Introduction

Health concerns and government restrictions have caused many people around the world to work from home during the COVID-19 pandemic, resulting in a sharp increase in telecommuting (i.e., doing paid work facilitated by information and communication technologies (ICT) from a location that is not the regular workplace, mostly home and possibly an alternate worksite). The term “telecommuting” is attributed to Jack Nilles, who proposed it in the 1970s as a way to reduce energy consumption and alleviate traffic congestion (Allen et al. 2015).

In addition to reducing vehicle miles traveled (VMT), decreasing energy use, and lowering emissions of air pollutants and greenhouse gases (GHG), proponents of telecommuting argue that it offers numerous co-benefits, including increasing the worker pool, generating transportation time and money savings, providing more time flexibility, improving work-life balance, and decreasing stress (Gajendran and Harrison 2007). It may also stimulate non-motorized and active modes (e.g., see Haddad et al. 2009; Lachapelle et al. 2018).

However, telecommuting may also affect promotion opportunities and ties with colleagues (Baert et al. 2020), physical health (Kubo et al. 2021), mental health (Escudero-Castillo et al. 2021), work-life balance for families with children (since child-care and school facility did not operate normally during the pandemic) (Sousa-Uva et al. 2021; Tavares et al. 2021), and even work productivity (Farooq and Sultana 2021). Moreover, the work travel savings from telecommuting may stimulate additional non-work travel (Elldér 2020). Telecommuting may also increase commuting duration and length (Elldér 2020; Melo and de Abreu e Silva 2017; Mokhtarian et al. 2004) because telecommuters tend to live in more suburban areas, usually associated with lower levels of transit supply and a higher likelihood of car use.

A large body of literature has analyzed telecommuting before COVID-19 (e.g., see Allen et al. 2015, Elldér 2020, or Mokhtarian et al. 1998, and the references herein), and some very informative studies recently explored how COVID-19 impacted working from home, either at the county level in California from March to November 2020 (McNally et al. 2023), at the county level in the US during April to June 2020 (Rafiq et al. 2022), or at the state level in the US for March 2020-March 2021 (Rafiq et al. 2023). However, to the best of our knowledge, a multivariate analysis of how COVID-19 impacted telecommuting in California at the individual level and how telecommuting frequency may change after the pandemic are still missing from the literature. The purpose of this paper is to start filling this gap.

California, which ranks as the fifth largest economy in the world, is an important state to study telecommuting because it is at the forefront of environmental and social issues in the US, and telecommuting can help address congestion, air pollution, and global climate change. California is also home to some of the largest and most innovative technology companies in the World (e.g., Google, Apple, Facebook). These companies, and others such as Zoom, have played a significant role in driving the growth of telecommuting and remote work, and have also been leaders in implementing and refining the technologies that make telecommuting possible.

In this paper, we analyzed how the frequency of telecommuting changed in California during the pandemic compared to before, and how it may evolve after COVID-19. Whereas most papers on telecommuting during the pandemic relied on non-random samples, our dataset was collected for us at the end of May 2021 by Ipsos, which randomly sampled Californian members of KnowledgePanel© (KP), the oldest and largest probability-based online panel in the US, so our results can be generalized to the California population. Moreover, our analysis covers a longer stretch of the pandemic (March 2020 to May 2021) than many other studies published until the end of 2022, and our structural equation models jointly explain car ownership, housing costs (thus accounting for residential self-selection), and telecommuting frequency.

Quantifying changes in telecommuting is important for updating sustainable community strategies and gauging telecommuting’s likely contribution to meeting California’s GHG reduction targets (Kallerman and Weinberg 2016). Moreover, our analysis of telecommuting frequency for different socio-economic groups and occupations should help policymakers concerned about the impacts of the pandemic on different segments of the labor market (Albanesi and Kim 2021).

In the next section, we review selected papers to inform our modeling choices. We then describe our data and present our methodology, before discussing our results. In the last section, we summarize our conclusions, discuss some policy implications, and suggest alternatives for future research.

Literature review

Background on telecommuting

As noted by Allen et al. (2015), although researchers have been studying telecommuting for decades, they have used various terminologies (e.g., distance work, flexplace, remote work, telework, or virtual work) and conceptualizations when reporting their results, which has hindered our understanding of telecommuting by making results difficult to compare across studies.

Telecommuting has been of interest to transport researchers since Nilles (1973) proposed it to reduce traffic congestion, sprawl, and the use of scarce non-renewable resources. Although telecommuting emerged during the 1973 OPEC crisis, it started developing in earnest in the US in the 1990s, first with the 1992 Interagency Telecommuting Pilot Project, which aimed to popularize external telecenters for government agencies in Washington D.C., and then with the 1996 National Telecommuting Initiative, which attempted to reenergize telecommuting among federal workers (Joice 2000). It also received a boost in the US from the 1990 Clean Air Act Amendments, which motivated large businesses to offer telecommuting to their employees. In the last two decades, telecommuting has been proposed as a possible travel demand management strategy thanks to developments in ICT. In 2019, almost a quarter of American workers did some of their work at home (Coate 2021). Globally, the percentage of employed persons who telecommute varied widely pre-pandemic. The corresponding percentages in the EU-27 and in Japan were only 9% and 10% in 2019, respectively, possibly because car commuting, which is one of the main reasons for promoting telecommuting in the US, is less prevalent in the EU and in Japan than in the US (European Union 2020; OECD 2021). Moreover, we note that jobs that can be performed with a computer and an internet connection are much more likely to lead to remote work arrangements.

Unfortunately, past efforts to increase telecommuting to deal with urban congestion have often yielded disappointing results (Noonan and Glass 2012). One reason might be that work trips are only 17.4% of all trips, which rises to 37.5% during morning and evening peaks (US Department of Transportation 2018). Another reason could be the low adoption rate of telecommuting (Coate 2021). Telecommuting may also be ineffective for reducing VMT because it often stimulates additional non-work travel (Elldér 2020).

Empirical findings show that various factors influence telecommuting. In addition to worker preferences, they include personal and household characteristics, and land use characteristics around home (Sener and Bhat 2011; Singh et al. 2013; Zhang et al. 2020). We examine them in turn below.

