ISSN 1403-2473 (Print)  
ISSN 1403-2465 (Online) 
 
 
 
 
 
 
 
 
 
Working Paper in Economics No. 743  
 
 
Experimental and non-experimental 
evidence on limited attention and present 
bias at the gym 
 
 
Paul Muller, Wolfgang Habla 
 
 
Department of Economics, October 2018 
 
 
Experimental and non-experimental evidence on
limited attention and present bias at the gym
Paul Muller∗ Wolfgang Habla∗∗
September 21, 2018
Abstract
We show that limited attention and present bias contribute to low levels of exercise. First, in
a large randomized experiment, we find that email reminders increase gym visits by 13 % and
that they benefit nearly all types of individuals. Limited attention can explain these effects.
Second, using a novel dataset, we find that many bookings for gym classes are canceled, and that
bookings are made even for classes that never have a waiting list. Comparing these findings to
the predictions of a dynamic discrete choice model, we conclude that many gym members use
bookings to commit themselves to future attendance.
Keywords: health behavior, randomized experiment, reminders, nudging, habit formation, lim-
ited attention, time inconsistency
JEL-code: C93, D91, I12
∗Department of Economics, Gothenburg University. Vasagatan 1, 405 30 Gothenburg. Paul.muller@gu.se
∗∗Centre for European Economic Research (ZEW) and University of Mannheim. L7 1, 68161 Mannheim.
habla@zew.de
We are grateful to Fysiken for their cooperation. In particular, we thank Maria Ljungqvist for agree-
ing to work together and granting us access to Fysiken’s data; Hanna Mattson for all her help in accessing
the data; and Sara Ha¨gge for her assistance with designing and translating the email reminders. We would
also like to thank Anna Bindler, Fredrik Carlsson, Andreas Dzemski, Martin Dufwenberg, Christina Gravert,
Mitesh Kataria, Raphael Krause and Joseph Vecchi for their many useful comments and suggestions. The
paper also greatly benefited from discussions with participants at the internal ZEW Econometrics Seminar,
the Labour/Applied Micro Economics Seminar at the University of Gothenburg and the Nordic Conference on
Behavioral and Experimental Economics in Gothenburg. This study has been registered with the AEA RCT
registry (AEARCTR-0001847).
1 Introduction
Recent research shows that the prevalence of insufficient physical exercise is over 40 % in
high-income Western countries and has been increasing over the past two decades (Guthold
et al., 2018). Given the overwhelming evidence on the long-term health benefits of regular
exercise and the related improvement in well-being, such low levels of physical activity are
both surprising and worrisome. Since exercising is costly in terms of time and effort, low levels
of activity can be interpreted to be a result of individual optimization, where immediate costs
are weighted against delayed benefits. There are, however, a number of reasons from both
a societal and an individual perspective why the prevalence of exercise might be below its
optimal level.
First, positive externalities of regular exercise have been documented, including reduced
sick leave and higher work productivity (Kool et al., 2004; Van Amelsvoort et al., 2006).
Moreover, most people are (to some degree) insured against medical expenses such that their
behavior is prone to moral hazard. Both of these phenomena may lead to underinvestment in
preventive activities such as regular exercise. Second, behavioral anomalies and biases may
play an important role in investment decisions that involve immediate costs and long-term
(delayed) benefits. One such investment is healthy behavior, and in this context, present
bias and limited attention have been identified as the most likely causes of underinvestment
(DellaVigna and Malmendier, 2006; Calzolari and Nardotto, 2017). As a result, individuals
may not be able to achieve levels of exercise that they find individually optimal in the long
run.
In this paper, we investigate the importance of these two popular behavioral explanations
in the repeated decision to exercise, using very detailed data from a gym. First, we evaluate
a randomized experiment among almost 2,500 gym members in which we sent weekly email
reminders to half of the sample. We find a robust positive effect of these reminders on the
number of gym visits, which we argue is most naturally explained by the prevalence of limited
attention among gym members. Our contribution lies in our use of a large sample size, which
allows for precise estimates, a sample of gym members who are arguably more representative of
the general population than those used in previous studies, and a gym that offers a wide range
