Does Working from Home Improve Employees’ Productivity? Empirical Evidence from a Meta-Analysis

Working from Home and Productivity

A large literature has examined the impacts of working from home (WFH) on employees. Teleworking can improve employee well-being by reducing stress, increasing happiness, and enhancing job satisfaction (Shamir and Salomon 1985; Anderson, Kaplan, and Vega 2015; Kazekami 2020), making it an increasingly desirable option in modern labor markets (He, Neumark, and Weng 2021). From the employee’s perspective, WFH often fosters a stronger sense of autonomy (Diab-Bahman and Al-Enzi 2020) and reduces turnover intention (Choi 2020; Tsen, Gu, Tan, and Goh 2021).

However, while many employees welcome WFH opportunities, the arrangement is not without challenges. Critics point to psychosocial downsides, such as reduced informal interaction with colleagues, which can lead to loneliness (Walz, Kensbock, De Jong, and Kunze 2024). WFH can also increase domestic tensions. For instance, overlapping demands from children or other family members may create stress that undermines the benefits of WFH autonomy (Xu et al. 2022). WFH can also lead to a decline in mentoring and professional development opportunities, especially for early-career workers (Yarberry and Sims 2021) and can reduce the likelihood of promotion and impair career advancement (Bloom, Liang, Roberts, and Ying 2015). More broadly, research indicates that blurred boundaries between work and home responsibilities can have detrimental effects on life satisfaction and mental health (Senik et al. 2024; Bertoni, Cavapozzi, Pasini, and Pavese 2025).

WFH participation was increasing even before the pandemic, with growing uptake observed across European economies (Brenke 2014; Vilhelmson and Thulin 2016) and in the United States (Lister and Harnish 2011). However, the onset of COVID-19 created a dramatic inflection point when millions of workers worldwide were asked, or required, to perform their jobs from home (see, e.g., Garcia 2022; Scheibe, De Bloom, and Modderman 2022). Unlike pre-pandemic adoption, this transition was often involuntary and universal, creating what some scholars have called a forced experiment in remote work. It disrupted long-held assumptions about workplace presence and fueled global debates about the future of work.

Post-pandemic, WFH has become a permanent feature of the employment landscape in many countries. While some firms have encouraged or mandated a return to the office, others have embraced hybrid or fully remote models. According to Parker, Horowitz, and Minkin (2022), a large share of employees favor continuing some form of WFH, citing improved work–life balance and reduced commuting costs. In addition, Emanuel and Harrington (2024) noted that workers have made both physical and human capital investment when setting up a home office and adapting to remote work that they may be unwilling to abandon. Meanwhile, employers are engaged in active discussions about how much flexibility to retain (Dennison 2024), especially in light of uneven productivity outcomes during the pandemic years.

From an organizational standpoint, teleworking offers tangible operational benefits. Reducing physical office space can lower rent, utilities, and maintenance expenditures (Morgan 2004; El Hajal 2022). Remote work policies also enable firms to access broader talent pools, allowing firms to recruit candidates from faraway regions when the local labor market does not offer a requisite skill set (Choudhury 2022). This decoupling of employment from geography has major implications for global hiring practices and wage convergence across locations. Nonetheless, telework also introduces managerial and coordination challenges. Informal knowledge-sharing—the hallway conversation or impromptu brainstorming session—becomes rarer, and some firms report declines in innovation speed or team cohesion (Taskin and Bridoux 2010).

Trust and oversight also emerge as key managerial concerns. Parker, Knight, and Keller (2020) noted that the shift to WFH has required new norms for accountability and evaluation. In traditional settings, performance could be inferred through visibility and interaction; remote contexts demand alternative performance metrics. The lack of physical oversight raises fears of shirking, though empirical evidence on this issue remains mixed.

Ultimately, the viability of WFH hinges on its effect on productivity. If productivity increases or remains stable, then the cost savings and recruitment advantages should lead to net gains for firms. But if output declines due to coordination delays, communication inefficiencies, or employee disengagement, then the model becomes financially and strategically questionable. For this reason, scholars have sought to empirically assess whether WFH enhances or hinders productivity. The results, however, remain inconclusive.

Several studies have proposed psychological and logistical channels through which remote work may enhance performance. Remote workers frequently report greater intrinsic motivation and job satisfaction (Bailey and Kurland 2002; Giménez-Nadal, Molina, and Velilla 2018), and these affective states have been linked to higher productivity (Oswald, Proto, and Sgroi 2015). WFH arrangements often afford employees the opportunity to work during peak energy times, avoid workplace distractions, and manage tasks in a sequence that best suits them. Kazekami (2020) and Metselaar, den Dulk, and Vermeeren (2023) argued that task autonomy and flexible scheduling reduce stress and increase output. Furthermore, avoiding daily commutes saves time and energy, which can then be redirected to core job functions (Arntz, Ben Yahmed, and Berlingieri 2022). These mechanisms are particularly salient in urban settings with long commuting times or unreliable transit.

Still, there are countervailing channels that may suppress productivity. The most frequently cited is reduced monitoring. Classic organizational theory posits that in the absence of supervision, workers may shirk (Causer and Jones 1996; Pyöriä 2011). While technology now enables some degree of monitoring (e.g., keystroke tracking or virtual check-ins), such practices can erode trust and trigger resistance. In parallel, household distractions, from children, pets, or shared spaces, can hinder focus and task completion (Mann and Holdsworth 2003; Delanoeije, Verbruggen, and Germeys 2019; Yang, Kelly, Kubzansky, and Berkman 2023). These issues are particularly relevant in multigenerational households or smaller living quarters.

Reduced in-person collaboration can also lead to communication lags and weaker problem-solving capabilities. Complex projects requiring real-time feedback or creative brainstorming may suffer in remote-only environments (Fonner and Roloff 2010; Sarker, Sarker, Xiao, and Ahuja 2012). Baruch (2000) highlighted the erosion of informal mentoring and team identity when work becomes distributed. This concern is echoed in studies of younger workers, who often rely more heavily on informal learning and feedback loops than their more experienced peers.

Empirically Assessing WFH and Productivity

Because WFH introduces both productivity-enabling and productivity-constraining mechanisms, its overall impact remains theoretically ambiguous. Consequently, researchers have pursued both qualitative and quantitative approaches to assess the net effect. Some rely on interviews, reflective surveys, or case studies (Bosua, Kurnia, Gloet, and Mendoza 2017; Korkeakunnas, Heiden, Lohela-Karlsson, and Rambaree 2023), while others estimate empirical models of the form:

P r o d u c t i v i t y i = β W F H i + γ X i + ε i ,

(1)

where productivity differences among employees (i) are modeled as a function of WFH status, other observable factors affecting productivity ( X i , discussed below), and an error term ε i . However, the literature has failed to reach a consensus on the size and direction of β . Some studies found that WFH enhances productivity (e.g., Neirotti, Paolucci, and Raguseo 2012; Bergeaud, Cette, and Drapala 2023); others concluded that it reduces it (Giménez-Nadal et al. 2020; Kitagawa, Kuroda, Okudaira, and Owan 2021a). Still others found no statistically significant effect or results that vary depending on specification and sample (Arntz et al. 2022; Gibbs, Mengel, and Siemroth 2023; Blind, von Mandach, Kotosaka, and Chiavacci 2024). This lack of consensus underscores the need for a systematic meta-analytic review.

