Further Analysis from the Tech Remote Work Survey

Building on the early findings from a survey of over 900 tech industry employees, five key insights have been derived through a series of regression-based analyses. They answer long-standing questions such as “what will make workers finally return to the office?” and “is there really a mass exodus happening from Silicon Valley?”. Warning: this post is heavy on the math, but I’d appreciate any feedback if you happen to be a stats nerd!

Get comfortable in your home office, these findings might surprise you!

Office Trips Will Be Employer-Driven

Despite widespread reporting on the popularity of remote work among tech employees, it seems that most see themselves commuting into an office more after COVID-19 is no longer a threat. 65% predict that they will be commuting at least one day per week with the plurality of respondents (28%) settling on a hybrid schedule of 3 days per week. A paired T-test of difference between mean days in the office currently versus in the future shows a definite positive trend.

P-Value< 0.00001
Effect Size (Cohen’s d)1.028
Difference Between Averages1.36
95% Confidence Interval of Difference1.23 to 1.49
Paired T-Test of Days in Office (After) – Days in Office (Current)

Interestingly, the clustering around a 3-day in-office schedule roughly aligns with employers’ stated policies, with 73% of those with onsite work requirements setting them at 3 days per week. Looking closer at these two datapoints from the survey (commute frequency and employer’s remote work policy), it seems they are closely correlated. In fact, the number of employer-mandated days in the office accounts for 86% of the estimated days in the office after COVID-19 in a linear regression model including employee count, age, and commute length as input variables. An increase of 1 employer-mandated day is associated with an increase of 0.62 in the estimated days in the office after the pandemic.

VariableRelative ImportanceCoefficientP-Value
Employer-Mandated Days in Office86%0.61589< 0.00001
Employee Count12%7.84e-50.0917
Age1%-0.019050.0812
Commute Length1%-0.009060.0887
Linear regression model for number of days in the office after COVID-19 (R2 = 0.61, P-Value < 0.00001, AIC = 348)

When comparing this model to previous studies that used binary random parameters logit models to determine workers’ likelihood of continuing to work-from-home after COVID-19, age seems to be a common variable. However, number of children in the household is not – but this could be due to the relatively low numbers of children in general within the tech worker population.

Overall, this strongly suggests that employer policies are what will primarily drive future commute patterns. The more that employers enforce mandates to be in offices, the more employees see themselves as commuting regularly whether they want to or not. Politicians hoping to restore vibrancy back to central business districts seem to understand this, going so far as to work with business groups to encourage more mandates.

Commute Length Determines Mode Choices

Employees’ choice of travel mode for work trips, as well as their proclivity for using various vehicle types more generally, seems highly dependent on their commute length, among other variables.

For example, there is a strong correlation between a respondent’s commute length and whether they drove a car alone to work before COVID-19. A chi-squared test of commute length (as a categorical variable) and whether “drive alone” was selected as a commute mode showed that 80.3% of those with a commute less than 5 miles did not drive, while 51.4% of those with a commute between 10 and 25 miles did drive.

P-Value< 0.00001
Effect Size (Cramér’s V)0.267
Sample Size425
Chi-Squared Test for Commute Length and “Drive Alone” (Before COVID-19)

These choices extended into the future as well. A binary logit model was built for whether a survey respondent selected bicycle, bike share, scooter, or other micromobility as an expected commute mode after COVID-19 is no longer a threat. The output variable was named bActiveCommuterAfter, and 45% of it was accounted for by commute length. Controlling for other variables in the model, the odds of a survey respondent being an active commuter in the future was 0.92 times as high with every 10-mile increase of commute length.

VariableRelative ImportanceCoefficientP-Value
Commute Length45%-0.00800.000110
Number of Cars32%-0.10580.00142
Is White16%0.13240.00972
Employer-Mandated Days in Office7%0.05270.0259
Binary logit model for whether active commute modes will be used after COVID-19 (McFadden’s R2 = 0.331, N = 163, AICc = 120)

Furthermore, both commute length and the binary variable bActiveCommuterAfter accounted for a significant portion of cars owned in a separate linear regression. A 10-mile increase of commute length is associated with an increase of 0.10 cars owned, while an increase of 0.10 in bActiveCommuterAfter is associated with a decrease of 0.033 cars owned.

However, it should be noted that other factors such as age and number of children had higher relative importance in this model.

VariableRelative ImportanceCoefficientP-Value
Number of Children29%0.98580.000818
Age27%0.01840.00284
Commute Length23%0.01010.00235
bActiveCommuterAfter19%-0.32590.00528
Income2%2.99e-70.649
Linear regression model for number of cars owned currently (R2 = 0.152, P-Value < 0.00001, AIC = 951)

This makes intuitive sense, and also suggests that mode shifts towards or away from driving individual automobiles are possible depending on relocation decisions discussed further below. If tech workers who had been living close to their offices decide to move more than 10 miles away but still need to commute, that commute trip will more likely be completed by driving and potentially lead to a net reduction in sustainable trips across the region.

