersey with moist polar air hadan average decrease in ridership of 1.0%. On weekends, however,the same air mass resulted in an average decrease of 9.9%(Fig. 3c). Similarly, dry tropical air in Northern New Jersey resultedin daily ridership increases of 7.0% on weekends, while the sameair mass had little impact on weekdays. A similar day-of-weektrend is also observed in both Chicago and the Bay Area, but thevariations in ridership for these two cities were smaller than thosein Hudson–Bergen. For example, in Chicago, ridership during moist

Table 8Sensitivity analysis for 7, 9, and 11-term running means

Sub-sample

F-statistica

Degrees offreedomb

7-Term movingaverage



CTA 2040 9.459 10.733 11.113BART 2270 4.585 5.326 5.534NJT 1271 6.550 6.336 6.881

a All results significant at p = .0005 or better.b Each sensitivity analysis was performed over the same set of days in order to

keep the degrees of freedom constant and thus make a fair comparison among theridership residuals based on the 7-term, 9-term, and 11-term moving averages. Thedays included are determined by the buffer periods needed to calculate the 11-termmoving average, for which we removed the first and last five weeks of the ridershiptime series for each city, as well as five weeks around any of the days on which anew station opened in a city (once for BART and four times for NJT). Because hol-idays and special events sometimes fall on one of the removed weeks, however, thedegrees of freedom for CTA/BART/NJT in Table 7 are not exactly 14/28/70 daysfewer than in the equivalent rows of Tables 3–5.

Fig. 3. Daily ridership residual vs. air mass type for weekends and weekdays for each rail system.

(a). CTA (weekend DT days were removed from the graph because there were fewerthan 10 days).

(b). BART (weekend DP and MT+ days were removed from the graph because therewere fewer than 10 days in those categories).

(c). NJT Hudson–Bergen line.


tropical days increased by 2.7% on weekends, but showed littlechange on weekdays (Fig. 3a). Similarly, transitional days resultedin a 3.6% decrease on weekends but only a 0.7% decrease on week-days. In the Bay Area, ridership increased or decreased by a largerpercentage on weekends than on weekdays for all air mass typeswith at least ten observations (Fig. 3b).

The ANOVA results for weekdays are significant for all three cit-ies and for the combined dataset at .001 or better. The weekend re-sults, perhaps due to smaller sample sizes, are significant at .002 orbetter in all four data sets.

4.2. Seasonal effects

Considering that the meteorological characteristics of each airmass change throughout the year (i.e., warmer in summer, colderin winter), one would expect that seasonality would have an im-pact on the relationships between air mass type and daily rail rid-ership in certain cases (Fig. 4). For example, the dry polar air massin Chicago is associated with above-average daily ridership in thesummer, when that air mass would be cool and dry (Fig. 4a). Thesame air mass, however, decreases ridership in the winter, whenthe dry polar air mass in Chicago is extremely cold. To test thesedifferences among the seasons, we ran ANOVAs for each air masstype and rail system using the season as the categorical indepen-dent variable and the ridership residuals as the dependent variable(Tables 9–11). While the differences between seasons often lookedsubstantial in Fig. 4, only three of the 24 ANOVAs were statisticallysignificant. In addition to the case of dry polar air masses in Chi-

cago given above, moist polar days in Chicago and moist tropicaldays in the Bay Area also varied significantly between seasons.For 10 of the air mass/city combinations, the mean ridership resid-uals were directionally similar (+ or �) in every season (Fig. 4), andthe ANOVAs were not significant at the .05 level. In these 10 cases,air masses tend to have a similar effect on ridership in all seasons.In the remaining 14 cases, a given air mass might have a positiveeffect on ridership in some seasons and a negative effect in others,but because of the sample sizes and/or the magnitude of the effect,the differences were generally not significant at the .05 level.

We also ran ANOVAs on subsamples consisting of each seasonin each city, once again using air mass type as the categorical inde-pendent variable. The purpose of these analyses is to determine ifthere are any particular seasons in which air mass type does notsignificantly affect ridership. Despite the smaller sample sizeswithin each season, most of the differences in ridership betweenair mass types are statistically significant at the .05 level. For Chi-cago, when each season is examined separately, the differences inridership among air mass types are significant at the .004 level orbetter. For Hudson–Bergen, only the fall is not significant at the .05level, while for BART all seasons are significant at the .10 level, andwinter and spring are significant at the .05 level.