These discrepancies may reflect differences in research design. Observational studies often use ordinary least squares (OLS) regression to estimate WFH impacts (e.g., Pabilonia and Vernon 2022). Some researchers argue that selection into telework is endogenous (Emanuel and Harrington 2024) and therefore attempt to correct for selection bias using quasi-experimental techniques such as instrumental variables or propensity score matching (Pigini and Staffolani 2019; Braun, Cyronek, and Rupert 2021; Kawaguchi et al. 2022). More recently, a few studies have employed randomized controlled trials (RCTs) in which workers are randomly assigned to either WFH or on-site arrangements (Atkin, Schoar, and Shinde 2023; Choudhury, Khanna, Makridis, and Schirmann 2024). These studies provide cleaner identification but face practical and generalizability limitations because they each take place within a single firm, so it is not clear to what extent the outcomes may be driven by unique management, corporate culture, or other structural characteristics of the specific organization. Our meta-analysis will evaluate various research designs and investigate whether estimates differ systematically across them.

Another potential cause of the lack of consensus may be that many studies assume a homogeneous productivity effect across all types of workers. Research suggests, however, that the WFH–productivity relationship is moderated by several factors. First, productivity may depend on the intensity of telework. Hybrid workers often show better outcomes than fully remote workers (Kazekami 2018; Giménez-Nadal and Velilla 2024), perhaps because they benefit from both autonomy and team interaction.

Second, effects may vary by sex or family status. Pabilonia and Vernon (2022) reported larger productivity gains for males than females, while Angelici and Profeta (2020) found that parents, especially those with young children, gain more from teleworking. These differences may reflect variations in caregiving roles, work–life conflict, or household support structures.

Third, effects may depend on sector and occupation. Most studies pool sectors, but some focus on information and communications technology (ICT) (Gibbs et al. 2023), manufacturing (Kitagawa, Kuroda, Okudaira, and Owan 2021b), services (Aksoy et al. 2025), or public employment (Choudhury, Foroughi, and Larson 2021). WFH may be more productive in knowledge-intensive or digital roles and less suitable in client-facing or manufacturing settings (Kazekami 2018; Rupietta and Beckmann 2018). However, few studies conduct cross-sectoral comparisons. The nature of the job also plays a role. Some occupations have lower teleworkability due to physical labor or high interdependence among workers; others (e.g., software engineering, writing, data analysis) can more easily be performed at home. Unfortunately, most studies fail to consider teleworkability, making it difficult to generalize findings with respect to occupations.

Fourth, national context may matter. Most studies analyze a single country, such as France (Bergeaud et al. 2023), Germany (Arntz et al. 2022), Italy (Neirotti et al. 2012), Japan (Kawaguchi and Motegi 2021), the United Kingdom (Deole, Deter, and Huang 2023), or the United States (Gibbs et al. 2023). Baseline productivity levels, digital infrastructure, and institutional settings vary widely across countries (Bernard and Jones 1996; Naor, Linderman, and Schroeder 2010). If such differences moderate the WFH–productivity relationship, that may account for some of the heterogeneity observed across study outcomes. A few studies pool data across countries (e.g., Ono and Mori 2021; Barczyk, Spychała, Urbanowicz, and Ziomek 2022), but none to date have attempted to make comparisons across countries.

Methodological Differences across Studies

The empirical literature examining the impact of WFH on productivity has employed a range of methods. Productivity measurement is inconsistent across studies and may account for some of the variation in observed results.

Some studies measure objective task output in a specific organization (Choudhury et al. 2021; Ko and Baek 2025), which accurately reflects the employee’s true production, but limits generalizability since productivity outcomes may reflect unique institutional or managerial factors, and tasks vary across occupations and firms.

Others use revenue (Monteiro, Straume, and Valente 2021; Kawaguchi et al. 2022), which is easy to observe and comparable for employees across many firms. However, revenue analyses rely on an assumption that the price of output remains constant and is the same across firms so that differences in revenue reflect differences in output only; the studies generally do not have data on prices to confirm the validity of this assumption.

Alternatively, many scholars have instead used either hourly earnings (e.g., Pabilonia and Vernon 2022; Birch and Preston 2025;) or total income (Kawaguchi and Motegi 2021; Deole et al. 2023). Measuring productivity using income is justified using the neo-classical assumption that employees are paid based on their marginal product. It has a major advantage in that data are often available from large representative surveys conducted by national statistics agencies and is easily comparable across jobs and industries. However, there are multiple disadvantages to measuring productivity using income. First, income can reflect market conditions and employee bargaining power, so is not a perfect proxy for output. Second, with respect to hourly earnings specifically, this measure does not account for the fact that teleworkers may work a different number of hours than non-teleworkers (see, e.g., Rupietta and Beckman 2018; Lyttelton, Zang, and Musick 2022) so that differences in hourly earnings may not fully represent differences in overall output between teleworkers and non-teleworkers. The use of total income should correct for this problem because it captures both hourly differences and differences in hours worked in a single measure; however, it obscures whether an earnings difference results from hourly productivity, working hours, or both.

Therefore, some scholars simply examine the change in total working hours as an indicator of productivity (e.g., Giménez-Nadal et al. 2020; Pabilonia and Vernon 2022). This measure requires the assumption that productivity per hour is either unchanged or at least changes in a way that leads to greater overall output when hours of work increase. Given the negative productivity channels described above, however, it is possible that an increase in hours worked could simply be a response to lower productivity per hour when teleworking, so that employees’ actual output is unchanged or possibly lower despite the extra working time, something this measurement method does not capture, and leads back to the importance of also considering hourly earnings and total income as alternative measures.

Last, many studies use a Likert scale (Kitigawa et al. 2021a; Chen et al. 2023; Deole et al. 2023), asking respondents to rate their productivity. This method has two advantages. First, it is a holistic measure that avoids all the specific disadvantages of the other methods listed above. Second, several national statistics agencies have adopted this measure in their surveys so it is available for large representative samples; the recent adoption by national statistics agencies also explains why it is commonly employed in the most recent works on this topic. However, it has its own disadvantages in that it provides a less exact measure of productivity changes and may suffer from measurement error if respondents perceive productivity differently from each other. Overall, the literature has used many distinct measures of productivity, each with its own advantages and disadvantages; we will investigate whether the measurement method affects the estimation of the WFH–productivity relationship.

Measurement of WFH itself also varies. A common approach is to use a binary indicator for participation (e.g., Pigini and Staffolani 2019), but this approach fails to capture variation in WFH frequency or duration. A better alternative is to measure the proportion of time spent teleworking (Kazekami 2020), which allows researchers to estimate the marginal effect of WFH intensity. Our meta-analysis compares results across both binary and continuous WFH measures.