Only a handful of empirical studies have controlled for residential self-selection (namely the fact that households tend to choose their residential location based on their abilities, needs, and preferences for travel; see Mokhtarian and Cao 2008) and accounted for the endogeneity of car ownership (de Abreu e Silva and Melo 2018b; de Abreu e Silva and Melo 2018a). It is well-known that ignoring self-selection and the endogeneity of explanatory variables may bias estimates of model coefficients, which in this context would alter our understanding of the determinants of telecommuting (He et al. 2015).

Personal and household characteristics

The telecommuting literature suggests that socio-economic characteristics are important to characterize telecommuters, but their impact is not always clear. For example, Mokhtarian et al. (1998) found that women tend to telecommute more than men because of their extra family duties, but Bonacini et al. (2021); Fu et al. (2012); Pouri and Bhat (2003) found the opposite.

Age is another factor associated with telecommuting. Generally, middle-aged people (i.e., 36–50 years old) tend to telecommute less than workers under 35 because they are more likely to have managerial responsibilities that require their presence in the workplace (Sener and Reeder 2012; Singh et al. 2013). Conversely, employees with more experience working independently are more likely to telecommute (Peters et al. 2004; Sener and Bhat 2011; Zhang et al. 2020). In fact, some studies argue that workers over 50 years tend to prefer telecommuting (Bonacini et al. 2021; Fu et al. 2012).

Workers with higher education levels may be more likely to telecommute because they are in a better position to negotiate with their employers (Singh et al. 2013; Zhang et al. 2020), whereas less-educated workers tend to have jobs where telecommuting is not possible (Dey et al. 2020). Likewise, workers with a higher income may have greater access to the tools necessary for telecommuting, and thus may be more inclined to do so (Bonacini et al. 2021; Loo and Wang 2018; Zhang et al. 2020). Conversely, many low-income workers have blue-collar jobs in sectors where telecommuting is unfeasible (He and Hu 2015; Sener and Bhat 2011; Singh et al. 2013).

Occupation thus matters for telecommuting. People in services are more likely to telecommute than those working in sales, manufacturing, trade, transport and communication (Sener and Bhat 2011; Singh et al. 2013; Zhang et al. 2020). Moreover, part time jobs may be more flexible so they may be more conducive to telecommuting (Felstead and Henseke 2017), although Abendroth and Reimann (2018) found the opposite.

Married people tend to favor telecommuting (Fu et al. 2012), and so do households with children (Bhuiyan et al. 2020; Fu et al. 2012; Singh et al. 2013) because of the added time flexibility it provides. Likewise, household size matters for telecommuting because workers in larger households tend to have more responsibilities which require them to spend more time at home (Fu et al. 2012; Yen 2000). However, it may be more challenging (e.g., finding a quiet place) for workers in larger households to work from home (Bhuiyan et al. 2020; Zhang et al. 2020).

Finally, African American and Hispanic workers tend to telecommute less since they are more likely to be in jobs for which telecommuting is not feasible (Dey et al. 2020).

Land use characteristics

Specific land uses may be more conducive to telecommuting because telecommuters tend to be located in suburban areas (Kim et al. 2012). As a result, they often have longer commutes than other workers (Zhu 2013, 2012). Telecommuting can be seen as a coping strategy, at least in the short term (Elldér 2020).

To organize our brief discussion of land use variables, we focus on density, diversity, design, destination/job accessibility, and distance to transit stops (Cervero et al. 2009; Cervero and Kockelman 1997; Ewing and Cervero 2010).

Density usually refers to the number of homes, people, or jobs per unit area. It is negatively associated with car ownership, and telecommuting frequency (Van Acker and Witlox 2011).

Diversity, or more specifically “land-use diversity,” measures the degree of proximity of various land uses. One example is the job-housing balance, which refers to the spatial relationship between the number of jobs and the number of housing units. An area is considered balanced if resident workers can obtain a job locally, and if available housing can serve the needs of a variety of workers (Giuliano 1991). A better job-housing balance is believed to lower car ownership, and reduce telecommuting (Fu et al. 2012; Ma and Chen 2013; Van Acker and Witlox 2011).

Design refers to road connectivity, which is the degree of connectivity towards destinations. Better road connectivity promotes the use of driving for commuting which should decrease the probability of telecommuting (Bhuiyan et al. 2020; Fu et al. 2012).

Destination/job accessibility refers to the ability of reaching activities or locations (Geurs and van Wee 2004). Ewing and Cervero (2010), and Kockelman (1997) argued that good accessibility can significantly reduce commuting times, which may discourage telecommuting.

Finally, distance to the nearest transit stop may also matter for telecommuting because good transit accessibility favors commuting via transit and may thus reduce telecommuting (Caulfield 2015; Mouratidis and Peters 2022).

Like Islam and Saphores (2022) for commuting, we also expect the cost of housing to play a role in the decision to telecommute because unaffordable housing increases the length of commutes (Sultana 2002), and longer commutes is one of the determining factors of telecommuting (de Abreu e Silva and Melo 2018a, 2018b).

Telecommuting during COVID-19

Stay-at-home restrictions due to COVID-19, better ICT, and an increasing emphasis on reducing VMT to decrease GHG emissions under SB 375 have made telecommuting a popular approach for addressing global health risks while allowing economic activity to continue (Nguyen 2021). As more data become available, the number of studies concerned with the impacts of the pandemic on telecommuting is growing. We review below some selected studies, starting with US studies.

Based on a nationally representative sample of around 50,000 respondents, Brynjolfsson et al. (2020) found that younger people are more likely to switch to telecommuting. Moreover, telecommuting is more prevalent in states with a higher share of ICT jobs.

Bick et al. (2021) reached slightly different conclusions from their analysis of data collected during an online nationwide survey with 46,450 respondents: for them, women, older, better educated, and higher income workers are more likely to telecommute. Moreover, while the share of workers who only telecommute jumped from 7.6% to 31.4% between February and May 2020, it declined back to 20.4% by the end of 2020.

Jiao and Azimian (2021) examined the relationship between socio-economic characteristics and telecommuting using Household Pulse Survey data collected between April 23, 2020, and March 1, 2021, by the US Census Bureau. They found that adults 35 or older are less likely to telecommute than those under 35, and that the likelihood to telecommute is higher in larger households and for people with an individual annual income over $100,000. Conversely, males, Whites, and workers without graduate degrees are less likely to telecommute.

