Measuring the Impact of Remote Knowledge Work Across the United States

Alan Kwan (HKU)

Ben Matthies (University of Notre Dame)

​

April 2023

Overview

• What we do:
• Construct a firm-level economy-wide measure of remote work using data on employee internet activity at each firm
• Firm-month panel of remote work share for hundreds of thousands of firms in the U.S.

​

• We join to confidential tax filing data to deliver first firm-level estimates of the relationship between remote work and output

​

• Present stylized facts about the determinants and consequences of remote work
• Determinants: Relationship to pandemic risk, monitoring cost proxies
• Consequences: Firm financials, micro estimates of worker behavior from the the internet-activity dataset, flow of workers from in-person to remote work firms

​

​

​

2

Literature positioning

• Lots of work on whether remote work "works"
• Erik Brynjolfsson et al. (2022), Cevat Giray Aksoy et al. (2022), Peter Lambert et al. (2022) Carl Benedikt Frey and Giorgio Presidente (2022)
• Natalia Emanuel and Harrington( 2022), Dalton, Dey and Lowenstein (2022, BLS)

​

• Methodologies of how to measure remote work
• Prithwiraj (Raj) Choudhury et al. (2022), Ruobing Han et al. (2022) Aruna Ranganathan and Aayan Das (2022), Piotr Lewandowski et al. (2022)

​

• Some work on housing demand and consequences of remote work
• John Mondragon and Johannes Wieland (2022), Matthew J. Delventhal and Andrii Parkhomenko (2022) Gupta, Mittal and Van Niewerburgh (2022)

​

Literature positioning

• Lots of work on whether remote work "works"
• Erik Brynjolfsson et al. (2022), Cevat Giray Aksoy et al. (2022), Peter Lambert et al. (2022) Carl Benedikt Frey and Giorgio Presidente (2022)
• Natalia Emanuel and Harrington( 2022), Dalton, Dey and Lowenstein (2022, BLS)
• First firm-level study with comprehensive administrative data!
• Methodologies of how to measure remote work
• Prithwiraj (Raj) Choudhury et al. (2022), Ruobing Han et al. (2022) Aruna Ranganathan and Aayan Das (2022), Piotr Lewandowski et al. (2022)

While not a drop-in replacement, we are the first to propose a real-time longitudinal measure of remote work

• Some work on housing demand and consequences of remote work
• John Mondragon and Johannes Wieland (2022), Matthew J. Delventhal and Andrii Parkhomenko (2022) Gupta, Mittal and Van Niewerburgh (2022)
• For now, nothing to say here! But we might be able to merge in real estate-related outcome variables.

​

Roadmap

• Introduction
• Data
• Tests
• Conclusion

​

​

​

Data sources

• Our Data Partner operates a cooperative of thousands of media publishers
• Collect data via cookies (longitudinally tracking users across IPs)
• 1 billion events per day, we posit can be used to create a remote work measure

​

• Combine with confidential U.S. govt filings to estimate firm output
• Not ready for dissemination. I thought it would be today, should be very soon

Internet activity mapped to firms

• Our Data Partner operates cooperative of thousands of media publishers
• Collect data via cookies (longitudinally tracking users across IPs)

1 billion events per day

​

Many household names in media as well as less well known sites

​

​

Publisher

A

Visitor

Visitor

Publisher

B

Data Vendor

Visitor 🡪 firm

Content 🡪 topic

Publisher

C

Scale of content is pretty large

Scale

Wide range of types of sites

Both obscure and well-known publishers

Inferred topic distribution is wide

Remote work measure – based on prepandemic info

• The goal is to classify IP addresses as {vpn, residential, business, mobile}
• >760 million in total
• Step #1: manual classification
• Rule-sets based on our understanding of the internet
• e.g. vpn2fa.hku.hk is probably a VPN
• Coverage of about 60% of internet activity

​

• Step #2: Pick up the rest with a model
• Lots of data sources
• Reverse DNS data from Rapid 7, Censys, Luminati – residential proxy data “RpaaS: residential proxies as a service”, IAB classification data from an ensemble of Webshrinker, BuiltWith, Dmoz, MaxMind connection type database, Shodan VPN list

​

​

​

Manual rule-set (30%+ of IPs, 60% of traffic)