Finally, model specification, that is, which control variables are included, may also matter. Some studies control for a worker’s commute distance (e.g., Gibbs, Mengel, and Siemroth 2021), arguing that productivity gains from avoiding commuting should be greater for workers with longer commutes. However, others do not include commute distance (e.g., Pigini and Staffolani 2019; Monteiro et al. 2021). Some control for firm size (e.g., Rupietta and Beckmann 2018; Arntz et al. 2022) since firm size is often related to employee productivity and employees in larger firms are more likely to participate in WFH (Morikawa 2021); however, others do not account for firm size (e.g., Angelici and Profeta 2020; Giménez-Nadal et al. 2020). Some include employees’ education level (e.g., Pigini and Staffolani 2019; Kazekami 2020), since productivity can differ by educational attainment; however, others do not consider an employee’s education level (e.g., Choudhury et al. 2021). Last, some control for whether the employee has time flexibility in their job (e.g., van der Lippe and Lippényi 2020; Okubo, Inoue, and Sekijima 2021), that is, the ability to perform tasks at a time best for them rather than being forced to work a fixed schedule; those with time flexibility should be more productive because they can optimize the timing of their work. However, other studies do not control for whether an employee has time flexibility (e.g., Arntz et al. 2022; Deole et al. 2023). If any of these factors moderates the productivity effect of telework, their exclusion from the model specification would bias the estimate of the WFH–productivity relationship and could account for differences in results across studies. We will evaluate the potential for omitted variable bias below.

In summary, the WFH–productivity literature is broad and diverse with respect to sampled populations, study methodologies, and, ultimately, results. We therefore undertake a meta-analysis to provide a clearer picture of when, where, and for whom WFH improves or diminishes productivity.

Publication Bias

Because a meta-analysis relies on observing the results of other scholars’ work to calculate an average effect, when researchers choose not to report findings containing statistically insignificant results, it can lead to a problem known as publication bias. In that case, our inability to observe some estimates could lead to a bias in the calculation of the averages reported in the previous section. In particular, if scholars suppress many negative but insignificant results, it will inflate the average observed value of the relationship. Conversely, suppressing many positive but insignificant results could result in either an underestimation or overestimation of the relationship depending on the magnitude of the effect sizes being excluded.

Publication bias is often assessed using a funnel plot, in which the PCC for each of the k = 1, 2, . . ., K collected estimates is shown on the x-axis and its precision (1/SE) on the y-axis. By the property of a normally distributed sampling distribution, the individual estimates should be clustered around the true population effect. At high levels of precision, that is, low standard errors, the estimates should be tightly clustered. As the precision level drops, that is, the standard error gets larger, the range of the observed estimates should expand. In the absence of publication bias, the plot should therefore appear as a symmetrical inverted funnel, hence the name “funnel plot.” When publication bias is present, though, some observations will be missing and so the plot would appear asymmetric; in particular, observations with small values but high standard errors are more likely to be missing as they would be statistically insignificant and therefore more likely to be suppressed by primary researchers. Figure 2 shows the funnel plot for the pre-pandemic and (post-)pandemic samples. A visual inspection finds each to be slightly asymmetric.

OPEN IN VIEWER

Figure 2.

Funnel Plot of Partial Correlation Coefficients (PCC)

We can empirically test for publication bias using the funnel asymmetry test (FAT) (see Stanley and Doucouliagos 2014):

PC C k = β 0 + γ S E k + ν k

(6)

The FAT is the statistical test of the slope coefficient γ . In the absence of publication bias, there should be no relationship between the PCC values and their standard errors, that is, at each level of precision, estimates should cluster around the same overall mean. However, when the estimate of γ is statistically significant, that indicates the average value of PCC differs by precision level, meaning that some observations are missing from the data that would be required to ensure the average PCC remained constant at all levels of precision. These missing observations are the sign that publication bias is present. A positive γ indicates that scholars are more likely to publish estimates exceeding the true average PCC, while suppressing weaker/insignificant results. By contrast, a negative γ indicates that scholars prefer estimates with values below the true average PCC.

Equation (6) can also be used to perform the precision effect test (PET). The PET is the statistical test of the intercept term. When the intercept is statistically significant, we can conclude that the WFH–productivity relationship is non-zero. The estimated coefficient for β 0 can be interpreted as the population average relationship, since it is the estimated PCC when SE = 0, that is, when there is perfect precision (Imai, Rutter, and Camerer 2021). However, the relationship modeled in Equation (6) is linear and the relationship between the PCC and SE is likely to be non-linear (Doucouliagos, De Haan, and Sturm 2022), leading to a downward bias in the estimation of the population mean effect β 0 (Stanley 2008). We therefore modify Equation (6) to adjust for non-linearity as shown in Equation (7). This method is known as the precision effect estimate with standard error (PEESE) (Stanley and Doucouliagos 2014):

PC C k = β 0 + γ S E k 2 + ν k

(7)

One issue of concern is the impact of the COVID pandemic on researchers’ interest in the topic of WFH and productivity. It is possible that researchers, referees, and editors in the (post-)pandemic period may have become less selective given the increased desire from the scholarly community for any information on this topic. This concern is not relevant to studies employing (post-)pandemic data since, by definition, they must have been published in the (post-)pandemic period. However, some studies employing pre-pandemic data were published pre-pandemic while others employing pre-pandemic data have been published more recently. Stanley and Doucouliagos (2014) demonstrated that it is possible that a subset of estimates may be subject to a different level of selectivity and publication bias, so we therefore employ Equation (8) to test for a difference in publication bias within the pre-pandemic data based on date of publication:

PC C k = β 0 + γ S E k 2 + θ ( S E k 2 × P a n d e m i c P u b l i c a t i o o ) + ν k

(8)

where the square of the standard error is interacted with an indicator variable identifying studies published in 2020 or later years. A statistically significant θ will indicate that, among studies employing pre-pandemic data, the degree of publication bias differs between studies published up until 2019 and studies published from 2020 onward.

