Climate change and global warming are severe threats that currently affect the daily lives of the world population. One of the human behaviors that can be most affected by weather conditions is that of personal travel, including commuting, an activity that millions of workers worldwide engage in daily. Within this framework, we analyze the relationships between weather conditions and daily commuting in the US. To that end, we use twenty nationally representative American Time Use Surveys, combined with daily weather data at the county level, spanning the period from 2003 to 2023. The analysis reveals significant relationships between daily weather conditions and commuting mode choices. Specifically, rainy days, high temperatures, and elevated snowfall are positively associated with an increased probability of using cars as the primary commuting mode. In contrast, these weather conditions exhibit a negative relationship with alternative modes of transport, such as public transit or walking. Further findings suggest that these estimates are predominantly driven by days characterized by extremely high temperatures, heavy precipitation, and light snowfall. Finally, our results suggest adaptation to higher temperatures in warmer regions. These results are important for the design of policies aimed at mitigating the mobility consequences of climate change.

1 Introduction

In this paper, we analyze the daily commuting patterns of workers in the United States, focusing on how daily weather conditions relate to the time workers spend commuting per day and the use of modes of transport for traveling to and from work. Millions of individuals travel every day, and commuting to and from work has increased over time in many developed countries (Gimenez-Nadal et al. 2018a, 2022a; UK Department for Transport 2024) and plays a big role in the daily life of most workers worldwide (European Commission 2020; Giménez-Nadal et al. 2020, 2024).

Commuting patterns, including both the time spent on daily commutes and the choice of transport mode to travel to and from work, have important implications for workers, economies, and the environment. Longer commute times have been linked to adverse outcomes for workers and firms, such as productivity losses (Grinza and Rycx 2020), increased sickness absence (van Ommeren and Gutiérrez-i-Puigarnau 2011; Gimenez-Nadal et al. 2022b), shirking behavior (Ross and Zenou 2008; Gimenez-Nadal et al. 2018b), lower levels of well-being and life satisfaction (Stutzer and Frey 2008; Gottholmseder et al. 2009; Wener and Evans 2011), and poorer overall health (Roberts et al. 2011; Künn-Nelen 2016). Furthermore, commuting and the chosen mode of transport have significant environmental consequences, contributing to pollution, congestion, and traffic accidents (Chapman 2007; Buehler 2011).

One factor that affects the commuting behavior of workers is that of weather conditions, as an activity inevitably exposed to the outdoor environment. Transportation and weather conditions are intrinsically linked according to a large body of research (for extensive reviews of this literature, see Koetse and Rietveld 2009; Böcker et al. 2013a; Dijst et al. 2013; Liu et al. 2017; Gössling et al. 2023), and weather is an important determinant of travel behavior, influencing practically every aspect of travel. Prior studies have found reductions in active and public transport use on rainy days (Guo et al. 2007; Arana et al. 2014; Liu et al. 2015; An et al. 2019; Wu and Liao 2020), hotter conditions negatively influence activities such as walking, cycling and bus ridership (Heinen et al. 2011; Böcker et al. 2013b; Liu et al. 2015; Tao et al. 2018), and air wind speed negatively affects public transit use and cycling (Rietveld and Daniel 2004; Aaheim and Hauge 2005; Guo et al. 2007; An et al. 2019). Adverse weather conditions, such as cold temperatures, snowfall, or precipitation, also impact traffic speed (Keay and Simmonds 2005; Akin et al. 2011; Call 2011; Sabir et al. 2011; Zhao et al. 2012; Hooper et al. 2014) and road safety (Seeherman and Liu 2015; Filomena and Picchio 2024; Wang and Zhang 2024).

The influence of weather conditions on the commuting behavior of workers has been analyzed by Aaheim and Hauge (2005), Cools et al. (2010), and Liu et al. (2015). Aaheim and Hauge (2005) use data for the Bergen area of Norway during March-May 2000 and show that walking and cycling for commuting increase with higher temperatures, while higher precipitation decreases walking and cycling to work, and increases private transport. Cools et al. (2010) conducted a stated adaptation experiment to study how 586 respondents in Flanders, a region of Belgium, respond to weather events, and show that snow has the largest impact on commuting behavior, whereas extreme temperatures, both warn and cold, have the least impact. Finally, Liu et al. (2015) collect travel data from the Swedish National Transport Survey and show that the share of commuting time by bicycle increases substantially in warm months, in contrast to the share of walking. In addition, rain is negatively related to the share of commuting time by bicycle. All in all, such data sources adopted in previous studies involve specific geographical areas within a country during a year (Aaheim and Hauge 2005; Cools et al. 2010) or cannot cover the whole picture of commuting behavior (Liu et al. 2015).1

Considering the predicted, and already observed, increase in the frequency and intensity of extreme weather events, in terms of conditions and temperatures due to climate change (Stott 2016; Intergovernmental Panel on Climate Change IPCC 2023), there is a compelling need to deepen our understanding of the relationship between weather conditions and workers’ daily travel behavior. Within this framework, this paper addresses the question of how weather conditions relate to the daily commuting behavior of workers in the US. To that end, we use nationally representative time diary data from the American Time Use Survey (ATUS) over the period 2003–2023, which contains very rich information on the daily activities performed by respondents, merged with weather data at the daily-county level in the US, taken from a total of 23,664 weather stations located across the country. Thus, our analysis covers a broad geographical region, with wide variations in weather across the country, and in different seasons, for a 20-year period.

Our findings reveal significant associations between weather conditions and daily commuting behavior. We find that high temperatures, increased snowfall, and rainy days are associated with a higher probability of using cars as the primary mode of commuting, while these conditions exhibit a negative relationship with alternative transportation modes such as public transit and walking. These findings suggest that cars are the preferred option for minimizing exposure to unfavorable environmental conditions during commutes. We also find that the relationships between weather conditions and commuting mode choices are nonlinear, with these relationships being particularly pronounced on days characterized by extreme heat, heavy precipitation, and light snowfall. Finally, a heterogeneous analysis highlights regional variations in these relationships, suggesting that workers in historically warmer climates appear better adapted to high temperatures. All these findings suggest that exposure to weather conditions is a key factor driving the results.

This paper makes three important contributions to the literature. First, we provide a comprehensive analysis of the relationship between weather conditions and daily commuting time for US workers.2 Commuting, a compulsory daily activity for many workers, is closely linked to productivity losses, monetary costs, time constraints, and broader environmental and public health issues. Previous research has primarily concentrated on small geographic regions, limited temporal spans, and weather data derived from a restricted number of stations, raising concerns about the generalizability and external validity of the findings. To address these limitations, we use twenty nationally representative time use surveys covering a wide time span (2003–2023) across the whole of the US, with its wide variety of climates, in stark contrast to other countries with more uniform weather patterns.

Second, we examine how specific weather conditions relate to the choice of various commuting modes of transport, focusing on the substitution effects between transportation modes. Unlike most previous studies, which focus on a single travel mode in the context of generalized travel activities due to data limitations (e.g., Sabir et al. 2011; Arana et al. 2014; Nosal and Miranda-Moreno 2014; Tao et al. 2018; Zhao et al. 2018; An et al. 2019; Ngo 2019; Abad et al. 2020; Wu and Liao 2020; Bo et al. 2021; de Kruijf et al. 2021; Jiang and Cai 2023; Ngo and Bashar 2024), our analysis considers multiple commuting mode choices simultaneously. In time-use surveys, such as the ATUS, respondents provide detailed accounts of their daily activities, reporting minute-by-minute time allocation. This includes comprehensive information on the time spent using different modes of transportation during their commute, enabling the construction of precise measures of daily commuting time by distinct transportation modes.

Finally, we analyze the role of adaptation and acclimation to extreme temperatures as a potential determinant of our findings. The adaptation and acclimation hypotheses suggest that the impact of high temperatures may be more pronounced in colder climates, as individuals in historically warmer regions are more accustomed to such conditions. However, to the best of our knowledge, these hypotheses have not yet been thoroughly explored for mobility patterns. In this context, the nationally representative scope of our datasets allows us to explore heterogeneity in the relationship between weather conditions and commuting behavior across diverse climatic regions. Our analysis reveals distinct patterns in the relationship between weather conditions and daily commuting behavior across different climates, indicating that workers in warmer areas are better adapted to extreme heat. These results align with evidence of worker adaptation to local climates and emphasize the need for region-specific climate adaptation policies in the transportation sector to address weather-related commuting responsiveness.

The remainder of the paper is organized as follows. Section 2 describes the data sources, sample selection, and variables used in our analysis, and provides some descriptive analyses. Section 3 presents the econometric strategy. Section 4 discusses the regression results. Section 5 details heterogeneity results. Finally, Section 6 concludes.

2 Data Sources and Variables

2.1 The American Time Use Data

We use nationally representative time diary data from the American Time Use Survey (ATUS) for the period 2003–2023, excluding the year 2020, which is not representative of the whole year.3 The ATUS database is an annual, nationally representative time use survey, considered the official time use survey of the US, sponsored by the Bureau of Labor Statistics (BLS) and conducted continuously since January 2003 by the US Census Bureau. It is the largest source of time diary data collected anywhere and is considered the state-of-the-art in time-use surveys (Aguiar et al. 2012). The respondents (a single individual from a unique household over age 15, previously interviewed in the Current Population Survey (CPS) randomly chosen to answer the questionnaire) fill in a time use diary, where they report their primary activities done for each minute of the 24 h of a single survey day (from 4:00 a.m. on the previous day to 3:59 a.m. on the interview day). Consequently, the ATUS collects one detailed time diary day per household. The advantage of self-reported time diary data over those from recall questions asking respondents about usual time spent or time spent over the last week is that diary-based estimates of time use are more precise and reliable (Bonke 2005; Kan 2008).

