Year 2005 UCD— ITS— RR— 05— 03
An Empirical Analysis of Causality in the Relationship
Between Telecommuting and Residential and Job Relocation
David T. Ory
Patricia L. Mokhtarian
Institute of Transportation Studies ◊ University of California, Davis
One Shields Avenue ◊ Davis, California 95616
PHONE: ( 530) 752- 6548 ◊ FAX: ( 530) 752- 6572
WEB: http:// its. ucdavis. edu/
An Empirical Analysis of Causality in the Relationship between Telecommuting
and Residential and Job Relocation
David T. Ory
Department of Civil and Environmental Engineering
and
Institute of Transportation Studies
University of California, Davis
Davis, CA 95616
voice: ( 415) 378- 9102
fax: ( 530) 752- 6572
e- mail: dtory@ ucdavis. edu
and
Patricia L. Mokhtarian
Department of Civil and Environmental Engineering
and
Institute of Transportation Studies
University of California, Davis
Davis, CA 95616
voice: ( 530) 752- 7062
fax: ( 530) 752- 7872
e- mail: plmokhtarian@ ucdavis. edu
February 2005
This research is funded by the National Science Foundation Integrative Graduate Education and
Research Traineeships ( IGERT) program and the University of California Transportation Center
( UCTC)
ii
Table of Contents
List of Tables......................................................................................................................... ...... iii
Executive Summary ..................................................................................................................... iv
1. Introduction ............................................................................................................................... 1
2. Empirical Setting and Available Data........................................................................................ 4
3. Telecommuter Definitions ......................................................................................................... 6
4. Relocation Patterns of Telecommuters and Non- telecommuters ............................................. 8
5. Timing of Relocations and Telecommuting Engagement ....................................................... 12
6. Stated Importance of Telecommuting ..................................................................................... 17
7. Modeling the Change in Commute Characteristics................................................................. 19
8. Conclusion .............................................................................................................................. 22
References..................................................................................................................... ............ 23
iii
List of Tables
Table 1: Key Characteristics of the Sample.................................................................................. 5
Table 2: Summary of Residential- Only Relocations by Telecommuters and Non- telecommuters9
Table 3: Differences in Residential Move Commute Characteristics among Current,
Former/ Future, and Never Telecommuters ........................................................................... 9
Table 4: Summary of Job- Only Relocations by Telecommuters and Non- telecommuters.......... 11
Table 5: Differences in Job Move Commute Characteristics among Current, Former/ Future, and
Never Telecommuters ......................................................................................................... 12
Table 6: Causal Inferences Based on Timing of Moves and Telecommuting Engagement ....... 13
Table 7: Changes in Commute Characteristics of Residential Movers Segmented by
Telecommuting Causality Categories .................................................................................. 14
Table 8: Temporal Incidence of Telecommuting Episodes and Residential Relocations ........... 15
Table 9: Changes in Commute Characteristics of Residential Movers Segmented by
Telecommuting Causality Categories with Manual Coding Changes .................................. 16
Table 10: Changes in Commute Characteristics of Job Movers Segmented by Telecommuting
Causality Categories............................................................................................................ 17
Table 11: Stated Importance of Telecommuting in Residential Relocation Decision.................. 18
Table 12: Linear Regression Model of Change in Commute Length following a Residential
Relocation ( N= 96)................................................................................................................ 21
Table 13: Linear Regression Model of Change in Commute Minutes following a Residential
Relocation ( N= 94)................................................................................................................ 21
iv
Acknowledgements
The initial survey design, data collection, and analysis for this project were funded by BMW AG
and performed by Dr. Carsten Gertz and Vikram Thairani together with the second author of this
report. Funding from the University of California Transportation Center permitted further analysis
of the data by Gustavo Collantes. This previous work was invaluable in preparing the data for
the analysis presented here, which is funded by a National Science Foundation Integrative
Graduate Education and Research Traineeships ( IGERT) fellowship. The assistance of Colby
Brown and valuable comments of Ilan Salomon are gratefully acknowledged. We also
acknowledge with appreciation the assistance of the State of California staff who facilitated the
data collection effort, including David Fleming, Department of Personnel Administration; Nan
Powers, California Energy Commission; Sue Teranishi, Franchise Tax Board; Elaine Scanlon,
Department of Motor Vehicles; Dorothy Carter, California Youth Authority; and Tom Benson,
Department of Social Services.
Executive Summary
Researchers have questioned whether the ability to telecommute is encouraging workers to
relocate to more desirable residences farther from work, and in doing so, exacerbate sprawl and
increase their net vehicle- miles traveled. The research presented here directly asks, is
telecommuting a “ friend or foe” of travel- reducing policies? Given that telecommuters tend to
have longer commutes than non- telecommuters, is the ability to telecommute prompting
workers to move farther away from work? Or, does the ability to telecommute allow those who
for other reasons have already chosen, or would in any case choose, to live in more distant
locations to commute less frequently? These questions are addressed using data collected from
more than 200 State of California workers, including current, former, and non- telecommuters.
The survey inquired retrospectively about their residential and job relocations, as well as their
telecommuting engagements, over a ten- year period.
The results indicate that, as expected, residential and job moves that are temporally associated
with telecommuting episodes tend to increase commute time and length compared to other
moves – though the evidence is not statistically significant. Analyzing the temporal order of
telecommuting engagement and residential relocation, the data show that those who are
telecommuting and then move actually tend to relocate closer to their workplace, whereas those
who begin telecommuting following a residential relocation tended to have moved much farther
from their workplace. For job relocations, the results differed slightly. Here, both key casual
groups ( those who are inferred to have their relocation caused by telecommuting and those who
are inferred to have their telecommuting engagement caused by relocation) relocate, on
average, to jobs farther from home. Analysis of the stated importance of telecommuting to
specific residential relocations did not show a convincing effect toward more distant moves.
Linear regression models of the change in one- way commute length following a residential
relocation confirm that the beginning of a telecommuting engagement following the move is
associated with increases in commute length, whereas engagement before the move is not.
Thus, the evidence more strongly supports the positive view of telecommuting, that it is
ameliorating the negative transportation impacts of moves that occur for other reasons.
1
1. Introduction
Telecommuting has found a niche in U. S. transportation policy. In response to a number of
convincing empirical studies ( see, e. g., Hamer, et al., 1991; Mokhtarian, et al., 1995) that have
found net reductions in the vehicle- miles traveled by telecommuters ( on days they
telecommute), policy makers at the state and national level have begun to incorporate
telecommuting alternatives into law. United States Public Law 106- 346 ( Sec. 359, 2001), for
example, states, “ each executive agency [ of the Federal government] shall establish a policy
under which eligible employees of the agency may participate in telecommuting to the maximum
extent possible without diminished employee performance.”
The problem with the empirical evidence to date is that it tracks travel behavior over a very short
period of time after the adoption of telecommuting – typically one to two years. The long- term
impacts of telecommuting on travel, specifically commute travel, are not fully known. Certainly
the evidence so far indicates telecommuting may reduce overall travel, but some researchers
( see, e. g. Janelle, 1986 for an early source of this idea) have suggested that the ability to
telecommute could, over time, prompt individuals to move farther away from their jobs to
cheaper or higher- amenity residential locations. Such moves would require longer, though less
frequent, commutes that may lead to a net increase in vehicle- miles traveled ( VMT). To the
extent such behavior occurs, telecommuting alternatives could accelerate urban sprawl, and
may not serve their intended role of reducing travel. Suspicion of telecommuting deepens with
the observation of several studies ( see e. g., Mokhtarian, et al., 1995; Gareis, 2003; Shaw, 2004)
that telecommuters tend to have significantly longer commutes than non- telecommuters.
On the other hand, the latter observation is not conclusive because the causal relationship could
be in the opposite direction: people who have already moved far from work would naturally be
more motivated to adopt telecommuting after the fact, as a solution to, rather than a cause of,
their longer commute.
Thus, we ask: is telecommuting a friend or foe of travel reduction policies? Is the ability to
telecommute prompting workers to move farther away from work and, hence, perhaps to
increase their overall commute distance? Or, does the ability to telecommute allow those, who
for other reasons have already chosen ( or would in any case choose) to live in more distant
locations, to commute less frequently and, hence, to reduce their total commute amount from
what it would have been without telecommuting, even if it is still more than it was before the
move?