Asfaw (2022) confirmed racial differences in telecommuting after analyzing data from the Current Population Survey (May 2020–July 2021): the odds of telecommuting for Black and Hispanic workers were 35% and 55% lower respectively than for White workers, and 44% higher for Asian workers.

After analyzing survey data from 4045 residents of the greater Los Angeles region collected in the Fall of 2020, Malik et al. (2023) reported that non-telecommuters are more likely to be non-White, younger, and with a lower household income than telecommuters. Moreover, their use of motor vehicles and active travel modes increased for non-work travel.

In the US, we also need to mention McNally et al. (2023), who explored changes in working from home and the resulting activity-travel behavior from March to November 2020 using descriptive statistics, Rafiq et al. (2022) who analyzed how COVID-19 impacted working from home and travel for eight weeks in April-June 2020 using structural equation modeling, and Rafiq et al. (2023), who studied telecommuting and travel during the first year of the pandemic using clustering methods, but they all considered data aggregated by county (McNally et al. 2023; Rafiq et al. 2022) or state (Rafiq et al. 2023).

A few studies from other regions are also directly relevant to our work. Astroza et al. (2020) analyzed data from 4395 adults from Chile collected via an online survey in March 2020. They reported that workers from high-income households, with more education, and women are more likely to telecommute whereas workers from larger households and essential services are not.

In Australia, Beck et al. (2020) examined the frequency of telecommuting based on 2020 data. They concluded that a higher household income and living in large metropolitan areas increases the probability of telecommuting while working in some technical and trade occupations reduces it. In a related paper, after analyzing the impact of working from home on modal commuting in 2020 in two large Australian cities, Hensher et al. (2022) reported an increase in many types of non-commuting trips.

In Oslo, Norway, and the surrounding Viken region, Mouratidis and Peters (2022) found an increase in several teleactivities during the pandemic based on data collected in March–May 2020. While the increase in telework and virtual meetings was more pronounced in denser neighborhoods, lower density neighborhoods saw a sharper increase in online learning. Similarly, in Germany, Ecke et al. (2022) showed that public transport has lost importance for commuting, and that people with more education and a higher income are more likely to telecommute, which confirmed findings from another Germany study by Reiffer et al. (2022).

Two telecommuting studies covered multiple countries. The first one (4,628 observations from online panel surveys conducted between August and December 2020) covered eight countries (Balbontin et al. 2021). The authors reported that older people and women tend to telecommute more often in South America. A higher income increases telecommuting in Australia and Chile, and so does commuting time in Australia and South Africa. As expected, car availability has a negative impact on the number of telecommuting days. The second study analyzed data collected between March 23 and May 12, 2020 in fourteen countries (Shibayama et al. 2021). Results indicate that in workplaces with essential workers, the shift to telecommuting does not typically exceed 30%, whereas in workplaces compatible with telecommuting it reaches 60–80%.

Another strand of the literature explored whether telecommuting changes will stick after the pandemic. Our paper selection emphasizes US studies but we also mention below some studies from Canada and the EU. Barrero et al. (2021) examined data from 28,597 Americans collected by the Survey of Working Arrangements and Attitudes. They found that employees with higher earnings and a better education tend to telecommute more, which reduces spending in major city centers but could increase productivity by 5% post-pandemic relative to before.

Mohammadi et al. (2022) analyzed two waves of survey data collected between April and October 2020, and from November 2020–May 2021. They reported a shift in preferences for telecommuting post-pandemic for millennials, employees with long commutes, high-income earners, and highly educated workers.

Likewise, Javadinasr et al. (2022) analyzed data from a longitudinal two-wave panel survey conducted in the US between April 2020 and May 2021. They found that 48% of workers anticipate having the option to telecommute after the pandemic. These workers are mostly younger, with a higher education and/or a higher income, and they tend to be more concerned about the environment. As a result, car and transit commuting may drop by 9% and 31% after the pandemic compared to before.

Based on an online survey with 1028 respondents conducted in South Florida in May 2020, Asgari et al. (2022) reported substantial heterogeneity in preferences for telework across many variables. Before COVID-19, males, full-time students, people with PhDs, and those with a higher income were more likely to have jobs with a telework option. They were also more likely to be pro-technology, pro-online education, workaholic, and pro-telework. During the pandemic, workers with professional/managerial/technical jobs and with lower physical-proximity measures had the highest telework frequency. The authors also concluded that teleworking during the pandemic reshaped preferences for teleworking in the future.

Changes in telecommuting were also explored in other parts of the world. In the Canadian province of British Columbia, for example, Rahman Fatmi et al. (2022) found that part-time female workers, mid-age individuals, full-time workers with children, and full-time workers with longer commutes have a significantly higher probability of telecommuting every day after the pandemic. In the Netherlands, Olde Kalter et al. (2021) reported that office workers and teaching staffs were more likely to telecommute during the lockdown, but that after the lockdown, only office workers expected to experience increases in telecommuting. Also in the Netherlands, de Haas et al. (2020) reported that 27% of home-workers expect to telecommute more often in the future. However, In Padova, Italy, Ceccato et al. (2022) concluded that the end of COVID-19 could see a rebound effect with shifts towards non-sustainable modes (e.g., driving personal vehicles).

Based on this literature review, a quantitative assessment based on a random survey of how COVID-19 impacted telecommuting in California, and how telecommuting frequency may evolve after the pandemic are still missing from the literature.

Data

Our dataset was collected in late May 2021 by Ipsos, which surveyed for us California members of KP. With approximately 60,000 members, KP is the oldest and largest probability-based online US panel. Owing to its size and the way its participants were recruited, its subset of Californian panelists is representative of the state’s population.

Conducting surveys using KP offers several advantages (Ipsos 2021). First, it helps to overcome non-response bias because survey cooperation rates (i.e., the ratio of panelists who take the survey to the number of panelists invited) typically exceed 70%. Second, it reduces survey fatigue because panelists are only required to participate in two to three surveys per month on average. Third, this approach helps overcome the self-selection bias inherent in online surveys because KP members are recruited using address-based mail sampling based on the Delivery Sequence File of the US Postal Service. Special care is taken to recruit harder-to-reach groups, such as African Americans, Latinos, Veterans, Americans with disabilities, LGBTQ and non-binary people, rural residents, and non-internet households. Upon enrolling, the latter receive a tablet with a mobile data plan. The socio-economic characteristics of panel members are recorded when they enroll and updated annually so they do not need to be collected during surveys.