The ATUS collects the exact start and stop times of activities, allowing us to define the time devoted to any given activity such as commuting or market work, the main time use categories in our analysis. For most activities, the ATUS also collects information about where the activities took place, the mode of transport, and who, if anyone, was present during the activity, except for activities that are generally done alone, such as sleeping, grooming, and certain other personal activities. Furthermore, the ATUS also collects information about a range of respondent and household characteristics through individual and household interviews.

We restrict the ATUS sample to workers between 16 and 65 years old (inclusive), who completed their diaries on working days, defined as days when respondents report working for at least 1 h, excluding commuting time (Gimenez-Nadal et al. 2018a, 2018b). We omit workers who filled in their diaries during holidays to avoid atypical or unusual days that do not reflect the usual commuting behavior of workers, and we exclude self-employed workers since they are more likely to work from home and generally have different commuting patterns than employees (van Ommeren and van der Straaten 2008; Roberts et al. 2011; Albert et al. 2019; Giménez-Nadal et al. 2018a, 2020, 2024). Finally, we exclude zero-commuters (van Ommeren and van der Straaten 2008). From these restrictions, our final ATUS sample is composed of 52,335 workers from the original 236,357 ATUS sample size for the period between 2003 and 2023, excluding the survey year 2020.

2.2 Weather Data

We integrate data from the ATUS with daily weather records sourced from a total of 23,664 weather stations around the US obtained from the National Climatic Data Center (NCDC) of the National Oceanic and Atmospheric Administration (NOAA), wherein station-level data are averaged to generate county-level measures.4 Utilizing the detailed data provided by the ATUS, which includes the day, month, and year of respondents’ interviews along with their county of residence, allows us to accurately match weather information to the specific county at the time of each interview. Importantly, the county of residence represents the most granular geographic detail available for ATUS respondents, facilitating a precise analysis of localized weather effects (Graff Zivin and Neidell 2014). We link the ATUS sample with the weather information by the county of residence of the respondent and the survey date, resulting in a final sample of just over 28,114 workers with valid weather data for the econometric analysis.

The main weather variables used here include daily total precipitation (in inches), snowfall (in inches), and maximum air temperature (in degrees Fahrenheit, °F). Following Graff Zivin and Neidell (2014), Krüger and Neugart (2018) or Dillender (2021), we focus on maximum temperature, rather than daily average temperature, because most individuals are indoors for a significant period, such as when they are sleeping, when minimum temperatures often occur. Maximum temperature is also likely to be highly correlated with other relevant temperature measurements throughout the day, making it a reasonable proxy for individual exposure. Based on the precipitation data, we define a dummy variable for rainy days that takes value 1 if daily precipitation was at least 0.10 inches, 0 otherwise (Connolly 2008).

It is possible that weather conditions affect the response rate and generate a nonresponse bias. For instance, prior research has examined the degree of nonresponse bias in the ATUS with respect to socio-demographic characteristics (Abraham et al. 2006). A related concern in our study is the potential influence of weather conditions on response rates. However, Appendix Table A.1 demonstrates that there are no large sample biases in our data. Our sample closely aligns with the mean socio-demographic characteristics of individuals reported in the most recent release of the 2020 U.S. Census, though male respondents are slightly oversampled.5 Moreover, mirroring the approach of Graff Zivin and Neidell (2014), Appendix Figure A.1 demonstrates that the temperature distribution for county dates included in our final sample closely mirrors that of the same counties across the entire analysis period, including days without time use data. These findings suggest that any nonresponse bias related to temperature is minimal.

2.3 Variables

We define the time spent commuting per day based on the activity code 180501, labeled as “travel to/from work”, by summing the total minutes allocated to this activity per day. Besides, importantly for our analysis, the time use diaries of the ATUS also report a number of transportation-related mode categories asking the question “Where were you?” (description “Location of activity”). This enables us to construct precise measures of the time workers allocate to commuting via different modes of transport per day, which is particularly valuable for analyzing the choice of commuting mode in our study.

There are twenty-six different locational coding categories, but we focus on the following answers: ‘car, truck, or motorcycle (as driver or passenger)’, ‘walking’, ‘bus’, ‘subway/train’, ‘bicycle’, ‘boat/ferry’, ‘taxi/limousine service’, and ‘airplane’. The travel modes in this analysis are grouped, following Giménez-Nadal and Molina (2019), into the private vehicle (car, truck, or motorcycle, both as driver or passenger), public transit (bus, subway/train, boat/ferry, taxi/limousine service, or airplane), walking, and cycling (bicycle). We calculate the total commuting time using each mode of transport, in minutes per day, as the sum of all episodes reported by each worker throughout the diary day, and then obtain the percentage of daily commuting time done by car, public transit, walking, and cycling.

The ATUS also contains rich individual and household information that allows us to define several variables to control for the socio-demographic characteristics of individuals that have been found to affect the daily time spent commuting, aimed at accounting for the observed heterogeneity of workers, both at the individual and household level. We first consider gender, defined as a dummy that takes values 1 for males, and 0 for females. We also include the age of the respondent, measured in years. The native status is defined through a dummy variable that takes value 1 for native citizens (being a naturalized US citizen), 0 otherwise. The highest educational attainment is separated into three dummy variables indicating if the respondent has achieved primary education, secondary education, or University education. We also take into account the full/part-time status and employment sectors of workers. The full/part-time status is included through a dummy variable that takes value 1 if the employee is a full-time worker and 0 otherwise. We also include a control for the employment sector and define a dummy variable that takes value 1 for employees in the public sector, 0 otherwise. Finally, household composition is controlled by three variables: a dummy variable that takes value 1 for the presence of a partner (either married or cohabiting) in the household, 0 otherwise, and two continuous variables indicating the number of children in the household (aged 17 or under) and the family size, respectively.

2.4 Descriptive Statistics

Table 1 presents the weighted summary statistics of the minutes workers spend commuting per day and the percentage of daily commuting time using each mode of transport, after imposing the above restrictions and merging with the weather data. We use the demographic weights provided by the survey to make the summary statistics nationally representative. The average commuting time in the sample is 45.81 min per day, with a standard deviation of 38.54 min. The car is the preferred mode of transport in our sample, with an average of 93.05% of daily commuting time done by private vehicle, as either the driver or as a passenger. The percentage of time spent commuting by public transit is, on average, 3.12% per day, while the percentage of daily commuting time by walking and bicycle is 3.21% and 0.61%, respectively. Most of our sample choose the car as the primary mode of transportation to commute, with an average of 90.80% of workers doing at least 95% of their daily commuting time by car. For other transportation modes, 1.57%, 2.15%, and 0.54% of workers allocated at least 95% of their daily commuting time to public transit, walking, and bicycling, respectively.6 In terms of weather conditions, the average daily precipitation in our sample is 10.42 inches per day, the percentage of rainy days is 50.43%, the average daily snowfall is 0.54 inches, and the average daily maximum temperature is 67.93 °F, respectively.

Table 1. Summary statistics.

	Mean	Std. Dev.
Dependent variables
Total commuting	45.8089	38.5412
% car	93.0482	24.3059
– % car ≥ 95%	90.7990	28.9045
% public	3.1245	15.9471
– % public ≥ 95%	1.5698	12.4306
% walking	3.2126	15.5991
– % walking ≥ 95%	2.1539	14.5175
% bicycle	0.6147	7.6701
– % bicycle ≥ 95%	0.5364	7.3043
Independent variables
Precipitation (inches)	10.4253	27.8334
Rainy day (Precipitation ≥ 0.1 inches)	0.5043	0.5000
Snowfall (inches)	0.5443	4.1396
Maximum temperature (°F)	67.9354	18.7653
Being male	0.5453	0.4980
Age	39.5379	12.7857
Native citizen	0.7723	0.4193
Primary education	0.0934	0.2910
Secondary education	0.2559	0.4364
University education	0.6507	0.4768
Full-time worker	0.8482	0.3588
Public sector worker	0.1546	0.3615
Live in couple	0.5980	0.4903
Number of children	0.8018	1.1186
Family size	3.1558	1.5443

Notes: Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Statistics computed using ATUS sampling demographic weights.

Table 1 also presents summary statistics of the set of socio-demographic characteristics for our sample. Around 54.5% of the sample are males, and workers are, on average, 39 years old. Further, 77.2% of workers are native citizens born in the US. In terms of the maximum level of education achieved, approximately 9.3% of workers have primary education, 25.6% have secondary education, and 65.1% have attained at least some college. In addition, 84.8% of respondents are full-time workers, and 15.5% are public sector employees. Last, 59.8% of workers have a spouse or partner in the household, the number of children is 0.80, and the average family size is 3 members.

Daily commuting involves significant time spent outdoors, making it inherently exposed to weather conditions. However, these conditions can vary significantly across geographical regions, potentially influencing the dynamics between weather and daily commuting patterns. Specifically, workers in historically warmer areas may be more accustomed to heat and therefore less sensitive to high temperatures, while those in colder regions may experience a greater impact from warmer days. To investigate these regional differences, Table 2 presents summary statistics segmented by two broad geographical regions with distinct climate characteristics.

Table 2. Summary statistics by geographical area.

	Warmer regions		Colder regions
	Mean	Std. Dev.	Mean	Std. Dev.	Difference
Dependent variables
Total commuting	45.3163	37.5986	46.1994	39.2691	−0.8831*
% car	95.5276	19.5212	91.0828	27.3552	4.4448***
% public	1.4136	10.9439	4.4808	18.8896	−3.0672***
% walking	2.3625	13.6100	3.8865	16.9816	−1.5240***
% bicycle	0.6963	8.1588	0.5499	7.2590	0.1464
Weather variables
Precipitation (inches)	8.7559	28.6286	11.7486	27.1146	−2.9927***
Rany day (Precipitation ≥ 0.1 inches)	0.4122	0.4922	0.5773	0.4940	−0.1651***
Snowfall (inches)	0.0634*	1.1214	0.9256	5.4218	−0.8622***
Maximum temperature (°F)	76.5460**	14.0933***	61.1097	19.1958	15.4363***
Individuals	12,432		15,682

Notes: Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working, excluding commuting. Self-employed workers and zero-commuters are excluded. Statistics computed using ATUS sampling demographic weights.
* p < 0.1
** p < 0.05
*** p < 0.01.