• Residential → in the same residential block as a residential proxy
• Worked with Lumniati to ping their blocks, any IP in the same /24 block (a.b.c.octet4) is likely residential

​

• Mobile IP
• At least 200 mobile phones, majority device types are mobile phones, and either (reverse DNS shows registration to a mobile company MaxMind infers connection type)

​

• Business IP
• If the company has classified this as a business IP consistently throughout 2019 through bi-weekly exercises

​

• VPN: reverse DNS is
• VPN 🡪 vpn2fa.hku.hk
• Shodan VPN list

​

​

​

​

Machine learning model

• Labeled set: ~10k cleanly labeled IP addresses
• Fortune 1000 known office IPs + some others
• Residential blocks for Comcast
• Some mobile networks (where there are at least 200 mobile phones daily)
• VPNs -- the reverse DNS says VPN in the name
• Residential proxy networks (consumer VPNs)
• Feature sets
• Geographic dispersion The number of unique (a mobile phone, then a VPN will have geographically disperse)
• Scale of content How much business relevant content you read
• Composition of content How much social content read; How much adult content read
• Temporal patterns
• % domination by a top user / top company

​

• Standard random forest model

The classified IP addresses “make sense”

Machine learning metrics

Remote work measure

15

2020 remote share: across all IPs in the United States

16

By type of IP (normalized to same scale)

18

County-level regression vs SafeGraph

By state and industry

Roadmap

• Introduction
• Data
• Tests
• Conclusion

​

​

​

What drives firm decisions to work remotely?

• Coordination: certain types of work benefit from discussion and collaboration

​

• Moral hazard: distance reduces monitoring (e.g. Dessein, Galleoti, Santos (2016), Gumpert, Steimer Antoni (2021))

​

• Infrastructure: IT systems for remote work (communication, collaboration, monitoring); Managerial hiring

​

​

​

​

​

​

​

​

​

​

​

21

• Commute distance: firms with longer commute distances benefit more from remote work

​

• Labor market: remote work firms may attract and retain stronger talent given strong worker preferences

​

• Reduced distraction/malaise

Costs

Benefits

* In some cases remote work is not possible (fast food, service); this type of variation should be absorbed in industry FE

Baseline determinants of remote work

IT drives remote work decisions sensibly

More managers 🡪 afford remote work cost

Two stage least squares regression

Outcome variables

Summary of results – corporate financials

• Cannot disclose any actual numbers
• Overall, the effects of remote work are positive for firm output
• Sales/assets, sales/wages, etc.
• Net profit/assets, net profit/wages
• Sensible variaton across tradeables/non-tradeables

27

Reading as a measure of work?

28

Elusive evidence of the afternoon siesta

Reading as a measure of work?

29

Firm behavior

• How does remote work impact firm investment into:
• Information Technologies? (increase monitoring and collaboration)
• Managerial layers? (increase monitoring)

​

• Measures
• IT Index: count of the number of unique technologies at the firm
• Constructed using data from Aberdeen Computer Intelligence Database (underlying data provider comes from the Dunn and Bradstreet database which underpins the National Establishment Time Series used in academic studies)
• Infrastructure, general data, internal data, communication, security
• Change in percent of job postings related to a managerial role (from 2019 to 2020)
• Job postings data from Burning Glass Technologies (near-universe of job postings)
• Results, an exogenous increase in remote work:
• Firms invest in IT technologies (particularly related to communications and monitoring)
• Firms tilt hiring activity towards managerial roles

​

​

​

​

​

​

​

30

Investments in labor

31

Labor effects

• Does remote work provide an advantage in the labor market?

​

• Measures
• Map every profile on LinkedIn to firms. For each person, compare where they worked just prior to COVID to where they worked post 2020

​

​

​

​

​

• Results
• Workers flow from low remote work firms to high remote work firms

​

​

32

Labor preference

33

Labor preference

34

Labor preference

35

Conclusion / what’s next

• Hopefully convinced that …
• This is a real remote work measure
• And there are material effects of remote work on organizational resources (IT, managers) and firm decisions
• Plausibly exogenous increases in remote work are associated with increases in firm output (although not yet something we can call “productivity”)
• Next steps:
• Estimate production function parameters explicitly
• Study the impact of remote work on firm investment and innovation [expenditures], data on compensation and labor flows across firms using (anonymized) individual-level tax data

​

​

36

​

​

Thank you

Commute distance

38