5. Discussion

Our results show that the impacts of air mass type on daily railridership are greater on weekends, when many more riders makediscretionary trips for recreational purposes and for one-time

Table 10ANOVA results for BART: daily ridership residual (DV) vs. season (IV)


Dry moderate 704 2.552 .055Dry polar 16 0.547 .659Dry tropical 70 2.038 .117Moist moderate 947 1.884 .131Moist polar 201 .723 .539Moist tropical 147 2.741 .046Transition 166 .695 .556Moist tropical + 10 .000 .986

Table 11ANOVA results for NJT: daily ridership residual (DV) vs. season (IV)


Dry moderate 344 .253 .859Dry polar 185 1.729 .163Dry tropical 81 1.135 .340Moist moderate 265 .524 .666Moist polar 111 .630 .597Moist tropical 161 .840 474Transition 127 .375 .771Moist tropical + 34 2.070 .125

Fig. 4. Daily ridership residual vs. air mass type for each season for (a) Chicago, (b) the Bay Area, and (c) Hudson–Bergen.

(a). DT, MT, and MT+ were removed for the winter, DT was removed for thesummer, and MT+ was removed for the fall due to a sample size less than 10 days.

(b). DP and MT+ were removed for all seasons, DT was removed for the winter, andMT was removed for the summer due to a sample size less than 10 days.

(c). DT and MT were removed for the winter, and MT+ was removed for the winter,spring, and fall due to a sample size less than 10 days.

Table 9ANOVA results for CTA: daily ridership residual (DV) vs. season (IV)


Dry moderate 506 1.052 .369Dry polar 351 6.771 .000Dry tropical 58 .990 .404Moist moderate 274 2.274 .080Moist polar 329 3.592 .014Moist tropical 197 1.897 .131Transition 196 .934 .425Moist tropical+ 80 .683 .565


events. Considering that bad weather can cancel outdoor plans, it isnot surprising that variations in ridership on weekends vary con-siderably with air mass type. As a result, uncomfortable air masseslead to a sharp decline in weekend ridership, as in Hudson–Bergen,

where ridership decreased by nearly 10% on weekends with moistpolar air in place.

On weekdays, when the majority of riders using the rail systemare commuting to and from work, air masses have a smaller effect.There are two possibilities why this might be occurring. First,weekday riders are mostly commuters traveling to work and, toa lesser extent, students traveling to school. Most workers and stu-dents cannot stay home in bad weather, and they may have littleflexibility to change modes. Many will have monthly transit passes,which may predispose them to take the train under all circum-stances. Thus, the first possibility is that comfortable or uncomfort-able air masses have little impact on a commuter’s or student’sdecision whether or not to take the rail system, and thus daily rid-ership on weekdays fluctuates only slightly with the weather. Sec-ond, it is possible that there are several effects that are largelycanceling each other out. Some automobile commuters may takepublic transportation on days with inclement weather to avoid


congested roadways. On the other hand, some regular transit com-muters, especially those with long walks to the nearest station,may choose to go by car instead. Any net increase in commuterand student ridership on inclement or uncomfortable weekdaysmight also be offset by a decrease in riders who are traveling forrecreational purposes.

While the same air mass type can feel different in different sea-sons, the variations in ridership among seasons for the same airmass in the same city were not significant, with a few notableexceptions. Although these results were surprising, they are con-sistent with previous research, which finds little seasonality in rid-ership for temperature, precipitation, and wind speed (Guo et al.2007). Thus, it is clear that ridership is dependent upon relativeconditions, suggesting that a significant number of individualschoose their mode of transportation based upon weather’s depar-ture from average in a given time of year. On the other hand, whenwe looked separately at each season in each city, the ridership dif-ferences among air mass types were usually significant, again witha few notable exceptions. The conclusion that transit agenciescould draw from the seasonal analysis is that the effect of air masstypes on ridership is fairly similar across all four seasons, and airmass type significantly affects ridership levels during most timesof the year.