To implement this geographical segmentation, we classify warmer regions as counties with an average maximum temperature exceeding the sample's mean maximum temperature during the analysis period (67.173 °F). Conversely, colder regions are identified as counties with an average maximum temperature at or below this threshold during the analysis period (Cosaert et al. 2024a, 2024b).7 This classification provides a useful framework for examining potential heterogeneous relationships between weather conditions and commuting patterns and exploring long-run adaptation in econometric analysis.

Table 2 presents summary statistics for commuting and weather variables, disaggregated by these two broad geographical regions. It also reports mean differences in the average values between regions along with the corresponding p-values from t-tests of equality of means. The results reveal significant regional variations in commuting patterns. Workers residing in warmer regions exhibit a higher prevalence of commuting by car, with a percentage that exceeds that in colder regions by 4.44 percentage points. In contrast, public transit use and walking are more prevalent in colder regions, with these percentages surpassing warmer areas by 3.07 and 1.52 percentage points, respectively. These average differences are statistically significant at standard levels (p < 0.01). As expected, weather variables also show marked disparities between the two regions. Warmer areas experience significantly less total precipitation, with an average reduction of 2.99 inches of precipitation and 16.51 fewer percentage points of rainy days. Snowfall is also significantly lower in warmer regions, averaging 0.86 fewer inches compared to colder regions, on average. Finally, maximum temperatures are considerably higher in warmer regions, with an average increase of 15.44 °F relative to maximum temperatures in colder regions.

3 Econometric Strategy

To examine the relationship between weather conditions and daily commuting time, we estimate the following linear regression model by Ordinary Least Squares (OLS):

C_{i, jt} = α + {Weather'}_{jt} β_{W} + {X'}_{i, jt} β_{X} + {Mode}_{i, jt}^{'} β_{M} + {FE}_{t} + {FE}_{w} + ε_{i, jt}

()

where

i

refers to individual “i”,

j

refers to county “j”, and

t

refers to survey day “t”. The dependent variable,

C_{i, jt}

, represents the daily minutes spent on commuting by individual

i

in county

j

at time

t

{Weather'}_{jt}

is a vector of weather conditions of county

j

at time

t

, including the dummy variable for rainy days, daily snowfall (in inches), and maximum air temperature (°F).

${X'}_{i, jt}$ represents a number of observable characteristics of individual $i$ correlated with commuting time, including gender; age, and age squared to allow for nonlinear effects; native status (1 if native citizen, 0 otherwise); education level (indicators for secondary and University education, ref.: primary education); full-time status (1 if full-time worker, 0 otherwise); type of worker (1 if public sector worker, 0 otherwise); cohabitation status (1 if living with a married/unmarried partner, 0 otherwise); the number of children aged 17 years or less in the household; and the total family size.

Recognizing that commuting behaviors, particularly daily commuting time, are strongly influenced by the mode of transport, we also include a vector denoting the primary commuting mode, represented by ${Mode}_{i, jt}$ . For example, reliance on public transport significantly increases commuting time compared to private vehicle use. This vector comprises four dummy variables, indicating whether at least 95% of the daily commuting time was spent using a car, public transit, walking, or bicycling, respectively. The reference category represents combined trips without a predominant transport mode (i.e., other means).

${FE}_{t}$ are dummies for the year of the interview (2023 is the reference survey year), month (December is the reference month), and day of the week (Sunday is the reference weekday), to capture possible changes in commuting time throughout the week (workdays Monday through Saturday), months (any seasonality in commuting) and years (specific macroeconomic circumstances or survey issues). The vector ${FE}_{w}$ comprises occupation-fixed effects to account for worker characteristics, with Transportation and Material Moving serving as the reference or omitted occupation. Finally, $ε_{i, jt}$ is the error term capturing unmeasured factors in the model.

Besides this, to analyze the relationship between weather conditions and the daily commuting mode choice, we use a multinomial logit (MNL) model to study the probability of choosing a primary mode of transport for commuting. Specifically, we estimate models with the following specification to study the probability that the individual

i

in county

j

at time

t

chose a primary mode of transport

k

for the daily commute:

Pr (M_{i, jt} = k) = {Weather'}_{jt} β_{W}^{k} + {X'}_{i, jt} β_{X}^{k} + {FE}_{t}^{k} + {FE}_{w}^{k} + ε_{i, jt}^{k} k = Car; Public; Walking; Bicycle; Other

()

where in Equation (2) the dependent variable,

M_{i, jt}

, represents the primary mode of transportation chosen

k

. This is a non-ordinal categorical variable that can fall into one of five possible mutually exclusive alternatives (i.e., from 1 to 5), corresponding to the predominant mode of transport during the commuting (Car, Public, Walking, Bicycle, or No predominant commuting mode). Finally,

Pr

represents the probability that worker

i

choose the

k

th alternative,

k = 1,2, \dots, 5

These two distinct specifications permit us to examine the relationship between weather conditions, on the one hand, and the daily commuting time and primary commuting mode choice, on the other. All regressions employ robust standard errors clustered at the state-month level (Neidell et al. 2021; Cosaert et al. 2024a, 2024b) to account for potential heteroscedasticity and address concerns related to spatial and serial correlations across states within a given month and within a state across months (Cameron and Miller 2015).9 Besides, we present estimates of the coefficients of Equation (2) as average marginal effects (AMEs) to clearly identify the effect of each weather variable on the probability of choosing a specific primary mode of transport for the commute.10 Finally, all estimates are weighted at the individual level using survey demographic weights provided by the ATUS.11 Our parameters of interest that relate weather conditions to commuting time, $β_{W}$ , are identified from cross-county and daily variations in weather.

One important limitation of the analysis presented in Equations (1) and (2) is that the data set consists of a cross-section of respondents. This limitation prevents us from controlling for unobserved heterogeneity among workers in our sample, thereby restricting our ability to establish causal relationships. Consequently, the results should be interpreted as conditional correlations reflecting workers’ commuting behavior in response to daily weather conditions, while accounting for observed worker characteristics and time factors. Nonetheless, the analysis leverages highly detailed datasets from time-use surveys, which provide precise information at the regional and survey-date levels. This enables us, for the first time in the literature, to explore the relationship between weather conditions and daily commuting in the US, considering its diverse range of climates.

4 Results

4.1 Baseline Results

Table 3 presents the results of estimating Equation (1) in Column (1) for the relationship between weather conditions and daily commuting time, and Columns (2-5) present the results from estimating Equation (2) for the AMEs of weather conditions on primary commuting mode choices, accounting for the socio-demographic characteristics of workers in our sample and time factors, respectively. We find that the coefficients for weather conditions display no statistically significant effects on the daily time devoted to commuting by US workers, but we do obtain significant relationships between weather conditions and the primary commuting mode choice.

Table 3. Baseline results.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1755	0.0142***	−0.0030	−0.0049*	−0.0017
	(0.5831)	(0.051)	(0.0020)	(0.0026)	(0.0013)
Snowfall	−0.0201	0.0033***	−0.0005	0.0002	−0.0014*
	(0.0494)	(0.0010)	(0.0004)	(0.0002)	(0.0008)
Maximum temperature	0.0067	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0248)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Being male	5.8740***	−0.0110**	−0.0024	0.0052**	0.0052***
	(0.6766)	(0.0053)	(0.0021)	(0.0026)	(0.0014)
Age	0.7020***	0.0006	0.0012	0.0003	0.0003
	(0.1744)	(0.0014)	(0.0006)	(0.0006)	(0.0003)
Age squared	−0.6663***	−0.0002	−0.0014*	−0.0007	−0.0004
	(0.2098)	(0.0017)	(0.0008)	(0.0008)	(0.0004)
Native citizen	−4.8162***	0.0466***	−0.0089***	−0.0115***	0.0007
	(0.8017)	(0.0064)	(0.0025)	(0.0024)	(0.0012)
Secondary education	−0.9051	0.0226**	−0.0094**	−0.0058	−0.0037
	(1.1293)	(0.0093)	(0.0036)	(0.0046)	(0.0027)
University education	1.7366	0.0031	−0.0062*	−0.0100**	−0.0037
	(1.1250)	(0.0094)	(0.0036)	(0.0042)	(0.0027)
Full-time worker	5.0859***	0.0100	0.0007	−0.0192***	−0.0010
	(0.8007)	(0.0073)	(0.0031)	(0.033)	(0.0017)
Public sector worker	−3.8546***	−0.0130**	−0.0011	0.0047	−0.0004
	(0.7634)	(0.0060)	(0.0033)	(0.0034)	(0.0018)
Live in couple	2.2482***	0.0389***	−0.0102***	−0.0115***	−0.0001
	(0.6157)	(0.0061)	(0.0030)	(0.0032)	(0.0015)
Number of children	−1.6185***	0.0034	−0.0004	−0.0010	−0.0007
	(0.5053)	(0.0038)	(0.0015)	(0.0018)	(0.0009)
Family size	1.4144***	0.0060**	0.0000	−0.0052***	−0.0007
	(0.4009)	(0.0030)	(0.0014)	(0.0015)	(0.0006)
At least 95% by car	−25.0962***	—	—	—	—
	(1.6053)
At least 95% by public	12.1748***	—	—	—	—
	(4.2683)
At least 95% walking	−44.7116***	—	—	—	—
	(1.9575)
At least 95% by bicycle	−29.5605***	—	—	—	—
	(3.1005)
Constant	39.1543***	—	—	—	—
	(4.7084)
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,114	28,114	28,114	28,114	28,114
(Pseudo) R-squared	0.1009	0.0686	0.0686	0.0686	0.0686

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y determines whether or not the specifications include controls for fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

For the primary commuting mode of transport, Table 3 reveals that at higher temperatures and during rainy days and days with higher snowfall, workers seem to shift towards the private car, against alternative transportation modes such as public transit, walking, and bicycling. Ultimately, rising temperatures, snowfall, and rainy days are associated with a greater predominance of the car as the primary mode of transport for the commute, which suggests that workers tend to seek greater comfort in their daily commuting mode choices and try to minimize their exposure to uncomfortable weather conditions during the commute.

