1. Background
Recently, scholars and policy makers have shown increasing interest in climate change adaptation and mitigation. Although large uncertainties and regional variations exist, climate changes (i.e., overall warmer temperatures and changes in precipitation patterns) have been projected and can already be observed today (e.g., Stocker et al. 2013). These weather changes are expected to have an impact on daily lives, especially those aspects that take place outdoors. One sector that may be particularly affected is the transportation sector (Koetse and Rietveld 2009). Of particular interest are the effects on cycling, which provide large societal benefits in terms of CO2 emission reduction and improved accessibility, liveability, and health. Intuitively, cycling should be highly affected by weather due to cyclists’ direct exposure. Yet, despite many recent contributions, scientific knowledge about this relationship is still somewhat limited.
Existing studies demonstrate the effects of daily weather [exceptions include, e.g., Sabir (2011), Tin Tin et al. (2012), and Gebhart and Noland (2013), who analyze weather also on an hourly level] on a wide range of travel behaviors, including transport mode choices, trip generation, and to a lesser extent distances traveled [for detailed literature reviews, see Koetse and Rietveld (2009), Heinen et al. (2010), and Böcker et al. (2013)]. Findings generally indicate that warm sunny or dry weather enhances walking and cycling, while cold, wet, or windy weather has the opposite effects—the effects typically being larger for recreational than for utilitarian purposes (e.g., Hanson and Hanson 1977; Nankervis 1999; Bergström and Magnusson 2003; Aaheim and Hauge 2005; Gallop et al. 2012; Sabir 2011; Flynn et al. 2012; Sears et al. 2012; Thomas et al. 2013). In contrast, studies pay less attention to time spent traveling. However, the effects of weather may be inherently related to time. People may tolerate traveling in the open air during weather conditions like heat, cold, or heavy rain for a short period of time, but may not want to expose themselves for too long.
Studies on weather and travel behavior show two important meteorological shortcomings. First, studies often assume linear weather effects on travel behaviors, whereas in reality these effects may be nonlinear through optimums or thresholds. However, some recent studies demonstrate that not only lower but also high temperatures may negatively affect cycling, for instance above air temperatures of 24°C in Portland (Ahmed et al. 2012), 28°C in Melbourne (Phung and Rose 2008) and Montreal (Miranda-Moreno and Nosal 2011), and 32.2°C in Boulder, Colorado (Lewin 2011). Second, except for a few studies demonstrating the effects of weather on cycling using weather ratings (e.g., Nankervis 1999) or thermal indices (e.g., Brandenburg and Arnberger 2001; Brandenburg et al. 2004), weather and climate change are solely described and discussed in terms of absolute values and trends—in either averages or extremes—of individual meteorological variables, such as air temperature, wind speed, radiation, or precipitation. However, in order to evaluate the impact of weather on people’s thermal comfort, health, and well-being, as well as travel behaviors, it is necessary to analyze the combined effects (e.g., Thorsson et al. 2011; Böcker et al. 2013).
To address these shortcomings in this paper we analyze and compare the impact of precipitation and the single as well as combined (via the physiological equivalent temperature index) impacts of air temperature, mean radiant temperature (radiant heat exchange between humans and environment), air humidity, and wind speed on cycling frequencies, time spent cycling, and the exchange between cycling and other transport modes. This work results from interdisciplinary research collaboration between transportation and climatology scientists with the ambition to integrate and expand on knowledge about the complex link between weather and cycling. This knowledge could be used to give guidance on how to plan and design sustainable, healthy, accessible, and attractive urban areas at present and in a changing climate.
2. Data and methods
Selected for this study was the greater Rotterdam area situated on the west coast of the Netherlands. The area is part of the Randstad, the densely populated economic heart of the Netherlands. Typical for the Netherlands, cycling shares are relatively large (around 20%). The region is characterized by a warm-temperate (maritime) climate (Geiger and Pohl 1954) with mild winters (average lows of 1°C and highs of 6°C), warm summers (average lows of 12°C and highs of 21°C), and relatively stable year-round precipitation patterns (KNMI 2013). The rationale for selecting Rotterdam is threefold. First, it is a metropolitan region with a wide variety of population categories (see section 2b). Second, it offers a wide variety of high-density as well as lower-density residential environments (Fig. 1). Third, the region demonstrates an active policy on sustainable transportation and climate change adaptation.
Greater Rotterdam study area.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
a. Meteorological attributes
Hourly meteorological data (air temperature, air humidity, wind speed, global radiation, and precipitation) for greater Rotterdam were obtained from the Dutch Meteorological Institute (KNMI 2013) from August 2012 through February 2013. The meteorological station is located at the northern edge of Rotterdam (51°57′N, 4°27′E) approximately 20 km from the sea (Fig. 1). Figure 2 describes observed daily maximum air temperature Ta(max), average wind speed Ws(avg), precipitation sum P(sum), and the occurrence of snow cover on the ground during the study period, along with normal air temperatures for the period 1980–2010. After a wet July month (not included in this research), warmer-than-average sunny weather occurred in late August [with a peak in Ta(max) of 34°C; 10°C above average] as well as early September and late October. After a warmer and much wetter than average December month, mid-January provided unusually cold winter weather (up to 10°C below average) with continuous snow cover and frost even during daytime.