We estimate Equations (7) and (8) using four estimators. First, we employ pooled OLS. Second, we add study fixed-effects which account for any unique characteristics of each study that may affect all estimates drawn from that study; however, a drawback to the fixed-effects estimator is that both the slope and intercept are calculated by demeaning the observations by study, which can lead to an inflated estimate of the publication-bias-adjusted effect. Therefore, third, we employ a between-effects estimator to compare the average from each study rather than the individual estimates from each study. While the fixed-effects estimator controls for each study’s unique characteristics but otherwise gives each collected estimate the same weight in the regression, the between-effects estimator gives each study the same weight in the regression, meaning that primary studies providing larger numbers of observations will not more heavily influence the results of the analysis; the between-effects estimator also does not inflate the publication-bias-adjusted estimate in the way the fixed-effects estimator does. Fourth, we employ the more recently developed meta-analysis instrumental variables estimator (MAIVE) (see Irsova, Bom, Havranek, and Rachinger 2023), which uses the inverse of the square-root of the number of observations from the primary study as an instrument for the standard error of PCC. The MAIVE estimator is designed to reduce the likelihood of spurious results that may occur because precision is estimated within the primary studies and can be affected by the modeling choices of the researchers. By addressing this issue, the MAIVE can provide a less biased estimate of β 0 than the other estimators in cases in which primary researchers have engaged in p-hacking. ¹

Estimates of Equation (7) for the pre-pandemic and (post-)pandemic data sets are shown in Table 2, panels A and C, respectively. Estimates of Equation (8) for the pre-pandemic data set are shown in panel B. In all models, standard errors for the coefficients are clustered by study. Three of the four estimators indicate the presence of publication bias in the pre-pandemic data; however, we do not find sufficient evidence to conclude that publication bias differs based on whether the studies themselves were published before or during/after the pandemic. All estimators indicate publication bias in the (post-)pandemic data. Overall, we find that the publication-bias-corrected estimates of the WFH–productivity relationship are positive and statistically significant, but small for both data sets.

OPEN IN VIEWER

Table 2.

Analysis of Publication Bias (PB)

	OLS	FE	BE	MAIVE
Panel A. PET-PEESE analysis of Pre-pandemic data
Presence of PB	6.058*** (1.653)	−1.342 (3.285)	13.970** (5.523)	7.009*** (2.489)
Estimate of the effect size	0.014*** (0.003)	0.019*** (0.002)	0.0002 (0.009)	0.013*** (0.004)
Number of observations	2,228	2,228	2,228	2,228
Number of studies	46	46	46	46
Panel B. PET-PEESE analysis of Pre-pandemic data
Presence of PB	6.877 (6.353)	0.161 (2.112)	57.315*** (20.784)	3.366 (3.922)
Presence of PB × Published during pandemic	−0.893 (6.459)	−2.537 (5.646)	−43.917** (20.361)	3.819 (4.516)
Estimate of the effect size	0.014*** (0.003)	0.020*** (0.003)	−0.007 (0.009)	0.014*** (0.004)
Number of observations	2,228	2,228	2,228	2,228
Number of studies	46	46	46	46
Panel C. PET-PEESE analysis of (Post)-Pandemic data
Presence of PB	25.039*** (6.169)	8.374* (4.186)	29.778** (12.828)	29.970* (17.270)
Estimate of the effect size	0.025*** (0.006)	0.032*** (0.002)	0.022* (0.011)	0.023** (0.009)
Number of observations	1,346	1,346	1,346	1,346
Number of studies	43	43	43	43

Notes: Standard errors (SEs) are shown in parentheses. For a full description of the precision effect test- precision effect estimate with standard error (PET-PEESE), please refer to Stanley and Doucouliagos (2014). In panels A and C, we estimate the model PCC_i = β₀+β₁(SE_i)²+v_i using four methods: 1) ordinary least squares (OLS), 2) fixed-effects (FE), 3) between-effects (BE), and 4) the meta-analysis instrumental variable estimator (MAIVE), in which the inverse of the number of observations from the primary study is used as an instrument for the standard error. In each estimation, we cluster standard errors by study. In panel B we consider whether publication bias might differ among the studies using pre-pandemic data based on date of publication by interacting SE² with an indicator variable denoting whether a study was published during the pandemic, that is in 2020 or later.

***

, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

The analysis presented in this section assumes that publication bias can be modeled as a continuous function of the standard error; however, that may not necessarily be the most appropriate specification. In Online Appendix C, we therefore consider alternative methods for estimating a publication-bias-adjusted WFH–productivity relationship under various assumptions about the nature of the bias. The results for these alternative estimates are similar to our main estimates for the pre-pandemic period, and slightly stronger than our main estimates for the (post-)pandemic period. While the techniques reported in this section and in Appendix C adjust for publication bias, they do not account for heterogeneity in study methods, which may also affect the estimation of the relationship. We therefore employ meta-regression analysis to produce estimates of the WFH–productivity relationship that are adjusted for both publication bias and methodological differences.

Meta-Regression Analysis

Although the results of the previous section control for the presence of publication bias, they do not address heterogeneity resulting from differences in study methods. We therefore employ meta-regression analysis (MRA), which is commonly used to identify systematic differences among collected estimates. To address methodological differences across studies, we can modify Equation (8) to include additional factors (Z) that may affect the estimation of the WFH–productivity relationship, as shown in Equation (9):

PC C k = β 0 + γ S E k 2 + θ ( S E k 2 × P a n d e m i c P u b l i c a t i o o ) + β 1 Z k + ν k

(9)

The inclusion of S E 2 in Equation (9) corrects for publication bias as described above while the inclusion of Z will produce an estimate of the WFH–productivity relationship that is less biased than the methods described in the previous section (see Kvarven, Strømland, and Johannesson 2020; Doucouliagos et al. 2022). Although we did not find a statistically significant relationship for the interaction between S E 2 and the date the study was published in the previous section, we nevertheless retain the interaction term here in case there is a relationship that we have not yet been able to identify.

A common problem in meta-analysis is that we face uncertainty with respect to the factors that should be included in Z. Simply including all potential factors can lead to multicollinearity problems and does not necessarily produce estimates that are superior to parsimonious specifications. General-to-specific (e.g., Minviel and Latruffe 2016; Bijlsma, Kool, and Non 2018) and alternative-specification strategies (e.g., Neisser 2021) that use trial-and-error to identify an effective model specification have been used to address the model uncertainty problem; however, these strategies may exclude an important variable at some stage. Instead, we employ Bayesian model averaging (BMA) to estimate Equation (9) (see Online Appendix E for a technical description of BMA). Rather than relying on a single model for statistical inference, BMA involves estimating multiple candidate models using various combinations of potential explanatory variables. Each candidate model is assigned a posterior model probability (PMP) indicating the likelihood that it represents the true structural relationship. The PMPs are used as weights to construct a weighted average—or posterior mean—of the estimated coefficients for each variable. The PMPs from each model in which a variable appears are also summed to produce a posterior inclusion probability (PIP) for that variable. A PIP greater than 0.5 indicates that the variable likely belongs in the true structural model.

The derivation of PMPs and corresponding PIPs requires prior assumptions about the distribution of the parameters and candidate models. For the parameters, we employ the commonly used unit-information g-prior (UIP) (see Eicher, Papageorgiou, and Raftery 2011) that offers the prior the same weight as one observation of the data (Kass and Wasserman 1995; Raftery 1995). A commonly used prior assumption for the distribution of the models is the uniform prior, a naïve assumption that all potential models have equal likelihood of being the correct one. We test the uniform assumption in Appendix E, but, for the main results presented below, we adopt a more advanced prior assumption, the dilution prior, that better addresses potential collinearity among the variables, which can lead to model space redundancy in which there are too many similar models composed of correlated variables. Unlike the uniform prior that gives all potential models equal weight, the dilution prior re-weights the PMPs proportionately to the determinant of the correlation matrix of the included variables (George 2010). For candidate models with little collinearity among included variables, the determinant is near 1 and the PMP changes very little. However, for candidate models with high collinearity, the determinant is closer to 0 and the individual model is down weighted. Because the calculation of the PMPs and corresponding PIPs is influenced by our chosen prior assumptions, we present robustness checks employing different priors in Appendix E.