Questionnaire

Our questionnaire had two parts. In Part I, we inquired about commuting and telecommuting before, during, and potentially after the COVID-19 pandemic. In Part II, we explored how Californians shopped for groceries and prepared meals during the same periods.

Our questionnaire was first written in English and pre-tested by graduate students. Ipsos then conducted a pilot study with 25 California members of KP in early May 2021. We modified our questionnaire to include the feedback received. The median completion time was 12 min.

To include Californians more comfortable with Spanish (according to the US Census Bureau, ~ 30% of Californians speak Spanish at home, and 55% English), we translated our survey in Spanish and pre-tested it with native speakers. Both versions of the survey were administered in late May 2021. Data collection was stopped after receiving answers from 1026 respondents.

KP members must remain anonymous, but we were allowed to ask for their residential and work (when applicable) ZIP codes, which allowed us to derive land use variables and add them to our models. Figure 1 shows the residential locations of employed Californians in our sample. As expected, more respondents (the size of a green dot is proportional to the number of workers who live in a ZIP code) reside in more populated areas (e.g., Los Angeles and the Bay area), but we also have employed respondents in central California, which is much more thinly populated.

Fig. 1
figure 1

Home location of employed respondents in each ZIP code. ZIP code areas for our respondents range from 0.28 to 1224.05 sq mile with a mean of 48.51 sq. mile

Survey timing and COVID-19

To contextualize the timing of this study, it is useful to look back at the evolution of the pandemic in California (see Fig. 2).

Fig. 2
figure 2

Evolution of the pandemic in California and survey timing. Data source: https://covid19.ca.gov/state-dashboard/#county-statewide

From Fig. 2 and Table 1, we see that our survey was conducted after the main wave of deaths (December 2020 to February 2021) had subsided, at a time when Californians were hoping for life to get back to normal as vaccinations were ramping up (the Pfizer–BioNTech and the Moderna vaccines were granted emergency use on December 11 and 18, 2020, respectively).

Table 1 Key COVID-19 policy action in California between March 2020 and June 2021

Dependent variables

In this paper, we analyzed data collected in Part I of our survey, where we asked about (tele) commuting before, during (between the March 2020 stay-at-home order from Governor Newsom and May 2021), and potentially after the pandemic. We characterized the latter as a time when there would be no more cases in the US, which in retrospect was overly optimistic since COVID-19 is likely to stay in the background like the flu. Our starting hypothesis was that the pandemic will increase post-COVID telecommuting, although much less than the high levels experienced during the pandemic. Our goal was to quantify these changes and understand who will be impacted most. To test that hypothesis, we estimated three models.

In our first and second model, the dependent variable is the average number of days per week an employed respondent telecommuted before and during the pandemic respectively (i.e., a number between 0 and 7). In our third model, the dependent variable is the number of days per week an employed respondent is expecting to telecommute after the pandemic.

Explanatory variables

Personal and household characteristics

We considered a wide range of personal and household variables that characterize households and workers based on the data that Ipsos collects annually from KP members.

For simplicity, we reclassified the seven income groups received from Ipsos into four groups. To reflect the presence of children in the household, we defined three binary variables: no child, one child, and two or more children. Like Dai et al. (2016), Ding et al. (2017), and Van Acker and Witlox (2011), we included household size as a count variable.

To capture generational effects, we defined binary variables for the age of employed respondents based on definitions from the Pew Research Center (2019). We combined members from Generation Z (18–24 years) with Millennials (25–40 years), and members from the Silent Generation (76–96 years) with Baby Boomers (57–75 years) because the numbers of workers from Generation Z and Silent Generation respondents are small.

To create a model with a manageable number of explanatory variables, we also reclassified the 25 categories of occupations into 9 groups based on the North American Industry Classification System (NAICS). We merged ‘primary industry’ and ‘art/ entertainment/recreation’ with ‘others,’ since their percentages are very small.

For ethnicity, we lumped ethnicities different from White, African American, and Asian into “Other” because of their relatively small numbers. We did not change the education variables. Moreover, we included in our models the gender of the respondent, Hispanic status, whether the survey was taken in Spanish, marital status, and full-time/part-time work status.

Finally, we relied on factor analysis to summarize the twelve variables that represent attitudes toward communication technology because telecommuting requires some familiarity with ICT.

Land use characteristics

Most empirical studies of commuting include land use characteristics around residences since commuting trips originate from home (Manaugh et al. 2010; Sun et al. 2017). Van Acker and Witlox (2011) showed that land-use around workplaces significantly influences car ownership, and telecommuting frequency but since 46% of workers in our sample worked fully from home during the pandemic, they did not have a work location to report (Ipsos 2021). To characterize land use patterns around residences, we relied on the following variables: job density, intersection density, distance to the nearest transit stop and to the nearest employment center, a measure of the job-housing balance. We also considered median home values.

For density, we considered job density but not population density since the former is more influential for explaining commuting (Van Acker and Witlox 2011). We obtained job density from Zip Code Tabulation Areas (ZCTAs) from the 2019 Longitudinal Employer-Household Dynamics (LEHD).

To calculate intersection density, we used a measure of road connectivity that characterizes intersections with three or more links in each ZCTA (Cervero et al. 2010). Our road network data come from the 2020 TIGER/Line shapefiles from the US Census.

The only location information we have for our respondents is their ZIP codes, so we approximated the location of a residence by its ZIP Code centroid. We then computed the network distance to the nearest transit stop using transit stops data from the General Transit Feed Specification (GTFS) dataset (https://gtfs.org).

To estimate job accessibility, we relied on distance to the nearest employment center. Following Giuliano and Small (1991), we found 45 subcenters in California using a ‘90%-10k’ approach applied to the 2019 LEHD data at the ZCTA level. These 45 subcenters offer a total of 5,846,238 jobs over 413,102.44 acres in 167 ZCTAs.

The simplest and most common measure of the job-housing balance in each ZCTA is the ratio of the number of jobs to the number of resident workers (Cervero 1989). Finally, we obtained median home values at the ZCTA level from the 2015–2019 American Community Survey.

