Work from Home and Productivity: Evidence from Personnel and Analytics Data on Information Technology Professionals

A.  Main Outcome Variables

HCL uses two systems, Sapience Analytics and IDMS (Internal Data Management System), to track employee activity and performance. Our three key outcome variables, Input, Output, and Productivity, derive from these. Managers use analytics reports based on these data to support decisions, goal setting, monitoring, and performance evaluation. Because of its corporate-wide significance, the company has devoted substantial effort to making sure that the data are meaningful and reliable. Such a “data-driven” approach is common in technology firms (Brynjolfsson and McAfee 2012).

We obtained data for all employees from the R&D part of the firm who are managed via these systems. Data cover April 2019 through August 2020, resulting in a panel data set with 10,384 unique employees observed over 17 months. The company moved abruptly to WFH in March 2020 as COVID-19 became serious in India.

Sapience Analytics software is installed on employee work devices. It records the time that an employee is working by tracking applications or websites used, and whether the employee is active (i.e., using the keyboard or mouse). If an employee procrastinates on a social media platform, and this is not part of the job, that would not be recorded as work time. If a program from a predefined list of relevant tools (programming, collaboration, document production, communication, etc.) is active, that is recorded as work time. It includes meetings entered in the Outlook calendar, which is used throughout the company, even if these are face-to-face. Sapience therefore records effective work hours. If an employee sits at his or her desk for 8 hours a day, Sapience might only record 5 hours, because breaks or surfing the web are not effective work time. Employees are aware that the software is used in this way.

Based on these data, we calculate the outcome variable Input, equal to average working hours per working day that month. That is, we take the total time worked that month and divide it by the number of working days (taking into account weekends and local holidays). In section II.C below, we describe other input measures that generate similar qualitative results; our findings are robust to details of the definition of hours worked.

IDMS is a proprietary system used to track employee performance, including the primary performance measure. That measure is normalized to make different jobs and roles comparable.⁵ For example, key measures might be the number of code segments, code reviews, or reports delivered per month (subject to quality and customer satisfaction standards). These are divided by the monthly goal, multiplied by 100, to scale as percentage achievement of the goal. As explained above, goals were not changed due to WFH.

It is possible to complete more tasks than are assigned, so the outcome variable Output can take values in ${ℝ}_0^{+}$, but is typically between 0 and 100. The most common value is 100, meaning that employees continue working longer hours until they meet their targets, but we also see employees falling short of or exceeding their targets.

Finally, our outcome variable Productivity is calculated by dividing Output by Input. Differences in productivity will often be a consequence of differences in Input needed to reach the goal, but they can also come from employees falling short of or exceeding targets. Table 1 displays summary statistics before and during WFH. The number of observations under the two regimes differs, because we have more pre-WFH months.

B.  Employee/HR Variables

We obtained information on employee characteristics, collected as of March 20, 2020 (roughly the date on which WFH was implemented). Summary statistics are in table 2. For some employees some variables are missing. One reason is that HR data are deleted if an employee leaves the company or transfers to a branch outside India.

Age, company tenure, and industry experience (collected at hiring and updated to the current date) are all measured in years. For each we generate median splits: HighAge, HighTenure, and HighExperience. Mean age is quite young, which is not unusual in the IT sector. Mean tenure is low at about 4 years, as is expected since employee turnover is high in the IT sector. As in tech companies around the world, men are a significant majority.

The variable NumChildren is the number of children up to age 21 who are covered under the company’s employee health insurance plan. The company believes that the vast majority of employees who have dependent children insure them via the company, because of its relatively generous health insurance coverage. However, some might instead be insured through a partner’s employer. Hence, a zero means that there are either no children at home, or there are but they have not been declared. The dummy Children equals 1 if and only if NumChildren is positive.

CommuteTime is an estimate of the time in hours needed to get from the home address to the office (during WFO), one-way. The company calculated this based on home and office addresses, using the Google Maps application programming interface to incorporate factors such as traffic and not merely distance. Thus, it is an estimate of the usual time taken, assuming that the employee commutes by car.⁶ Address data are often incomplete, so there are more missing values than for other variables. We discarded extreme values (larger than 2 hours). According to the company these are cases where commute time is unreliable; for example, an employee actually worked at a client office closer to home, not the company office where his or her team is located.