Regarding rainy days, in the first row of Table 3 we find that rainy days are positively related to the probability of using the car as the primary mode of transport for the commute, whereas the relationship is found to be negative for walking. Specifically, a rainy day is associated with an increase of 1.42 percentage points in the probability of using the car as the primary mode choice for the commute, whereas it is related to a decrease of 0.49 percentage points in the probability of choosing walking as the primary mode of transport for the commute (p = 0.061). As a result, on rainy days, American workers adjust their commuting mode choices, shifting from walking to private car usage.

Turning to other weather variables, we also observe that daily snowfall is positively associated with the probability of using the car as the main mode of transport to commute, whereas it is negatively related to the probability of using the bicycle as the main mode of transport. The coefficients displayed suggest that each additional inch of snowfall is associated with an increase of 0.33 percentage points in the probability of using the car as the primary commuting mode and a decrease of 0.14 percentage points for the bicycle (p = 0.097).

Finally, with respect to daily maximum temperature, we find that each additional degree Fahrenheit in maximum temperature is associated with a 0.13 percentage point increase in the probability of using the car as the primary mode of commuting. Conversely, each additional degree Fahrenheit of maximum temperature is linked to decreases of 0.04 and 0.03 percentage points in the probability of using public transit and walking, respectively, as primary commuting mode choices. These results suggest that workers substitute their primary commuting mode choices, from public transit and walking to the car, as daily temperatures rise.

All in all, the observed patterns suggest that rainy days, snowfall, and high temperatures are associated with a decreased probability of choosing public transit, walking, and bicycling as primary commuting modes. Conversely, these weather conditions are positively correlated with an increased probability of using private vehicles as the primary mode of transportation among workers. This observation aligns with the hypothesis that private vehicles are an optimal travel choice for minimizing exposure to adverse weather conditions, as they offer a controlled microclimate, increased privacy, and enhanced security, making them particularly appealing for protection against inclement weather (Böcker et al. 2013b; Gatersleben 2014; Graff Zivin and Neidell 2014; Böcker et al. 2015).

We must emphasize that the R-squared statistics shown in all these models are quite low. All this suggests that many non-observable characteristics, such as attitudes towards commuting and working from home, or distance between home and workplace, may condition commuting behaviors. Within this context, many studies have reported similar explanatory power in their estimates using the ATUS (Gimenez-Nadal et al. 2018a, 2018b) or other datasets (van Ommeren and van der Straaten 2008; Gimenez-Nadal et al. 2022a; Echeverría et al. 2022). Thus, our results are in line with prior literature from this point of view and suggest that several aspects of commuting time require further research attention.

4.2 Nonlinear Relationships

We extend our econometric specifications to investigate potential nonlinear relationships between weather conditions and daily commuting patterns, incorporating precipitation, snowfall, and temperature bins to capture these dynamics. While our initial analysis focused on linear associations, it is important to consider the possibility of nonlinear relationships between weather and daily travel behaviors, as suggested by prior research (Phung and Rose 2008; Miranda-Moreno and Nosal 2011; Wu and Liao 2020; de Kruijf et al. 2021).

To capture potential non-linearities in the relationships between weather conditions and daily commuting, we adopt a flexible approach by including a set of dummy variables for distinct weather bins, covering the full distribution of our weather variables rather than their continuous values. Specifically, we control for varying levels of precipitation, snowfall, and maximum temperature, using weather categories aligned with prior research in the US context (Connolly 2013; Graff Zivin and Neidell 2014; Dillender 2021; Jiao et al. 2021; Liu and Hirsch 2021). This methodology strengthens the robustness of our analysis and offers a more nuanced understanding of the complex relationships between weather conditions and daily commuting behavior.

To further explore the connection between daily temperatures and commuting, we replicate the works of Graff Zivin and Neidell (2014) and Jiao et al. (2021). We modify previous Equations (1) and (2) by adding a series of indicators for every 5-degree temperature increment, rather than relying on the daily maximum air temperature. Specifically, the coldest temperature bin covers maximum temperatures equal to or below 30 °F, whereas the hottest bin covers temperatures above 100 °F. We then define 16 dummy variables that take value 1 if the daily maximum temperature in county $j$ at time $t$ is within this range and 0 otherwise, including these explanatory variables in the models previously estimated, plus the snowfall and precipitation variables. For comparability with existing research using the ATUS (Connolly 2013; Graff Zivin and Neidell 2014; Jiao et al. 2021) or German time-use data (Krüger and Neugart 2018), we omit the 76–80 °F temperature bin, which involves temperature levels where most humans feel comfortable (Picchio and Van Ours 2024).

Table 4 presents the results derived from estimating Equations (1) and (2) for these temperature bins and indicates that the relationship between temperature and daily commuting mode choice is nonlinear. Specifically, lower temperatures are negatively (positively) related to the probability of using the car and bicycle (public transit and walking) as the primary mode of transport for the commute, whereas higher temperatures are related to a greater (lower) predominance of the car (walking). This suggests that workers find it easier to protect themselves from low temperatures when using these modes of travel, whereas shielding against high temperatures often requires minimizing exposure altogether. Quantitatively, a daily maximum temperature exceeding 100 °F, defined as an extremely hot day in our study, significantly increases the probability of using a car as the primary commuting mode by 5.56 percentage points compared to a day with a maximum temperature in the 76–80 °F range. In contrast, an extremely cold day corresponds to a 1.42 percentage point increase in the probability of using public transit and a 7.65 percentage point decrease in the probability of using a bicycle as the primary commuting mode, relative to a day with a maximum temperature of 76–80 °F.

Table 4. Additional results: maximum temperature.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1752	0.0151***	−0.0035*	−0.0045*	−0.0022
	(0.5834)	(0.0052)	(0.0021)	(0.0027)	(0.0014)
Snowfall	−0.0129	0.0024***	−0.0004	0.0001	−0.0006
	(0.0528)	(0.0007)	(0.0003)	(0.0002)	(0.0005)
Maximum temperature (≤ 30 °F)	−1.2543	0.0209	0.0142**	0.0100	−0.0765***
	(2.0569)	(0.0172)	(0.0063)	(0.0074)	(0.0086)
Maximum temperature (31–35 °F)	0.1763	0.0058	0.0268***	0.0280***	−0.0777***
	(2.3067)	(0.0205)	(0.0067)	(0.0092)	(0.0087)
Maximum temperature (36–40 °F)	−2.7358	−0.0312*	0.0216***	0.0051	0.0015
	(1.8477)	(0.0168)	(0.0068)	(0.0081)	(0.0039)
Maximum temperature (41–45 °F)	−2.1297	−0.0489***	0.0257***	0.0087	−0.0030
	(1.8148)	(0.0143)	(0.0061)	(0.0056)	(0.0038)
Maximum temperature (46–50 °F)	−1.4646	−0.0402***	0.0154***	0.0027	−0.0004
	(1.6784)	(0.0153)	(0.0059)	(0.0075)	(0.0029)
Maximum temperature (51–55 °F)	−0.0820	−0.0205*	0.0124**	0.0040	−0.0001
	(1.9292)	(0.0124)	(0.0055)	(0.0053)	(0.0040)
Maximum temperature (56–60 °F)	−0.1310	−0.0265**	0.0122**	0.0042	0.0016
	(1.5839)	(0.0112)	(0.0051)	(0.0054)	(0.0031)
Maximum temperature (61–65 °F)	−2.1153	−0.0174	0.0116**	−0.0035	0.0029
	(1.4729)	(0.0115)	(0.0049)	(0.0051)	(0.0025)
Maximum temperature (66–70 °F)	1.6768	−0.0195*	0.0064	0.0028	0.0036
	(1.3762)	(0.0102)	(0.0044)	(0.0046)	(0.0024)
Maximum temperature (71–75 °F)	−1.0547	0.0082	0.0017	0.0030	0.0014
	(1.2840)	(0.0109)	(0.0053)	(0.0044)	(0.0022)
Maximum temperature (81–85 °F)	−2.0631*	0.0090	0.0046	−0.0024	−0.0044*
	(1.1210)	(0.0103)	(0.0045)	(0.0049)	(0.0026)
Maximum temperature (86–90 °F)	−0.5977	0.0208**	0.0024	−0.0044	−0.0029
	(1.2910)	(0.0105)	(0.0045)	(0.0047)	(0.0023)
Maximum temperature (91–95 °F)	−1.1022	0.0392***	−0.0012	−0.0119**	0.0000
	(1.3757)	(0.0142)	(0.0058)	(0.0057)	(0.0021)
Maximum temperature (96–100 °F)	−1.3642	0.0463**	−0.0116	−0.0101	−0.0027
	(3.3757)	(0.0216)	(0.0102)	(0.0106)	(0.0044)
Maximum temperature (> 100 °F)	0.6768	0.0556**	−0.0060	−0.0399*	−0.0023
	(1.6729)	(0.0261)	(0.0084)	(0.0218)	(0.0054)
Socio-demographics	Y	Y	Y	Y	Y
Mode choice	Y	N	N	N	N
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,114	28,114	28,114	28,114	28,114
(Pseudo) R-squared	0.1017	0.0741	0.0741	0.0741	0.0741

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode choice, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects. Full model results are available from the authors on request.
* p < 0.1
** p < 0.05
*** p < 0.01.