It is important to keep in mind that telecommuters would not move farther from work just
because they are able to telecommute. Rather, in the “ foe” scenario, telecommuting simply
makes possible a move that is desired for other reasons – to accommodate household needs,
to obtain a larger and/ or cheaper home, to be near scenic locations or other amenities, and so
on. Thus, in this scenario, telecommuting can be viewed as a facilitator, not an actual driver, of
such moves. Nevertheless, to the extent telecommuting is responsible for releasing a constraint
that was preventing further decentralization ( and to the extent such decentralization is
considered socially undesirable), telecommuting is arguably as culpable as if it were a driver in
its own right. In reality, of course, it could be rather difficult to disentangle the relative roles of
drivers and facilitators, to determine the extent to which one factor could be singled out as being
responsible for the move, as opposed to a number of factors acting in concert, with indivisible
impact.
2
Several studies have examined the long- term impacts of telecommuting on residential location,
and hence urban form, using economic theory. Higano and Orishimo ( 1990), Lund and
Mokhtarian ( 1994), Kim ( 1997), and Safirova ( 2001) assumed a monocentric city; Stough and
Paelinck ( 1996) and Shen ( 2000) relax the monocentric assumption in different ways. Each of
these studies found, not surprisingly, that telecommuting can increase the rate of
decentralization of a city. However, none of these theoretical models has been empirically
verified. Unfortunately, the ideal data set, which would track telecommuters and comparable
non- telecommuters over a substantial period of time ( e. g. 10 years), does not exist. As a result,
it remains difficult to confirm or refute the alternative conjectures as to the causal relationship
between telecommuting and residential ( or job) relocation.
In lieu of the ideal data set described above, the present study analyzes a single cross- sectional
survey of current, former, and non- telecommuters – inquiring retrospectively about their
residential and job relocations, as well as their telecommuting engagement patterns, over a ten-year
period ( from the fall of 1988 to the fall of 1998). Though such data are obviously subject to
recall error, the changing of residence or job location is probably an important enough life event
to be recalled with the desired accuracy – the unit of analysis being a quarter- year or 3- month
period ( for a more detailed discussion of the use of retrospective surveys for collecting
residential relocation data, see Hollingworth and Miller, 1996). Telecommuting engagement is
probably a less significant life event, though perhaps memorable enough that respondents
seemed to have little trouble answering the questions in the survey ( relatively few missing or
unclear data). To our knowledge, this study ( including other analyses of the same data)
constitutes the first attempt to empirically examine relationships among telecommuting,
residential location, job location, and total commute travel over a ten- year period.
A previous analysis of these data ( Mokhtarian, et al., forthcoming) confirmed that in this sample,
too, telecommuters lived farther from work than non- telecommuters. However, it also found that
the telecommuters telecommuted often enough to more than compensate for this: the total per
capita commute distance traveled by telecommuters was smaller than for non- telecommuters. A
later extension of that analysis ( Ory and Mokhtarian, 2004) found that telecommuters also
commuted for shorter durations than non- telecommuters, with the majority of that time savings
due to faster travel speeds. Those studies, however, did not address whether the longer one-way
commute length of telecommuters was a cause, or an effect, of their telecommuting. In
contrast, the focus of the present analysis will be on attempting to ascertain causality.
Attributing causality in the relationship between residential/ job relocation and telecommuting is
difficult because individuals move closer to and farther from work all the time and take jobs that
are closer to and farther from their homes all the time, for a variety of reasons that most often
have nothing to do with telecommuting. The behavior of telecommuters can theoretically be
assessed against that of comparable non- telecommuters, but in practice it can be a challenge to
identify an appropriately comparable control group.
In the present study, we will use four distinct methods of analysis. First, we will compare the
changes in one- way commute length following residential or job relocations for telecommuters
and non- telecommuters ( defined in various ways), to answer the question: are telecommuters
moving farther from work/ home than non- telecommuters? If they are not ( as was found in an
earlier, short- term, study of State of California workers conducted by Nilles, 1991), the
conceptual basis for inferring that telecommuting is stimulating further decentralization is
undermined. If they are, on the other hand, it still remains to ascertain whether the increased
commute length is caused by telecommuting, or whether telecommuting results from the move.
Thus, this analysis does not address the causality question directly: if telecommuters are found
3
to be moving home and job farther apart than non- telecommuters, we still do not know if that is
an effect or cause of telecommuting. It does, however, constitute a necessary benchmark
against which to gauge the rest of the investigation.
In the second method of analysis, the timing of the moves in relation to the timing of
telecommuting engagement is used to infer causality: temporal precedence of X over Y is
generally viewed ( where there is also conceptual and statistical justification) as strengthening
the inference that X caused Y rather than the reverse ( see, e. g., Baumrind, 1983; Holland,
1986). However, temporal precedence is not completely definitive, of course. For one thing, the
timing might simply be coincidence, and neither event be related to the other. After all, people
move closer to and farther from work all the time, for a variety of reasons that most often have
nothing to do with telecommuting. Conversely, people telecommute for reasons having nothing
to do with reducing travel, such as increasing their concentration ( one- way commute distances
as short as three miles have been reported by telecommuters in some studies).
Even if a relationship is present, it might not be the obvious one. If Y follows X, did X cause Y,
or did the anticipation of Y cause X? Though the implementation of telecommuting may follow a
residential move in time, the actual choice to telecommute may have preceded and prompted
the move, and conversely, if a residential move followed a telecommuting episode, it is possible
that the decision to move preceded and prompted the telecommuting.
To help address these kinds of ambiguities, one may say, “ Why not just ask whether the move
took place because of telecommuting?” The survey instrument used in this study does, in fact,
directly ask respondents to rate the importance that telecommuting played in their relocation
decision, and analyzing these responses is the third analysis approach we use. But these
answers may not be reliable ( though still worth examining) because respondents completing a
survey regarding their telecommuting engagement may consciously or unconsciously report
telecommuting as being more important than it really is, due to their heightened awareness of it.
The final method of analysis estimates linear regression models of the change in one- way
commute duration and length following a residential relocation. The models consider the impact
of socio- demographics, stated reasons for moving, telecommuting status, and work schedule,
among other variables. The purpose of the modeling is to determine the relative importance of
telecommuting status ( before and after the move) while controlling for measures more
traditionally associated with relocation decisions ( e. g. the presence of small children), and to
ascertain if the timing of the telecommuting engagement impacts the resulting commute
characteristics. These models are less than ideal in that only the most recent residential
relocations are considered – a necessary restriction as socio- demographic traits only available
for the time of data collection are used to predict these relocation decisions made in the past.
These four methods – comparing the resulting commute characteristics following the relocation
of telecommuters and non- telecommuters, assessing the timing of moves, examining the stated
causality for the moves, and modeling the change in one- way commute length and duration –
though each flawed in different ways, together should offer useful insight into the causal
relationship between telecommuting and residential and job relocation. To the extent the
multiple methods point to similar conclusions, our confidence in the result increases.
The organization of this report is as follows. The next section discusses the data in greater
detail. Section 3 provides a detailed discussion of how the respondents are labeled as
telecommuters ( which is needed for comparing telecommuters to non- telecommuters) through
time. Section 4 presents the first analysis: comparing the relocation patterns ( both employment
4
and residential) of telecommuters and non- telecommuters. Section 5 addresses the issue of
causality by investigating the timing of relocations in relation to the timing of telecommuting
engagements. Section 6 examines the stated importance of telecommuting in the decision to
make a residential move, as recorded in the survey instrument, and Section 7 presents linear
regression models of changes in commute duration and length following a residential move. A
concluding section ends the report.
2. Empirical Setting and Available Data
From 1988 to 1990, the State of California conducted one of the best- known early
telecommuting pilot programs for its employees, involving around 150 telecommuters in 14 state
government agencies ( JALA Assoc., 1990; Kitamura et al., 1990). Through the years,
telecommuting has continued to thrive in some of these agencies, offering an opportunity to
explore the long- term relationships of interest in this study.
To gather the desired data, a 16- page self- administered survey was designed and distributed in
November 1998 to employees of the following six California state agencies: California Energy
Commission, Department of Personnel Administration, Franchise Tax Board, California Youth
Authority, Department of Motor Vehicles, and Department of Social Services. Each of these
agencies has kept their telecommuting programs active since the pilot implementation in 1988.
The survey was distributed to those who responded to an initial broadcast email message, sent
to key divisions or groups within each agency. The message stressed the need for participation
from telecommuters, non- telecommuters, and former telecommuters, and offered a drawing for
cash prizes of $ 250, $ 150, and $ 100. Due to the intended approach of enriching the sample
with telecommuters, the data are not representative of any general population – i. e., the ratio of
telecommuters to non- telecommuters in the sample is higher than in the population as a whole.
However, to the extent that each subsample is representative of the population from which it is
drawn, comparisons of average behavior across subsamples will be valid even if the share of
the sample in each group is not itself representative.