Rating is the supervisor’s subjective evaluation of the employee on an integer scale of 1 to 5, where 1 is the best rating. We have the most recent rating from May/June 2020. The outcome measures discussed above are predictive of performance ratings: mean input and mean productivity in months prior to the rating significantly improve that rating (see table A.1). Figure B.1 plots kernel density estimates of subjective ratings for different levels of Output. Ratings generally rise with Output, but start to decrease once Output substantially exceeds the target. A possible interpretation is that such an employee gave too much emphasis to meeting the goal, and the supervisor gives a lower subjective rating to balance multitask incentives. Another is that the goal was too easy to achieve. Overall, this is more evidence that the outcome measures introduced above are meaningful and reflected in subjective performance evaluations.

C.  Workplace Analytics Data

Microsoft Workplace Analytics (WPA) is a tool that many companies use to track and analyze various aspects of their workforces. For example, it can be used to study collaboration or professional networking activity by using data on emails, calendar appointments, amount of time spent in meetings, and so forth. WPA data have been used in several organizational studies (Brynjolfsson and McAfee 2012; Hoffmann, Lesser, and Ringo 2012; Levenson 2018).

The company had been considering adoption of this tool. For the purposes of this study they purchased 914 licenses to apply to a subset of employees in our full sample. Table A.2 compares those in the WPA sample to those not in the WPA sample. Those in the WPA group are slightly younger, have lower tenure, and are less productive, but are overall quite similar.

Table 3 summarizes variables obtained from WPA. WPA data were collected at the weekly level (in contrast to Sapience data, which are collected at the monthly level). We have 10 weeks of data before WFH (starting January 1, 2020) and 24 weeks of data during WFH (ending September 6, 2020). The switch to WFH happened in the week starting March 16.

Variables fall into several categories. Working Hours measures overall time worked by the employee. It is best viewed as capturing time “at work” (at home or in the office), not every minute of which is necessarily spent working; for example, it will count small breaks in between emails as work time. It can detect longer work hours due to additional email traffic, meetings in the calendar or online, and other activity involving Microsoft services. In contrast, Sapience Input captures time “effectively working,” excluding breaks or nonwork activities.

During WFO, employees spent on average 44 Working Hours per week (table 3), which is close to the contractual 40 hours per week. After Hours measures the number of weekly hours worked outside regular office hours. Working Hours and After Hours are mechanically related as the latter is part of the former.

A second group of variables (Focus Hours, Collaboration Hours, Meetings Manager, Meetings 1∶1, Coaching Meets, and MS Teams Calls) relate to meetings. Focus Hours measures time blocks of 2 hours or more that are uninterrupted by meetings, calls, or emails. It is thus a measure of the amount of time the employee can concentrate on tasks. Collaboration Hours is the total time spent in various forms of meetings. The latter four variables measure time in meetings by structure and purpose. Meetings Manager is the number of meetings the employee attends that involve their manager, and Meetings 1∶1 are personal meetings between the employee and manager. Coaching Meets is the number of meetings involving the employee, his or her manager, and the manager’s direct reports. MS Teams Calls is the number of calls using MS Teams (which hosts video meetings similar to Zoom or Skype).

Table A.3 shows (pre-WFH) pairwise correlations between these meeting-related variables. As expected, all correlate negatively with Focus Hours, especially Collaboration Hours and Meetings Manager. All pairwise correlations are statistically significant at the 1% level. The different types of meetings are positively related among each other, but with smaller coefficients. These correlations are positive both across employees (some job roles involve more meetings than others) and across time (some periods involve more meetings of all types).

The third group of variables (Internal NW, External NW, NW EXT, NW ORG, Emails) relates to networking with colleagues and clients more explicitly. The first two measure the number of individuals (inside and outside the company, respectively) with whom the employee had contact. The next two measure the number of business units (e.g., teams) involved in those contacts. These measure the breadth of the employee’s communications and networking contacts. The last is the number of emails sent by the employees that week. Table A.4 shows the (pre-WFH) pairwise correlation between these networking-related variables. All correlations are positive and highly statistically significant, across employees as well as across time.