In Table 5, we show the results of including additional regressors in the main specifications by focusing here on the relationship between precipitation and daily commuting. We conduct additional analyses regarding precipitation, using a precipitation-bin approach that allows for a more flexible specification of precipitation. Specifically, we incorporate indicators for various levels of rainfall, with “no rain” serving as the reference or omitted category. The categories include 0–0.1 inches (considered as light rainfall), 0.1–0.5 inches, 0.5–1 inches, 1–2 inches, and > 2 inches (considered as heavy rainfall), ensuring ease of comparability with Connolly (2013) and Dillender (2021). The results indicate that heavy rainy days are associated with a 1.66 percentage point increase in the probability of choosing the car as the primary mode of transport, while the probability of using the bicycle decreases by 0.37 percentage points.

Table 5. Additional results: rainy days.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day (0–0.1 inches)	−1.1124	0.0227	−0.0105*	−0.0005	−0.0059*
	(1.3664)	(0.0140)	(0.0055)	(0.0081)	(0.0031)
Rainy day (0.1–0.5 inches)	−0.9489	0.0197*	−0.0077*	−0.0060	−0.0033
	(1.0433)	(0.0106)	(0.0043)	(0.0046)	(0.0026)
Rainy day (0.5–1 inches)	−1.0013	0.0124	−0.0079*	−0.0039	0.0012
	(1.2905)	(0.0114)	(0.0042)	(0.0062)	(0.0025)
Rainy day (1–2 inches)	−2.1643	0.0123	0.0005	−0.0038	0.0034
	(1.4034)	(0.0117)	(0.0053)	(0.0063)	(0.0022)
Rainy day (more than 2 inches)	0.2244	0.0166***	−0.0030	−0.0050*	−0.0037**
	(0.7113)	(0.0055)	(0.0021)	(0.0029)	(0.0016)
Snowfall	−0.0334	0.0035***	−0.0006	0.0002	−0.0015
	(0.0498)	(0.0012)	(0.0004)	(0.0002)	(0.0011)
Maximum temperature	0.0068	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0248)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Socio-demographics	Y	Y	Y	Y	Y
Mode choice	Y	N	N	N	N
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,114	28,114	28,114	28,114	28,114
(Pseudo) R-squared	0.1011	0.0701	0.0701	0.0701	0.0701

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode choice, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects. Full model results are available from the authors on request.
* p < 0.1
** p < 0.05
*** p < 0.01.

Finally, we focus on the relationship between snowfall and commuting behavior by categorizing snowfall into 6 bins, similar to the approach for rain. Rather than analyzing the actual snowfall amounts, we use days with no snowfall as the reference category, following the approach employed by Liu and Hirsch (2021) with the CPS data set over the 2004–2014 period. The results, summarized in Table 6, reveal that on light snowfall days (less than 0.1 inches), the probability of using a car as the primary commuting mode significantly increases by 8.88 percentage points while reducing the probability of using a bicycle by 8.29 percentage points. Similarly, snowfall between 0.5 and 1 inch is associated with an 8.76 percentage point rise in car usage and an 8.21 percentage point decline in bicycle usage. On extreme snowfall days, defined as those with more than 2 inches of snow, car usage increases by 2.88 percentage points, while bicycle usage decreases by 1.34 percentage points.12

Table 6. Additional results: snowfall.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1911	0.0138***	−0.0027	−0.0054*	−0.0018
	(0.5892)	(0.0052)	(0.0020)	(0.0028)	(0.0013)
Snowfall (0–0.1 inches)	−4.1286	0.0888**	−0.0168	−0.0077	−0.0829***
	(3.5079)	(0.0438)	(0.0156)	(0.0208)	(0.0094)
Snowfall (0.1–0.5 inches)	2.1318	0.0379	−0.0067***	0.0089	−0.0070
	(3.6076)	(0.0278)	(0.007)	(0.0093)	(0.0046)
Snowfall (0.5–1 inches)	−0.7268	0.0876***	−0.0033	0.0138	−0.0821***
	(2.3748)	(0.0252)	(0.0107)	(0.0131)	(0.0092)
Snowfall (1–2 inches)	−1.4264	0.0346	−0.0070	−0.0125	0.0017
	(2.1575)	(0.0231)	(0.0094)	(0.0116)	(0.0056)
Snowfall (more than 2 inches)	0.2589	0.0288**	−0.0078	0.0081	−0.0134**
	(1.4575)	(0.0146)	(0.0058)	(0.0078)	(0.0058)
Maximum temperature	0.0084	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0256)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Socio-demographics	Y	Y	Y	Y	Y
Mode choice	Y	N	N	N	N
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,114	28,114	28,114	28,114	28,114
(Pseudo) R-squared	0.1009	0.0692	0.0692	0.0692	0.0692

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode choice, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects. Full model results are available from the authors on request.
* p < 0.1
** p < 0.05
*** p < 0.01.

In summary, these additional results emphasize the nonlinear relationships between weather conditions and daily commuting mode choices, consistent with previous research on general time allocation (Graff Zivin and Neidell 2014; Liu and Hirsch 2021; Jiao et al. 2021). Extreme temperatures—both hot and cold—along with light and heavy snowfall days and heavy rain days, significantly influence commuting mode choices, highlighting the complexity of the relationship between weather conditions and commuting behaviors. Specifically, extreme heat, heavy rainfall, and substantial snowfall are associated with an increased probability of private car usage. In contrast, lower maximum temperatures are associated with a higher probability of using public transit as the primary commuting mode, whereas bicycles show a negative relationship with extremely cold days. Furthermore, light snowfall days are linked to a reduction in the probability of choosing the bicycle for commuting, while simultaneously increasing the probability of using the car as the primary mode of transport.

4.3 Robustness Checks

In this subsection, we assess the robustness of our work through a comprehensive examination of various sample constraints, methods of estimation, and model specifications. Initially, we exclude part-time workers and public workers from the final sample, providing analogous results for the distinct subsamples comprising full-time workers and private workers in Tables A.2 and A.3. We also include self-employed workers, together with a dummy variable for self-employed workers, and find analogous results in Table A.4. In another iteration, we replace the daily maximum temperature with the average temperature (computed as the mean of the daily minimum and maximum temperatures), resulting in similar outcomes in Table A.5. However, the outcomes indicate that the impact is most pronounced for maximum temperatures, implying that temperature during commuting hours is more influential than nighttime temperatures, as expected.

The COVID-19 pandemic dramatically changed the daily travel behavior of individuals, so we exclude the post-COVID-19 years (2021–2023) and find similar results, which are detailed in Table A.6. Table A.7 also shows results of estimating Equations (1) and (2) after including the 2020 survey wave. Finally, we focus on the percentage of daily commuting time done by each mode of transport and regress it using the Seemingly Unrelated Regression (SUR) estimator in Table A.8, to properly address substitution effects across modes of transport and account for the dependence across these compositional variables. We also evaluate the robustness of our results to alternative estimators and use a Poisson estimator for Equation (1), after including zero-commuters. The results, reported in Table A.9, suggest consistency too.13

As a concluding measure, considering our analysis has primarily concentrated on contemporaneous relationships, we introduce lagged weather conditions into our model. Surprisingly, coefficients on lagged weather conditions prove to be insignificant, while coefficients on contemporaneous weather conditions remain largely unchanged. This suggests that workers do not strategically plan their commute based on past weather conditions. For those interested, detailed results of these robustness checks are available upon request from the authors.

5 Heterogeneity and Long-Run Adaptation

We have primarily examined the overall relationship between daily weather conditions and commuting. However, Table 2 highlights spatial heterogeneity in commuting and weather variables between historically warmer and colder areas, which may influence the estimated relationships due to geographical differences in climate. Specifically, we observe heterogeneity among commuters in warmer versus colder regions, suggesting that long-term adaptation to local climate could be a potential channel behind our estimates.

According to adaptation and acclimation hypotheses, the observed relationships between extreme high temperatures and commuting patterns are expected to be more pronounced in colder areas, where such temperatures are less frequent. Conversely, people in warmer areas have had extended periods to adapt to these conditions. To investigate these long-term adjustments, we estimate Equations (1) and (2) separately for warmer and colder regions.

Table 7 shows the estimates for these two subgroups of the population to probe into the potential heterogeneity in the estimated relationships between daily weather conditions and commuting mode choice across geographical areas, focusing on distinct climate regions. We observe distinct relationships between weather and commuting mode choice depending on the local climate, as anticipated. Furthermore, our findings support the adaptation and acclimation hypotheses, with commuting mode choices showing only responses to high temperatures in colder regions.