To correct for publication bias, we include the square of the standard error of the PCC in all candidate models. In addition, we consider 33 candidate variables (34 including the SE–publication date interaction for the pre-pandemic data) that may affect the WFH–productivity relationship. These include differences in sampled populations such as frequency of working from home, sex, parental status, sector, and country; and methodological differences such as the measures of productivity and WFH, study design, and characteristics of the data collection process. We also include the logarithm of the h-index of the publication in which the study appears to assess whether results differ between high-quality outlets, low-quality outlets, and working papers, and thus whether referees and editors at higher ranked journals are more selective. Finally, we include the year the study’s data were collected to examine whether the relationship has been strengthening over time as WFH technology has improved. Table A.3 in Online Appendix D provides definitions and summary statistics for each of the variables considered. We have 2³³ = 8.6 billion possible model combinations to consider. To simplify the calculations, we use a Markov chain Monte Carlo (MCMC) algorithm to identify the best candidate models, as in Bajzik, Havranek, Irsova, and Scharz (2020) and Zigraiova, Havranek, Irsova, and Novak (2021). We calculate 5 million potential models and use the best 10,000 to construct averages for each variable’s coefficient and PIP.

Figure 3 provides an overview of the model space employed in BMA. Panel A presents results for the pre-pandemic data and panel B presents results for the (post-)pandemic data. Each column represents one candidate model ordered by PMP. We show S E 2 in the first row because it is included in all models. The candidate variables are then shown by row in order by PIP. A blank cell indicates that the variable is not included in a particular model. A blue cell (darker cell in grayscale) indicates a positive coefficient in the estimated model while a red cell (lighter cell in grayscale) indicates a negative coefficient. Most variables appear in most models and are shown in only one color, indicating that the direction of their estimated effect on PCC remains constant across the models. The variables that appear near the bottom of the figure are only present in models with a low likelihood of representing the true structural relationship; the signs of some of these variables flip, as demonstrated by the color changes, showing an inconsistency in their estimated effect on the PCC.

OPEN IN VIEWER

Figure 3.

Model Inclusion in Bayesian Model Averaging for the WFH–Productivity Relationship

The BMA results are reported in Table 3 with panel A displaying results for the pre-pandemic data and panel B displaying results for the (post-)pandemic data. The first, second, and third columns of each panel show the posterior mean, posterior standard deviation, and PIP values, respectively. For the pre-pandemic data, BMA identifies 22 of the candidate variables as having PIPs greater than 0.5. We therefore construct a single model containing S E 2 and these identified variables as BMA indicates this specification is most likely to represent the correct model explaining the observed PCC values. We then estimate the model using OLS for three reasons. First, as a frequentist technique, its result does not rely on making prior assumptions about the distribution of the data in the way that BMA does; rather, its result is determined solely by the data. Second, it allows us to perform standard hypothesis tests of the statistical significance of the estimated coefficients. Third, it is simple to employ, and many researchers understand how to interpret its results. For the (post-)pandemic data, BMA identifies 16 variables with PIPs higher than 0.5. We conduct an OLS analysis of this specification also. For both OLS estimations, we cluster standard errors by study and report the p-values.

OPEN IN VIEWER

Table 3.

Analysis of Working from Home (WFH) and Productivity

Outcome variable: PCC	A. Pre-pandemic data					B. (Post-)Pandemic data
	BMA			OLS		BMA			OLS
Meta-explanatory variables	Mean	SD	PIP	Coeff.	p-value	Mean	SD	PIP	Coeff.	p-value
Publication selection bias
SE 2	4.581	1.055	1.000	4.216	0.132	7.253	3.450	1.000	5.871	0.220
SE2 x Published during pandemic	−0.003	0.478	0.049
WFH frequency (reference: WFH any amount)
Hybrid	0.025	0.004	1.000	0.025	0.023	−0.019	0.004	1.000	−0.019	0.003
Remote	−0.020	0.008	0.956	−0.019	0.056	−0.008	0.006	0.710	−0.010	0.300
Gender (reference: studies of males)
Studies of females	−0.012	0.002	1.000	−0.012	0.001	−0.0002	0.001	0.060
Studies combining males and females	0.018	0.004	1.000	0.017	0.048	0.003	0.004	0.424
Parental Status (reference: studies of non-parents)
Studies of parents	0.022	0.003	1.000	0.022	0.000	−0.020	0.012	0.831	−0.028	0.004
Studies combining parents and non-parents	0.026	0.004	1.000	0.025	0.006	−0.012	0.010	0.672	−0.017	0.002
Sector (reference: studies of multiple sectors)
Studies of the ICT sector	0.002	0.006	0.113			0.004	0.009	0.257
Studies of the manufacturing sector	0.00003	0.003	0.046			−0.026	0.011	0.912	−0.028	0.117
Studies of the public sector	0.001	0.004	0.081			0.008	0.019	0.170
Studies of the services sector (excluding ICT)	0.001	0.004	0.108			−0.002	0.005	0.119
Location (reference: United States)
Japan	−0.034	0.007	1.000	−0.032	0.018	−0.048	0.007	1.000	−0.047	0.000
United Kingdom	−0.101	0.008	1.000	−0.102	0.000	−0.022	0.007	0.975	−0.025	0.009
Northern/Western Europe	0.033	0.007	1.000	0.035	0.013	−0.001	0.003	0.082
Southern Europe	−0.003	0.007	0.186			−0.002	0.007	0.130
Other high-income countries	−0.031	0.007	0.998	−0.032	0.073	0.026	0.008	0.983	0.024	0.293
Middle-income countries	−0.005	0.010	0.238			0.001	0.003	0.065
Productivity measure (reference: total income)
Tasks performed	−0.031	0.009	0.984	−0.031	0.045	−0.021	0.006	0.984	−0.019	0.368
Revenue	−0.072	0.010	1.000	−0.076	0.000	−0.0002	0.002	0.031
Hours worked	−0.019	0.003	1.000	−0.027	0.009	−0.00001	0.001	0.026
Hourly wages	−0.027	0.003	1.000	−0.019	0.031	−0.0001	0.001	0.031
Likert scale	−0.056	0.007	1.000	−0.056	0.012	0.019	0.004	0.998	0.020	0.081
WFH measure (reference: binary measure of participation in WFH)
Continuous measure of WFH time	0.015	0.008	0.830	0.017	0.280	0.027	0.006	0.998	0.026	0.039
Selection of controls
Commute distance	0.010	0.004	0.916	0.011	0.062	0.001	0.003	0.150
Firm size	−0.014	0.007	0.848	−0.016	0.116	0.020	0.006	0.976	0.020	0.011
Time flexibility	−0.004	0.007	0.275			−0.005	0.007	0.353
Worker education	−0.019	0.004	1.000	−0.019	0.090	−0.001	0.003	0.104
Assessment of productivity (reference: employee self-assessed)
Employer assessed	−0.004	0.007	0.289			0.000	0.003	0.063
Study Design (reference: observational study)
Quasi-experimental	0.0001	0.001	0.051			−0.031	0.006	1.000	−0.032	0.001
Randomized-control trial	−0.025	0.013	0.837	−0.031	0.062	0.052	0.026	0.844	0.061	0.005
Data type (reference: cross-section data)
Panel data	−0.012	0.006	0.853	−0.015	0.139	−0.026	0.005	1.000	−0.025	0.004
Study characteristics
h-index	0.000004	0.0001	0.048			−0.004	0.001	0.997	−0.004	0.025
Data collection date	0.002	0.0004	1.000	0.002	0.030	0.00001	0.001	0.029
Intercept	−3.587	—	1.000	−3.376	0.031	0.043	—	1.000	0.072	0.000
Number of observations	2,228			2,228		1,346			1,346
Number of studies	46			46		43			43