Thus, more important, and independent from the representativeness of the telecommuter/ non-telecommuter
ratio, is the question of whether the telecommuters in the sample are
representative of the general population of telecommuters. Unfortunately, there are no reliable
data on the demographics and other characteristics of telecommuters in the population, making
any comparison of our sample ( in terms of gender, income, etc) to the population of
telecommuters impossible. We speculate, however, that over the 10- year retrospective period
covered by the survey, the sample of telecommuters may be less representative the farther
back in time we go. That is, we suspect that from our sample, collected entirely in 1998, the
subset identified as telecommuting in Fall 1988 could be less representative of telecommuters
at that point in time than is the subset identified as telecommuting in Summer 1998. Specifically,
we speculate that people who remember telecommuting 10 years ago well enough to report it
on a survey a decade later may have had more extreme circumstances than the typical
telecommuter of that time ( see Mokhtarian, et al., forthcoming for further discussion on this
issue, and on whether the non- telecommuters serve as an appropriate control group). An
additional potential for bias is the sample being composed solely of state employees. The
attitudes, opinions, and behaviors of state employees may differ in a systematic, or biased, way
from those employed in the private sector ( though no particular hypothesis regarding the current
analysis presents itself). Important socio- demographic characteristics of the sample are
included in Table 1.
5
Table 1: Key Characteristics of the Sample
Total sample Current
telecommuters
Former
telecommuters
Never-
Telecommuters
Categorical/ ordinal variables: Frequency ( column %)
Gender
Female 150 ( 68.8) 45 ( 72.6) 26 ( 74.3) 79 ( 65.3)
Male 68 ( 31.2) 17 ( 27.4) 9 ( 25.7) 42 ( 34.7)
Total 218 ( 100.0) 62 ( 100.0) 35 ( 100.0) 121 ( 100.0)
Occupation
Manager/ Administrator 29 ( 13.3) 6 ( 9.7) 6 ( 17.1) 17 ( 14.0)
Professional/ Technical 161 ( 73.9) 53 ( 85.5) 24 ( 68.6) 84 ( 69.4)
Administrative support 23 ( 10.6) 2 ( 3.2) 5 ( 14.3) 16 ( 13.2)
Other 3 ( 1.4) 1 ( 1.6) - (-) 2 ( 1.7)
Total 216 ( 99.1) 62 ( 100.0) 35 ( 100.0) 119 ( 98.3)
Education
Some grade or high school 2 (. 9) - (-) 1 ( 2.9) 1 (. 8)
High school graduate 3 ( 1.4) - (-) - (-) 3 ( 2.5)
Some college 86 ( 39.4) 20 ( 32.3) 14 ( 40.0) 51 ( 43.0)
Four- year college degree 62 ( 28.4) 22 ( 35.5) 6 ( 17.1) 34 ( 28.1)
Some graduate school 32 ( 14.7) 10 ( 16.1) 5 ( 14.3) 17 ( 14.0)
Completed graduate degree 31 ( 14.2) 10 ( 16.1) 8 ( 22.9) 13 ( 10.7)
Total 216 ( 99.1) 62 ( 100.0) 34 ( 97.1) 120 ( 99.2)
Annual personal income
Less than $ 15,000 3 ( 1.4) 1 ( 1.6) - (-) 2 ( 1.7)
$ 15,000 to 34,999 40 ( 18.3) 5 ( 8.1) 4 ( 11.4) 31 ( 25.6)
$ 35,000 to 54,999 93 ( 42.7) 27 ( 43.5) 21 ( 60.0) 45 ( 37.2)
$ 55,000 to 74,999 49 ( 22.5) 18 ( 29.0) 3 ( 8.6) 28 ( 23.1)
$ 75,000 to 94,999 18 ( 8.3) 5 ( 8.1) 4 ( 11.4) 9 ( 7.4)
$ 95,000 or more 15 ( 6.9) 6 ( 9.7) 3 ( 8.6) 6 ( 5.0)
Total 218 ( 100.0) 62 ( 100.0) 35 ( 100.0) 121 ( 100.0)
Residential area type
Large city 91 ( 41.7) 21 ( 33.9) 15 ( 42.9) 55 ( 45.5)
Suburb of large city 59 ( 27.1) 18 ( 29.0) 11 ( 31.4) 30 ( 24.8)
Medium- size city 25 ( 11.5) 10 ( 16.1) 5 ( 14.3) 10 ( 8.3)
Small city 17 ( 7.8) 2 ( 3.2) 3 ( 8.6) 12 ( 9.9)
Town or village 6 ( 2.8) 1 ( 1.6) - (-) 5 ( 4.1)
Countryside 18 ( 8.3) 9 ( 14.5) 1 ( 2.9) 8 ( 6.6)
Total 216 ( 99.1) 61 ( 98.4) 35 ( 100.0) 120 ( 99.2)
Continuous variables: Mean ( s. d., N)
Age 43.2 ( 8.7, 217) 43.2 ( 8.6, 62) 44.7 ( 8.9, 35) 42.8 ( 8.8, 120)
Household size 2.8 ( 1.4, 218) 2.7 ( 1.3, 62) 3.0 ( 1.1, 35) 2.9 ( 1.4, 121)
Source: Mokhtarian, et al. ( 2004)
6
The survey instrument contained two 10- year timelines ( segmented into quarter- years) that
captured the key data for this analysis. On the first timeline, current and former telecommuters
indicated all the periods of time during which they telecommuted regularly, the frequency with
which they telecommuted during each of those periods, and reasons for quitting or changing
frequency in each case. (“ Regular” telecommuting was defined as “ at least two days a month on
average, for at least three consecutive months”). On the second timeline, all respondents
recorded their job and residential relocations that took place during the 10- year span and, for
each job- residence location pair ( including the pair at the beginning of the 10- year period),
indicated their commute length, time, and mode. The two timelines were separated by three
pages, so although respondents could have made a conscious effort to “ match them up”, they
were not particularly led to do so by the survey design.
In addition to the timelines, the survey collected data on attitudes toward telecommuting;
reasons for, and other characteristics of, their three most recent residential and two most recent
job relocations; impacts of telecommuting on relocations and vice versa; impacts of
telecommuting on frequency and destination for a number of trip purposes; general
transportation- related choices; job characteristics, home characteristics, and standard socio-demographic
characteristics. The current study focuses on the information available from the
two timelines as well as the questions about the role of telecommuting in relocations;
preliminary analysis of other parts of the survey can be found in Gertz and Mokhtarian ( 1999).
This study analyzes the 218 individuals having essentially complete timeline responses. Over
the 10- year period, these 218 individuals changed job locations only ( i. e. no accompanying
residential change) 171 times, changed residential locations only 198 times, and changed both
job and residential locations in the same quarter 76 times. The focus of the analysis presented
here will be on the changing commute characteristics of each move, comparing those
characteristics between telecommuters and non- telecommuters, and the timing of those moves
in relation to telecommuting engagement. Of course, as indicated earlier, the job and residential
relocations examined could be taking place for any number of reasons, and may not even be
the decision of the individual reporting the change ( e. g. the spouse may be driving the
residential move or a layoff may cause the job change). However, to the extent these varying
circumstances are present in both the telecommuter and non- telecommuter samples, the
differences between the two groups can be suggestively attributed to telecommuting.
3. Telecommuter Definitions
When comparing the behavior of telecommuters to non- telecommuters in the first part of our
analysis, the means of labeling each individual, or more specifically labeling each individual in
each time period, as a telecommuter or a non- telecommuter is very important. Various plausible
definitions of telecommuter represent various expectations or hypotheses regarding behavioral
differences. For example, it may be that those who are currently telecommuting behave
differently than those who are currently not telecommuting. Or, it may be that those who have
telecommuted in the past behave differently than those who have not. Or, it may be that those
who at some point become telecommuters differ from those who never do.
To investigate these different potential behavioral influences of telecommuting on travel, we
defined each individual during each time period ( quarter- year) as a telecommuter or non-telecommuter
in four distinct ways. Here, we present and briefly motivate each of the four
definitions.
7
Definition 1: Current Telecommuter
The first definition labels an individual during each quarter- year as a telecommuter if she is
actively engaged in a “ regular” telecommuting episode ( again, defining “ regular” as “ at least two
days a month on average, for a period of at least three consecutive months”) during that
quarter. This definition allows for straightforward analysis: do those who are currently
telecommuting behave differently, in terms of travel and relocation behavior, than those who are
not?
Definition 2: After Telecommuter
The second definition can be re- stated, more directly, as, “ once a telecommuter, always a
telecommuter.” After an individual has started telecommuting regularly, she is considered to be
a telecommuter for each remaining quarter- year in the 10- year span. This definition is based on
the concept that individuals, once exposed to telecommuting, may behave differently than those
who have not yet tried it, or will never try it. It may be that individuals, once they telecommute,
have a certain “ footloose” location choice set, even though they may not be telecommuting at
the time of their moves.