of activities and sports.1 All of these factors speak towards the strong external validity of our
analysis. Second, by exploiting novel data on booking, canceling and attending gym classes,
we investigate the importance of present bias. The possibility to book a spot in gym classes
provides us with information on the making and changing of plans regarding future visits to
the gym. We document a range of empirical observations and compare them to predictions
from a simple dynamic discrete choice model. We conclude that many gym members book
1We use data from a gym that offers activities ranging from regular weight-lifting to ball sports, yoga,
cycling, swimming and dancing classes. As a result, we are likely to observe a larger share of members’ exercise
activities than we would find at an average gym.
2
classes in advance simply to commit themselves to exercising in the future. Since commitment
problems stem from present bias, we infer that time-inconsistent preferences are an important
factor influencing the decision to exercise. Below, we describe the two parts of the paper in
more detail and briefly discuss the related literature.
Limited attention is defined slightly differently across studies (sometimes, it is called ‘lim-
ited memory‘). For instance, Taubinsky (2014) defines limited attention as the probability to
forget to take an action, while Ericson (2017) defines limited memory as the probability to
forget about the completion of planned tasks. Dean et al. (2017) model inattentive consumers
who tend to focus their attention on a small number of alternatives (the ‘attention set’ or
‘consideration set’) when faced with a large choice set. What these interpretations have in
common is that a decision maker, due to cognitive limitations, does not take the full set of
choices or actions into consideration. In the context of exercising, gym members may forget to
visit the gym on a particular day, forget to book a gym class in advance or forget to attend a
previously booked class. Reminders are a potential remedy for limited attention because they
can direct people’s attention to a particular choice or action and thus make the execution of
this choice or action more likely. To test the importance of limited attention, we performed a
large-scale randomized experiment in which we sent weekly email reminders to gym members
and observe these members’ subsequent gym attendance. The setup is related to the reminder
study by Calzolari and Nardotto (2017) but is applied to a much larger and more representa-
tive sample and a much more diverse set of activities and individuals.2 The reminders were
sent during a three-month period to gym members who had recently purchased an annual
membership. While the content varied across weeks, the emails were always short, containing
only text encouraging regular visits to the gym.3
We find that email reminders lead to a substantial and statistically significant increase in
gym visits of 13 %. The effect appears to be somewhat larger for class training (17 %) than
for free training (12 %), although the difference is not significant. These estimates follow
from a difference-in-differences (DiD) model in which we control for a slight pre-experiment
difference in attendance between the control and treatment groups, but we find similar effects
when using simple differences. These findings are robust to a range of different specifications.
2The gym members we study exercise roughly once per week, which is arguably closer to typical attendance
rates than those documented in previous studies. For example, Calzolari and Nardotto (2017) consider only
students that volunteered to participate in the study, and they report an average attendance of almost two
visits per week, acknowledging that this figure is high compared to typical gym attendance.
3Reminders have been shown to be effective in a variety of domains. For example, they increase the amount
of charitable giving (Sonntag and Zizzo, 2015; Damgaard and Gravert, 2018), monthly savings (Karlan et al.,
2016a) and loan repayments (Karlan et al., 2016b); they improve rule compliance at the library (Apesteguia
et al., 2013) and in traffic (Chen et al., 2017); they also increase the payment of TV license fees (Fellner
et al., 2013), the claiming of tax credits (Bhargava and Manoli, 2015) and college enrollment (Castleman and
Page, 2015); they decrease misdemeanor defendants’ failure-to-appear rates in court (Bornstein et al., 2012)
and residential energy consumption (Allcott and Rogers, 2014). In contrast to these findings, Himmler et al.
(forthcoming) show that reminders are not effective when students study for, sign up for, or take exams.
3