Daily air mass conditions consistently have a more powerful ef-fect on ridership on NJT’s Hudson–Bergen line than on the CTA orBART systems. We speculate that this has more to do with the nat-ure of the rail systems than the climates or cities in question. Ifanything, people in the New York area, which has the highest tran-sit mode share in the country, and the lowest car ownership of thethree regions (Table 1), should be more predisposed to taking tran-sit in all conditions than people elsewhere in the country. Also,people in Northern New Jersey and Chicago should be more accus-tomed to changeable and extreme weather conditions than BayArea residents. Therefore, it seems more likely that Hudson–Ber-gen’s stronger response stems from the rail system itself.

The Hudson–Bergen line is a light-rail system, while CTA andBART operate heavy rail. Thus, the NJT stations are nearly all out-doors and more exposed to the weather. Similarly, the Hudson–Bergen line is the smallest of the three, and as a result, any modestvariation in raw ridership totals would correspond to relativelylarge percentage changes which were used in this analysis. Fur-thermore, the Hudson–Bergen system is new, and therefore rider-ship may not yet be as habitual as in Chicago and the Bay Area.Finally, the employment centers served directly by the Hudson–Bergen line are not nearly as large as those served by CTA andBART. To get to New York City, NJT riders are forced to transferto PATH and NY Waterway, and it could be that the transferringprocess is less appealing in uncomfortable weather, or that moreof the trips made to New York City are discretionary recreationalor shopping trips. More research needs to be conducted to under-stand the differential impact of weather on different rail systems –and on rail vs. bus – across the country and around the world.

6. Conclusions

This study adds to the scant literature relating transit ridershipto weather and climate by successfully linking air mass type withdaily rail ridership for three rail transit systems across the UnitesStates. Air mass data have the advantage of incorporating a num-ber of meteorological variables – such as temperature, humidity,cloud cover, and wind speed – into a single measure that describesthe large block of air over a region. Our findings generally supportthe conclusions of Changnon (1996), de Palma and Rochat (1999),and Guo et al. (2007) that precipitation tends to depress transit rid-ership. Going beyond measurable precipitation, our findings sug-

gest that uncomfortable air masses such as those associated withthe coldest wintertime temperatures, the hottest summertimetemperatures, as well as elevated likelihood of precipitation alllead to a decrease in rail ridership. Conversely, ridership increaseson days with air masses associated with mild winter temperaturesand cool summer temperatures. As mentioned in our introduction,these are the types of effects that Guo et al. (2007) were suggestinga different weather variable might be able to detect. Similar to Guoet al. (2007), changes in ridership are more pronounced on week-ends rather than weekdays, presumably due to the more discre-tionary nature of weekend trips. Additionally, it appears as ifseasonality only plays a small role in rail ridership, with specificair mass types eliciting similar responses throughout the year de-spite fluctuating meteorological conditions.

This study also extends the small climate literature on the effectof air masses. While previous studies have shown that air massescan have a large impact on human health, this is the first studyto our knowledge that relates air mass type with actual humanbehavioral responses.

Taken together with previous research, this study suggests thattransit planners take weather and climate into greater consider-ation as well as the more traditional criteria of size, density, econ-omy, demography, spatial structure, time of day, congestion, andspecial events. It may not be practical, however, for rail-systemoperators to use our findings to modify published transit sched-ules. At BART, for instance, any modifications to the schedule oftrains and number of cars per train is planned weekly in order toadapt to special events, sporting events, and scheduled mainte-nance [personal communication]. The weekly operations plan isused to schedule labor, equipment, and electricity purchases (BARTis one of the largest electricity consumers in the Bay Area).Although the impact of an air mass can increase or decrease dailyridership by 1–10%, which can equate to tens of thousands of addi-tional or fewer riders, the lead time required for setting schedulesmay be too long and the weather forecasts too unreliable for theweather–ridership relationships found here to be incorporated intoweekly operations planning.

There are several possible strategies for incorporating thesefindings in a more practical manner. For example, ridership fore-casts for new starts should take weather and climate into account.Based on the data provided in this and other studies, it may also bebeneficial for transit agencies to adjust service levels, possiblythrough increases or decreases in the number of cars, for weekendweather effects in particular. As proposed in Guo et al. (2007), thisstudy may provide important data for cost-benefit analyses ofweather-related improvements to outdoor stations.