Definition 3: Before or After Telecommuter
The third definition is an expanded or “ fuzzy” version of Definition 1. Here, an individual is
considered a telecommuter during quarters in which he telecommutes regularly, as well as the
single quarters before and after a telecommuting episode. This broader definition allows for the
fact that major decisions, such as job and residential relocations, may take place over a period
of time greater than a quarter- year; we want to allow the influence of telecommuting, if
engagement started or ended around this time, to be considered as part of that decision. It also
helps compensate for the use of discrete intervals to demarcate a timeline that is actually
continuous. In other words, an individual may end a telecommuting engagement in one quarter
and then change residences in the next, though the events could take place only a day apart.
Definition 4: Ever Telecommuter
The final definition broadens Definition 2. Here, if an individual engages in regular
telecommuting at any point during the study period, she is considered to be a telecommuter for
all quarter- years. Thus, alone among the four definitions, this definition is irrelevant to the
timeline and holds at the individual ( rather than the individual- quarter) level. This definition could
be said to capture people who are prone to telecommute. It may be that some personality or
other characteristics that drive them to telecommute may also drive them to engage in different
types of travel and relocation behavior – independent of their current telecommuting
engagement status. Of course, this definition will unavoidably erroneously classify some
telecommuters as non- telecommuters, when their ( past or future) telecommuting engagement
takes place outside the 10- year study period.
Ironically, none of these definitions can be considered “ definitive.” Should one have to
telecommute longer than one quarter before being labeled a telecommuter? Should a
telecommuting episode have to overlap a specified length of time with a relocation before a
relationship can be inferred? For example, if one stops telecommuting the quarter after a move,
can it reasonably have been a cause of the move? What if a move prompts a telecommuting
episode, but not until six months later? Our definitions do not resolve all possible difficulties, but
given the arbitrary nature of any alternative requirements on telecommuting duration, the fact
that different durations may be relevant for different people, and the limitations of the small
8
sample size, we focus on the four alternatives defined above. They span the spectrum from
most conservative ( Definition 1: Current Telecommuter) to most liberal ( Definition 4: Ever
Telecommuter), and together should bracket the range of plausible relationships.
4. Relocation Patterns of Telecommuters and Non- telecommuters
In this first stage of analysis, the residential and job relocation patterns ( and, more importantly,
the commute characteristics) of telecommuters ( defined in the four separate ways presented in
Section 3) and non- telecommuters are compared. The goal is to determine if telecommuters
( both active and inactive), on average, select homes that are farther away from their
workplaces, or choose jobs that are farther away from their homes than non- telecommuters do.
Table 2 cross- tabulates the resulting commute characteristics of telecommuters and non-telecommuters
following residential moves made by respondents in the sample ( those making
“ simultaneous” job and residential relocations, i. e. within the same quarter, are not considered in
this analysis on the assumption – supported by Clark, et al., 2003 – that such moves could differ
importantly from residential- only moves, but without enough of them – only 2 of the 76 moves
were made by telecommuters – to analyze separately). The table shows an overall average
increase in one- way commute length and duration of about 2.02 miles and 2.40 minutes,
respectively. This result is expected and indicates that the sample as a whole is increasing its
commute length and duration following a residential move. Since the sample contains a minority
of telecommuters, this result is a useful reminder that decentralization is by no means confined
to telecommuters. The fact that the average changes in commute time and length are positive
for the sample overall, and for the non- telecommuters as well as telecommuters, reinforces the
observation that ongoing decentralization is occurring independent of telecommuting, and
complicates the effort to separate any effects of telecommuting specifically, from this general
trend affecting everyone.
When segmenting the sample based on telecommuting status, however, it becomes evident that
telecommuters, on average, tend to move farther away from their workplaces than non-telecommuters.
In each of the four definitions, telecommuters’ moves result in greater commute
durations than do non- telecommuters’; in all but Definition 2: After Telecommuter, a longer-distance
commute emerges. The differences in commute characteristics are most pronounced
using Definition 3: Before/ After Telecommuter. However, this definition is perhaps the one most
likely to confound cause and effect. Since the individual would be classified as a telecommuter
the quarter before she actually started, if the residential relocation occurred in that prior quarter,
it is more likely the move that prompted the telecommuting rather than the converse.
While the results do show a general trend of telecommuters relocating farther from their jobs
than non- telecommuters, it should be noted that the differences between these groups are not
statistically significant at the 95 percent confidence level – a result of the small sample size and
relatively large variances in length and duration changes.
A more detailed cross tabulation of changes in commute distance following a residential move is
presented in Table 3. Here, the sample is first segmented into Ever Telecommuters ( those who
at some point in the ten- year study period engaged in regular telecommuting) and Never
Telecommuters ( those who did not engage). Ever Telecommuters are then segmented further
by the Current Telecommuter definition ( as in Table 2) to distinguish between telecommuting
and non- telecommuting quarters. This segmentation allows for a comparison among three
groups: those who are currently telecommuting in a quarter of residential relocation, those who
are not currently telecommuting but have telecommuted in the past or will telecommute in the
9
future ( within the study period), and those who at no point in the study period telecommuted. We
further segment moves by whether they are closer to work, farther from work, or the same
distance away.
Table 2: Summary of Residential- Only Relocations by Telecommuters and Non- telecommuters
Telecommuter
Definition Status Number of
Moves1
Avg. Change in
Cmt Length ( mi)
Avg. Change in
Cmt Duration ( min)
Everyone --- 183, 181 2.02 2.40
1: Current Telecommuter 19, 19 2.95 6.74
Telecommuter
Non- telecommuter 164, 162 1.91 1.89
Telecommuter 31, 30 1.06 3.00
2: After Telecommuter
Non- telecommuter 152, 151 2.22 2.27
3: Before/ After Telecommuter 29, 29 5.76 7.93
Telecommuter
Non- telecommuter 154, 152 1.32 1.34
Telecommuter 72, 71 3.52 5.17
4: Ever Telecommuter
Non- telecommuter 110, 109 1.02 0.70
1 The numbers shown are the number of moves ( not respondents) with valid commute length ( first number) and
duration data ( second number).
Table 3: Differences in Residential Move Commute Characteristics among Current, Former/ Future,
and Never Telecommuters
Ever Telecommuter
Current Telecommuter Not Current Telecommuter
Direction of Never Telecommuter
move in
relation to
job location N1 (%) Delta
length2 N (%) Delta length N (%) Delta length
Closer 9 ( 47.4) - 13.1 18 ( 34.0) - 14.3 36 ( 32.7) - 9.2
Zero 2 ( 10.5) 0.0 7 ( 13.2) 0.0 27 ( 24.5) 0.0
Farther 8 ( 42.1) 21.8 28 ( 52.8) 16.3 47 ( 42.7) 9.4
All 19 ( 100.0) 2.95 53 ( 100.0) 3.72 110 ( 100.0) 1.02
1 Number of valid commute length measures in each group; 2 Average change in one- way commute length,
in miles, following a residential relocation.
10
Table 3 indicates that those who are currently telecommuting behave differently from those who
never telecommute, but not that differently than those who telecommute at some point, yet are
not doing so at the current time. The difference in change in commute length for those who are
currently telecommuting and those who are not, but have in the past or will in the future, is
rather small. In contrast, those who never telecommute have a much smaller average change in
commute length than both of the other groups.
Table 3 shows that, ironically, for the two groups not telecommuting in the relocation quarter,
moves were more often farther from work than closer to it, whereas Current Telecommuters
moved closer to work about as often as they moved farther away. This picture shifts, however,
when changes in the commute length itself are considered. For the Never Telecommuters,
moves farther from work increased the commute by slightly more than moves closer to work
reduced it, and there are enough more of the former that the net average change in commute
length was one mile farther away. Similar but stronger patterns for the Not Current
Telecommuters result in a net average increase of 3.7 miles. For Current Telecommuters, on
the other hand, moves farther from work increased the commute substantially more than moves
closer to work reduced it, but since the numbers of moves in each category are nearly balanced,
the net increase is three miles, smaller than for the Not Current Telecommuters.
In sum, Current Telecommuters move farther from work relatively less often than do non-telecommuters,
but when they do move farther away, it tends to be substantially farther away
than non- telecommuters. The net change in commute length for Current Telecommuters falls
between that of Not Current Telecommuters and Never Telecommuters. Again, however, the t-statistics
comparing the overall means of these three groups show no statistical difference
between any pair of groups at the 95 percent confidence level. Similarly, chi- squared tests show
no statistical difference in the distribution among these groups at the 95 percent level ( the
moves were segmented within groups based on the distance of the moves, information not
shown in Table 3), though the Current Telecommuter group is distributed differently from the
Never Telecommuter group at the 90 percent level.