Furthermore, the positive impact of email reminders applies to nearly all types of individuals
(students, non-students, men, women, etc.). Only low attenders (based on the pre-experiment
period) seem to benefit substantially more than high attenders, which is not surprising, given
that high attenders typically have a lower scope to increase their exercising frequency. When
extending the observation period to three months after the last reminder was sent, we find
evidence that the positive effects of reminders persist (visits remain 13 % higher in the treated
group). We interpret this finding as a sign of habit formation, although it should be noted
that these results are less precise and less robust.
The positive impact of regular reminders on gym visits clearly suggests that limited atten-
tion does indeed play a major role in the context of gym attendance and that email reminders
are effective at overcoming such cognitive limitations.4
In the second part of the paper, we use a novel and detailed dataset on the booking,
canceling and attendance of gym classes to investigate whether intertemporal choices are
affected by time inconsistency and whether people use class bookings as a commitment device.
Time-inconsistent preferences imply that a decision maker’s preferences over current and
future choices may change over time, such that a preference for a certain choice or action
at one point in time can become undesirable at another point in time. More specifically,
individuals have been found to assign more weight to payoffs that are closer to the present
time when considering trade-offs between two future moments. This behavior is usually
referred to as present bias (O’Donoghue and Rabin, 1999a,b, 2001).5 The decision to visit the
gym is particularly vulnerable to present bias, as it entails immediate (time and effort) costs
and delayed (health) benefits.
Each class offered at the gym has a limited capacity, and gym members can enroll in
a class up to one week before the class. If a member changes his or her mind, he or she
must cancel his or her booking (or pay a fine). Using detailed data on the timing of each
booking and cancellation, we provide insights on how individuals make intertemporal choices,
i.e., how and when they make plans for future gym visits and how likely they are to follow
through on their plans. We first document a number of behavioral patterns in the data and
then establish a simple two-period discrete choice model to demonstrate how these decisions
4An alternative explanation for why reminders are effective could be guilt aversion (see, e.g., Battigalli and
Dufwenberg, 2009), which suggests that guilt-averse players experience a utility loss if they believe that they
let someone down. In our case, gym members could feel obliged to visit the gym when they receive a reminder
email because they believe that the gym – or, rather, the gym’s employees – expect them to come to the gym.
This explanation seems more plausible for less anonymous settings in which the gym facility is relatively small
and there is a small number of employees. The health club with which we cooperated is large, and therefore,
members are unlikely to see the same staff members at the gym every time they attend.
5Present bias is frequently modeled using (quasi-) hyperbolic (beta-delta) discounting. This type of temporal
discounting assumes that discount rates decline over time (Phelps and Pollak, 1968; Laibson, 1997). Hyperbolic
discounting induces an individual to consume more today and invest less in the future than she would like from
an ex ante perspective (see Frederick et al., 2002 for a thorough discussion of hyperbolic discounting and time
discounting, as well as time preferences more generally).
4
are affected by uncertainty about the future (in terms of the effort cost of attending the
class), time-inconsistent preferences (present bias) and limited attention. Comparing the
empirical findings with the predictions from the model, we conclude that many bookings
are made only to commit oneself to attending the class.6 Commitment devices are useful
only for individuals who have commitment problems, that is, individuals who have time-
inconsistent preferences. Thus, we show that gym members suffer from time inconsistency
when deciding (repeatedly) whether to visit the gym. These findings complement recent
studies that demonstrate the demand for and effectiveness of commitment contracts at the
stage of buying a gym membership (e.g., Goldhaber-Fiebert et al., 2010; Bhattacharya et al.,
2015; Royer et al., 2015). Note that our results suggest that not only time inconsistency but
also uncertainty about the future costs of attending a class play a large role in decisions with