Notes: Panels A and B display estimates for the pre-pandemic and (post-)pandemic samples, respectively. “Mean” is the posterior-mean in Bayesian model averaging (BMA). “SD” is the posterior-standard deviation in BMA. “PIP” is the posterior inclusion probability in BMA; a PIP > 0.5 indicates that the variable is likely to explain variation in partial correlation coefficient (PCC). We use ordinary least squares (OLS) to check the statistical significance of the variables selected for model inclusion by BMA, with standard errors clustered by study. We report the p-values here. See Online Appendix Table A.3 for a description of the variables.

We draw the following conclusions from Table 3. First, we identify differences in the WFH–productivity relationship between studies that group remote and hybrid teleworkers together and studies that consider remote and hybrid teleworkers separately. The result is similar to the findings of Giménez-Nadal and Velilla (2024) and Kazekami (2018), who identified productivity differences based on time spent teleworking. Our finding indicates that future researchers should separate remote employees from those in hybrid arrangements when estimating the WFH–productivity relationship.

We also find that the relationship differs between males and females in the pre-pandemic data. This result, which is similar to the findings of Giménez-Nadal et al. (2020) and Pabilonia and Vernon (2022), suggests that researchers should separate males from females when estimating the relationship. Nevertheless, we do not find evidence of a sex difference in the (post-)pandemic data, suggesting either that the differential has dissipated or that the uniqueness of the pandemic experience caused a temporary leveling of the productivity effect between males and females.

In the pre-pandemic data, we find that the relationship is higher for parents compared to non-parents. This finding is similar to Angelici and Profeta (2020) and Pabilonia and Vernon (2022), who argued that WFH allows parents to better balance their work and childcare responsibilities, which provides them with an extra productivity boost that non-parents do not receive. This result flips in the pandemic data, however, where non-parents are shown to experience larger productivity gains. The pandemic-era finding is not surprising, perhaps, as lockdowns also led to school closures, so many parents had to simultaneously telework and mind their children, impeding their productivity (Deole et al. 2023). Therefore, future researchers should separate parents from non-parents when estimating the WFH–productivity relationship.

In the pre-pandemic data, we do not find any evidence that the relationship differs by sector as the PIPs for all sector variables are well below the 50% threshold; however, we do find that studies of the manufacturing sector during the pandemic produced lower productivity estimates compared to multi-sector studies. Both Kazekami (2018) and Monteiro et al. (2021) also reported differences across sectors; however, given the limited attempts in the literature thus far to compare sectors and, therefore, the limited number of observations we have for each sector, more investigation of sectoral differences is needed.

Furthermore, we find differences in the size of the relationship across geographic regions. In the pre-pandemic data, studies of Northern/Western Europe report PCCs higher than studies of the United States. We do not identify statistical differences between PCCs from studies of Southern Europe or Middle-income countries and studies of the United States. By contrast, studies of Japan and the United Kingdom report smaller PCCs compared to the United States. In the (post-)pandemic data, studies of Japan and the United Kingdom continue to report lower PCCs than studies of the United States, whereas studies of Northern/Western Europe, Southern Europe, and Middle-income countries are not statistically different from the United States.

In addition to differences by subpopulations, we also find that differences in study methods affect the estimation of the WFH–productivity relationship. In the pre-pandemic data, studies employing total income report higher productivity gains than other measures. By contrast, among studies employing (post-)pandemic data, the Likert measure produces the highest PCCs. BMA does not indicate a difference between studies employing total income, studies employing hours worked, and studies employing hourly earnings. Studies employing tasks performed show a lower average PCC than studies employing total income, but the difference is not statistically significant. Overall, these results indicate that measurement method does impact the estimation of the relationship, so future researchers should attempt to employ multiple measures to check the robustness of their conclusions, as some scholars (e.g., Arntz et al. 2022; Kawaguchi et al. 2022) have done. We find that estimates from studies using a continuous measure of time spent working from home produce larger PCCs than those employing the binary measure, but the difference is only statistically significant in the (post-)pandemic data. In the pre-pandemic data, studies controlling for commute distance produce stronger PCCs than studies omitting this control, while studies controlling for an employee’s education level produce weaker PCCs than studies excluding this variable. In the (post-)pandemic data, studies controlling for firm size produce stronger PCCs than studies ignoring firm size. Therefore, we can conclude that omitting commute distance, firm size, and employee education level can cause omitted variable bias and we recommend future researchers control for all three in model specifications.

We do not find that studies using employer-provided data produce results that differ from studies using employee-provided data, suggesting that both employers and employees perceive productivity the same way. However, we do find some statistical differences between observational, quasi-experimental, and randomized-controlled trial studies. The results indicate that there likely is some selection bias inherent in observational studies, and therefore, WFH should be treated as endogenous. For the pre-pandemic data, we do not find a statistical difference between the use of cross-sectional data and the use of panel data. However, for the (post-)pandemic data, the use of panel data on average produces smaller PCCs compared to the use of cross-sectional data. Because the (post-)pandemic panel data in our sample followed employees through both lockdown and lockdown-relaxation periods, we cannot rule out the possibility that medium-term external factors related to the pandemic were causing productivity to vary, thus weakening the relationship between WFH and productivity in panel studies. We recommend that future research be conducted specifically using post-pandemic panel data to investigate the long-run effect in a non-pandemic setting.