Overall, then, for those making a residential relocation, indications are that telecommuters are
moving farther away from their workplaces than non- telecommuters, but the statistical evidence
is relatively weak.
Turning now to job relocations, a similar set of analyses is presented. As mentioned before,
individuals may have little to no control over a change in job location. However, to the extent
that circumstances independent of telecommuting apply to each of the population segments,
and we have no reason to believe they do not, we can infer that the differences in commute
characteristics are related to telecommuting.
Table 4 presents the change in commute characteristics following a job relocation for telecom-muters
and non- telecommuters. It shows a pattern similar to Table 2, with a general trend of
telecommuters changing to jobs which are located farther from their homes than non-telecommuters.
This pattern holds for changes in commute duration for all definitions save
Definition 4: Ever Telecommuter ( which is the loosest definition); it holds for changes in
commute length for Definition 1: Current Telecommuter ( where the magnitudes of the
differences are largest) and Definition 3: Before/ After Telecommuter. Though a trend does
appear, the differences between telecommuters and non- telecommuters within any definition
are not statistically significant at the 95 percent confidence level.
11
Table 4: Summary of Job- Only Relocations by Telecommuters and Non- telecommuters
Telecommuter
Definition Status Number of
Moves1
Avg. Change in
Cmt Length ( mi)
Avg. Change in
Cmt Duration ( min)
Everyone --- 153, 153 1.95 1.89
1: Current Telecommuter 16, 16 5.31 7.69
Telecommuter
Non- telecommuter 137, 137 1.55 1.21
Telecommuter 23, 23 1.57 3.83
2: After Telecommuter
Non- telecommuter 130, 130 2.02 1.55
3: Before/ After Telecommuter 21, 21 3.05 4.52
Telecommuter
Non- telecommuter 132, 132 1.77 1.47
Telecommuter 72, 72 1.55 0.96
4: Ever Telecommuter
Non- telecommuter 79, 79 2.25 2.59
1 The numbers shown are the number of moves ( not respondents) with valid commute length ( first number) and
duration data ( second number).
In addition to comparing commute changes across telecommuter definitions, a closer
examination of those who have engaged in telecommuting during the study period is presented
in Table 5. Similar to Table 3, the analysis presented in Table 5 disaggregates the behavior of
Ever Telecommuters by whether they are or are not engaged in a telecommuting episode during
the job move quarter.
The table does present a general trend that those who are telecommuting tend to be selecting
job locations farther away from their homes than those who are currently not telecommuting.
However, the sample size is really too small to make any type of valid statistical inference.
To summarize, this section has demonstrated that during telecommuting episodes,
telecommuters tend to select residential and job locations with more spatial and temporal
separation than they do when not telecommuting and than Never Telecommuters do.
12
Table 5: Differences in Job Move Commute Characteristics among Current, Former/ Future, and
Never Telecommuters
Ever Telecommuter
Current Telecommuter Not Current Telecommuter
Direction of Never Telecommuter
move in
relation to
job location N1 (%) Delta
length2 N (%) Delta length N (%) Delta length
Closer 5 ( 31.3) - 11.2 20 ( 35.7) - 10.4 30 ( 38.0) - 11.1
Zero 3 ( 18.8) 0.0 10 ( 17.9) 0.0 10 ( 12.7) 0.0
Farther 8 ( 50.0) 17.6 26 ( 46.4) 9.0 39 ( 49.4) 13.1
All 16 ( 100.0) 5.31 56 ( 100.0) 0.47 79 ( 100.0) 2.25
1 Number of valid commute length measures in each group; 2 Average change in one- way commute length,
in miles, following a residential relocation.
5. Timing of Relocations and Telecommuting Engagement
The aim of Section 4 was to determine whether telecommuters, on average, choose to live and
work farther apart than non- telecommuters, and it appears they do, although statistically the
differences are not substantial. This section gets to the heart of the report: is the ability to
telecommute causing this increase in length and duration ( the “ foe” scenario), or is telecom-muting
allowing less frequent commuting to take place for moves that would have happened
anyway ( the “ friend” scenario)?
The timing of the residential and job moves in relation to the timing of telecommuting
engagement is examined in an attempt to answer these questions. The causal inferences based
on the timing of the relocation and telecommuting are summarized in Table 6. The rows in Table
6 represent the status of telecommuting engagement the quarter immediately preceding the
relocation; the columns represent the telecommuting engagement status the quarter
immediately following the relocation ( we later informally examine periods of time around the
move longer than one quarter). For example, if an individual is regularly telecommuting the
quarter preceding a move ( i. e. a Definition 1: Current Telecommuter) she is considered to be
“ telecommuting” before the relocation. If she is also telecommuting regularly the quarter after
the relocation, she is considered to be telecommuting after the relocation, and would be placed
in the lower- right hand element of the matrix. The telecommuting status the same quarter of the
relocation is not considered, since it cannot be used to establish temporal precedence and since
one’s status during the periods before and after the move is arguably more relevant to the
question of causality.
Each cell of the matrix allows for a different inference to be made about the causal relationship
between telecommuting and relocation. Individuals who telecommute neither before nor after a
relocation ( the upper- left cell) presumably moved independently of telecommuting, and are
referred to as the control group ( i. e. non- telecommuters). Similarly, if a person telecommuted
before, but not after a move ( the lower- left cell), it is concluded that telecommuting did not
influence the move ( although the move may well have influenced the cessation of
telecommuting); these cases are referred to as “ telecommute quitters.”
13
Table 6: Causal Inferences Based on Timing of Moves and Telecommuting Engagement
After Relocation
Not Telecommuting Telecommuting
Not Telecommuting No relationship ( control) Relocation caused telecommuting
( beneficial)
Before Relocation
Telecommuting No influence of telecommuting
( telecommute quitters)
Telecommuting caused relocation
( detrimental)
The key groups of interest are present in the right- hand side of the table. Individuals who do not
telecommute before a move, but do telecommute after a move ( the upper- right cell) are
assumed ( for the sake of discussion – as mentioned in the Introduction, this evidence is not
definitive) to have their telecommuting engagement caused by the relocation. Conversely, those
in the lower- right cell are assumed to have their relocation prompted by the ability to
telecommute because they begin telecommuting, then relocate, and continue to telecommute.
After segmenting the sample into these four causal blocks, the number of moves and net
changes in commute characteristics of each block, following relocation, can be examined. This
summary for residential relocations is presented in Table 7. Each of the four blocks in the table
contains the number of moves in each group, the resulting average change in one- way
commute duration ( in minutes), and the resulting average change in one- way commute length
( in miles). Though the two key groups have small sample sizes ( 13 and 15 moves, respectively),
the results could hardly be more disparate. Those who are said to have their telecommuting
engagement caused by their residential relocation tend to move away from their workplace –
increasing their commute time by about 18 minutes and their commute length 15 miles. In this
scenario, an individual locates to a more distant residential location and presumably in response
to the increased commute, begins telecommuting. Here, the telecommuting is a “ friend” of
travel- reducing policies, as the longer commute is relieved, to some degree, by telecommuting.
On the other side of the causal argument, those who are said to have their relocation caused by
telecommuting actually tend to move closer to their workplace – an average reduction in
commute time of 3 minutes and commute length of nearly 4 miles. Thus, in this possible “ foe”
scenario, the negative changes in commute duration make telecommuting at worst neutral to,
and at best a friend of travel- reducing policies. It may be implausible to attribute this reduction
specifically to telecommuting, but it is consistent ( to the extent jobs are found in central
locations) with the stereotype of the young, tech- savvy, telecommuter choosing to live in a more
urban, central location ( Ellen and Hempstead, 2002).
Table 8 presents a graphical representation of the residential relocation and telecommuting
patterns of the sample, on which the results in Table 7 are based. The residential relocations
are denoted by a two letter symbol ( NN, YN, NY, YY), which correspond to the four causal
groups discussed previously. Telecommuting engagements are denoted by solid black squares
in each quarter where a residential relocation did not take place, and by a dark outline in the
14
quarters where a residential relocation did take place ( only responses with a move relating to
the two key causal groups are presented in the table).