D.  Empirical Strategy

As a first step, in section III.A we estimate the average WFH effect. Our main specification exploits differences in outcomes for each employee, when working from home compared to working in the office, controlling for employee and customer team fixed effects. The unit of observation is the employee-month. Index the employee by i and the month by $t=1$, 2, … , 17. For outcome variable y_it, we estimate by OLS:

$$\begin{array}{}\text{(1)}& y_{it}={\alpha }_i+{\beta \mathrm{WFH}}_t+\sum _j{\gamma }_j{\mathrm{CustomerTeam}}_{jit}+\sum _s{\delta }_s{\mathrm{Month}}_{st}+{\epsilon }_{it},\end{array}$$

where α_i is the employee fixed effect, WFH is a dummy variable indicating months working from home, and CustomerTeam_jit is a dummy variable equal to 1 if and only if employee i in month t was part of team j. Month_st is a month (not month-year) dummy variable, so that ${\mathrm{Month}}_{1t}=1$ if and only if t is January, ${\mathrm{Month}}_{2t}=1$ if and only if t is February, and so forth. In addition, we report an alternative specification controlling for a linear rather than seasonal time trend:

$$\begin{array}{}\text{(2)}& y_{it}={\alpha }_i+{\beta \mathrm{WFH}}_t+\sum _j{\gamma }_j{\mathrm{CustomerTeam}}_{jit}+\delta t+{\epsilon }_{it}.\end{array}$$

Once the average WFH effect is established, we analyze variation in this effect. Section III.B studies how effects vary with employee characteristics. We interact the WFH dummy in the previous specifications with additional explanatory variables X_1i, X_2i, … . Because the X_ji variables are employee specific but time invariant, we do not separately control for them, as that is already achieved by the employee fixed effects.⁷

We exclude March 2020 ($t=12$) from regressions because our main outcome variables are collected at the monthly level, and working from home started in mid-March 2020.⁸ Thus, this month is neither purely WFO nor WFH. Moreover, it is likely that WFH increased in the days prior to the official WFH start, so the switch date was not clear-cut. An implication of excluding March 2020 is that teething problems and short-term adaptation effects are not reflected in our estimate.

Section III.C analyzes the effects of mechanisms such as time spent in communication and focus time. For these we rely on the WPA data described above. Here our empirical strategy is identical to the one described in equation (2), except that we control for weekly instead of monthly time trends, as these data are available weekly. Hence, in these regressions $t=1,\dots ,34$ represents weeks. For all analyses we cluster standard errors at the employee level.

A.  Average WFH Effect

Before proceeding to the regression analysis, figure 1 plots the three main outcomes by month to get an intuitive idea about the WFH effect. This will also help us understand which of the econometric models (1) or (2) seems most appropriate for controlling for time trends.

According to figure 1A, Input, employees provide about 5–5.5 hours of daily input, that is, time in which they are actively using their software or programming tools, or in meetings or communications. There is relatively little variation in average input before WFH, with a slight upward trend. Hence, a linear time trend as in model (2) might be more appropriate for this outcome measure. From the first month of WFH, there is a large and persistent jump in input by more than 1.5 hours per day.⁹

For Output, figure 1B, there is no visible monotonic or linear trend, so a seasonal time correction might be more appropriate here. Moreover, average output appears to be slightly lower during WFH.

For Productivity, figure 1C, the graph is more volatile, which is not surprising for a ratio. There is no clear linear time trend before WFH, but some variation from month to month, so a seasonal correction might be more appropriate. Productivity drops visibly during WFH. Finally, figure 1D plots the log of Productivity, which drops considerably after the start of WFH.

To quantify the WFH effect, and to control for employee and team time-invariant variables (via employee and team fixed effects), we now turn to the regression analyses. Informally, the estimates give us average differences in outcomes before and during WFH for the same employee, controlling for team effects (since employees sometimes switch teams) and time trends.

Table 4 reports WFH effect estimates based on OLS regressions for all three outcome variables, plus the natural logarithm of Productivity, in each case with linear and seasonal time trend corrections. All estimates are in line with the visible effects in the raw data in figure 1.¹⁰

According to the estimates, WFH increased the time worked per day by roughly 2.1 hours (based on a seasonal time trend) or by 1.6 hours (with a linear time trend). Both estimates are economically very meaningful, and statistically significant at all conventional levels.¹¹ Since both figure 1A and the regression indicate a linear time trend, we prefer column 2. The estimate of the WFH effect in column 1 is larger, because it does not take the pre-WFH time trend into account in the same way.

Columns 3 and 4 show that both WFH effect estimates on Output are slightly negative, but only the estimate with the seasonal time trend is significantly different from zero. While the regression indicates a significantly negative linear time trend, we prefer specification 3, since a linear trend does not reflect the raw data in figure 1B very well. According to this specification, Output changed by −0.5 percentage points during WFH (recall that fulfilling all monthly tasks implies an output of 100%). Hence, we conclude that WFH had no or only a small negative effect on Output.