Rotterdam weather conditions during survey period. Source: Helbich et al. (2014). Based on KNMI’s (2013) publically available weather data. Key: Ta(max), observed = observed daily maximum air temperature; Ta(max), normal = daily maximum air temperature averages for the period 1980–2010.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
Weather effects on mobility were analyzed using daily meteorological variables aggregated from hourly meteorological data over the 18-h period between 0600 LT and 2400 LT, the time interval in which the majority of mobility takes place (LT is local time). The reason for using daily over hourly variables is threefold. First, some of our analyses (cycling frequencies and total cycling durations) can only be analyzed on the daily level. For uniformity, interpretability, and model comparison it is preferred to also perform the other analyses on the daily level. Second, transport mode choices are often made on the daily rather than instantaneous level. For instance, when deciding in the morning to take the bicycle to work, one probably also keeps in mind the predicted weather later that day when returning home. Third, usage of daily aggregates will allow for better comparison with existing studies, as well as better compatibility with climate change projections. We are aware that weather characteristics aggregated over the course of a day may not always correctly describe the weather experienced at a specific moment. For mode choices, we also tested an analysis with weather variables on the hourly level. This hourly model revealed the same picture and performed roughly the same; hence the daily level model was preferred.
For Ws(avg) the daily average (in m s−1) was aggregated; for P(sum) the daily sum is in millimeters.1 Thermal conditions were investigated using daily maximums rather than averages because these better describe the type of day. For instance, a clear day with cold (at night) and hot conditions (during the day) may result in a similar temperature average as a cloudy day with mild conditions at night as well as during the day. Thermal conditions were analyzed in three ways: by means of maximum daily air temperature [Ta(max)], mean radiant temperature [Tmrt(max)] and physiological equivalent temperature [PET(max)].
Mean radiant temperature (Tmrt) is one of the most important meteorological variables governing the human energy balance and outdoor thermal comfort (heat load), especially during warm and sunny days (Mayer and Höppe 1987). It is a sum of the exposure to all shortwave and longwave radiation fluxes (direct, diffuse, reflected, and emitted) in a given surrounding area. The value of Tmrt is directly influenced by the surface geometry (buildings, vegetation, and topography) and surface materials, which also makes it a good measure to identify built geometries causing increased risks. Compared to Ta, Tmrt shows large spatial variations over a short distance during the day (e.g., Thorsson et al. 2011; Lindberg and Grimmond 2011a). The calculation of Tmrt was done using SOLWEIG1D (Lindberg 2012), a subversion of SOLWEIG version 2.3 (Lindberg et al. 2008; Lindberg and Grimmond 2011b). SOLWEIG1D calculates the Tmrt for a person residing in a specific site (point) and requires observation data on air temperature, air humidity, and solar radiation (global and diffuse components). In addition, geographical information (latitude, longitude, and height above sea level) and information about surface characteristics (sky view factor, albedo, and emissivity) along with information about the proportion of radiation received by the human body in each direction (angular factors) and absorption coefficients for shortwave and longwave radiation are needed. In this study the Tmrt was calculated for a person standing/walking in a constantly sun-exposed environment with a sky view factor (SVF) of 0.6 (selected to represent the general SVF for the area). Angular factors were set to 0.22 for radiation fluxes from the four cardinal points (east, west, north, and south) and 0.06 for radiation fluxes from above and below. Standard values of absorption coefficients for shortwave and longwave radiation were used (i.e., 0.7 and 0.97, respectively) (Höppe 1992; VDI 1998). Albedo and emissivity for buildings and vegetation were set to 0.20 and 0.95, respectively according to Oke (1987). For detailed equations and model validation, see Lindberg et al. (2008) and Lindberg and Grimmond (2011b).
Mayer and Höppe’s (1987) physiological equivalent temperature index quantifies the combined effect of Ta, Ws, Tmrt, and air humidity. A PET value of, for instance, 30°C means that the human body perceives the combined outdoor thermal conditions as a fictive indoor environment with an air temperature of 30°C (Mayer and Höppe 1987). In calculating the PET, characteristics of human beings are set as constants: 80 Watts for internal heat production, 0.9 clo (1 clo = 0.155 m2 °C W−1) for clothing heat transfer resistance. PET values between 18° and 23°C indicate comfortable conditions. Higher (lower) values indicate increasing probability of heat (cold) stress (Matzarakis and Mayer 1996).
b. Mobility attributes and modeling techniques
This research draws on travel diary data from a panel study among 945 Greater Rotterdam residents, who were selected with the assistance of a questionnaire agency from an existing Internet panel. Respondents aged 18 years and older were randomly selected from four different types of residential environments in the greater Rotterdam area: inner city, outer center, green/suburban, and rural (see Fig. 1). We oversampled for nonnative Dutch originating from hotter climates (mainly Surinamese, Dutch-Antillean, Indonesian, Turkish, and Moroccan) and older age groups (aged 65 or higher) because of generally lower response rates for these groups in research (Adler et al. 2002), as well as to account for specific weather preferences (e.g., potentially lower heat or heavy wind tolerances among elderly or higher heat tolerances for ethnic groups originating from hotter climates).
Respondents were assigned randomly to six sample days—two in summer (August–September), two in autumn (October–November), and two in winter (December–February)—in which they completed records of all their travel and destination activities. As a result, all days during the data period were covered, with the single exception of the Christmas holiday. Records were only registered during periods that were typically regular working weeks for the respondent, including weekends but excluding holidays. Extra diaries were distributed during hot summer days and cold (snow covered) winter days in order to obtain sufficient response numbers during these interesting but rarely occurring weather extremes. Sample attrition was 13.1% between waves 1 and 2 and 13.0% between waves 2 and 3 and appeared to be random rather than biased when various key sociodemographics were considered. For this reason we also included in our final sample respondents who only completed one or two seasons. Table 1 specifies the composition of the final sample of 945 respondents after cleaning the data for unreliable records and records featuring trips outside the Netherlands subject to different weather conditions. Generally the sample is well balanced with substantial shares for various sociodemographic groups. Additionally, the sample is reasonably representative when compared to the general population in this region, with the exception of lower educated and nonnative Dutch being somewhat underrepresented, which is quite typical for Internet surveys (Adler et al. 2002).