Methodologically, this study introduced a moving-averagemethod for filtering out the ridership fluctuations due to day ofweek and time of year, leaving a signal that appears to be relatedto the daily weather. By computing a moving average for onlythe same day of the week, over four weeks on either side of theday in question, we were able to compare the ridership on theday in question with an average of similar days of the week andtimes of the year. This modified moving-average method – whichproduced statistically significant results using seven, nine, or ele-ven terms – may prove useful to the study of other daily transpor-tation phenomena.

Although this study offers some compelling evidence that airmasses influence rail usage, it does not explain why riders respondas they do. The aggregate effects that we found are the net result oftens or hundreds of thousands of individuals deciding whether ornot to make a given trip, where to go, and which mode to take.While some people may switch from auto to rail to avoid the addi-tional road congestion that often accompanies inclement weather,such behavior appears to be more than offset by a reduction intrip-making. Based on personal communication, transit officials


were anecdotally aware of some weather effects but unsure of con-ditions and causes. Further research is needed to determine thecomplex tradeoffs involved in each individual’s decision-makingprocess and to build on the disaggregate research by Aaheim andHauge (2005). Future research should assess variations in air masseffects based on gender, age, income, type of trip, and other factors.Research is also needed into which component of a trip is most im-pacted by the weather: the walk to/from the station, the wait at thestations, the activity at the destination, or perhaps the effect on thecompeting modes. Guo et al. (2007) also observed this deficiency inthe literature and noted the possible link between weather-basedridership changes and trip length, flexibility, and purpose.

Due to the significance of air mass effects on system-wide tran-sit ridership, future research should assess air mass effects at indi-vidual stations. The differences in station characteristics may offerinsight into some of the aforementioned issues (i.e., why riders re-spond the way they do). Finally, the decision to use these rail sys-tems is likely impacted by weather forecasts, in addition to theactual meteorological conditions as examined here. Thus, researchis needed that focuses on weather forecast/ridership relationships.Ultimately, these studies will assist transit agencies in developing amore comprehensive evaluation of system performance and prior-itizing system upgrades.

Acknowledgements

This research was developed in an interdisciplinary seminarunder Arizona State University’s Integrative Graduate Education,Research and Training (IGERT) program in Urban Ecology, spon-sored by National Science Foundation Grant #9987612. We wouldlike to thank ASU’s Global Institute of Sustainability for use of theirfacilities. We would like to thank Carol Atkinson-Palombo, now atUniversity of Connecticut, for organizing the seminar, and theother students for their feedback and support: Matthew Lord,Jong-Geun Kim, James Miller, Brent Hedquist, Donna Hartz,Nayoung Ryu, and Christopher Upchurch. Scott Sheridan of theDepartment of Geography at Kent State University provided theair mass data. Finally, we thank those at CTA, BART, and NJT whokindly provided us with daily rail ridership data.

References

Aaheim, H.A., Hauge, K.E., 2005. Impacts of climate change on travel habits: anational assessment based on individual choices. Center for InternationalClimate and Environmental Research Report 2005:07. <http://www.cicero.uio.no>.

American Public Transit Association, 2007. <http://www.apta.com/transit_by_mode/>.Andrey, J., Mills, B., Leahy, M., Suggett, J., 2003. Weather as a chronic hazard for road

transportation in Canadian cities. Natural Hazards 28, 319–343.Arlinghaus, S.L., Nystuen, J.D., 1985. Climatic effects on bus durability. University of

Michigan Transportation Research Institute. Ann Arbor, MI.Cervero, R., 1993. Ridership Impacts of Transit Focused Development in California.

University of California, Berkeley, CA.

Changnon, S., 1996. Effects of summer precipitation on urban transportation.Climate Change 32, 481–494.

Chicago Transit Authority, 2006. <http://www.transitchicago.com>.de Palma, A., Rochat, D., 1999. Understanding individual travel decisions: results

from a commuters survey in Geneva. Transportation 26, 263–281.Demery Jr., L.W., 2002. Urban transit disinformation: a recent history. <http://

www.publictransit.us/cadocs/politicaldebates/transitdisinfo.htm> (4.12.04).Fielding, G.J., 1995. Transit in American cities. In: Hanson, S. (Ed.), The Geography of

Urban Transportation, second ed. Guilford Press, New York, pp. 287–304.Filion, P., 2001. Suburban mixed-use centres and urban dispersion: what difference

do they make? Environment and Planning-A 33, 141–160.Flyvbjerg, B., Skamris Holm, M.K., Søren, L.B., 2005. How (in)accurate are demand

forecasts in public works projects? The case of transportation. Journal of theAmerican Planning Association 71, 131–146.