Columns 5 and 6 show that both WFH effect estimates on Productivity are negative, but only the estimate with seasonal time trend is significantly different from zero. We prefer that specification, since both the plot and the linear time trend coefficient indicate that a linear trend is not as appropriate.¹² According to this specification, productivity decreased by 0.26 output percentage points per hour worked. Given an average WFO productivity of 1.36, this estimate corresponds to a 19% drop in output per hour worked. This is economically significant: if employees worked a fixed 40 hours per week, this would imply a drop in output of 10.2 output percentage points in a week. In other words, if employees had not increased time worked during WFH, on average they would have completed only 90 of 100 assigned tasks.

Columns 7 and 8 explain the log of Productivity, which strongly increases the fit of the regression. The WFH effect is negative and significantly different from zero at all significance levels, irrespective of time controls.

In summary, this evidence indicates that employees worked longer but less productively, with output remaining about the same or dropping slightly. Our interpretation of these patterns is that employees were less productive during WFH, but still aimed to reach the same output or goals, and hence worked longer until the same output was reached. In the next sections, additional results will suggest that productivity decreased due to increased distractions and coordination costs.

A potential alternative explanation for the jump in work time during WFH might be that employees “gamed” the Input numbers. This is unlikely for several reasons. First, Sapience time measurement is sophisticated and designed to be resilient to manipulation attempts. Merely keeping the computer on for longer or watching videos instead of working does not increase Input. Rather, it would require having the relevant work software as the active window, and giving continuous mouse or keyboard input. Second, gaming time measurement in Sapience would not translate into increases in WPA’s time measurement.¹³ WPA time recording is from activity in MS Outlook, MS Teams, and so forth, rather than programming tools or similar, and is not dependent on mouse/keyboard activity. Yet the WFH effect we see with this alternative time measurement is also positive; see section III.C. Third, employees are not paid by the hour. To impress superiors, time is better spent generating output. Fourth, Sapience was in use long before the switch to WFH, so this potential concern cannot explain the WFH effect well. Fifth, the additional WFH effects that we find from WPA activity (sec. III.C), such as more time spent on conference calls and fewer focus hours, cannot be explained by “gaming.”

Another potential concern might be that the increase in Input is due to measurement error, as spontaneous (unscheduled) face-to-face meetings during WFO might not be counted as working time, while spontaneous online meetings during WFH are counted. This is not backed up by the data. We are able to reproduce the findings in this section using other measures of working time, available for a subsample of employees (sec. III.C). We also find adverse effects of WFH on direct measures of networking and training; in particular, the number of some types of meetings decreases during WFH. Moreover, employees with children, and those working in roles in which networking with people outside the company is more important, are more adversely affected by the switch to WFH. This is difficult to explain with unscheduled off-line meetings, but is consistent with increased disruptions and communication costs. Last, we observe when an employee starts and ends his or her working day. The length of the workday measured in this way increases from 7.64 hours to 9.17 hours during the WFH period. Taken together, these findings suggest that spontaneous off-line meetings explain at most a small part of our overall results.

If employees used their own devices to work from home, we would not track that activity (in either period). However, HCL did not allow employees to do so, to insure integrity of confidential client information. Any employees who worked at home in both periods had to use a company-owned laptop computer and phone, both of which included the tracking software used to collect our variables, with one small exception. During the transition to WFH, a small percentage of employees who did not yet have a company-owned laptop computer were briefly allowed to use personal devices to perform some work. This lasted only until the company could provide them with laptops. In such cases Sapience would not track their work, so our Input variable has missing values. This affects a very small percentage of our data.

Another possible explanation for the increase in time worked is that pandemic lockdown measures closed restaurants, cinemas, and so forth, thereby reducing the value of leisure time. Under this explanation, however, we would expect Output to increase and Productivity to remain approximately constant, which is not what we observe. Figure B.2 shows that while we see a slight dip in working hours after every stage of lockdown easing, the effect is small and, more importantly, only temporary. We also do not find evidence that productivity or other outcomes covary with national or regional indicators of the severity of COVID, such as deaths or case rates.

Finally, we are able to measure the number of days the employee worked relative to the number of work days in a month. According to table A.6, the number of sick days decreased significantly. This suggests that absences or sickness were not a driver of the decrease in productivity.