Sample composition and representativeness. A dash indicates that data are not publicly available.
Our multivariate analysis considered the marginal effects of weather on mode choices, cycling frequencies, and cycling durations, as well as its relative contribution compared to various background variables. These include classical individual and household background variables (e.g., age, gender, income, etc.), spatiotemporal attributes (e.g., residential environment, weekend/weekday, time of the day, etc.), and attitudes that may intermediate the role of weather effects on mobility (e.g., favorite season, environmental concern, and whether a person is more urban or countryside oriented). These are listed in Fig. 3. Mode choices are analyzed on the trip level, while cycling frequencies and durations are analyzed per person per day. The trip-level attributes of trip purpose and type, as well as the spatiotemporal attributes related to time of the day, are only included in the trip-level mode choice analysis. All other attributes are included in all three analyses.
Overview of all variables used in the models. Key: Ta(max) = maximum daily air temperature; Tmrt(max) = maximum daily mean radiant temperature; PET = maximum daily physiological equivalent temperature; Ws(avg) = daily average wind speed; P(sum) = daily precipitation sum.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
To test the effects of weather on mode choices a multinomial LOGIT model was used. This model, based on utility maximization, is similar to binary logistic regression, except that it allows a categorical variable with more than two alternatives. The reference category is the car, allowing for comparison with earlier studies. For daily cycling frequencies and durations, negative binomial and TOBIT models (respectively) were used. These were preferred over standard Poisson (count data) and ordinary least squares (OLS) (interval/ratio data) regression, as they better handle the dependent variables’ absence of negative values and excess of zeros due to people not making cycling trips on a day (Cameron and Trivedi 1998; Greene 1997). Because this is a panel study and each respondent performs several trips during several days, trips performed by one respondent may not be strictly independent. To relax the usual requirement of independent observations, all statistical analyses in this paper were performed with the Stata software package “vce-cluster” command (clustered by respondent ID). The estimation of equal robust standard errors per respondent corrects for intragroup correlation (Wooldridge 2002).
For all analyses we estimated separate models for the different trip purposes of work, errands, social visits, and leisure, because existing literature indicates that the weather effects are smaller for utilitarian trips than for leisure (e.g., Aaheim and Hauge 2005; Sabir 2011). However, for the overview and focus of the paper these separate trip purpose models will not be presented but only discussed in text. Additionally, separate models were estimated for Ta(max), Tmrt(max), and PET(max), which because of multicolinearity (Pearson r, Ta(max), Tmrt(max) = 0.81; Ta(max), PET(max) = 0.96; Tmrt(max), PET(max) = 0.90) are analyzed one at a time. Finally, to check for potential nonlinear effects, for each of the three thermal variables, we estimated models in two different ways: untransformed (describing a linear relationship) and transformed (describing a bell-shaped relationship). To model the bell-shaped effect, use is made of a Gaussian function (see Fig. 4), in which f(x) indicates the transformed thermal variable as a function of its original, a is the position of the center of the peak (optimal thermal condition for cycling), and b is the full width at half maximum (FWHM) (see Fig. 4). For each model, various values for a and b were tested until a maximum model log likelihood was reached.
3. Results
a. Weather and cycling: Descriptive results
Figures 5, 6, and 7 provide schematic descriptive overviews of the effects of weather on mode choices and cycling frequencies and durations, calculated for the full Greater Rotterdam sample (all trip purposes combined). The modal splits in Fig. 5 show that, overall, around half of all trips are made by car. As is typical for the Netherlands, cycling also has considerable shares (around 20%). Weather conditions generally seem to have a clear effect on mode choices. Of all transport modes, cycling appears to be the most sensitive to weather.
Weather effects on modal split. Key is as in Fig. 3.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
Weather effects on number of cycling trips per person per day. Key is as in Fig. 3.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
Weather effects on time spent cycling per person per day. Key is as in Fig. 3.
Citation: Weather, Climate, and Society 6, 4; 10.1175/WCAS-D-13-00066.1
The three thermal variables have an initially positive effect on cycling shares until a certain thermal optimum. Thereafter the effect is negative. This is congruent with the—sometimes described as parabolic (e.g., Phung and Rose 2008)—effect of Ta on cycling, often indicated but less often empirically demonstrated in the literature. Cycling shares seem to peak on days with Ta(max) values between 20° and 25°C, Tmrt(max) values between 55° and 60°C, and PET(max) values between 23° and 29°C, which according to the literature should be classified as “slightly warm” (Matzarakis and Mayer 1996). Note that here we refer to maximum temperatures. Cycling in the morning or evening those days may occur during more comfortable thermal conditions. Where cycling peaks, a dip in car usage is observed and to a lesser extent walking, while public transport is affected to a minimum during hot days.