Online Shopping Reduces Trips for Remote Workers

The widespread adoption of online shopping, especially for groceries and food delivery, may actually prevent the future upsurge in trips historically predicted as rebound effects of remote work. This runs counter to our original hypothesis that non-work trips would increase with less commuting.

Online grocery delivery, in particular, is a relatively new service even for those in the tech industry. Only 8% of survey respondents used it before the pandemic, but 17% are using it now and 14% expect to continue using it even after COVID-19 is no longer a threat. A Fisher’s Exact Test comparing whether a respondent started grocery delivery during the pandemic (i.e., didn’t use it before but using it currently) with their prediction of whether they’d use it in the future showed a strong statistically significant relationship. 67% of those who started grocery delivery would continue, and similar figures exist for restaurant delivery (71%).

P-Value< 0.00001
Effect Size (Cramér’s V)0.528
Sample Size647
Fisher’s Exact Test for Started Grocery Delivery and “Order groceries online” selected (After COVID-19)
P-Value< 0.00001
Effect Size (Cramér’s V)0.302
Sample Size647
Fisher’s Exact Test for Started Restaurant Delivery and “Order food for delivery from a restaurant” selected (After COVID-19)

This suggests that these services are very sticky (i.e., have high retention rates among their customers). If more remote workers try grocery and food delivery services, they are likely to get hooked on the convenience and reduce the number of trips they take to grocery stores and restaurants themselves in the long-run. This would likely reduce the overall vehicle miles travelled by remote workers, but it would also increase the number of trips made by delivery drivers. Global vehicle miles traveled may only be reduced if sequences of deliveries are optimized (i.e., a single driver can make multiple deliveries in one trip), which fortunately aligns with operational goals of these service providers.

Relocations Mirror Existing Trends

The long arc of American history, especially in the postindustrial era, bends towards suburbanization – for better or worse. Despite the glimmer of an “urban renaissance” in the early aughts, the majority of population (and even employment) growth in the last decade has occurred in suburban rings as opposed to inner cities. Regardless of whether this is due to housing costs, shifting employment opportunities, or cultural preferences, this trend doesn’t appear to be changing anytime soon – and the COVID-19 pandemic has almost certainly accelerated it.

Relocation interest from survey participants was quite high, with 53.6% considering a move or had already completed one. (64.2% of that group were motivated by remote work options specifically.) While 32% of respondents lived in a suburb before COVID-19, 43% would choose a suburb after the pandemic. 54% of those moving would choose a stand-alone home over an apartment or other housing type. A median relocation distance of 19.37 miles also suggests city-to-suburb movement as opposed to movement between major cities. “More space” is also one of the top reasons cited for relocation, suggesting a suburban value.

The factors that contribute to one’s desire to relocate are also not too surprising. In a binary logit model for whether a participant had interest in relocation (including whether they already moved), age and number of children were meaningful drivers. However, the most significant driver was the number of days expected in the office after the pandemic. In fact, controlling for the other input variables, every additional day in the office reduced the odds of relocation interest by 3%.

VariableRelative ImportanceCoefficientP-Value
Days in Office (After)32%-0.03200.000451
Age31%-0.00530.00140
Number of Children22%-0.05470.0235
Employee Count15%-6.89e-60.0433
Binary logit model for having interest in relocation (McFadden’s R2 = 0.097, N = 312, AICc = 401)

Once again, this paints a picture of tech employers and their remote work policies as controlling factors in employees’ major life decisions. If employers mandate more days in the office, employees will likely stay close and reconsider relocation plans. If those employers are in city centers, this may actually result in a slowing of suburbanization – but many employers are also on the move as well and might draw their employees into the suburbs with them.

Public Transit Recovery Will Remain Slow

Published ridership data from public transit agencies in the San Francisco Bay Area can be combined with findings from the survey in order to produce some “back of the envelope” predictions about ridership recovery in the long-term. In particular, matching ridership levels with estimates of average days in the office and total transit trips taken at two timepoints (before the pandemic and “currently” – i.e., November/December 2021) can enable a basic linear projection of ridership into the future.

Weekday Ridership

For example, survey responses suggested that tech workers were commuting to offices an average of 4.5 days per week before COVID-19 and only 0.6 days per week “currently” at the time of the survey. Data from SFMTA shows an average of 679,968 boardings per weekday between November 2019 and February 2020 versus only 334,635 boardings per weekday between November and December 2021. If the relationship between days in the office and average weekday boardings stays constant, it can be described by this linear equation:

SFMTA avg. weekday boardings =
8546.92307692308*(avg. days in office) + 281506.8461538461

Hence, future ridership levels can be projected based on the estimated days in the office once COVID-19 is no longer a threat, as reported by survey participants. The result is a prediction that SFMTA weekday ridership after COVID-19 will be about 67% of pre-pandemic levels for the foreseeable future, if nothing else changes.