Last, we do not find evidence that studies published in higher ranked journals provide results that differ from those published in lower ranked journals or from working papers in the pre-pandemic data. In the (post-)pandemic sample, however, higher ranked journals have, on average, published weaker results. This finding indicates that referees and editors at higher ranked journals have not been more selective in terms of the estimates they choose to accept when evaluating studies employing pre-pandemic data, which, as we previously showed in Table 1, have typically demonstrated smaller effects. But, referees and editors at higher ranked journals appear more skeptical of (post-)pandemic data demonstrating larger effects, perhaps because these larger effects are not consistent with the previous pre-pandemic literature. In the pre-pandemic data, we demonstrate that the relationship strengthens each year, likely because of technology improvements. While we do not find statistical evidence that this trend continues in the (post-)pandemic data, that may simply be the result of having too few years post-pandemic to ascertain a meaningful difference over time.

Estimating the Work-from-Home–Productivity Relationship

In the previous sections, we have identified both publication bias and methodological differences as factors influencing the outcomes reported in primary studies. Using the meta-regression results, we can correct for these factors when estimating the WFH–productivity relationship. To do so, we follow the example of Elminejad, Havranek, Horvath, and Irsova (2023) and use fitted values from the OLS estimations reported in Table 3 to construct conditional point estimates of the WFH–productivity relationship, shown in Table 4, corrected for both publication bias and the use of best practices in research design. Fitted values are calculated by substituting the following values for each of the explanatory variables into the equations from Table 3. First, to correct for publication bias, we use SE = 0, that is, perfect precision. Second, for individual effects (i.e., WFH frequency, sex, and so on) we substitute 1 for the characteristic of interest and 0 for other characteristics from the same variable group. Third, we assume that estimates are collected using a continuous measure of time spent teleworking and set that variable equal to 1. Fourth, we calculate the estimates assuming the use of either a randomized-controlled trial or a quasi-experimental design to address endogeneity. Fifth, we set the relevant control variables identified by BMA as potentially leading to omitted variable bias (that is, commute distance, firm size, worker education) to 1 to avoid that bias. Sixth, because we have demonstrated that the WFH–productivity relationship strengthened over time in the pre-pandemic data, we calculate fitted values for the pre-pandemic data for the year 2019, the final year in that data series. For all other variables, we substitute their means, providing the meta-average, corrected for publication bias and best methodological practice, with results reported in Table 4.

OPEN IN VIEWER

Table 4.

Conditional Point Estimates

	Pre-pandemic data		(Post-)Pandemic data
	Estimate	[90% Interval]	Estimate	[90% Interval]
All studies	0.024	[−0.012; 0.061]	0.019	[0.008; 0.031]
WFH frequency
Hybrid	0.040	[0.005; 0.074]	0.005	[−0.010; 0.020]
Remote	−0.005	[−0.046; 0.036]	0.015	[−0.003; 0.033]
Sex
Females	0.008	[−0.030; 0.046]
Males	0.020	[−0.019; 0.060]
Parental status
Parents	0.025	[−0.010; 0.061]	0.009	[−0.004; 0.022]
Non-parents	0.003	[−0.032; 0.038]	0.037	[0.022; 0.053]
Sector
Manufacturing			−0.007	[−0.033; 0.019]
Other (Including Multi-sector estimates)			0.020	[0.008; 0.032]
Country
Japan	−0.009	[−0.045; 0.026]	−0.023	[−0.036; −0.010]
Northern/Western Europe	0.058	[0.016; 0.099]
United Kingdom	−0.079	[−0.135; −0.024]	0.012	[0.002; 0.023]
United States, Southern Europe, and Middle-income countries	0.022	[−0.016; 0.060]
United States, Northern/Western Europe, Southern Europe, and Middle-income countries			0.037	[0.019; 0.055]

Notes: We obtain point estimates from the fitted values of the ordinary least squares (OLS) estimation reported in Table 3 employing the variables selected by Bayesian model averaging (BMA) having a posterior inclusion probability (PIP) > 0.5. Fitted values are calculated using the averages of the included variables with the following exceptions. First, we set standard error (SE) equal to 0 (i.e., perfect precision). Second, when estimating for subgroups (i.e., by sex, parental status, sector, and country), the characteristic of interest is set to 1 and any other variables within the same characteristic group are set to 0. Third, to apply best methodological practices, we calculate the estimate assuming a randomized-controlled experiment (i.e., no endogeneity concern) and including all identified control variables (i.e., no omitted variable bias). Fourth, because the BMA and OLS models indicate that the WFH–productivity relationship increases in strength each year in the pre-pandemic sample, we calculate results for the year 2019, the final year before the pandemic. 90% confidence intervals are calculated using standard errors clustered by study. Refer to Online Appendix Table A.3 for a description of the variables and variable groups.

Based on these findings, we can draw the following conclusions. The overall effect of WFH on productivity was not statistically significant in the pre-pandemic period, but is positive and statistically significant, though economically small, in the (post-)pandemic period. Nevertheless, we find that pre-pandemic productivity gains were experienced by hybrid teleworkers; those who worked remotely before the pandemic had productivity levels that were at best the same and possibly lower than those who never teleworked. This finding complements the work of Giménez-Nadal and Velilla (2024) and Kazekami (2018) who found that hybrid teleworkers were more productive than remote teleworkers. During and after the pandemic, we estimate positive coefficients for both hybrid and remote workers, but the wide confidence intervals mean the results are not statistically significant, likely due to having too few observations and thus low precision.

Similarly, when examining teleworking parents and non-parents pre-pandemic, we have large confidence intervals and do not have statistically significant results. Nevertheless, during the pandemic, non-parents saw the largest productivity increases.

While we do not identify sectoral differences in the pre-pandemic sample, we find that productivity gains in the (post-)pandemic sample were concentrated in sectors other than manufacturing; manufacturing, by contrast, did not see a statistically significant difference in productivity between teleworkers and non-teleworkers.

Prior to the pandemic, the WFH–productivity relationship was strongest in Northern/Western Europe; not statistically significant in the United States, Southern Europe, Japan, and Middle-income countries; and actually negative in the United Kingdom. By contrast, in the (post-)pandemic period, teleworking has improved productivity in most places, although the gains in the United Kingdom have been smaller than elsewhere, and in Japan productivity among teleworkers has actually been lower than among office workers.

Overall, the results indicate that the WFH–productivity relationship is positive, at least in some contexts, but the effect is heterogenous across workers. Our analysis has relied on the use of the partial correlation coefficient as a standardization measure to combine the results of heterogeneous primary studies. However, a main drawback of the PCC is that it only indicates the strength of the relationship between the outcome and explanatory variables and does not provide an easily interpretable value for the size of the effect. Therefore, to provide a more useful economic interpretation, and to investigate the actual impact of WFH, in the next section, we meta-analyze subsets of our two samples in which the productivity and WFH measures employed are the same across estimates.