In line with the literature (e.g., Aaheim and Hauge 2005; Sabir 2011), P(sum) and Ws(avg) both negatively affect cycling shares in a more linear way, mostly at the cost of car usage. When it comes to the effects of Ws(avg), it should be noted that days with high Ws(avg) correlate to some extent with rainy (Pearson r = 0.36) and cloudy weather conditions, which may be additional contributors to the here observed decrease in cycling shares.
Figure 6 presents the effects of weather conditions on the number of bicycle trips per person per day. It represents actual bicycle frequencies (demand) rather than mode shares, discussed above. Generally, weather effects on cycling frequencies show a similar pattern as observed for cycling shares, but the effects seem to be a bit more pronounced. Congruent with the modal split descriptives, daily P(sum)—especially on days with over 5 mm of precipitation—and Ws(avg) show negative effects.
Regarding thermal conditions, cycling frequencies increase until the same optimum temperatures found earlier for the modal split shares are reached. But rather than the parabolic relationship indicated in the literature, thermal conditions seem to demonstrate a bell-shaped effect on cycling, which flattens out on the lower extreme. This is especially visible when looking at the graphs for Ta(max) and Tmrt(max). It seems that, at least within the wide range of thermal conditions observed during our study period, a core of weather-tolerant cyclists exists who keep cycling, regardless of the thermal conditions. Remarkably, the PET(max) graph reports a dip in cycling frequencies at days with PET(max) values between 18° and 23°C, which according to the literature should be classified as comfortable thermal conditions (Matzarakis and Mayer 1996). However, the exceptionally high cycling peak in the next class may indicate that the observed anomaly is related to arbitrarily chosen interval classes. Also, it should be noted that here we refer to descriptives that take no account of simultaneous effects of temperatures, precipitation, and wind speeds. These descriptive graphs should therefore be interpreted as indicators of general trends rather than exact representations of individual classes.
Figure 7 presents the effects of weather on daily cycling durations. A roughly comparable picture arises for the cycling frequencies, with more or less similar bell-shaped effects for the thermal variables and negative effects for P(sum) and Ws(avg). One exception is that the hottest thermal classes for Ta(max), Tmrt(max), and PET(max) no longer show a decline over the preceding classes. A closer look at the raw mobility data for these hot days revealed that these higher cycling durations are mostly related to distinct leisure activity patterns involving recreational cycling tours and cycling trips to more distant natural recreation areas, including beaches. Overall, the weather effects for cycling durations appear to be even more prominent than for cycling frequencies. This is, for instance, clearly demonstrated when comparing the slope in the graphs for Ws(avg). It seems that during calm, dry, and pleasant thermal conditions people cycle not only more frequently, but also longer.
b. Multivariate results
This section presents a multivariate analysis of the effects of weather on cycling. Tables 2, 3, and 4 summarize the effects of weather and weather-related parameters on mode choices, cycling frequencies, and cycling durations. To not distract from the main meteorological focus of this paper, the tables do not display the full underlying models, which also include all other independent background attributes2 listed in Fig. 3. These full models, as well as the separate models estimated for the different trip purposes,3 can be requested by sending an e-mail to the authors.
Weather effects on transport mode choice. One asterisk (*) indicates p < 0.10, two (**) p < 0.05, and three (***) p < 0.01. Ta(max) = maximum daily air temperature; Tmrt(max) = maximum daily mean radiant temperature; PET(max) = maximum daily physiological equivalent temperature; Ws(avg) = daily average wind speed; P(sum) = daily precipitation sum.
As in Table 2, but for weather effects on the number of hours spent cycling per person per day.
1) Weather effects on cycling shares compared to other transport modes
Table 2 presents a summary of the effects of meteorological variables on transport mode choices, as well as the effects of whether or not a trip was made in daylight and whether or not snow cover occurred on the ground. The parameter estimate (B) indicates the change in log odds of choosing a transport mode over the car for a one-unit parameter change. The z statistic indicates the ratio between the parameter estimate and the robust standard errors clustered per respondent. Statistically significant effects have been marked with asterisks. Overall, meteorological variables have a clear effect on the exchange between cycling and the car, even when corrected for various background characteristics, as well as snow cover on the ground and daylight. The exchange between the car and other transport modes is less affected by weather.
Regarding the thermal conditions, the bell-shaped transformed variables for Ta(max), Tmrt(max), and PET(max) (see section 2b) demonstrate more significant effects and provide better overall model performance than the untransformed originals. This indicates that the effect of thermal conditions on transport mode choice is bell-shaped rather than linear. The positive parameter estimates for the transformed variables imply that the effects of thermal conditions on cycling follow the bell-shaped curve. More precisely, the best model fit was reached for bell-shaped variables with an optimum of 24°C for Ta(max) (FWHM = 25°C), 52°C for Tmrt(max) (FWHM = 25°C), and 30°C for PET(max) (FWHM = 42°C). These optimums are roughly in line with the optimum intervals found in the descriptives. When comparing the three bell-shaped thermal parameters, Tmrt(max) and PET(max) provide higher significance levels and better or similar model fit—indicated by the Wald chi-square statistic4—than Ta(max). The only thermal variable that is significant on the 99% confidence interval is PET(max). This indicates that combining meteorological variables (Ta, Tmrt, humidity, and Ws) within PET seems to lead to a better explanation of the exchange between cycling and the car. However, Tmrt is the dominant component within the PET and the most important component of thermal comfort and heat load (Mayer and Höppe 1987). Moreover, the Tmrt model as a whole, including Ws(avg) as a separate meteorological variable, performs better (higher Wald chi-squared score) than the PET model, in which Ws(avg) has been integrated within the PET(max).