In addition to offering an easy visual representation of telecommuting engagement and
residential relocation through time, Table 8 also helps to validate the methodology used to
segment the sample into causal groups. Recall, the causal segmentation was based solely on
the telecommuting engagement status the quarter immediately preceding and immediately
following the residential relocation. If, for example, an individual telecommuted a full year before
the residential relocation, then only one quarter following the relocation, then had no more
episodes in the study period, our methodology would have assigned this person the causal
status of “ telecommuting caused relocation”, when, arguably, a more proper assignment would
have been the status of “ telecommute quitter”, because the individual essentially quit
telecommuting following the move. Instead of simply looking at the single preceding quarter and
the single following quarter, some thought was given to examining multiple quarters before and
after. Table 8 shows that doing so would not change a significant number of the causal
assignments. Perhaps three moves, found in records 5035, 1030, and 5005, would arguably
warrant having their causal statuses changed – to YY ( from YN), NY ( from YY), and YY ( from
NY), respectively. Changing these classifications would result in the outcomes shown in Table 9
– the substance of the findings remains intact.
Table 7: Changes in Commute Characteristics of Residential Movers Segmented by Telecom-muting
Causality Categories
After Relocation
Not Telecommuting ( N) Telecommuting ( Y)
Control Relocation causes telecommuting
148, 150 13, 13
1.23 minutes 17.69 minutes
Not Telecommuting
( N)
1.17 miles 14.54 miles
Telecommute quitters Telecommuting causes relocation
1, 1 15, 15
45.00 minutes - 3.00 minutes
Before Relocation
Telecommuting
( Y)
34.00 miles - 3.73 miles
Notes: N = no; Y = yes. The first row in each cell denotes the label given to the group; the second row
contains the number of observations – first for change in commute duration, then length – in each
segment; the third row contains the average change in one- way commute duration following a residential
move; and the fourth row contains the average change in one- way commute length following a residential
move.
15
Table 8: Temporal Incidence of Telecommuting Episodes and Residential Relocations
ID 4 1 129 893 4 1 129 903 4 1 129 913 4 1 129923 4 1 129933 4 1 129943 4 1 129953 4 1 129963 4 1 129973 4 1 129983 4
1030 􀂄 YY
􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
1034 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 Y
Y 􀂄
1056 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
1059 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 Y
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
2004 N
N 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 YY
􀂄 􀂄 􀂄 YY
􀂄 􀂄
3007 N
Y 􀂄 􀂄 􀂄 N
N N
Y 􀂄 􀂄 􀂄 YY
􀂄 􀂄 YY
􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 N
N
3010 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 Y
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
3020 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
4024 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5003 N
N N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5004 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5005 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5008 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 Y
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 YY
􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5035 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 Y
N 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5045 NY
􀂄 􀂄 􀂄 NY
􀂄 􀂄 􀂄 YY
􀂄
5047 N
N N
N N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5048 N
N N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5057 N
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5058 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 YY
􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5064 NY
􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5071 􀂄 􀂄 􀂄 Y
Y 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄 􀂄
5100 N
N N
N N
N N
N 􀂄 􀂄 YY
􀂄 􀂄 􀂄 YY
􀂄 􀂄 􀂄 􀂄 􀂄
Notes: YY = telecommuting engagement one quarter before and after move; YN = telecommuting engagement before, but not after move; NY = telecommuting engagement after, but not
before move; NN = not telecommuting before or after move; Dark box indicates telecommuting engagement the quarter of the move; solid square indicates telecommuting engagement in
non- move quarters.
16
Table 9: Changes in Commute Characteristics of Residential Movers Segmented by Telecom-muting
Causality Categories with Manual Coding Changes
After Relocation
Not Telecommuting ( N) Telecommuting ( Y)
Control Relocation causes telecommuting
148, 150 13, 13
1.23 minutes 18.46 minutes
Not Telecommuting
( N)
1.17 miles 14.31 miles
Telecommute quitters Telecommuting causes relocation
0, 0 16, 16
--- - 0.63 minutes
Before Relocation
Telecommuting
( Y)
--- - 1.19 miles
Notes: N = no; Y = yes. The first row in each cell denotes the label given to the group; the second row
contains the number of observations – first for change in commute duration, then length – in each
segment; the third row contains the average change in one- way commute duration following a residential
move; and the fourth row contains the average change in one- way commute length following a residential
move.
Turning back to job relocations, a similar causal analysis is presented; temporal precedence is
again used. Even though changing jobs is a fundamentally different decision than changing
homes ( i. e. it often has a weaker connection to household needs, other than financial, and, in
the case of layoffs or termination, may be done against the preferences of the individual), the
use of temporal precedence to assign causality in the relationship between telecommuting and
relocation is similarly applicable.
The job relocation data, segmented into the four causal groups, is presented in Table 10.
Similar to the results found in the analysis of residential relocations, the “ control” group ( those
not telecommuting before or after the relocation) are selecting jobs, on average, slightly farther,
in terms of length and duration, away from home. Also, a relatively small number of
observations are again found in the two key causal groups (“ relocation causes telecommuting”
and “ telecommuting causes relocation”). The key difference is that for the job relocations, both
causal groups have a larger increase in commute length and duration as compared to the
control group. As such, telecommuting, by our causal assumptions, is both alleviating vehicle-miles
traveled and potentially encouraging decentralization and increasing net travel amounts
( depending on the relative frequency of the before and after commutes).
17
Table 10: Changes in Commute Characteristics of Job Movers Segmented by Telecommuting
Causality Categories
After Relocation
Not Telecommuting ( N) Telecommuting ( Y)
Control Relocation causes telecommuting
124, 124 7, 7
1.38 minutes 5.00 minutes
Not Telecommuting
( N)
1.80 miles 2.00 miles
Telecommute quitters Telecommuting causes relocation
3, 3 10, 10
- 8.33 minutes 8.50 minutes
Before Relocation
Telecommuting
( Y)
- 7.67 miles 7.30 miles
Notes: N = no; Y = yes. The first row in each cell denotes the label given to the group; the second row
contains the number of observations – first for change in commute duration, then length – in each
segment; the third row contains the average change in one- way commute duration following a job move;
and the fourth row contains the average change in one- way commute length following a job move.
6. Stated Importance of Telecommuting
The third method of analysis examines the stated importance of telecommuting to residential
relocations. The survey instrument directly inquires about the importance of telecommuting to
residential moves ( but not job moves). Specifically, respondents could state that having the
ability to telecommute was “ most important”, “ important”, “ somewhat important”, or “ not at all
important” in their previous three residential relocations. We can then examine whether a high
reported importance of telecommuting translates into a residential location more distant from the
workplace. As discussed in the Introduction, these responses may be biased toward
exaggerating the importance of telecommuting.
Table 11 presents the responses for the stated importance of telecommuting to the residential
move decision, segmented by the four causal groups presented in Section 5. In addition to the
number of moves in each cell, the average change in commute length is presented.
Table 11 suggests that, despite the presumed bias in favor, telecommuting does not play a very
important role in residential relocation. Out of a total of 179 relocations, only 11 times did
respondents rate the role of telecommuting as “ the most important factor” ( 2 times), “ important”
( 5), or “ somewhat important” ( 4). Even among the 15 moves for which temporal precedence
suggests a causal role for telecommuting ( the “ Yes, Yes” column), it is given some degree of
importance to the move in fewer than half ( 7) of the cases – supporting the discussion in the
Introduction of the point that temporal precedence by itself does not establish causality.
However, those within that segment were considerably more likely to credit telecommuting ( 7 of
the 11 “ positive” responses mentioned above) than those who did not telecommute before the
move, but did so following the move ( 2 of the 11). This result is consistent with the assumed
direction of causality, based on the timing of the move and the telecommuting engagement,
18
made in the previous section. Also, the average changes in commute length for those
respondents reinforce the previous unexpected result, that those who relocated their residence
due to the ability to telecommute reduced their commute, on average. Of course, as noted
earlier, the temporal precedence rule is also flawed in the sense that the actual timing of the
decision ( as opposed to the action) to telecommute and/ or relocate is not known. As shown in
Table 11, the two respondents in the “ No, Yes” ( relocation causes telecommuting) column who
stated that telecommuting was “ important” to their moves ( even though they apparently did not
start telecommuting until after the move) may be erroneously labeled by the temporal
precedence methodology, which looks at the action ( not the decision) of telecommuting and
relocating. Because the average change in commute length for these two respondents is very
high ( 48.50 miles), reassigning them to the opposing causal group could significantly change
the results and present telecommuting as more of a potential “ foe” to travel reduction policies.