COVID-19 severity

The decision to telecommute during the pandemic was likely impacted by public health restrictions. To capture the impact of COVID-19 on telecommuting, we therefore included in our models a COVID-19 severity variable defined as the cumulative number of cases from March 2020 to May 2021 in a respondent’s county divided by county population (California Open Data Portal 2020). Most of the COVID-19 policies and restrictions were enacted at the state level. They include stay at home orders, non-essential business closures, bar and restaurant closures, mask mandates, gathering restrictions, and quarantine mandates (COVID19 StatePolicy 2022; Hale et al. 2021). We therefore did not reflect county-level restrictions in our models.

Sample sizes

Our telecommuting frequency model before the pandemic was estimated with 511 observations out of the 594 respondents who were employed because we lost 83 observations to missing variables. Summary statistics are provided in Fig. 3 and Table 2. Models of telecommuting frequency during and after the pandemic were estimated with 498 respondents (we lost 72 observations to missing variables).

Fig. 3
figure 3

Descriptive statistics for binary explanatory variables—before COVID-19 (N = 511)

Table 2 Descriptive statistics for non-binary explanatory variables—before COVID-19 (N = 511)

Methodology

Conceptual model

Our conceptual model is shown in Fig. 4. For an explanation of the symbols used, see the notes below Fig. 4 and refer to Kline (2015).

Fig. 4
figure 4

Conceptual model. Notes: Following standard practice in structural equation modeling (Kline 2015), a one-way arrow between variables implies that variables at the start of the arrow have a direct effect on the variable at the tip of the arrow. The symbols μ1, …,μ11, ε0, ε1, ε2 and ε3 are error terms. “Tech savviness” is an unobserved endogenous variable estimated via confirmatory factor analysis using the variables listed in the lower half of Table 2. We initially assumed that how much workers telecommute would depend on the length of their commute, but it was not the case, so we simplified our final model by omitting the commuting time equation

In our model, we assumed that the socio-economic characteristics of a worker and those of her/his household lead her/him to select a dwelling, whose characteristics (structural, locational, environmental) are reflected in its price, in accordance with microeconomics theory. For simplicity, we assumed that the other residential land use variables are exogenous. Like de Abreu e Silva and Melo (2018a), we assumed that telecommuting frequency is influenced by car ownership and land use characteristics around residences, because land use determines the presence and characteristics of other modes. We also assumed that the duration of the work commute would impact whether or not a worker telecommutes, but that relationship turned out to be not statistically significant, so we dropped it from our final model (this explains the dotted lines to and from commute duration).

To control for residential self-selection (namely the fact that households tend to choose their residential location based on their abilities, needs, and preferences for travel; see Mokhtarian and Cao 2008), personal and household characteristics explain median home value around the residence, which implies that personal and household characteristics can indirectly affect telecommuting behavior via residential median home values. To capture the impact of COVID-19, we included a COVID-19 severity variable in the models of telecommuting frequency during and after the pandemic. Finally, tech savviness—a latent factor- estimated from the eleven indicators listed in Table 2, influences telecommuting frequency.

SEM models

Structural Equation Models (SEM) can estimate the statistical relationships among a set of observed and unobserved variables represented as latent factors based on a theoretical model that reflects the influence of exogenous variables on endogenous variables, and the influence of endogenous variables on each other (Kline 2015). Each of our models is a system of simultaneous equations plus a latent factor (estimated via confirmatory factor analysis) that jointly reflect the causal paths shown in Fig. 4. Excluding commuting time, which was not significant, and the equations for the technology savviness factor, our model can be written:

Regression equation for residential home value:

$${\varvec{L}} = {\varvec{X}}_{1}{\varvec{\varGamma}}_{1} + {\varvec{\varepsilon}}_{1} ,$$
(1)

Censored (from below at zero) regression equation for car ownership:

$${ }{\varvec{C}} = {\text{Max}}\left( {0,{\varvec{C}}^{\user2{*}} = \beta_{21} {\varvec{L}} + {\varvec{X}}_{2}{\varvec{\varGamma}}_{2} + {\varvec{\varepsilon}}_{2} } \right),$$
(2)

Ordered probit equations for telecommuting frequency:

$$y_{i} = j \quad {\text{if}}\,\,\tau_{j} < y_{i}^{*} \le \tau_{j + 1}$$
(3a)

for j∈{0, …,7} and i∈{1, …, n}, where the corresponding latent variable is

$$y_{i}^{*} = \beta_{0} + \beta_{31} L_{{\varvec{i}}} + \beta_{32} C_{{\varvec{i}}} + {\varvec{X}}_{{{\varvec{i}}3}}{\varvec{\varGamma}}_{3} + \varepsilon_{3i} .$$
(3b)

In the above:

  • L is an n × 1 vector of residential median home values (in $100,000);

  • Xk (k ∈ {1,2}) is an n × pk matrix of explanatory variables (personal and household characteristics, land use characteristics, and COVID-19 severity), assumed to be exogenous; likewise, Xi3 (i ∈ {1, …,n}) is the 1 × p3 matrix of explanatory variables (personal and household characteristics, land use characteristics, and COVID-19 severity), for respondent i;

  • C is an n × 1 vector of numbers of household cars, and C* is the associated latent vector;

  • yi is the average number of days of telecommuting per week for respondent i, and \({y}_{i}^{*}\) is the associated latent variable for the ordered probit model;

  • Γ1, Γ2, and Γ3 are unknown pk × 1 vectors of model parameters to estimate jointly with the unknown scalar parameters β0, β21, β31 and β32, and ordered probit model thresholds τ1, …,τ7 (τ0 = −∞, τ8 =  + ∞); and

  • ε1 and ε2 are n × 1 error vectors, and ε3i is a scalar error, all with standard normal distributions.

L and C are endogenous. Since our model is recursive, it is identified (Kline 2015). Unknown model parameters were estimated by minimizing the difference between the sample covariance and the covariance predicted by the model (Bollen 1989).

SEM decomposes the impacts of exogenous and endogenous variables on the dependent variable into direct, indirect, and total effects. Direct effects quantify the impact of one variable on another without mediation. Indirect effects are mediated by at least one other variable. Finally, total effects are the sum of direct and indirect effects (Bollen 1989).

Exploratory factor analysis

Our SEM models include a latent factor designed to capture technology savviness based on answers to the questions listed in the bottom half of Table 2. Using exploratory factor analysis, we first explored the adequate number of factors needed to summarize these questions. Based on the Kaiser criterion (Fabrigar and Wegener 2012), we retained only one factor as only one eigenvalue is > 1. We then discarded question 10 because its loading was below 0.3 (de Abreu e Silva et al. 2012; Antipova et al. 2011).