Table 7. Heterogeneity results by climatic regions.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	Warmer region					Colder region
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Rainy day	−0.3677	0.0017	0.0017	−0.0012	−0.0018	−0.0041	0.0259***	−0.0078**	−0.0073**	−0.0023
	(0.7544)	(0.0057)	(0.0017)	(0.0037)	(0.0022)	(0.8425)	(0.0077)	(0.0032)	(0.0035)	(0.0014)
Snowfall	0.0751	0.0041	0.0004	−0.0005	0.0000	−0.0073	0.0024***	−0.0005	0.0002	−0.0006
	(0.1920)	(0.0033)	(0.0004)	(0.0006)	(0.0008)	(0.0531)	(0.0008)	(0.0004)	(0.0002)	(0.0005)
Maximum temperature (≤30 °F)	−7.5018	0.9914***	−0.1061***	−0.2443***	−0.0950***	−2.3443	0.0614***	−0.0073	0.0039	−0.0630***
	(8.1958)	(0.0562)	(0.0119)	(0.0248)	(0.0128)	(2.8966)	(0.0235)	(0.0095)	(0.0110)	(0.0108)
Maximum temperature (31–35 °F)	−3.9490	0.0695	0.0120	0.0243*	−0.0983***	−0.8848	0.0405	0.0113	0.0245*	−0.0637***
	(4.4441)	(0.0440)	(0.0082)	(0.0142)	(0.0124)	(3.0945)	(0.0265)	(0.0098)	(0.0129)	(0.0109)
Maximum temperature (36–40 °F)	−0.3067	0.2370***	−0.1094***	0.0128	−0.0962***	−3.8647	0.0171	0.0062	−0.0026	0.0066
	(8.9579)	(0.0409)	(0.0122)	(0.0136)	(0.0129)	(2.5970)	(0.0218)	(0.0092)	(0.0110)	(0.0049)
Maximum temperature (41–45 °F)	−5.5938	0.2104***	−0.1114***	0.0140	−0.0978***	−2.5672	−0.0134	0.0145*	0.0014	0.0021
	(3.7747)	(0.0284)	(0.0120)	(0.0105)	(0.0122)	(2.5553)	(0.0188)	(0.0087)	(0.0094)	(0.0031)
Maximum temperature (46–50 °F)	−5.2597**	−0.0068	0.0090	−0.0010	−0.0019	−1.2935	−0.0075	0.0016	−0.0036	0.0048
	(2.6361)	(0.0225)	(0.0067)	(0.0115)	(0.0047)	(2.2831)	(0.0203)	(0.0087)	(0.0111)	(0.0048)
Maximum temperature (51–55 °F)	−2.3931	−0.0004	0.0125***	−0.0043	0.0040	0.5973	0.0099	−0.0027	−0.0012	0.0018
	(3.3875)	(0.0166)	(0.0045)	(0.0085)	(0.0057)	(2.6881)	(0.0175)	(0.0087)	(0.0078)	(0.0054)
Maximum temperature (56–60 °F)	−1.1649	−0.0002	0.0039	0.0094	−0.0046	0.1227	−0.0078	0.0044	−0.0084	0.0062
	(2.3542)	(0.0123)	(0.0039)	(0.0058)	(0.0053)	(2.3082)	(0.0173)	(0.0082)	(0.0091)	(0.0041)
Maximum temperature (61–65 °F)	−2.6229	0.0163	−0.0029	−0.0074	0.0048*	−1.6404	−0.0175	0.0089	−0.0057	0.0034
	(1.6146)	(0.0141)	(0.0052)	(0.0069)	(0.0029)	(2.3477)	(0.0183)	(0.0075)	(0.0090)	(0.0036)
Maximum temperature (66–70 °F)	1.8714	−0.0112	0.0105***	0.0013	0.0038	1.5401	−0.0092	−0.0069	0.0005	0.0050
	(1.8440)	(0.0112)	(0.0036)	(0.0055)	(0.0026)	(2.1439)	(0.0165)	(0.0076)	(0.0081)	(0.0035)
Maximum temperature (71–75 °F)	−0.6381	0.0089	0.0044	0.0046	0.0017	−0.9912	0.0112	−0.0042	−0.0005	0.0027
	(1.8687)	(0.0138)	(0.0043)	(0.0043)	(0.0039)	(1.7976)	(0.0151)	(0.0087)	(0.0079)	(0.0030)
Maximum temperature (81–85 °F)	−2.4905*	0.0197*	0.0078**	−0.0072	−0.0082*	−1.4077	−0.0074	0.0028	0.0025	−0.0031
	(1.4317)	(0.0110)	(0.0037)	(0.0059)	(0.0045)	(1.6251)	(0.0155)	(0.0080)	(0.0073)	(0.0029)
Maximum temperature (86–90 °F)	−0.7941	0.0249**	0.0037	−0.0060	−0.0019	−0.5300	−0.0143	0.0082	0.0024	−0.0045
	(1.5824)	(0.0112)	(0.0039)	(0.0053)	(0.0035)	(1.9545)	(0.0157)	(0.0075)	(0.0082)	(0.0034)
Maximum temperature (91–95 °F)	−2.3677	0.0384***	−0.0030	−0.0158**	0.0013	2.5615	−0.0363*	0.0208*	0.0089	−0.0027
	(1.7471)	(0.0120)	(0.0039)	(0.0062)	(0.0026)	(3.2263)	(0.0190)	(0.0106)	(0.0097)	(0.0045)
Maximum temperature (96–100 °F)	0.5774	0.0148	−0.0069	−0.0021	−0.0008	−15.5610***	0.1144**	0.0120	−0.0257	−0.0751***
	(3.5056)	(0.0187)	(0.0062)	(0.0087)	(0.0052)	(4.0407)	(0.0475)	(0.0200)	(0.0221)	(0.0110)
Maximum temperature (> 100 °F)	0.4550	0.0208	0.0019	−0.0271	−0.0010	−6.9123	1.6269***	−0.3007***	−0.3695***	−0.0660***
	(1.9575)	(0.0181)	(0.0051)	(0.0171)	(0.0069)	(9.2874)	(0.1155)	(0.0356)	(0.0353)	(0.0104)
Socio-demographics	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Mode choice	Y	N	N	N	N	Y	N	N	N	N
Occupation F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Observations (workers)	12,432	12,432	12,432	12,432	12,432	15,682	15,682	15,682	15,682	15,682
(Pseudo) R-squared	0.0912	0.1116	0.1116	0.1116	0.1116	0.1184	0.0980	0.0980	0.0980	0.0980

Notes: OLS estimates on Columns (1) and (6), MNL estimates on Columns (2–5) and (7–10). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking and cycling as the primary commuting mode choice, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects. Full model results are available from the authors on request.
* p < 0.1
** p < 0.05
*** p < 0.01.

The results show that very hot places seem to be better adapted to higher temperatures, due to the greater frequency of these events. Specifically, we find that workers living in colder regions are more sensitive to warmer temperatures, in contrast to those residing in warmer regions, who may be able to acclimate to high temperatures. Conversely, workers in warmer areas are more responsive to cold temperatures. These findings are in line with those obtained in Graff Zivin and Neidell (2014), Barreca et al. (2016), Behrer and Park (2017), Alberto et al. (2021), Dillender (2021), Heutel et al. (2021), Jiao et al. (2021), Johnston et al. (2021), Nguyen et al. (2021, 2023) and Hua et al. (2023), in that they are consistent with the idea of long-run adaptation and acclimatization to local climate outlined above. On the other hand, rainy days and snowfall are only related to the commuting mode choice in colder places.14

6 Conclusions

In this paper, we examine how weather conditions relate to the daily commuting time and commuting mode choices of US workers, using weather and time use data for 20 entire years, over the period 2003–2023. Previous research has focused on generalized travel patterns, specific regions, or particular modes of transport, resulting in an incomplete understanding of the response to weather conditions among commuters. Within this framework, our study represents the first investigation into the relationship between commuting time, mode choices, and weather conditions using time use surveys on a nationally representative setting.

Our findings indicate that precipitation, snowfall, and heat are significantly associated with commuting mode choices. Specifically, these weather conditions are associated with an increased reliance on cars as the primary mode of transport, while they simultaneously reduce the probability of choosing public transport, walking or bicycling. This suggests that workers adjust their commuting patterns to prioritize comfort and reduce exposure to adverse weather, such as high temperatures, snowfall, or rainy days, by opting for private car use. Furthermore, we identify a nonlinear relationship between weather conditions and daily commuting mode choices. In particular, extreme high temperatures, light snowfall, and high precipitation are associated with a shift towards car usage, rather than alternative modes of travel such as public transit, walking, or cycling. Finally, workers in hotter regions appear to be better adapted to high temperatures, and the relationship between extremely hot temperatures and commuting mode choice is only observed in colder regions, and not in warmer areas. This suggests that commuting mode choices in colder regions may be more susceptible to high temperatures.

This study is not without limitations, which could be an inspiration for further research. Our data involve a cross-section of individuals, which constrains our results to conditional correlations subject to unobserved worker characteristics. Although weather is an external variable that cannot be influenced by individuals, at least in the short term, we acknowledge that our estimates are subject to unobserved heterogeneity, meaning that we cannot infer any causal interpretation from our results. For example, individuals may move within the country and select their residential location based on climatic conditions (Rappaport 2007), which can ultimately influence their commuting time and weather exposure. For the purposes of this manuscript, we are testing the relationship between weather conditions and daily commuting, and the estimates refer to short-term responses. We are not estimating long-term responses to weather conditions, such as the potential for individuals to move closer to their jobs if temperatures consistently rise.

One way to overcome this bias to some extent is to use panel data to solve the endogeneity issue caused by unobserved time-invariant heterogeneity across workers. Nevertheless, to the best of our knowledge, we are not aware of any other time-use survey that contains longitudinal data on commuting time, nor any such detailed information at the geographical scale. Similarly, the ATUS is a nationally representative time-use survey in the US. As a result, our analytical results do not automatically apply to other locations characterized by different climates or dominant modes of transport (e.g., cycling in the Netherlands, Denmark, or Germany; walking in Romania or Bulgaria; and public transit in Latin American countries or China). Besides, the time use data allows us to establish a relationship between weather conditions and commuting, but other data are needed to explore the underlying mechanisms, so it is therefore left for further analysis. Finally, future research should explore the relationship between weather conditions and more specific fuel types in private vehicles, with the appropriate data set. As the transition from conventional vehicles to electric vehicles remains in its early stages, it offers unique environmental implications (De Vos 2024).

Despite its limitations, this study addresses a critical research question and provides valuable insights for transportation planning in the US. In the current context of global warming, marked by the increasing frequency and intensity of extreme weather events, milder winters, and hotter summers, understanding the transportation implications of climate change is essential for improving planning and policy. By analyzing the relationship between weather conditions and daily commuting behavior, this research underscores the need for adaptive strategies to prepare for and mitigate the potential impacts of evolving climatic conditions on commuting. Our findings indicate that adverse weather conditions, including precipitation, snowfall, and extreme heat, lead to greater reliance on cars for commuting while reducing the choice of alternative modes of transport. These shifts contribute to heightened congestion in transportation systems, emphasizing the urgency of incorporating climate resilience into transport planning.