Table 11: Stated Importance of Telecommuting in Residential Relocation Decision
Stated role of
telecommuting in
decision to move
No, No
( Control)
Yes, No
( Quitters)
No, Yes
( Reloc-> TC)
Yes, Yes
( TC-> Reloc) Total
Not N 16 0 3 1 20
telecommuting
at the time Avg.
delta 4.25 --- - 9.67 5.00 2.20
N 16 0 7 7 30
Not at all
important Avg.
delta - 0.05 --- 4.00 0.00 0.80
N 0 0 0 4 4
Somewhat
important Avg.
delta --- --- --- - 11.50 - 11.50
N 0 1 2 2 5
Important Avg.
delta --- 34.00 48.50 - 4.00 24.60
The most N 1* 0 0 1 2
important
factor Avg.
delta - 20.00 --- --- - 7.00 - 13.50
All ( including N 150 1 13 15 179
never
telecommuter
responses**)
Avg.
delta 1.17 34.00 14.54 - 3.73 1.91
Notes: “ Yes, No” means telecommuting the quarter before the move and not telecommuting the quarter
after, and so on. “ Avg. delta” is the average change in one- way commute length ( miles) following the
move.
* We speculate that a telecommuting experience more than one quarter before the move influenced the
move.
** In addition to the Never Telecommuters, this category also includes those Ever Telecommuters who did
not answer the question on the role of telecommuting in their move ( which accounts for the differences in
this row and the sum across the preceding five rows).
19
Also of interest is the relationship between the change in average commute length and the
stated importance of telecommuting on residential relocations, independent of the
telecommuting status. The final column in Table 11 summarizes the average change in
commute length and number of responses across the four causal groups. No obvious
relationship exists between the average commute length change and the stated importance of
telecommuting to the move. Those who were not telecommuting at the time of their residential
relocation, or placed no importance on telecommuting to their move, increased their commute
length by about the overall average, and those who placed an importance on the ability to
telecommute did not uniformly move closer to or farther from work. Thus, whatever the stated
role of telecommuting may have been, its actual impact on residential location appears to be
neutral at worst and benign at best.
7. Modeling the Change in Commute Characteristics
To more formally quantify the role of telecommuting in residential relocation decisions, linear
regression models of the change in one- way commute length and duration following a move are
estimated. The purpose of the modeling is to determine the relative importance of
telecommuting while controlling for other variables in the dataset that may traditionally be
important predictors of change in commute length and duration ( e. g. presence of children,
desire for a bigger home) and to ascertain which of the telecommuting definitions ( i. e. Current,
After, Before/ After, or Ever) most strongly impacts these measures. By allowing the temporally
causal telecommuting definitions to enter the model, the causal relationships hypothesized in
Section 5 may ( depending on the results) be reinforced by the modeling.
One limitation of the modeling is that socio- demographic, work and home characteristic data
was only collected at the time of the retrospective survey administration. As such, the data in
these categories is not representative of the individual at the time of the relocation; it is
representative of the individual in 1998. To help ameliorate this shortcoming, only the most
recent residential relocation was considered in the modeling ( the average time of the most
recent move was 4.1 years before 1998, the standard deviation 2.8 years).
The variables included in the modeling fall into the following categories: telecommuting status,
work characteristics, home characteristics, travel characteristics, reasons for moving, and
general socio- demographics. A brief description of all the considered variables is as follows:
􀂃 Telecommute status:
o based on the four definitions of telecommuting presented in Section 3;
o based on whether telecommuting before the move or ( in a separate variable),
after the move;
o based on the four temporally causal groups presented in Section 5;
􀂃 Work characteristics:
o occupation type: management, professional, administrative, and other;
o work schedule type: part- time, conventional, flex- time, compressed work week,
and other;
o duration with current organization and supervisor;
􀂃 Home characteristics ( of the new residence):
o duration at current home ( how long ago the relocation took place);
20
o regional location in California ( Sacramento or not);
o second home ownership;
o housing type: single family home, condominium, apartment, duplex, and other;
o presence of dedicated telecommuting space at the home;
􀂃 Travel characteristics:
o current commute mode: automobile, non- motorized, or transit;
o commute mode at previous residential location;
o change in commute mode following the relocation ( 0- 1 variable);
􀂃 Reasons for moving:
o an important factor: my own job changed, job of household member changed,
household size changed, was dissatisfied with home, was dissatisfied with
neighborhood, wanted more room or different location for household activities,
wanted to get closer to work, was dissatisfied with school, and other;
o the most important factor ( among the preceding list);
􀂃 General socio- demographics:
o age, gender, household size, age of household members, number of workers,
educational background and personal income.
Our a priori expectations were that variables in the socio- demographic, reasons for moving, and
home characteristics would dominate the models of change in one- way commute length and
duration. Specifically, we expected that households with young children, those stating a desire
for “ more room”, and those moving to single family homes would move, on average, farther from
work in terms of length and time. Due to the relatively low t- statistics presented in Section 3, we
did not expect telecommuting variables to enter the models.
The model estimation results for the change in commute length and duration are presented in
Tables 12 and 13, respectively. The two models are quite similar. The dominant variable in each
of the models is the “ wanted to get closer to work” stated reason for moving dummy variable. In
both models, the coefficient on this variable is large and negative, indicating that those stating a
desire to move closer to work are, in fact, moving substantially closer to work ( in terms of time
and distance).
The second variable common to both models is the flex- time work schedule measure. Those
working a flex- time schedule ( as opposed to traditional, compressed, or part- time schedules)
tend to move farther from work than those working other schedules. This result suggests that
those working flex- time may be able to avoid peak- hour congestion, and, in doing so, be able to
live farther from work. Conversely, it may be that those who live farther from work may be
adopting flex- time work schedules. This is an intriguing result, which calls for an investigation
into the relationship between flex- time and residential relocation comparable to that being
conducted for telecommuting.
The final significant coefficient in the length model holds a positive sign on the temporally causal
telecommuting status variable, “ not telecommuting before the move; telecommuting after the
move.” This result is consistent with the findings in Section 5 of those in the “ relocation causes
telecommuting” segment to be moving farther from work than those not in this causal group.
21
Table 12: Linear Regression Model of Change in Commute Length following a Residential Reloca-tion
( N= 96)
Dependent Variable : Change in one- way commute length ( in miles) following a residential relocation [- 48,85]
Explanatory Variables Coefficient t- statistic Beta
Constant 3.663 2.24
Telecommute status
Not telecommuting before the move; telecommuting after the move [ 0,1] 10.796 2.29 0.205
Work characteristics
Flex- time employment type [ 0,1] 9.848 3.08 0.274
Stated reasons for moving
Wanted to get closer to work ( one reason) [ 0,1] - 23.606 - 5.60 - 0.510
[ ] = range of possible or observed responses
Adjusted R2 = 0.266 ( R2 = 0.289) F- statistic = 12.50 ( p = 0.000)
Table 13: Linear Regression Model of Change in Commute Minutes following a Residential Reloca-tion
( N= 94)
Dependent Variable : Change in one- way commute duration ( in minutes) following a residential relocation [- 50,75]
Explanatory Variables Coefficient t- statistic Beta
Constant 5.175 2.26
Telecommute status
Definition 3: Before/ after telecommuter [ 0,1] 8.495 1.95 0.173
Work characteristics
Flex- time employment type [ 0,1] 9.625 2.31 0.206
Stated reasons for moving
Wanted to get closer to work ( one reason) [ 0,1] - 32.475 - 6.01 - 0.539
[ ] = range of possible or observed responses
Adjusted R2 = 0.281 ( R2 = 0.304) F- statistic = 13.10 ( p = 0.000)
This finding supports the positive view of telecommuting. While taken at face value the model
seems to suggest that telecommuting plays a more important role in relocation decisions than
more expected variables, such as the presence of young children or a move to a single- family
home, it is likely that household considerations are implicitly included among the reasons for
wanting to get closer to work.
A similar result is found in the commute duration model. Here, a positive coefficient is estimated
( with marginal significance) on the Definition 3: Before or After telecommuter status variable.
This finding indicates that those who telecommute during the quarter of the move, or one
quarter before or after the move, tend to relocate farther from work, in terms of time, than those
who do not. Unfortunately, this finding does not help address the causal impact of
telecommuting in that the temporal order of telecommuting and the relocation is not known.
22
However, it is telling that a telecommuting variable enters the model at all, especially in the
presence of such a variety of other, more expected, variables.
8. Conclusion
This report presents an empirical examination of the causal relationship between telecommuting
and residential/ job relocation. The key question we attempt to answer is whether telecommuting
is a “ friend or foe” of travel reduction policies. In the friend scenario, telecommuting allows those
independently inclined to live in distant locations to commute less frequently to work. In the foe
scenario, telecommuting motivates those who would otherwise live closer to work to move to
more distant locations, possibly resulting in more vehicle- miles traveled than would have
occurred without the option of telecommuting.