To assess the adequacy of the resulting factor, we performed some common diagnostics. We calculated Cronbach's alpha (which indicates how well a set of variables measures a single underlying construct), conducted a Bartlett test for sphericity (which checks whether the correlation matrix of the variables differs significantly from the identity matrix; if not, the factor is inappropriate), and computed the Kaiser–Meyer–Olkin (KMO) statistic (which measures the proportion of the variance common to the variables considered for factorization; a lower proportion is better and leads to a higher KMO value) (Azevedo 2003; Kline 2015).

For our before-pandemic telecommuting model, alpha and KMO are 0.73 and 0.81 respectively, and for the telecommuting models during and after the pandemic, they are 0.74 and 0.82 respectively, which is adequate (Azevedo 2003; Kline 2015). For all three models, the Bartlett test (the null hypothesis was rejected) supported our “Tech-savviness” factor.

Results

COVID-19 and telecommuting in California

Before analyzing the results of our SEM models, let us briefly consider how working from home has changed and is likely to change in California because of COVID-19. To match our sample to the California population aged 18 and over, Ipsos calculated sample weights by raking the following distributions of Californians aged 18 and over from the 2019 American Community Survey: gender by age, race and Hispanic status, education, household income, and language proficiency (for English and Spanish). We used these weights to calculate the percentage of different telecommuting frequencies before, during, and after the pandemic shown on Fig. 5.

Fig. 5
figure 5

Changes in telecommuting frequency. Notes: the thickness of a flow line is proportional to the percentage of people who telecommuted at that frequency at the start of a period. Unemployed Californians are included because we are mapping our results to all Californians 18 and over

First, we see that the pandemic had a substantial impact on telecommuting: while 42.2% of Californians never telecommuted before, that percentage shrank to 22.8% during the pandemic. At the same time, the percentage of Californians who telecommuted some almost doubled to 36.0% (4.5% + 7.5% + 24.0%), up from 19.2% before. The frequency that increased the most is “5 + days a week,” which jumped from 9.3% pre-pandemic to 24.0%. We also note an uptick in the percentage of Californians who are not employed (which includes homemakers, retirees, and Californians seeking employment) at the time of our survey (40.6%), up from 37.9% before the pandemic.

Post-pandemic, our respondents expect the percentage of Californians who never telecommute to drop to 34.9% (down from 42.2% pre-pandemic). Two telecommuting frequencies are also expected to increase: “1–2 days per week” (to 8.7%, up from 5.3% pre-pandemic), and “3–4 days per week” (to 5.8%, up from 4.6% pre-pandemic). Conversely, “5 + days per week” could go down to 8.9% (from 9.3%), which echoes findings from a 2020 survey that only 12% of American workers want to work from home full-time (Gensler Research Institute 2020). Totaling the percentage of Californians expecting to telecommute for these three frequencies, the net gain would be 4.2%, which is substantial but not as large as might have been expected. Moreover, these changes did not/will not uniformly affect all Californians, as shown by our multivariate models.

Telecommuting by occupation category

Figure 6 shows the weighted percentage by telecommuting frequency for different occupations of our working respondents with a known occupation before, during, and potentially after the pandemic. For each occupation and time period (before, during, and after COVID-19), four frequencies are considered: (1) never; (2) 1–2 times a week; (3) 3–4 times a week; and (4) 5 or more times a week. For each time period, frequencies over all occupations sum to 100%.

Fig. 6
figure 6

Telecommuting frequency of Californian workers before, during, and after COVID-19

First, we see that the percentage of workers who never telecommute decreased for all occupation categories (the lone exception is health care, which is unchanged) during the pandemic compared to before. That percentage decreased most (in relative terms) for engineering, architecture, law, and social sciences, and for education. Apart from health care, categories where this percentage changed the least include social and government services (which include a number of first responders), trades, transport, construction, installation (either because of essential workers or the impossibility to work remotely), and sales and services (also because of the impossibility to work remotely).

Second, the change in intermediate telecommuting frequencies depends on the occupation category. For example, the percentage of workers telecommuting 1–2 times a week decreased for “social and government services,” “education,” “trades, transport, construction, installation,” and “engineering, law, social sciences,” and increased for other categories. However, the percentage of workers telecommuting 3–4 times a week increased for all occupation categories, except for “trades, transport, construction, installation.”

Third, although the percentage of California workers who never telecommute is expected to go down after the pandemic (see Fig. 5), that change varies by occupation, and the percentage of workers who never telecommute in “Education”, “Sales and services” and “Trades, transport, construction, installation” may actually increase slightly.

SEM results

Overview

After preparing our dataset with Stata 17.0, we estimated our models using Mplus 8.9 because it offers more SEM tools than Stata. We relied on the weighted least squares mean and variance adjusted (WLSMV) estimator to account for non-normally distributed variables (Muthén and Muthén 2017), since many of our explanatory variables are binary and our telecommuting frequency variable is categorical. For the models discussed below, the maximum value of the variance inflation factors (VIF) for our explanatory variables is under 3.5, so, multicollinearity is not a problem here.

We explored several model specifications, and used common fit statistics (χ2, the Root Mean Square Error of Approximation (RMSEA), the Comparative Fit Index (CFI), and the Tucker-Lewis Index (TLI)) (Kline 2015) to select our preferred models, which are presented in Table 3. For conciseness, only significant coefficients are reported. Cut-off criteria for these fit statistics are: χ2 with p-value > 0.05, CFI > 0.50, and TLI > 0.50, where 1 represents the best fit, RMSEA < 0.05 where smaller values indicate a better model fit (Kline 2015). All fit statistics for our models have acceptable values except the χ2 values because they increase with sample size, so models with larger sample sizes might be rejected even though the differences between the observed and model-based covariance matrices are small (Kline 2015).

Table 3 SEM results

Equation 1 is a linear regression model, so its interpretation is straightforward. Its dependent variable (median home value around the residence) is in hundreds of thousands of dollars so to obtain the impact of changing an explanatory variable by one unit, its coefficient needs to be multiplied by 100 to get results in thousands of dollars ($1k).