regard to booking, canceling and attending classes. Finally, the fact that a substantial share
of bookings results in ‘no-shows’ (members who do not attend and do not cancel), further
supports the hypothesis that limited attention is an important factor.
The rest of this paper is structured as follows. Section 2 describes the setup for the
reminder experiment. In Section 3, we analyze the effects of reminders on the number of gym
visits and consider robustness and heterogeneous effects. Section 4 presents the booking and
canceling data, establishes a simple dynamic discrete choice model, and compares empirical
findings and model predictions. Section 5 concludes.
2 Experimental setup
Below, we briefly describe the health club at which our study was performed, the population
of members who were part of it, and the setup and design of the reminder experiment.
2.1 The health club
The study was carried out in collaboration with one of the largest health clubs in Gothenburg
(Sweden). The club is a non-profit organization, has four facilities and records over 800,000
visits per year. Three facilities offer free workout areas, as well as a large range of group
exercise classes led by fitness instructors (more than 200 per week). The fourth facility is a
large climbing hall. Contracts can be bought for access to a single facility, or to all facilities
(‘Multi card’). Access to each of the four facilities requires the scanning of an electronic
membership card. Participation in a group exercise class requires an additional scan. Gym
6The intuition behind this finding is the following. We find that even for class types that are unpopular and
never fully booked, most attendees make a booking prior to the class. This finding holds even when selecting
only experienced members who have attended the same class several times and are therefore certainly aware
of the unpopularity of the class. Since such bookings are clearly unnecessary in order to secure a spot, this
behavior can be explained only by gym members using bookings as a commitment device. See our model in
Section 4.2 for further discussion.
5
Figure 1: Gym attendance (July 1, 2016 - June 30, 2017)
0 20,000 40,000 60,000frequency
Facility 4
Facility 3
Facility 2
Facility 1
(a) Visits per location
0 2,000 4,000 6,000frequency
dance
ballsport
core
body and mind
cycle/run/swim
endurance and strength
(b) Visits per category (only class training)
attendance is thus almost perfectly observed. Because many classes are at risk of being
crowded, each class has a participant limit, and members can sign up for classes in advance.
We discuss the club’s booking and cancellation policy in Section 3.4.
From the health club, we received all data on class attendance, class bookings, class book-
ing cancellations and class no-shows for the time period from July 1, 2016, to June 30, 2017,
for all individuals who purchased a 12-month contract between July 1, 2016, and December
1, 2016 (this is the experimental sample, see Section 2.2). In Figure 1, we show descriptive
statistics. Altogether, we observe 115,232 gym visits. Panel (a) shows the distribution of
free training sessions (95,791) across the four facilities. Facility 1 is the most popular facility,
with 50 % of all visits, while the other 50 % of the visits are spread across the other three
facilities. We observe 19,441 class attendances (approximately 17 % of all visits). More than
37 different types of classes were offered, and we split them into six main categories (based
on the classification used by the health club): endurance and strength, body and mind, cy-
cle/run/swim, core, ball sports, and dance.7 The distribution of participation across these six
categories is presented in panel (b) of Figure 1, while the distribution of visits across the 37
specific class types is shown in Figure 6 in the appendix. The most popular classes are indoor
cycling (spinning), ‘krafttag’ (muscle strength training using free weights) and yoga.
7The actual number of different classes is much larger if we account for, e.g., differences in length, focus,
instruction language etc. Classes included in the 6 main categories are: 1 (endurance and strength): Body-
combat, Bodystep, Cirkelgym, CrossFit, Fight, Fys, Gymclass, Gympa, Krafttag, Powercage, Powercircuit,
Tabata; 2 (body and mind): Bodybalance, Mindfulness, Pilates, Triggerstretch, Yoga; 3 (cycle/run/swim):
Spinning, Indoor walking, Swim-training; 4 (core): Cxwork, Functional training, Q-performance, TRX; 5 (ball
sports): Badminton, Basketball, Football, Indoor hockey, ‘Spo¨kboll’, Volleyball; 6 (dance): Dancing classes.
6
2.2 Sample selection, randomization and balance
As mentioned above, we focus on all individuals who purchased a 12-month contract (rather