B.  Who Copes Better with WFH? Heterogeneous WFH Effects

We now explore what drives the WFH effect in more depth, and which subgroups are most affected by the shift to WFH. Table 5 displays estimates for all outcome variables, separately by whether employees have children at home and by gender, using our preferred time control from the previous section. The number of observations is slightly reduced since the additional explanatory variables are missing for some employees. We converted the explanatory variables (except commute time) into dummy variables for easier interpretation.

In India, all schools closed in March 2020 during the COVID-19 pandemic, so working from home was presumably an even greater challenge for some parents, as children needed to be supervised and perhaps taught. Hence, we investigate whether having children at home changed an employee’s WFH effect. An important qualification is that under normal conditions, children would attend school and adverse effects of WFH on productivity of parents might be lower.

Column 1 shows that employees who have at least one child at home (as measured by company health insurance coverage) increased work time more during WFH than did their counterparts without children. Possibly, this is due to the fact that employees with children get distracted more often during WFH and compensate by working longer hours. Employees with children at home work almost a third of an hour more per working day during WFH than those without children, who themselves still work 1.4 hours more during WFH. These effects are highly significant. Column 3 reveals no significant change in the WFH effect on output with children at home. However, column 5 shows that the increased working time implies a larger drop in productivity when there are children at home. Consequently, the patterns seen for the average employee are exacerbated for employees with children.

The even columns in table 5 investigate whether there was a gender difference in how outcomes changed during WFH, conditional on whether there were children at home.

The WFH × Male interaction represents the difference in the WFH effect between male and female employees without children. Male employees without children increased working time by about 0.2 hours less per day than did female employees without children, a significant effect at the 5% level. They also suffered a significantly smaller decline in output.

The WFH × Children interaction represents the difference in the WFH effect between female employees with and without children. Female employees with children did not significantly increase working time during WFH compared to female employees without children, nor did their output or productivity significantly differ.

Finally, the sum of WFH × Male and WFH × Male × Children is the difference in the WFH effect between male and female employees with children. This difference is roughly zero for all outcome measures, so there is no gender difference among employees with children at home.

Thus, female employees are more adversely affected by WFH, but this is not due to child care responsibilities. The latter finding contrasts with much of the narrative in Western countries, where child care responsibilities are given as a main reason why women are more adversely affected by WFH (Financial Times 2021a). We conjecture that this is due to greater expectations placed on women by parents and parents-in-law in the Indian domestic setting.

Next, we investigate whether employees with more industry experience or company tenure were affected differently by WFH. One reason this could be the case is that they have greater institutional knowledge and social capital, and are less reliant on help from colleagues or find it relatively easier to obtain during WFH. To investigate the effects of age, company tenure, and industry experience (at the company or elsewhere), we generate dummy variables with a median split. Since these variables are highly correlated, we estimate their effect on the WFH estimate jointly in table 6.¹⁴

Column 1 shows that work experience has the largest and only significant impact on the WFH effect for time worked. During WFH, experienced employees worked roughly a quarter-hour more per day compared to less experienced employees, holding age and company tenure constant.¹⁵ Our interpretation is that more experienced employees have more managerial duties. The increased costs of coordination (also see next section) during WFH are therefore borne by these experienced employees, who have to put in more time to coordinate their team. Additionally, it is likely that the lion’s share of managing the WFH transition falls on experienced employees with more responsibility.¹⁶

The change in Output during WFH is roughly 0.5 percentage points larger per hour for employees with longer company tenure, holding age and experience constant. Thus, while low-tenure employees’ Output falls, high-tenure employees keep meeting their goals. The other characteristics do not show a significant effect. It appears that employees who had worked at the company for longer were able to adapt more effectively to the WFH shock, and that this was more important than general industry experience. This finding suggests that greater firm-specific human capital in the form of familiarity with company procedures, or more fully developed networks and working relationships with colleagues and clients, were helpful during WFH. Alternatively, those with greater experience or tenure might be in positions with more responsibility, and so responded more to the shift to WFH. For the last outcome measure, Productivity, there is no significant difference in the WFH effect among these employee groups.

The even columns in table 6 estimate the WFH effect by employee commute time (when working from the office). The effect does not significantly differ by commute time for any outcome measure. Hence, the finding that WFH increased hours worked is not merely due to more available time. Rather, it supports the interpretation that productivity fell during WFH, and employees worked more to compensate.