Although alternative modes of travel are more vulnerable to weather conditions than private cars, transportation planners must focus on implementing targeted measures to influence daily commuting behavior and incentivize the use of alternative travel modes. Since weather itself cannot be controlled, these efforts should prioritize creating resilient and appealing transportation options that encourage shifts away from car dependency, even in the face of adverse weather conditions. For example, enhancing infrastructure by installing ventilation systems, air conditioning, covered pathways, and insulation. Additionally, improving the timing and reliability of public transit services can reduce commuters’ exposure to outdoor weather conditions and increase the predictability of these modes. Policymakers might also consider adopting dynamic transit schedules informed by weather forecasts, increasing service frequency during adverse weather events to reduce overcrowding, and upgrading transit infrastructure such as bus shelters and access points to offer better protection against inclement weather. These improvements can make alternative modes of transport more attractive and accessible.

Similarly, the rise in telework adoption following the COVID-19 pandemic offers an opportunity to reduce weather-related exposure during commutes and enhance adaptation to climate change. Employers and policymakers should consider expanding working-from-home policies and providing schedule flexibility during extreme weather events, where feasible. Finally, public campaigns could promote the safety and convenience of alternative modes of transport. When implementing these strategies, planners should consider the regional context to address regional differences, as individuals appear to be more sensitive to unfamiliar weather events.

Acknowledgments

We are grateful to the Editor, Edward Coulson, and two anonymous reviewers of this journal for their invaluable comments and suggestions, which greatly contributed to improving the manuscript. We further thank participants at the 47th Symposium of the Spanish Economic Association (SAEe 2022, Valencia) for their comments and suggestions on an earlier version of this paper.

Conflicts of Interest

The authors declare no conflicts of interest.

Endnotes

1 Liu et al. (2015) limit their analysis to total travel time and the shares of walking and cycling observed in their data set. On the other hand, Aaheim and Hauge (2005) only use data from a single weather station.

2 In general, we contribute to the literature on time use and weather conditions (Connolly 2008; Graff Zivin and Neidell 2014; Krüger and Neugart 2018; Garg et al. 2020; Alberto et al. 2021; Jiao et al. 2021; Liu and Hirsch 2021; Neidell et al. 2021; Nguyen et al. 2021, 2023; Hajdu 2024).

3 The survey year 2020 is not representative of the whole year, as the survey procedure was temporally suspended from mid-March to mid-May 2020 due to the COVID-19 pandemic. Consequently, the main sample excludes that year. In a robustness check, we report the findings of a sample that includes part of the 2020 data.

4 Data is available at https://www.ncdc.noaa.gov/cdo-web/#t=secondTabLink (accessed in November 2024).

5 This discrepacy of our sample mimics recent research using the ATUS to analyze the daily time spent commuting, which indicates that men comprise slightly over half of the sample (Belloc et al. 2024).

6 These latter percentages do not add up to 100 due to other trips with no predominant mode of transport (4.94% in our sample, on average).

7 We use the information from counties included in the sample during the entire observation period (2003-2023, excluding 2020 year), regardless of time use data availability, to group counties into colder and warmer regions.

8 We opt for a multinomial logistic regression model to estimate the mode choice directly, so we focus on the primary mode choice rather than the percentage of commuting time spent in each mode.

9 The standard errors in the main analysis are clustered at the state-month level, as recommended by Neidell et al. (2021). This approach accounts for potential within-cluster correlation and heteroscedasticity. Additionally, we verified the robustness of our results to alternative clustering procedures, such as clustering at the county level or using two-way clustering at the state and month levels. The findings are qualitatively similar under these specifications.

10 In the estimates, we do not report the AMEs for the residual category for the sake of simplicity.

11 These weights account for specific elements of the survey design, including the intentional oversampling of certain demographic groups (e.g., households with a Hispanic or non-Hispanic Black householder, or households with children) and days of the week (e.g., 50% of the sample reports on Saturdays and Sundays). The weights also adjust for nonresponse rates across demographic groups and days of the week.

12 The marginal effects from the MNL estimates do not necessarily sum up to zero because of the omission of the AMEs for the category denoting no primary mode choice.

13 These coefficients should be interpreted in terms of percentage changes in the dependent variable related to a specific weather condition, as the Poisson estimator imposes a log-transformation in the dependent variables. Consequently, the estimates for non-log-transformed independent variables refer to semi-elasticities and should be multiplied by 100 to obtain the percentage change in daily commuting time.

14 The lack of a relationship between snowfall and commuting mode choice in warmer regions may be attributed to the near absence of snowfall in these areas. Besides, if we use fourteen-degree temperature bins and set the hottest (coldest) bin to maximum temperatures higher than 95 °F (equal to or lower than 35 °F), to improve precision due to limited number of observations of warm (cold) temperatures in historically cold (warmer) counties, we only find statistically significant estimates for extremely hot temperatures in colder areas. These results are displayed in Table A.10.

Appendix

(See Tables A1–A10 and Figure A1)

Table A.1. Comparison of sample with 2020 US census.

Socio-demographic characteristics	Sample		2020 US Census
Socio-demographic characteristics	Mean	Std. Dev.	2020 US Census
Being male	0.545	0.498	0.491
Age	39.538	12.786	39.200
Native citizen	0.772	0.419	0.757
Primary education	0.093	0.291	0.103
Secondary education	0.256	0.436	0.259
University education	0.651	0.477	0.638
Full-time worker	0.848	0.359	0.891
Public sector worker	0.155	0.362	0.144
Living in couple	0.598	0.490	0.583
Number of children	0.802	1.119	—
Family size	3.156	1.544	3.090
Number of individuals	28,114

Notes: Data from the 2020 US Census is taken from https://data.census.gov/profile (accessed in November 2024).

Table A.2. Robustness check: Excluding part-time workers.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1207	0.0145***	−0.0034	−0.0049*	−0.0022
	(0.6615)	(0.0052)	(0.0021)	(0.0025)	(0.0015)
Snowfall	−0.0054	0.0031***	−0.0006	0.0002	−0.0011
	(0.0529)	(0.0009)	(0.0004)	(0.0002)	(0.0007)
Maximum temperature	0.0094	0.0014***	−0.0004***	−0.0004***	−0.0001
	(0.0284)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	23,994	23,994	23,994	23,994	23,994
(Pseudo) R-squared	0.0873	0.0757	0.0757	0.0757	0.0757

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers, part-time workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.3. Robustness check: Excluding public sector workers.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	0.1113	0.0145***	−0.0048**	−0.0043	−0.0014
	(0.6574)	(0.0055)	(0.0022)	(0.0028)	(0.0014)
Snowfall	−0.0400	0.0033***	−0.0006	0.0001	−0.0014
	(0.0547)	(0.0011)	(0.0004)	(0.0002)	(0.0010)
Maximum temperature	0.0121	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0267)	(0.0002)	(0.0001)	(0.0001)	(0.0001)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	23,339	23,339	23,339	23,339	23,339
(Pseudo) R-squared	0.1016	0.0711	0.0711	0.0711	0.0711

Notes: OLS estimates on Column (1), MNL estimates on Columns (2-5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16-65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers, public sector workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.4. Robustness check: Including self-employed workers.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1817	0.0151***	−0.0029	−0.0052**	−0.0017
	(0.6005)	(0.0049)	(0.0018)	(0.0025)	(0.0013)
Snowfall	−0.0458	0.0033***	−0.0005	0.0002	−0.0014*
	(0.0465)	(0.0010)	(0.0003)	(0.0002)	(0.0008)
Maximum temperature	0.0131	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0240)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Self-employed worker	−1.8607	0.0183**	−0.0022	0.0012	−0.0024
	(1.2289)	(0.0090)	(0.0039)	(0.0044)	(0.0025)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	30,873	30,873	30,873	30,873	30,873
(Pseudo) R-squared	0.0934	0.0686	0.0686	0.0686	0.0686

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.5. Robustness check: Including average temperature.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.1753	0.0104**	−0.0019	−0.0040	−0.0017
	(0.5777)	(0.0049)	(0.0019)	(0.0026)	(0.0013)
Snowfall	−0.0108	0.0031***	−0.0004	0.0002	−0.0014*
	(0.0497)	(0.0010)	(0.0003)	(0.0002)	(0.0008)
Average temperature	0.0237	0.0010***	−0.0003***	−0.0002**	0.0000
	(0.0269)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,109	28,109	28,109	28,109	28,109
(Pseudo) R-squared	0.1009	0.0666	0.0666	0.0666	0.0666

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.6. Robustness check: Excluding post-covid years (2021–2023).

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	0.1409	0.0148***	−0.0052**	−0.0043	−0.0020
	(0.6237)	(0.0052)	(0.0021)	(0.0027)	(0.0014)
Snowfall	−0.0544	0.0037***	−0.0004	−0.0002	−0.0014*
	(0.0488)	(0.0011)	(0.0003)	(0.0002)	(0.0009)
Average temperature	0.0050	0.0015***	−0.0004***	−0.0004***	0.0000
	(0.0257)	(0.0003)	(0.0001)	(0.0001)	(0.0000)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	25,720	25,720	25,720	25,720	25,720
(Pseudo) R-squared	0.1011	0.0695	0.0695	0.0695	0.0695

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2019. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.7. Robustness check: including 2020 survey year.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	(1)	(2)	(3)	(4)	(5)
Rainy day	−0.2386	0.0128**	−0.0033*	−0.0040	−0.0016
	(0.5761)	(0.0049)	(0.0019)	(0.0026)	(0.0013)
Snowfall	−0.0346	0.0035***	−0.0006	0.0001	−0.0014*
	(0.0488)	(0.0010)	(0.0004)	(0.0002)	(0.0008)
Average temperature	0.0081	0.0013***	−0.0004***	−0.0003***	0.0000
	(0.0248)	(0.0002)	(0.0001)	(0.0001)	(0.0000)
Mode choice	Y	N	N	N	N
Socio-demographics	Y	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y
Observations (workers)	28,898	28,898	28,898	28,898	28,898
(Pseudo) R-squared	0.1025	0.0701	0.0701	0.0701	0.0701

Notes: OLS estimates on Column (1), MNL estimates on Columns (2–5). ATUS data observation period from 2003 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking, and cycling as the primary commuting mode, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.8. Percentage of commuting time by modes of transport, SUR estimates.