TimeframeAvg. Days in OfficeSFMTA avg. weekday boardings
Before COVID-194.5679,968
Nov/Dec 20210.6334,635
After COVID-19 (est.)2.0458,601
Linear projection of SFMTA weekday ridership summary

The same analysis can be done with ridership data from BART, which results in a prediction that weekday ridership will only rebound to 53% of pre-pandemic levels based on this equation:

BART avg. weekday trips =
75063.07692307692*(avg. days in office) + 57519.15384615385

TimeframeAvg. Days in OfficeBART avg. weekday trips
Before COVID-194.5395,303
Nov/Dec 20210.6102,557
After COVID-19 (est.)2.0207,645
Linear projection of BART weekday ridership summary

The results are even more dire with Caltrain (data from their monthly Board of Directors meetings), with a prediction that ridership will only reach 46% of pre-pandemic levels based on this equation:

Caltrain avg. weekday riders =
14177.179487179488*(avg. days in office) + 2230.6923076923067

Caltrain’s ridership dipped the most during the pandemic, likely due to this system’s historical focus on white-collar downtown commuters who were overwhelmingly remote-eligible. Also, the Caltrain service corridor, running along the peninsula from San Francisco to San Jose, is aligned with the vast majority of software companies in Silicon Valley.

TimeframeAvg. Days in OfficeCaltrain avg. weekday riders
Before COVID-194.566,028
Nov/Dec 20210.610,737
After COVID-19 (est.)2.030,585
Linear projection of Caltrain weekday ridership summary

Total Monthly Ridership

A similar but broader analysis also considers non-work trips (e.g., for shopping, social, or entertainment purposes) and looks at whether public transit might be the mode of choice used by the survey participants. An estimate of total public transit trips per week was calculated for each participant based on trip types that were taken once a month or more where public transit was a selected mode, and averages for the same three time periods were matched with total monthly ridership numbers from the National Transit Database.

Survey respondents took an estimated 14.2 transit trips per week before the pandemic, only 1.9 trips per week at the time of the survey (November/December 2021), and only expected 3.4 trips per week in the future. Matching this with SFMTA monthly unlinked passenger trips from the NTD resulted in this equation and projections:

SFMTA monthly unlinked passenger trips =
681982.2764227643*(transit trips per week) + 7839306.674796748

TimeframeTransit Trips per WeekSFMTA monthly trips
Before COVID-1914.217,523,455
Nov/Dec 20211.99,135,073
After COVID-19 (est.)3.410,158,046
Linear projection of SFMTA monthly ridership summary

This method suggests that SFMTA will only recover 58% of pre-pandemic ridership for the foreseeable future, slightly worse than just the weekday/commute analysis found. This reflects a combination of fewer non-work trips overall as well as significant opting-out of public transit for more trips generally.

A similar pattern can be seen with BART, which is only projected to recover 37% of pre-pandemic ridership:

BART monthly unlinked passenger trips =
578160.7317073172*(transit trips per week) + 1654782.6097560974

TimeframeTransit Trips per WeekBART monthly trips
Before COVID-1914.29,864,665
Nov/Dec 20211.92,753,288
After COVID-19 (est.)3.43,620,529
Linear projection of SFMTA monthly ridership summary

Finally, Caltrain again is in the direst situation, projected to barely reach 30% of pre-pandemic ridership in the long-term:

Caltrain monthly unlinked passenger trips =
97447.56097560977*(transit trips per week) + 129475.63414634144

TimeframeTransit Trips per WeekCaltrain monthly trips
Before COVID-1914.21,513,231
Nov/Dec 20211.9314,626
After COVID-19 (est.)3.4460,797
Linear projection of SFMTA monthly ridership summary

There is a major caveat with all of these projections, however. They assume the behaviors of remote-eligible workers (as represented by tech workers) explain all of a transit agency’s ridership numbers, while in reality there are many other people who ride transit for both work and non-work reasons. In fact, most “essential” workers who did not have the option to work remotely (i.e., those holding the other 52% of jobs in the San Francisco Bay Area) have been riding transit throughout the pandemic and will likely continue doing so into the future. However, the distribution of remote-eligible versus “essential” workers is difficult to predict between various transit agencies. As seen above, some transit agencies seem to have more regular non-remote-eligible riders than others (i.e., SFMTA versus Caltrain). Further research into each agency’s ridership demographics would be necessary to produce more accurate ridership projections.


What this all means for urban/transportation planners will be discussed in a future post, along with some concluding remarks and ideas for future research. Until then, feel free to reach out if you have any suggestions or feedback on the analysis so far!

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.