Figure 2 shows the mean and 95% confidence intervals of the estimated effect of WFH on productivity separately for different customer teams. Those are teams of employees who work for the same customer, but not necessarily on the same project. In the figure teams are sorted by size (number of team members) in increasing order from left to right. The effect is negative for almost all teams. This illustrates that the overall negative effect on productivity (illustrated by the dashed line in fig. 2) is not driven by a few customers or teams. Rather, WFH productivity declined broadly across the group studied.

For 43 teams the negative effect is statistically significant. Five teams, by contrast, show a statistically significant positive effect on productivity. With the exception of one team, those are all teams in the bottom half of the size distribution. In smaller teams it might be easier to solve coordination problems during WFH. Still, even among the smaller teams, the effect on productivity is negative for the vast majority.

C.  Mechanisms: What Contributes to Lower Productivity?

To better understand the mechanisms behind the decrease in productivity, we study the subsample of 914 employees for which WPA data were obtained (see table A.2). Using these data we document three patterns: an overall increase in working time; a shift away from performing work tasks and toward spending time on meetings, calls, or answering emails; and reduced time networking with others or meeting one-to-one with one’s manager.

1.  Working Hours

Figure 3 illustrates the shift in working patterns after the start of WFH. In line with the evidence in section III.A, the WPA data show that overall hours worked increased, including those after regular office hours. Panels C and D show an interesting pattern. Employees spend more time in meetings or calls and have less “focus time” (as described above, this measures the amount of work time that was not interrupted by various types of meetings or communications). The increased time spent in meetings and its persistence after the initial WFH transition phase suggest substantial and ongoing coordination costs with WFH, which negatively impact time available to work in a productive manner. The fact that employees have less time to work in a focused or uninterrupted manner could also explain why they have been found to multitask more during meetings (Cao et al. 2021).

The technological shift is illustrated in Figure B.4. In particular there is a drastic increase in the number of hours spent on virtual meetings using MS Teams. Interestingly, the number of such meetings continues rising almost 6 months after the switch to WFH. The increase in number of meetings is clearly much larger than the increase in total time spent in meetings. Therefore, employees are interrupted more frequently by having to attend more meetings of shorter duration.

Table 7 shows regressions estimating the WFH effect on these outcomes. Both overall working hours and working hours outside regular office time increase during WFH. In fact, comparing the size of coefficients in columns 1 and 3 we see that the increase in overall working hours takes place almost entirely outside normal working hours. The table also confirms the increase in time spent in meetings and on calls, with a corresponding decrease in uninterrupted work time (focus hours). In all cases, the WFH effect persists and remains highly statistically significant when we include a linear time trend. The estimated effect size is smaller in these cases but remains substantial. After controlling for time trends, employees work 2.7 hours more per week, out of which 1.8 are spent working outside regular office hours. However, they also spend 1.4 hours less working in a focused or uninterrupted manner. Since working hours are the sum of focus hours, collaboration hours, and “unfocused hours” (i.e., hours working but getting interrupted), the table also shows that WFH replaced focus hours by many more unfocused work hours. These shifts in working patterns could explain why productivity decreases under the WFH regime.¹⁷

	Working Hours		After Hours		Focus Hours		Collaboration Hours
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
WFH	4.431***	2.743***	4.212***	1.822***	−2.665***	−1.417***	1.251***	.704***
	(.161)	(.218)	(.288)	(.320)	(.245)	(.297)	(.122)	(.137)
Linear weekly trend		.096***		.137***		−.071***		.031***
		(.009)		(.014)		(.012)		(.005)
Employee fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Team fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R2	.703	.709	.743	.747	.717	.719	.745	.746
Observations	25,893	25,893	25,893	25,893	25,893	25,893	25,893	25,893
Clusters	914	914	914	914	914	914	914	914

We conjectured that some of the increase in working hours and decrease in productivity is due to increased costs of communication and coordination within teams. If this is the case, then roles that are characterized by more interaction and networking prior to WFH are more affected by the switch to WFH. If we median split by how many internal networking contacts (variable Internal NW; see table 3) an employee had prior to WFH, working hours in specification 1 of table 7 increase by 6.01 hours for those with above-median contacts and by 3.515 hours for those with below-median contacts (both results are significant at the 0.1% level). The difference is highly statistically significant (t-test $p<0.0001$), suggesting that coordination costs are indeed an important factor behind the decline in productivity during WFH.