Likewise, Eq. 2 is a censored regression model, so it can be interpreted as a linear regression model for values of the dependent variable that are greater than zero.

Equation 3 is an ordered probit model, so we simply report estimated coefficients. Providing a simple quantitative interpretation of estimated coefficients is not possible here (Train 2009), but we know from Eqs. (3a-b) that augmenting the variable corresponding to a positive (negative) estimated coefficient would potentially increase (decrease) the average weekly number of telework days.

The last two columns of Table 3 report total effects. For conciseness, indirect effects are not shown since they can be simply calculated by subtracting direct effects from total effects. We discuss total effects for variables with significant indirect effects, which are shaded in Table 3.

Since residence home value is not statistically significant in any of the telecommuting frequency equations, we only discuss results for Eqs. 2 (car ownership) and 3 (telecommuting frequency) before and during the pandemic. After the pandemic, we only discuss Eq. 3 (telecommuting frequency) since car ownership is not significant in Eq. 3. Finally, we note that the tech-savviness factor is highly significant for all three time periods.

Before COVID-19

Car ownership (Eq. 2; Column I).

Starting with worker characteristics, we see that married people (0.310‡) are more likely to own more cars than unmarried people, which is expected. Workers with a graduate degree (− 0.465†) tend to own fewer cars but a higher household income has the opposite effect (0.588‡, 0.698‡ and 0.857‡ for incomes of $50k to $100k, $100k to $150k and > $150k, respectively). Conversely, household size is positively associated (0.385‡) with car ownership although the presence of children (− 0.588‡ for one child; − 1.009‡ for two or more) acts as a correction.

For land use characteristics around residences, median home value (− 0.055‡) shows a mild negative association with car ownership which is unexpected. Finally, car ownership slightly rises with job-housing ratio (0.095*) while job density shows the opposite effect (− 0.072‡).

Telecommuting frequency (Eq. 3; Column II).

Starting with worker characteristics, we see that Generation X members were more likely to telecommute (0.334†) before the pandemic because they probably have more experience working independently than younger people (Peters et al. 2004; Sener and Bhat 2011; Zhang et al. 2020).

Full-time workers (− 0.656‡) telecommuted less before the pandemic compared to part-time workers, possibly because part-time jobs are often more flexible (Felstead and Henseke 2017).

Among household characteristics, we note that owning more cars had a negative impact on telecommuting (− 0.097*), as commuting often requires access to more motor vehicles.

In addition, the “car ownership” equation created a number of indirect effects (as indicated by shaded cells in Column VI) for the “telecommuting” equation. First, households with higher incomes (− 0.057*, − 0.068*, and − 0.083* for incomes $50k to $100k, $100k to $150k, and > $150k respectively) were less likely to telecommute, possibly because their higher income comes with management responsibilities that require them to work on-site. Second, household size had a negative impact on telecommuting (− 0.037*), possibly because of the difficulty to find a quiet space to work in many larger households. Conversely, households with children prefer telecommuting (0.057* and 0.098* for households with one child, two or more children respectively) possibly as they seek a balance between work and childcare. Third, households who can afford more expensive neighborhoods (0.005*) tend to telecommute more overall, which agrees with urban economic theory (for which households select their residential locations after considering trade-offs between commuting and housing costs). Finally, telecommuting rises slightly with the job density (0.007*) around residences which is unexpected.

During COVID-19

Car ownership (Eq. 2; Column III).

The determinants of car ownership during COVID-19 are similar to those before the pandemic, so we do not discuss them further.

Telecommuting frequency (Eq. 3; Column IV).

Contrasting Columns IV and II shows that the pandemic had a substantial impact on the determinants of telecommuting.

Starting with worker characteristics, we see that respondents who took our survey in Spanish were more likely to telecommute (1.282‡) compared to respondents who took it in English. However, the Hispanic indicator is not statistically significant.

Whereas before COVID-19 education is not statistically significant, during the pandemic more educated Californians became more likely to telecommute (0.465† for a bachelor’s degree; 0.630‡ for a graduate degree) possibly because they were in a better position to negotiate with their employers the option to telecommute (Singh et al. 2013; Zhang et al. 2020).

The occupation variables also saw substantial changes. Unlike before COVID-19, during the pandemic workers in Education (0.486*) and Engineers/Architects/Lawyers/Social Scientists (0.416*) were more likely to work from home, whereas heath care workers had to disproportionately go to work (− 0.985‡), despite a shift to telemedicine (Friedman et al. 2021). The greater adoption of telecommuting was made possible by more tech savviness (0.223†), which did not play a role in explaining telecommuting before the pandemic.

Looking at household characteristics, we see that the importance of car ownership for telecommuting waned during the pandemic (− 0.136‡) compared to before (− 0.097*) as Californians worked more from home. Surprisingly, COVID-19 severity around residences was not significant in our models, possibly because many restrictions were statewide.

Finally, indirect effects (via the car ownership variable) played an important role in the “telecommuting” equation (Column VII). Unlike before COVID-19, they slightly increased the impact of education (0.703‡ for workers with graduate/professional degrees), mitigated the impact of marital status (− 0.035*) and dampened the impact of full-time work status (− 0.008*).

After COVID-19 (Expectations about telecommuting post-pandemic)

Telecommuting frequency (Eq. 3; Column V).

Interestingly, generation, race, or Hispanic status do not impact expectations about telecommuting post-pandemic. Education does, however, with more educated workers expecting to telecommute more (0.383* for bachelor’s degree) than workers with a high school education or less. This effect depends on occupation, however, with workers in healthcare (− 0.710†), education (− 1.023‡), trades/transport/construction/installation/repair (− 0.754†), and sales and services (− 0.567†) expecting to telecommute less than managers (our baseline). Moreover, full-time workers expect to telecommute less (− 0.426‡) after the pandemic compared to part-time workers, likely because many part-time jobs offer more flexibility, and are thus more conducive to telecommuting (Felstead and Henseke 2017).

Tech savviness again comes into play for telecommuting (0.256†) post-pandemic. Conversely, household size plays a negative impact on telecommuting (-0.126†), possibly because finding a quiet space to work in larger households can be difficult.

There were no indirect effects for this equation, so total effects equal direct effects here.