Guo, Z., Wilson, N.H.M., Rahbee, A., 2007. Weather impact on transit ridership inChicago. Transportation Research Record 2034, 3–10.

Kalkstein, L.S., 1991. A new approach to evaluate the impact of climate on humanmortality. Environmental Health Perspectives 96, 145–150.

Kalkstein, L.S., Barthel, C.D., Greene, J.S., Nichols, M.C., 1996. A new spatial synopticclassification: application to air mass analysis. International Journal ofClimatology 16, 938–1004.

Keay, K., Simmonds, I., 2005. The association of rainfall and other weather variableswith road traffic volume in Melbourne, Australia. Accident Analysis andPrevention 37, 109–124.

Khattak, A., 1991. Driver response to unexpected travel conditions: effect of trafficinformation and other factors. Ph.D. dissertation, Department of CivilEngineering, Springer, Evanston, Illinois.

Khattak, A.J., de Palma, A.D., 1997. The impact of adverse weather conditions on thepropensity to change travel decisions: a survey of Brussels commuters.Transportation Research A 3, 181–203.

Khattak, A., Polydoropoulou, A., Ben-Akiva, M., 1995. Commuters’ normal and shiftdecisions in unexpected congestion: pre-trip response to advanced travelerinformation systems. Transportation Research Record 1537, 46–54.

Knapp, K.K., Smithson, L.D., 2000. Winter storm event volume impact analysis usingmultiple-source archived monitoring data. Transportation Research Record1700, 10–16.

Kuby, M., Upchurch, C., Barranda, A., 2004. Factors influencing light rail stationboardings in the United States. Transportation Research Part A 38, 223–247.

Levine, L., 1990. Phoenix bus riders don’t swelter in the shelter. METRO MagazineJuly/August, pp. 38–39.

Maze, T.H., Agarwal, M., Burchett, G., 2006. Whether weather matters to trafficdemand, traffic safety, and traffic operations and flow. Transportation ResearchRecord 1948, 170–176.

New Jersey Transit, 2006. <http://www.njtransit.com/>.Parsons Brinckerhoff Quade and Douglas, Inc, 1996. Transit and Urban Form,

National, Academy Press, Washington, DC.Pickrell, D., 1992. A desire named streetcar: fantasy and fact in rail transit planning.

Journal of the American Planning Association 58, 158–176.Pucher, J., 2004. Public transportation. In: Hanson, S., Giuliano, G. (Eds.), The

Geography of Urban Transportation, third ed. Guilford, New York, pp. 199–236.Pushkarev, B.S., Zupan, J.M., 1982. Where transit works: urban densities for public

transportation. In: Levinson, H.S., Weant, R.A. (Eds.), Urban TransportationPerspectives and Prospects, Eno Foundation, Westport, CT, p. 341.

Ruth, M., Kirshen, P., 2001. Integrated impacts of climate change uponinfrastructure systems and services in the Boston Metropolitan Area. WorldResources Review 13, 106–122.

San Francisco Bay Area Rapid Transit District, 2006. <http://www.bart.gov/>.Sheridan, S., 2002. The redevelopment of a weather-type classification scheme for

North America. International Journal of Climatology 22, 51–68.Sheridan, S., 2008. Spatial synoptic classification. Department of Geography. Kent

State University. <http://sheridan.geog.kent.edu/ssc.html>.Smoyer, K.E., Kalkstein, L.S., Greene, J.S., Ye, H., 2000. The impacts of weather and

pollution on human mortality in Birmingham, Alabama and Philadelphia,Pennsylvania. International Journal of Climatology 20, 881–897.

http://www.cicero.uio.no

http://www.apta.com/transit_by_mode/

http://www.transitchicago.com

http://www.publictransit.us/cadocs/politicaldebates/transitdisinfo.htm

http://www.publictransit.us/cadocs/politicaldebates/transitdisinfo.htm

http://www.njtransit.com

http://www.bart.gov

http://sheridan.geog.kent.edu

An analysis of air mass effects on rail ridership in three US cities

Documents

Transcript of An analysis of air mass effects on rail ridership in three US cities