Dependent variable:	% car	% public	% walking	% bicycle
	(1)	(2)	(3)	(4)
Rainy day	1.3335***	−0.5344**	−0.6417**	−0.1573
	(0.4584)	(0.2711)	(0.2873)	(0.1371)
Snowfall	0.1066***	−0.0870***	−0.0129	−0.0066
	(0.0296)	(0.0179)	(0.0236)	(0.0065)
Average temperature	0.1289***	−0.0792***	−0.0482***	−0.0014
	(0.0215)	(0.0144)	(0.0111)	(0.0045)
Socio-demographics	Y	Y	Y	Y
Occupation F.E.	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y
Observations (workers)	28,114	28,114	28,114	28,114
R-squared	0.0350	0.0238	0.0206	0.0089

Notes: SUR estimates on the percentage of daily commuting time done by modes of transport. ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the percentage of daily commuting time by modes of transport. Y determines whether or not the specifications include controls for socio-demographics and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.9. Alternative methods of estimation, commuting time.

Dependent variable:	Total commuting, no commuters	Total commuting, commuters
	(1)	(2)
Rainy days	−0.0084	−0.0033
	(0.0134)	(0.0127)
Snowfall	−0.0028**	−0.0004
	(0.0012)	(0.0011)
Maximum temperature	−0.0006	0.0002
	(0.0006)	(0.0005)
Mode choice	Y	Y
Socio-demographics	Y	Y
Occupation F.E.	Y	Y
Weekday F.E.	Y	Y
Month F.E.	Y	Y
Year F.E.	Y	Y
Observations (workers)	36,459	28,114
R-squared	0.203	0.101

Notes: Poisson estimates. ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers are excluded in Columns (1) and (2), and zero-commuters in Column (2). Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day). Y determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects.
* p < 0.1
** p < 0.05
*** p < 0.01.

Table A.10. Heterogeneity results by climatic regions, alternative temperature intervals.

Dependent variable:	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)	Total commuting	Pr(Car)	Pr(Public)	Pr(Walking)	Pr(Bicycle)
	Warmer region					Colder region
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
Rainy day	−0.3657	0.0016	0.0017	−0.0009	−0.0018	−0.0081	0.0258***	−0.0077**	−0.0073**	−0.0023
	(0.7495)	(0.0058)	(0.0017)	(0.0037)	(0.0022)	(0.8432)	(0.0077)	(0.0032)	(0.0035)	(0.0014)
Snowfall	0.0713	0.0043	0.0003	−0.0005	0.0000	−0.0097	0.0024***	−0.0005	0.0002	−0.0006
	(0.1915)	(0.0034)	(0.0004)	(0.0006)	(0.0008)	(0.0544)	(0.0008)	(0.0004)	(0.0002)	(0.0005)
Maximum temperature (≤35 °F)	−5.4515	0.0988**	0.0090	0.0151	−0.0998***	−1.6662	0.0475*	0.0029	0.0157	−0.0633***
	(4.6388)	(0.0404)	(0.0078)	(0.0135)	(0.0129)	(2.8071)	(0.0234)	(0.0086)	(0.0117)	(0.0108)
Maximum temperature (36–40 °F)	−0.2979	0.2364***	−0.1088***	0.0128	−0.0963***	−3.8389	0.0168	0.0063	−0.0023	0.0066
	(8.9582)	(0.0409)	(0.0119)	(0.0136)	(0.0129)	(2.5983)	(0.0219)	(0.0092)	(0.0111)	(0.0049)
Maximum temperature (41–45 °F)	−5.5848	0.2102***	−0.1111***	0.0139	−0.0978***	−2.5532	−0.0136	0.0146*	0.0016	0.0021
	(3.7750)	(0.0284)	(0.0120)	(0.0104)	(0.0122)	(2.5552)	(0.0188)	(0.0087)	(0.0094)	(0.0031)
Maximum temperature (46–50 °F)	−5.2552**	−0.0069	0.0091	−0.0011	−0.0019	−1.2848	−0.0076	0.0016	−0.0035	0.0048
	(2.6341)	(0.0225)	(0.0068)	(0.0115)	(0.0047)	(2.2826)	(0.0203)	(0.0087)	(0.0111)	(0.0048)
Maximum temperature (51–55 °F)	−2.3917	−0.0004	0.0126***	−0.0045	0.0040	0.6020	0.0098	−0.0027	−0.0011	0.0018
	(3.3864)	(0.0166)	(0.0045)	(0.0085)	(0.0057)	(2.6880)	(0.0175)	(0.0087)	(0.0079)	(0.0054)
Maximum temperature (56–60 °F)	−1.1642	−0.0002	0.0040	0.0093	−0.0046	0.1236	−0.0079	0.0044	−0.0084	0.0062
	(2.3525)	(0.0123)	(0.0039)	(0.0057)	(0.0053)	(2.3080)	(0.0173)	(0.0082)	(0.0091)	(0.0041)
Maximum temperature (61–65 °F)	−2.6228	0.0163	−0.0028	−0.0076	0.0048*	−1.6389	−0.0175	0.0089	−0.0057	0.0034
	(1.6129)	(0.0140)	(0.0052)	(0.0069)	(0.0029)	(2.3477)	(0.0183)	(0.0075)	(0.0090)	(0.0036)
Maximum temperature (66–70 °F)	1.8717	−0.0112	0.0105***	0.0013	0.0038	1.5402	−0.0092	−0.0069	0.0005	0.0050
	(1.8433)	(0.0112)	(0.0036)	(0.0055)	(0.0026)	(2.1439)	(0.0165)	(0.0076)	(0.0081)	(0.0035)
Maximum temperature (71–75 °F)	−0.6381	0.0089	0.0044	0.0045	0.0017	−0.9907	0.0112	−0.0042	−0.0004	0.0027
	(1.8683)	(0.0138)	(0.0043)	(0.0043)	(0.0039)	(1.7974)	(0.0151)	(0.0087)	(0.0079)	(0.0030)
Maximum temperature (81–85 °F)	−2.4906*	0.0197*	0.0078**	−0.0072	−0.0082*	−1.4082	−0.0075	0.0028	0.0025	−0.0031
	(1.4316)	(0.0110)	(0.0037)	(0.0059)	(0.0045)	(1.6252)	(0.0155)	(0.0080)	(0.0073)	(0.0029)
Maximum temperature (86–90 °F)	−0.7941	0.0249**	0.0037	−0.0060	−0.0019	−0.5319	−0.0144	0.0082	0.0025	−0.0045
	(1.5840)	(0.0112)	(0.0039)	(0.0053)	(0.0035)	(1.9548)	(0.0157)	(0.0075)	(0.0082)	(0.0034)
Maximum temperature (91–95 °F)	−2.3681	0.0386***	−0.0030	−0.0158**	0.0013	2.5565	−0.0365*	0.0209*	0.0090	−0.0027
	(1.7431)	(0.0120)	(0.0038)	(0.0062)	(0.0026)	(3.2267)	(0.0190)	(0.0106)	(0.0097)	(0.0045)
Maximum temperature (>95 °F)	0.5314	0.0134	−0.0025	−0.0089	−0.0009	−15.0926***	0.1228**	0.0099	−0.0272	−0.0750***
	(2.2896)	(0.0133)	(0.0044)	(0.0079)	(0.0043)	(3.8970)	(0.0486)	(0.0205)	(0.0222)	(0.0110)
Socio-demographics	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Mode choice	Y	N	N	N	N	Y	N	N	N	N
Occupation F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Weekday F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Month F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Year F.E.	Y	Y	Y	Y	Y	Y	Y	Y	Y	Y
Observations (workers)	12,432	12,432	12,432	12,432	12,432	15,682	15,682	15,682	15,682	15,682
(Pseudo) R-squared	0.0912	0.1105	0.1105	0.1105	0.1105	0.1184	0.0970	0.0970	0.0970	0.0970

Notes: OLS estimates on Columns (1) and (6), MNL estimates on Columns (2–5) and (7–10). ATUS data observation period from 2003 to 2019, and from 2021 to 2023. Sample is restricted to workers aged 16–65 on their working days, defined as days workers spend at least 60 min working excluding commuting. Self-employed workers and zero-commuters are excluded. Estimates computed using ATUS sampling demographic weights. Robust standard errors, clustered at the state-month level, are shown in parentheses. Dependent variables are the total commuting time (in minutes per day), and the probability of using car, public transit, walking and cycling as the primary commuting mode choice, respectively. Y (N) determines whether or not the specifications include controls for socio-demographics, mode choices, and fixed effects. Full model results are available from the authors on request.
* p < 0.1
** p < 0.05
*** p < 0.01.

Details are in the caption following the image — **Figure A.1**
Open in figure viewer PowerPoint

*Histogram of maximum temperature*. *Notes*: Authors’ calculations. The figure shows the distribution of daily maximum temperatures (°F) by 16-degree bins during the observation period 2003–2019, 2021–2023. Maximum temperature bins on horizontal axis. Frequency in percent on vertical axis. Each bar depicts the frequency of days in each temperature interval. ‘ATUS sample’ refers to the distribution of maximum temperatures for those county-dates from which we have observations in our final sample, while ‘Population’ represents the maximum temperatures observed for those county-dates during the same period, but also on days without time using diary data.

Open Research

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

References

Citing Literature

Volume65, Issue3

June 2025

Pages 818-842

Weather Conditions and Daily Commuting

ABSTRACT

1 Introduction