In line with the descriptives, Ws(avg) and P(sum) negatively affect cycling shares as compared to the car. The Ws(avg) effect is significant only in the Ta model (with 90% confidence) and not in the Tmrt model. This may be explained by the higher negative correlation between Ws(avg) and Tmrt(max) (Pearson r = −0.29) than between Ws(avg) and Ta(max) (Pearson r = −0.17). The significant negative effect of Ws(avg) on cycling found in the Ta model may be partially associated with potentially unpleasant or threatening cloudier skies, while in the Tmrt model these cloudier skies are better captured by Tmrt(max). For P(sum) the negative effects on cycling shares are significant (with a confidence interval of 95% or higher) in all models. Snow cover on the ground seems to affect cycling negatively, potentially related to slippery conditions. However, the effect is not statistically significant, possibly because slippery conditions also negatively affect car usage (nonsignificant positive effects are shown for walking and public transport shares relative to the car). An especially notable observation is the strong positive effect of daylight on both cycling and walking shares in comparison to the car. It seems that after sunset, active transport modes become less attractive. This may partially be related to colder thermal conditions at night, but also to reduced visibility or potential safety issues.
2) Weather effects on cycling frequencies
Table 3 presents a summary of meteorological effects on the number of cycling trips per person per day. Rather than relative cycling shares, this is an analysis of cycling demand in absolute terms. The parameter estimate (B) indicates the change in the expected log count of the number of bicycle trips per person per day for a one-unit parameter change. Robust standard errors (S.E.) clustered per respondent and z statistics indicating the ratio between parameter estimates and standard errors are also presented.
Overall, weather seems to have a clear effect on cycling frequencies. In line with the descriptives and the effects on cycling mode shares, thermal conditions demonstrate a bell-shaped effect on cycling frequencies (all significant with 99% confidence). The highest cycling frequencies occur on days with values of 24°C for Ta(max), 52°C for Tmrt(max), and 33°C for PET(max). This 33°C for PET(max) is slightly higher than the optimums found in the descriptives and the mode choice model. It seems that although cycling mode shares start decreasing above 30°C PET, cycling frequencies in absolute terms start decreasing above 33°C PET—thermal conditions classifying as “warm,” two classes above “comfortable” in the literature (Matzarakis and Mayer 1996). Similar to the cycling shares, cycling frequencies Tmrt(max) and PET(max) seem to also be better indicators (higher z statistics as well as overall model Wald chi-squared) than Ta(max).
As for cycling mode shares, P(sum) shows significant linear negative effects on cycling frequencies (all with 99% confidence). Earlier, a negative effect of snow cover could not be statistically confirmed on cycling shares relative to the car. In contrast, when it comes to absolute cycling frequencies, snow cover demonstrates a highly significant negative effect in all models, suggesting an influence of slippery conditions. It seems that Ws(avg) negatively affects cycling frequencies, but these effects are not statistically significant.
3) Weather effects on cycling duration
Table 4 presents a summary of meteorological effects on total hours cycling per person per day. The parameter estimate (B), similar to standard OLS regression, indicates the predicted change in cycling hours per person per day for a one-unit parameter increase. Robust standard errors (S.E.) clustered per respondent and t statistics indicating the ratio between parameter estimates and standard errors are also presented.
Overall, weather strongly affects daily cycling durations. All included meteorological attributes show significant effects. Generally, the effects of weather on cycling durations show a similar picture as cycling shares and frequencies. Thermal conditions show bell-shaped effects on cycling durations (all significant with 99% confidence). Approximately in line with the descriptives, the longest daily cycling durations occur on days with maximum values of 24°C for Ta(max), 52°C for Tmrt(max), and 31°C for PET(max). Both P(sum) and snow cover on the ground demonstrate highly significant negative effects on cycling durations (all with 95% or 99% confidence). Also, Ws(avg) shows a negative effect on cycling durations, which, in contrast to the cycling frequencies, is significant (with 99% confidence in the Ta model and with 90% confidence in the Tmrt model). In line with the descriptives, it seems that Ws(avg) affects cycling durations more strongly than cycling frequencies. This indicates that people do still make bicycle trips on windy days but that those trips are shorter than on calm days.
4. Discussion and conclusions
This paper aims to investigate the complex link between weather and cycling. Drawing on travel diary data from a triple-wave panel study among 945 respondents from Greater Rotterdam, the Netherlands, and after controlling for various individual, household, and trip background characteristics, we analyzed the effects of aggregated daily weather conditions on cycling. Daily weather data produced results similar to those of hourly weather data and were preferred because of uniformity and compatibility and because they are more in line with the moment of transport mode decision-making. To provide a complete picture of cycling behavior, unlike most existing research, an analysis was made not only of cycling frequencies, but also of cycling durations and the exchange between cycling and other transport modes. The results demonstrate a clear overall effect of daily weather conditions on transport mode choices in general and cycling in particular. Congruent with the literature (e.g., Nankervis 1999; Bergström and Magnusson 2003; Aaheim and Hauge 2005; Sabir 2011), daily precipitation sum [P(sum)] and average wind speed [Ws(avg)] demonstrate linear negative effects on cycling, mostly benefiting car usage. Cycling durations seem to be more significantly affected by Ws(avg) than cycling shares and frequencies. This indicates that on windy days people not only cycle less often, but also for shorter durations. Thermal conditions affect cycling in a nonlinear bell-shaped way. Congruent with some recent studies (e.g., Phung and Rose 2008; Ahmed et al. 2012; Lewin 2011), a thermal optimum for cycling can be identified (days with maximum air temperatures around 24°C). Below and above this temperature cycling shares, frequencies, and durations decrease and car usage increases. Compared to the exposed bicycle, the car offers its user better protection from wind, precipitation, and cold as well as heat.