Estimating the Impact of WFH on Productivity Measures

Whereas the PCC analysis presented above is useful for investigating general variations across groups, as well as the impact of different methodological choices on the estimation of the WFH–productivity relationship, it does not provide an easily interpretable economic measure. Therefore, we consider the impact on WFH on specific measures of productivity identified in the literature. To be able to combine estimates from multiple studies while retaining their original units, the studies must use the same measures of productivity and WFH. For productivity measures, we are able to consider the natural logarithm of hourly earnings, the number of hours worked per week, and Likert scale responses for both the pre-pandemic and (post-)pandemic data. We are also able to consider the natural logarithm of total income for the pre-pandemic data, but we have an insufficient number of comparable observations to consider it with the (post-)pandemic data. In all cases, we are limited by sample size to considering estimates employing the binary measure of WFH participation. We follow the same procedure described above to perform the meta-regression analysis for each productivity measure. An explanation of the variables included is provided in Table A.3 in Online Appendix D while the MRA results are provided in Tables A.4 and A.5 in Online Appendix F for the pre-pandemic and (post-)pandemic samples, respectively. We follow the same procedure described above to derive the impact of WFH on each productivity measure, adjusted for publication bias and best practice in methodology, with results reported in Table 5.

OPEN IN VIEWER

Table 5.

Impact of Working from Home (WFH) on Productivity Measures

	ln(Hourly wage)		ln(Total earnings)		Hours worked per week		Likert scale a
	Estimate	[90% Interval]	Estimate	[90% Interval]	Estimate	[90% Interval]	Estimate	[90% Interval]
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
A. Pre-pandemic data
All Studies	0.050	[0.042; 0.057]	0.086	[0.071; 0.100]	2.394	[−0.259; 5.048]	0.081	[0.020; 0.143]
WFH frequency
Hybrid			0.118	[0.106; 0.130]	2.864	[0.164; 5.564]
Remote			0.018	[−0.006; 0.044]	0.750	[−2.543; 4.042]
Gender
Females	0.028	[0.008; 0.048]
Males	0.049	[0.028; 0.071]
Parental status
Parents	0.097	[0.070; 0.124	0.151	[0.117; 0.184]
Non-parents	0.034	[0.025; 0.045]	0.053	[0.038; 0.068]
B. (Post-)Pandemic data
All Studies	0.148	[0.086; 0.211]			9.567	[9.419; 9.715]	0.189	[0.132; 0.246]
WFH frequency
Hybrid							0.117	[0.033; 0.202]
Remote							0.237	[0.191; 0.283]
Gender
Females	0.094	[−0.00003; 0.187]
Males	0.217	[0.165; 0.269]
Parental status
Parents							−0.358	[−0.363; −0.353]
Non-parents							0.199	[0.142; 0.257]

Notes: We estimate Equation (9) using Bayesian Model Averaging (BMA) for each of the productivity measures separately. Employing the variables selected by BMA having a posterior inclusion probability (PIP) > 0.5, we use Ordinary Least Squares (OLS) to perform a frequentist check. The results are reported in Tables A.4 and A.5 in the Online Appendix F for the pre-pandemic and (post-)pandemic samples, respectively. Fitted values are then calculated using OLS results and employing the averages of the included variables with the following exceptions. First, we set standard error (SE) equal to 0 (i.e., perfect precision). Second, when estimating for subgroups (i.e., by sex, parental status), the characteristic of interest is set to 1 and any other variables within the same characteristic group are set to 0. Third, we include all identified control variables (i.e., no omitted variable bias) and assume a quasi-experimental design to address selection bias. Fourth, where BMA and OLS models indicate that the WFH–productivity relationship increases in strength each year, we calculate results for the year 2019 for the pre-pandemic models and the year 2022 for the (post-)pandemic models, as these are the final years in which data was collected for each sample, respectively. 90% confidence intervals are calculated using standard errors clustered by study.

a

5-point Likert Scale.

Pre-pandemic, teleworkers on average earned approximately 5.1% more per hour than non-teleworkers. Both male and female teleworkers earned more than their non-teleworking counterparts, but the difference was larger for males. However, teleworking parents earned 10.2% more per hour than non-teleworkers, whereas teleworking non-parents earned only 3.5% more per hour than non-teleworkers. WFH is shown to increase total earnings by 9.0% overall; however, that is driven by earnings for hybrid teleworkers, which rose by 12.5%. Total income for remote teleworkers did not differ statistically from non-teleworkers. Total income for teleworking parents was 17.0% higher than non-teleworkers while for teleworking non-parents the increase was 5.4%. The discrepancy between hourly and weekly earnings suggests hours worked plays an important role in overall productivity. These income figures indicate that, although WFH raises hourly productivity, teleworkers produce even more over the course of a week because they also work longer hours than non-teleworkers. In fact, that is exactly what column (5) shows.

As displayed in column (5), prior to the pandemic, teleworkers were, on average, working about 2.4 more hours per week than non-teleworkers. However, this result appears to be driven by hybrid teleworkers who, on average, worked 2.9 hours more per week than non-teleworkers. The difference in hours worked between remote workers and non-teleworkers was not statistically significant. This result suggests that hybrid teleworkers may have been using the WFH opportunity as a supplement to time spent in the office, rather than as a substitute for office time, performing additional tasks at home that were not completed at their employer’s premises. This finding suggests that some instances of working from home may in fact have a negative impact on work–life balance for hybrid teleworks, a subject that should be investigated further.

When considering Likert responses in the pre-pandemic sample, teleworkers rated their productivity 0.081 points (on a 5-point scale) higher than non-teleworkers.

Since the beginning of the COVID-19 pandemic, hourly earnings for teleworkers have been almost 16.0% higher than for non-teleworkers; for males the gap between teleworkers and non-teleworkers has been even higher at 24.2%, while female teleworkers still experienced a wage premium of 9.9%. Although we do not have enough estimates to meta-analyze the overall impact on total income post-pandemic, the increase in total income is likely higher than the increase in hourly wages because, as in the pre-pandemic period, hours worked for teleworkers are on average higher (9.6 more hours per week) compared to non-teleworkers. When considering Likert responses, teleworkers rated their own productivity 0.189 points (on a 5-point scale) higher than non-teleworkers. These results appear to be driven more by remote teleworkers who reported average scores 0.237 points higher than non-teleworkers, while hybrid teleworkers reported average scores 0.117 points higher than non-teleworkers. Non-parents rated their productivity 0.199 points higher than non-teleworkers; however, teleworking parents in contrast rated their productivity 0.358 points lower than non-teleworkers, potentially due to the impact of having had to care for children during school closures during the pandemic. We do not have enough observations to distinguish differences across industrial sectors in any of the models or to consider the full range of regional variables included in the PCC analysis above. Furthermore, as shown in Online Appendix Tables A.4 and A.5, we also either do not have sufficient observations to examine, or do not find sufficient evidence of, differences based on frequency of telework, sex, parental status, or other variables of interest in some model specifications. Therefore, further research should be undertaken to examine these issues. Furthermore, the large differences in both hourly earnings and hours worked between teleworkers and non-teleworkers during the pandemic may be influenced by the unique economic circumstances of the pandemic, and so further research should be undertaken post-pandemic to ascertain how much (if any) of that difference still remains.