Conclusions

In this paper, we estimated three structural equation models to assess the impact of COVID-19 on the frequency of telecommuting in California before, during, and potentially after COVID-19. Our dataset was collected in late May 2021 via a survey of Californians in KnowledgePanel© conducted for us by Ipsos. Compared to papers published until the end of 2022, our study covers a longer period of the pandemic (March 2020 to late May 2021), and our respondents are representative of the California population, which enables us to generalize our results to the whole state.

Our results show some generational impacts (for Generation X), but no gender and race effects. However, workers with more education started telecommuting more during the pandemic, a trend that is likely to continue post-pandemic. As expected, occupation type and full time work status matter (full time workers are less likely to telecommute). Although household income has no direct impact on telecommuting, it had significant indirect and total effects before and during the pandemic (higher income workers telecommuted less). Household size and the presence of children also matter, but their effect is complex. Finally, although some residential land use variables are significant, their impact is small, and so is the magnitude of residential self-selection.

The nature of an occupation plays a key role in telecommuting both during and potentially after the pandemic since ICT-supported jobs are suitable for telecommuting. Our results show that during the pandemic workers in Education and Engineers / Architects / Lawyers / Social Scientists were more likely to work from home, whereas heath care workers (e.g., nurses) had to be at work in person. However, after the pandemic workers in healthcare, education, trades / transport / construction / installation / repair, and sales and services are expecting to telecommute less since most of these jobs require their presence in the workplace.

Overall, our results suggest that an additional 4.2% of Californian workers could engage in some level of telecommuting post-pandemic which is substantial but much less than suggested by Conway et al. (2020), who analyzed data from a 2020 nonprobability US sample.

Although just over a third of the respondents to a 2022 Gallup poll conducted in the US expressed a preference for working fully remotely (versus 9% in 2019) (Wigert and Agrawal 2022), many firms (including at tech firms such as Salesforce, Google, and Meta) want their employees back in the office (Thier 2023), partly over concerns about training opportunities for new hires, worries about work coordination, and apprehension about productivity decreases.

US employers are not the only ones to backtrack on remote work. Mandated office returns are happening all over world, particularly in the US, Singapore, and Australia, although not as much in the United Kingdom and Ireland (Unispace 2023). A 2023 survey of 17 countries indicates that hybrid work is currently most common: in early 2023, employees spent on average 3.5 days per week in the office, with a high of 3.8 days in Hong Kong, although the percentage of employees who do not want to return full time is higher in Hong Kong, than in the UK, Singapore, and even the US (Unispace 2023). There is therefore a discrepancy between what employees would prefer (2.8 days per week in the office globally) and the current situation.

As argued by Lee and De Vos (2023), future frequencies of telecommuting depend on attitudes, past frequencies of telecommuting, and constraints such as the nature of an occupation, the support of employers, and the characteristics of their (potential) home office. To support a switch to (at least partial) telecommuting, employers may consider offering a mix of in-person and remote work, which would allow workers to maintain or create ties with colleagues while reducing their commuting expenses. Although employer decisions will play a major role in defining the future forms and adoption of telecommuting, employee preferences and constraints, such as access to appropriate technologies to work from home and the home environment are clearly important (Tahlyan et al. 2022).

In 2021, ~ 91% of California households had access to high-speed internet, and ~ 85% of California residents used a desktop, laptop, or tablet to connect to the internet, but income remains a digital gatekeeper as ~ 29% of households who earn under $40,000 a year have no internet connection or have internet access only through a smartphone (Mackovich-Rodriguez 2021). Although only around half of all jobs in California are suitable for telework (US Census Bureau 2019), the state should continue its efforts to give broadband access to all Californians (see EO N-73–20, the Governor’s 2020 “Broadband For All” Executive Order, and the December 2020 Broadband For All Action Plan), because, in addition to telework, fast access to the internet opens the door to telemedicine, cultural programs, education opportunities, and better online shopping.

As mentioned in the introduction, one of the initial motivations for promoting telecommuting was the desire to reduce traffic congestion. In California, the South Coast Air Quality Management District (the regulatory agency responsible for improving air quality in Los Angeles, Orange, Riverside, and San Bernardino counties) allows firms to use telecommuting as part of a menu of options to reduce VMT under Rule 2202, which applies to worksites with 250 or more employees. In 2021, California also published Statewide Telework Policy 0181, whose purpose is to provide a structure to establish effective telework programs that incorporate telecommuting as a work option (California Department of General Services 2021). Monetary incentive were also put in place, including stipends for represented state employees under the Telework Stipend Program (California Department of Human Resources 2022). Some lawmakers tried to go further and proposed in early 2022 an income tax credit (which was not adopted) funded by the Greenhouse Gas Reduction Fund of $1,000 for telecommuting at least 25 h per week. However, monetary inducements for telecommuting run counter state and local tax breaks granted to many large employers for locating some of their facilities in California to spur economic activity and increase the tax base by bringing in well paid jobs. Many of these agreements, which were concluded well before COVID-19, do not consider that a substantial percentage of the workforce at these worksites could work remotely, which would sharply limit the intended benefits of these tax breaks (Spring 2023). It would therefore make sense to revisit some of these agreements (and not just in California) to better reward firms that hire local telecommuters and discourage hiring out-of-state telecommuters who do not contribute to the local or state tax base.

One limitation of this study is that our dataset does not contain the exact residential location of our respondents (we just know their ZIP code). We also know the work ZIP code of only a subset of Californian workers in our dataset, although models estimated on that subset showed that commuting time to work was not statistically significant. A more extensive dataset that captures time use and travel behavior over several days is needed to better explore the impact of commuting time on the decision to telecommute.

There are multiple avenues for future research. First, ongoing analysis is needed as behaviors are still shifting as the pandemic is waning. For example, as a number of companies have allowed their workers to work from home (Howington 2023), some households moved to more affordable areas, possibly out of state (Walczak 2021) or even abroad (Masterson and Shine 2022) because they were attracted by relocation incentives. Capturing changes in residential location and travel to investigate the long-term impacts of the pandemic is of interest but will take a longer time frame. Second, it would be useful to examine the impact of attitudes and perceptions (related, for example, to productivity at home, or impacts of telecommuting on family life) on telecommuting since they often affect the decision to telecommute. Finally, we agree with Elldér (2020) about the value of conceptualizing telecommuting as a coping strategy for organizing everyday activities, which suggests that it should be analyzed in the context of daily activities.