In contrast with most existing studies on weather and transportation, when analyzing thermal conditions we not only take into account the daily maximums for air temperature [Ta(max)], but also for mean radiant temperature [Tmrt(max)], which combines the effects of air temperature (via emission of long-wave radiation from the sky and surrounding surfaces) and solar radiation (direct, diffuse, and reflected), and physiological equivalent temperature [PET(max)], which combines the effects of Ta, Tmrt, Ws, and humidity. Weather may be perceived as warmer during calms and sunshine than during cloudy and windy conditions, despite a similar Ta. The use of PET and Tmrt provides a more comprehensive account of weather conditions, although it should be mentioned that in order to calculate these indices, assumptions needed to be made regarding the urban geometry and human parameters such as shape, heat production, and clothing. Our results demonstrate that Tmrt(max) and PET(max) provide better explanations for cycling shares, frequencies, and durations than just Ta(max), confirming the importance of combined weather effects. Nevertheless, Ta(max), although somewhat weaker, also demonstrates significant effects. In general, Ta may therefore still be a powerful indicator as it is easily interpretable, free of assumptions, widely accessible, and easily compatible with weather forecasts and climate change scenarios.
Insights from this paper are relevant to policy makers in several ways. When analyzing cycling trends over time, as in when evaluating the effects of cycling policies, for example, the insights presented here could be used to disentangle cycling trends from the noise caused by short-term weather fluctuations. Similarly, the insights could be used to improve annual modal split estimations by adjusting for weather conditions—particularly when these are based on origin-destination survey data conducted over the course of less than a full year, or during years with abnormal weather conditions. Additionally, the effects of weather on cycling are relevant in the context of climate change. Most of the year, a warmer future climate may have positive effects on cycling. However, in summer heat may pose a threat. A brief calculation based on Dutch Meteorological Institute climate change projections for the Netherlands (KNMI 2009) reveals that the number of days with maximum air temperatures of 25°–30°C and 30°–35°C—days on which optimum conditions for cycling are exceeded—is expected to increase respectively from 24 currently to 30–47 in 2050 and from 4 currently to 7–14 in 2050, depending on whether mild (+1°C global temperature rise; no changing prevailing wind patterns) or severe (+2°C global temperature rise, changing prevailing wind patterns) climate change scenarios are utilized. A potential future increase in overall precipitation and extreme precipitation events plus changes in solar radiation/cloudiness may further (negatively) impact cycling. Policy makers are advised to consider (thermal) climate adaptations in urban design, for instance through usage of deciduous trees in urban areas and along cycling infrastructures, which provide cooling in summer when foliated—mainly through shading but also through transpiration—while retaining valued sunshine in winter when defoliated (e.g., Lindberg and Grimmond 2011a; Konarska et al. 2014). In addition to alleviating heat and solar radiation, trees can be used to regulate wind and humidity and to protect from sudden (summer) downpours. Wind barriers and precipitation shelter, in the form of trees or other, particularly along main cycling infrastructures, could be additional interventions to reduce cyclists’ weather exposures, which may lead to increased bicycle-over-car usage.
This research explores the complex integrated weather effects on travel behavior. Future research could further elaborate this topic into several directions. First, the notions of combined weather and nonlinearity explored in this study could further be investigated. Studies could further explore the effects of precipitation combined with thermal conditions, as well as the existence of (combined) precipitation, wind, or thermal thresholds triggering sudden travel behavior changes. Also, studies may explore the effects of weather variability (e.g., showers versus continuous rain; wind gusts; sunny spells during cloudy days), as well as lagged weather effects (e.g., first warm sunny spring day after a long winter). Second, research could investigate integrated weather effects on other related travel behavior decisions, such as choices for destinations, departure times, speeds, and routes. Such analyses may require other data collection techniques (e.g., smart GIS interfaces), as well as more precise spatial and temporal resolutions. Third, research on the integrated effects of weather on a wide range of (travel) behavioral activities may further focus on the development of new or the fine-tuning of existing meteorological indices to the specific the activities one wants to address. For instance the PET may be better recalibrated to cycling (or other outdoors activities). Fourth, research into the integrated effects of weather should be explored not only objectively, but also in terms of peoples’ subjective interpretations of weather in connection with emotional experiences during travel or outdoors activities. Such subjective accounts have been studied qualitatively (e.g., Spencer et al. 2013) but may also be integrated into quantitative travel surveys or diaries. Fifth, to better explore thermal optimums and potential negative effects of heat, especially in the light of climate change, future research is strongly recommended in hotter climate regimes where people experience heat more regularly. Finally, to better guide climate adaptation policies, studies could further investigate the effects of spatial design interventions, such as the usage of trees, wind barriers, and precipitation shelters, on the (experienced) weather exposure of cyclists.
Acknowledgments
This research is supported through the research programme “Climate and Environmental change and Sustainable Accessibility of the Randstad (CESAR)” funded by the Netherlands Organisation for Scientific Research (NWO). We thank Dr. Fredrik Lindberg (Gothenburg University) for his assistance with meteorological calculations, Prof. Dr. Martin Dijst (Utrecht University) for his constructive comments, and Julia McQuoid for her assistance with detecting typographical errors. We also thank Mervin van Veen and Peter de Vijver from questionnaire agency Intomart-GFK for their assistance with the data collection. We finally thank the editor Prof. Dr. Kirstin Dow, reviewer Brian Lynch, and one anonymous reviewer for their useful comments during the review process.