The examination presented here used data from a 10- year retrospective survey of 218 State of
California workers, including current, former, and non- telecommuters. We first confirmed that
telecommuters, defined in four separate ways, are, in fact, putting more distance ( and travel
time) between their homes and workplaces than are non- telecommuters. While these results are
not statistically robust due to the small sample size, they do provide suggestive evidence, and
support further investigation into the causal nature of those observed differences.
Three additional methods were then used to examine the causal relationship between
telecommuting engagement and residential/ job relocation. First, the timing of the moves in
relation to the timing of the telecommuting engagement was examined. Here, it is inferred that
the ability to telecommute caused a relocation if an individual telecommuted directly before and
after the move. Conversely, it is inferred that the relocation caused the telecommuting
engagement if an individual began telecommuting following relocation. Those who neither
telecommute before nor after the move are used as the control group, and those who
telecommute before relocation, but not afterwards ( labeled as telecommute quitters) are
assigned to neither causal group. This analysis revealed that those who are inferred to have
their residential relocation caused by telecommuting unexpectedly reduce their commute
duration and length, on average. Conversely, those who are inferred to have telecommuting
caused by their relocation substantially increase their one- way commute duration ( by
approximately 15 minutes) and length ( 18 miles). These results suggest that, on average, those
who are inspired to move because of telecommuting are moving closer to work ( or that the
temporal relationship is largely coincidental, although as discussed below, the result holds even
for moves in which telecommuting was directly reported to be important), while those who
relocate far from work are afterwards inspired to telecommute.
For job relocations, the results differed slightly. Here, both key casual groups ( those who are
inferred to have their relocation caused by telecommuting and those who are inferred to have
their telecommuting engagement caused by relocation) relocate, on average, to jobs farther
from home.
The second method of causal analysis examined the stated reasons for residential relocations
as captured by the survey instrument. In this analysis, the timing of the moves is again
examined, this time in relation to the stated importance of telecommuting to the relocation. Here,
though only small amounts of data are present, the results are consistent with the causal
inferences made based on temporal precedence, in that those who are inferred to have their
relocation caused by the ability to telecommute rated such an ability to be an “ important” ( to
some degree) factor in their residential move more than those who are inferred to have their
telecommuting caused by their residential move. Again, however, moves for the former group
23
tended to reduce commute lengths. There was no relationship between the “ importance” of
telecommuting in residential relocation decisions and the resulting change in commute length.
The final method of analysis estimated linear regression models of the change in commute
duration and length following a residential relocation. These models considered a variety of
variables, including telecommuting status, work characteristics, home characteristics, travel
characteristics, and general socio- demographics. Because certain variables were only present
at the time of data collection ( e. g. general socio- demographics), only each individual’s most
recent residential relocation was considered ( on average, 4.1 years before the time of data
collection). The model of change in commute length included a significant coefficient on the
telecommuting temporally causal status variable, “ not telecommuting before the move;
telecommuting after the move.” This result buttresses the findings in the temporal precedence
analysis, further supporting the idea that telecommuting is often triggered by distant relocations.
The results presented here more strongly support the hypothesis that telecommuting is a
“ friend” of travel reduction policies, with telecommuting more often following, rather than
preceding, the relocations that lengthen the commute. Together with the finding from previous
analysis of these data, that telecommuters’ commute person- miles traveled is no worse than
that of non- telecommuters’ despite their longer one- way commute lengths ( Mokhtarian, et al.,
forthcoming), it appears that ( at least for now) policymakers need not fear substantial “ rebound
effects” from telecommuting programs.
Given the limitations of our retrospective survey design and small sample, however, continued
monitoring of the long- term impacts of telecommuting is clearly desirable. Ideally, future studies
should perform a similar set of analyses on a larger sample of individuals, tracked over time,
rather than surveyed retrospectively. Having a full range of socio- demographic and attitudinal
data at each quarter, or other time interval, would allow for more complex and interesting
analyses, including the impacts of telecommuting engagement on transportation attitudes
through time; finer estimation of the relationship between telecommuting engagement and
commute characteristics ( i. e. after accounting for changes in income, lifestyle, etc., what role
does telecommuting play?); and a mapping of telecommuting patterns of individuals over time
( do they lose interest in telecommuting?). It would also be desirable to extend a similar analysis
to home- based business owners, and ( as mentioned earlier) to the relationship between flex-time
and residential location.
References
Baumrind, Diana ( 1983) Specious causal attributions in the social sciences: The reformulated
stepping- stone theory of heroin use as exemplar. Journal of Personality and Social Psychology
45( 6), 1289- 1298.
Ellen, Ingrid Gould and Katherine Hempstead ( 2002) Telecommuting and the demand for urban
living: A preliminary look at white- collar workers. Urban Studies 39( 4), 749- 766.
Gareis, Karsten ( 2003) Home- based versus mobile telework: The interrelationship between
different types of telework. Chapter 13 in Paul Jackson and Birger Rapp, eds. Organisation and
Work Beyond 2000. Berlin: Springer- Verlag, 171- 185.
Gertz, Carsten and Patricia L. Mokhtarian ( 1999) Endbericht: Die Langfristigen Auswirkungen
von Telearbeit auf das Verkehrsverhalten und Umzugsentscheidungen ( The Long- Term Effects
of Telecommuting on Travel Behavior and Residential Location). Report to BMW AG, Verkehrs-konzepte
Munchen, Germany, July ( in German).
24
Hamer, Rebecca, Eric Kroes, and Harry van Ooststroom ( 1991) Teleworking in the Netherlands:
An evaluation of changes in travel behaviour. Transportation 18, 365- 382.
Higano, Yoshiro and Isao Orishimo ( 1990) Impact of spatially separated work places on urban
residential location, consumption and time allocation. Papers of the Regional Science Associa-tion
68, 9- 21.
Holland, Paul W. ( 1986) Statistics and causal inference. Journal of the American Statistical
Association 81( 396), 945- 960 ( and comments by various authors immediately following).
Hollingworth, Brian J. and Eric J. Miller ( 1996) Retrospective interviewing and its application in
study of residential mobility. Transportation Research Record 1551, 74- 81.
JALA Associates ( 1990). The California Telecommuting Pilot Project Final Report. Department
of General Services, State of California. North Highlands, CA. June.
Janelle, Donald G. ( 1986) Metropolitan expansion and the communications - transportation
trade- off. In Susan Hanson, ed. The Geography of Urban Transportation. New York: The
Guilford Press, 357- 385.
Kim, Sun- Wung ( 1997) Impacts of Telecommuting Policies on Urban Spatial Structure and the
Environment: Home- based and Center- based Telecommuting. PhD Dissertation, Regional
Science, University of Pennsylvania.
Kitamura, Ryuichi, Jack Nilles, David Fleming, and Patrick Conroy ( 1990). Telecommuting as a
transportation planning measure: Initial results of State of California pilot project. Transportation
Research Record 1285, 98- 104.
Lund, Jay R. and Patricia L. Mokhtarian ( 1994) Telecommuting and residential location: Theory
and implications for commute travel in monocentric metropolis. Transportation Research
Record 1463, 10- 14.
Mokhtarian, Patricia L., Susan L. Handy and Ilan Salomon ( 1995) Methodological issues in the
estimation of travel, energy, and air quality impacts of telecommuting. Transportation Research
A 29A( 4), 283- 302.
Mokhtarian, Patricia L., Gustavo O. Collantes and Carsten Gertz ( 2004) Telecommuting,
residential location, and commute distance traveled: Evidence from State of California
employees. Environment and Planning A 36( 10), 1877- 1897
Nilles, Jack M. ( 1991) Telecommuting and urban sprawl: Mitigator or inciter? Transportation
18( 4), 411- 432.
Ory, David T. and Patricia L. Mokhtarian ( 2004) Does telecommuting really save commute time?
Time, distance, and speed evidence from State of California Workers. Paper presented at 2004
Time Competition Workshop, Utrecht, the Netherlands, June. Available from the authors.
Safirova, Elena ( 2001) A model of a residential economy with telecommuting. Manuscript
available from the author at safirova@ rff. org.
Shaw, Shih- Lung ( 2004) Spatial patterns of telecommuting in the Atlanta metropolitan area.
Presented at ICT: Mobilizing Persons, Places, and Spaces. Doorn, the Netherlands, November
4- 7.
Shen, Qing ( 2000) New telecommunications and residential location flexibility. Environment and
Planning A 32( 8), 1445- 1463.
25
Stough, Roger R. and Jean Paelinck ( 1996) Substitution and complementary effects of
information on regional travel and location behavior. Proceedings of the workshop on New
International Perspectives on Telework, Brunel University, West London, England, July 29-
August 2, pp. 380- 397.