REFERENCES
Aaheim, H. A., and Hauge K. E. , 2005: Impacts of climate change on travel habits: A national assessment based on individual choices. CICERO Rep. 07, Centre for International Climate and Environmental Research, Oslo, Norway, 35 pp.
Adler, T., Rimmer L. , and Carpenter D. , 2002: Use of Internet-based household travel diary survey instrument. Transp. Res. Rec., 1804, 134–143, doi:10.3141/1804-18.
Ahmed, F., Rose G. , Figliozzi M. , and Jakob C. , 2012: Commuter cyclist’s sensitivity to changes in weather: Insight from two cities with different climatic conditions. Proc. Transport Research Board 2012 Annual Meeting, Washington, DC, Transportation Research Board of the National Academies, 1–20.
Bergström, A., and Magnusson R. , 2003: Potential of transferring car trips to bicycle during winter. Transp. Res., 37A, 649–666, doi:10.1016/S0965-8564(03)00012-0.
Böcker, L., Dijst M. , and Prillwitz J. , 2013: Impact of everyday weather on individual daily travel behaviours in perspective: A literature review. Transp. Rev., 33, 71–91, doi:10.1080/01441647.2012.747114.
Brandenburg, C., and Arnberger A. , 2001: The influence of the weather upon recreation activities. Proc. First Int. Workshop on Climate, Tourism, and Recreation, Halkidiki, Greece, Neos Marmaras, 123–132.
Brandenburg, C., Matzarakis A. , and Arnberger A. , 2004: Visitor motivation and dependence on the weather of recreationists in Viennese recreation areas. Advances in tourism climatology, A. Matzarakis, C.R. De Freitas, and D. Scott, Eds., Berichte des Meteorologischen Instituts der Universität Freiburg, No. 12, 189–197. [Available online at http://www.urbanclimate.net/cctr/ws2/report.htm.]
Cameron, A. C., and Trivedi P. K. , 1998: Regression Analysis of Count Data. Cambridge University Press, 411 pp.
Cervero, R., and Seskin S. , 1995: An evaluation of the relationships between transit and urban form. TCRP Res. Results Dig., 7, Transportation Research Board, 55 pp. [Available online at http://onlinepubs.trb.org/onlinepubs/tcrp/tcrp_rrd_07.pdf.]
Dieleman, F. M., Dijst M. J. , and Burghouwt G. , 2002: Urban form and travel behavior: Micro-level household attributes and residential context. Urban Stud., 39 (3), 507–527, doi:10.1080/00420980220112801.
Flynn, B. S., Dana G. , Sears J. , and Aultman-Hall L. , 2012: Weather factor impacts on commuting to work by bicycle. Prev. Med., 54, 122–124, doi:10.1016/j.ypmed.2011.11.002.
Gallop, C., Tse C. , and Zhao J. A. , 2012: Seasonal autoregressive model of Vancouver bicycle traffic using weather variables. Proc. 91st Transportation Research Board Annual Meeting, Washington, D.C., Transportation Research Board, 17 pp. [Available online at http://docs.trb.org/prp/12-2119.pdf.]
Gebhart, K., and Noland R. B. , 2013: The impact of weather conditions on Capitol bikeshare trips. Proc. Transportation Research Board 92nd Annual Meeting, Washington, DC, Transportation Research Board, 24 pp. [Available online at http://amonline.trb.org/2013-1.264263/13-0563-1-1.274658?qr=1.]
Geiger, R., and Pohl W. , 1954: Revision of the Köppen–Geiger Klimakarte der Erde. Erdkunde,8, 58–61.
Greene, W. H., 1997: Econometric Analysis. 3rd ed. Prentice-Hall, 1075 pp.
Hanson, S., 1982: The determinants of daily travel-activity patterns: Relative location and sociodemographic factors. Urban Geogr., 3, 179–202, doi:10.2747/0272-3638.3.3.179.
Hanson, S., and Hanson P. , 1977: Effects of weather on bicycle travel. Transp. Res. Rec., 629, 43–48.
Heinen, E., Van Wee B. , and Maat K. , 2010: Commuting by bicycle: An overview of the literature. Transp. Rev., 30, 59–96, doi:10.1080/01441640903187001.
Helbich, M., Böcker L. , and Dijst M. J. , 2014: Geographic heterogeneity in cycling under various weather conditions: Evidence from Greater Rotterdam. J. Transport Geogr., 38, 38–47, doi:10.1016/j.jtrangeo.2014.05.009.
Höppe, P., 1992: Ein neues Verfahren zur Bestimmung der mittleren Strahlungstemperatur in Frien. Wetter Leben, 44, 147–151.
KNMI, 2009: Klimaatverandering in Nederland: Aanvullingen op de KNMI’06 klimaatscenario’s.KNMI. [Available online at http://www.knmi.nl/cms/content/67290/aanvullingen_op_de_knmi06_klimaatscenarios_2009.]
KNMI, cited 2013: Publically available weather records and climatological data. KNMI. [Available online at http://www.knmi.nl/klimatologie/.]
Koetse, M. J., and Rietveld P. , 2009: The impact of climate change and weather on transport: An overview of empirical findings. Transp. Res., 14D, 205–221, doi:10.1016/j.trd.2008.12.004.
Konarska, J., Lindberg F. , Larsson A. , Thorsson S , and Holmer B. , 2014: Transmissivity of solar radiation through crowns of single urban trees—Application for outdoor thermal climate modelling. Theor. Appl. Climatol.,doi:10.1007/s00704-013-1000-3, in press.
Lewin, A., 2011: Temporal and weather impacts on bicycle volumes. Transportation Research Board of the National Academies, 18 pp. [Available online at http://trid.trb.org/view.aspx?id=1092520.]
Lindberg, F., 2012: The SOLWEIG-model. Gothenburg University. [Available online at http://www.gvc.gu.se/english/staff/staff/thorsson-sofia/current-projects/the-solweig-model/.]
Lindberg, F., and Grimmond C. S. B. , 2011a: Nature of vegetation and building morphology characteristics across a city: Influence on shadow patterns and mean radiant temperatures in London. Urban Ecosyst., 14, 617–634, doi:10.1007/s11252-011-0184-5.
Lindberg, F., and Grimmond C. S. B. , 2011b: The influence of vegetation and building morphology on shadow patterns and mean radiant temperature in urban areas: Model development and evaluation. Theor. Appl. Climatol., 105, 311–323, doi:10.1007/s00704-010-0382-8.
Lindberg, F., Holmer B. , and Thorsson S. , 2008: SOLWEIG 1.0 – Modeling spatial variations of 3D radiant fluxes and mean radiant temperature in complex urban settings. Int. J. Biometeor., 52, 697–713, doi:10.1007/s00484-008-0162-7.
Matzarakis, A., and Mayer H. , 1996: Another kind of environmental stress: Thermal stress. WHO News, 18, 7–19.
Mayer, H., and Höppe P. , 1987: Thermal comfort of man in different urban environments. Theor. Appl. Climatol., 38, 43–49, doi:10.1007/BF00866252.
Miranda-Moreno, L., and Nosal T. , 2011: Weather or not to cycle: Temporal trends and impact of weather on cycling in an urban environment. Transp. Res. Rec., 2247, 42–52, doi:10.3141/2247-06.
Nankervis, M., 1999: The effect of weather and climate on bicycle commuting. Transp. Res., 33A, 417–431, doi:10.1016/S0965-8564(98)00022-6.
Oke, T. R., 1987: Boundary Layer Climates. Methuen, 435 pp.
Phung, J., and Rose G. , 2008: Temporal variations in Melbourne’s bike paths. Proc. 30th Australasian Transport Research Forum, Melbourne, Australia, 15 pp. [Available online at http://www.atrf.info/papers/2007/2007_Phung.pdf.]
Sabir, M., 2011: Weather and travel behaviour. Ph.D. dissertation, Vrije Universiteit, 138 pp. [Available online at http://trid.trb.org/view.aspx?id=1148919.]
Sears, J., Flynn B. S. , Aultman-Hall L. , and Dana G. S. , 2012: To bike or not to bike—Seasonal factors for bicycle commuting. Transp. Res. Rec., 2314, 105–111, doi:10.3141/2314-14.
Spencer, P. G., Watts R. A. , Vivanco L. , and Flynn B. S. , 2013: Effect of environmental factors on bicycle commuters in Vermont: Influences of a northern climate. J. Transp. Geogr., 31, 11–17, doi:10.1016/j.jtrangeo.2013.05.003.
Stocker, T. F., and Coauthors, Eds., 2013: Climate Change 2013: The Physical Science Basis. Cambridge University Press, 1535 pp.
Thomas, T., Jaarsma R. , and Tutert B. , 2013: Exploring temporal fluctuations of daily cycling demand on Dutch cycle paths: The influence of weather on cycling. Transportation, 40, 1–22, doi:10.1007/s11116-012-9398-5.
Thorsson, S., Lindberg F. , Björklund J. , Holmer B. , and Rayner D. P. , 2011: Potential changes in outdoor thermal comfort conditions in Gothenburg, Sweden due to climate change: The influence of urban geometry. Int. J. Climatol., 31, 324–335, doi:10.1002/joc.2231.
Tin Tin, S., Woodward A. , Robinson E. , and Ameratunga S. , 2012: Temporal, seasonal and weather effects on cycle volume: An ecological study. Environ. Health, 11 (12), doi:10.1186/1476-069X-11-12.
VDI, 1998: Methods for the human-biometeorological assessment of climate and air hygiene for urban and regional planning. Part I: Climate. VDI 3787, VDI/DIN-Handbuch Reinhaltung der Luft, Band 1b, Düsseldorf, 29 pp.
Wooldridge, J. M., 2002: Econometric Analysis of Cross Section and Panel Data. MIT Press, 752 pp.
For precipitation we also analyzed the effects of precipitation duration, but in the end total sum was preferred because of better performance in the models.
The respective effects of these sociodemographic, household, trip, and urban form attributes on mode choice outcomes are generally in line with the literature (e.g., Hanson 1982; Cervero and Seskin 1995; Dieleman et al. 2002).
In line with the literature (e.g., Hanson and Hanson 1977; Aaheim and Hauge 2005; Sabir 2011), it was shown that mode choices for leisure trips and social visits are more strongly affected by weather conditions than mode choices for utilitarian trips from/to work, study, and especially errands.
The Wald chi-square evaluates to what extent a chosen model performs better than a baseline model in which all estimated parameters are set to zero. If none of the parameters would add anything to the chosen model, the Wald chi-square would be zero. A higher Wald chi-square indicates a better model fit.