than a one- or three-month contract), between July 1, 2016, and December 1, 2016. This ap-
proach provides us with 2,881 individuals. We select individuals who remained active members
on December 1, 2016, which was true for almost all members, as an annual membership can-
not be canceled during its duration.8 Furthermore, we exclude members who did not provide
a valid email address when registering with the gym, leaving us with a sample of 2,463 gym
members.
We randomly assigned these individuals into treatment and control groups in the first week
of December 2016. The experiment, which we describe in detail below, began on January 9,
2017 (see Figure 2 for a timeline of our setup). We present descriptive statistics in Table 1,
which contain the mean, standard deviation, minimum and maximum for the control group
(columns 1-4) and for the treatment group (columns 5-8), and the p-value of a t-test for
equal means (column 9). The top panel presents individual characteristics, the second panel
membership and contract characteristics. In our sample, the average age is 30, and 58 %
of the gym members are males. In total, 52 % registered as students and 46 % had been
members of this gym previously. A total of 40 % bought a Multi card and 5 % bought a
membership that is valid only during daytime hours (a ‘daytime card’).9 The mean date of
purchasing the contract is September 23 for both groups (Figure 7 in the Appendix shows
that the full distributions of contract buying dates in the control and treatment groups are
very similar). Most variables do not differ significantly between control and treatment groups,
with the exception of a slightly lower share of recurring members and a slightly higher share
of members who bought a daytime card in the treatment group.
The five-week period of pre-experiment gym visit observations (December 2, 2016 - Jan-
uary 8, 2017) allows us to compare the control and treatment groups in terms of attendance.
In the lower panel of Table 1, we find that the treatment group has a significantly lower
number of weekly visits in the pre-experiment period, and this difference is caused by a lower
number of free training visits. In Section 3.1, we discuss how we control for these differences
by including individual fixed effects when estimating the impact of the reminders. Note that
the number of weekly visits is lower in this pre-experiment period than is typically the case
because this time period includes the Christmas holidays, during which gym attendance is
low. In Figure 8 in Appendix A, we compare the distributions of gym visits of the control
and treatment groups across facilities, type of class and number of visits, and we find that
the control and treatment groups are similar along these dimensions.
8A few exceptions occurred when newly enrolled members canceled their membership within a few days, or
due to, for example, serious injury or moving.
9With a daytime card, members can exercise before 3:30 pm on weekdays but not on weekends.
7
T
ab
le
1:
C
ha
ra
ct
er
is
ti
cs
of
th
e
tr
ea
tm
en
t
an
d
co
nt
ro
l
gr
ou
ps
C
on
tr
ol
gr
ou
p
T
re
at
m
en
t
gr
ou
p
m
ea
n
sd
m
in
m
ax
m
ea
n
sd
m
in
m
ax
p-
va
l.
t-
te
st
A
ge
30
10
14
79
30
11
16
80
0.
77
M
al
e
0.
59
0.
49
0.
58
0.
49
0.
50
St
ud
en
t
0.
51
0.
50
0.
54
0.
50
0.
10
R
ec
ur
ri
ng
m
em
b
er
a
0.
49
0.
50
0.
44
0.
50
0.
01
M
em
b
er
sh
ip
ty
p
e:
M
ul
ti
ca
rd
(a
nd
C
ro
ss
F
it
)b
0.
41
0.
49
0.
40
0.
49
0.
66
F
ac
ili
ty
1
0.
32
0.
47
0.
34
0.
47
0.
43
F
ac
ili
ty
2
0.
10
0.
30
0.
10
0.
31
0.
73
F
ac
ili
ty
3
0.
12
0.
32
0.
10
0.
30
0.
34
F
ac
ili
ty
4
0.
05
0.
23
0.
06
0.
23
0.
86
D
ay
ti
m
e
ca
rd
0.
04
0.
19
0.
06
0.
23
0.
05
D
at
e
of
pu
rc
ha
si
ng
co
nt
ra
ct
c
23
se
p2
01
6
38
01
ju
l2
01
6
01
de
c2
01
6
23
se
p2
01
6
39
01
ju
l2
01
6
01
de
c2
01
6
0.
66
W
ee
kl
y
(p
re
-e
xp
er
im
en
t)
at
te
nd
an
ce
:d
G
ym
vi
si
ts
0.
97
1
0
7
0.
88
0.
96
0
5.
8
0.
02
F
re
e
tr
ai
ni
ng
0.
83
0.
98
0
6.
8
0.
74
0.
90
0
5.
8
0.
03
C
la
ss
tr
ai
ni
ng
0.
15
0.
39
0
3.
4
0.
14
0.
38
0
3.
2
0.
51
O
bs
er
va
ti
on
s
12
32
12
31
a
A
‘R
ec
ur
ri
ng
m
em
b
er
’
is
so
m
eo
ne
w
ho
ha
d
an
ac
ti
ve
co
nt
ra
ct
at
so
m
e
p
oi
nt
du
ri
ng
th
e
si
x
m
on
th
s
pr
io
r
to
ou
r
sa
m
pl
e
in
flo
w
p
er
io
d
(J
an
ua
ry
1,
20
16
-
Ju
ne
30
,
20
16
).
b
A
C
ro
ss
F
it
m
em
b
er
sh
ip
is
si
m
ila
r
to
a
‘M
ul
ti
ca
rd
’
m
em
b
er
sh
ip
bu
t
al
so
al
lo
w
s
en
tr
y
to
C
ro
ss
F
it
.
c
F
or
th
e
di
st
ri
bu
ti
on
of
th
e
da
te
s
of
co
nt
ra
ct
pu
rc
ha
se
s
in
th
e
co
nt
ro
l
an
d
tr
ea
tm
en
t
gr
ou
ps
,
se
e
F
ig
ur
e
7
in
th
e
A
pp
en
di
x.
d
A
tt
en
da
nc
e
st
at
is
ti
cs
ar
e
co
m
pu
te
d
ov
er
th
e
p
er
io
d
fr
om
D
ec
em
b
er
2,
20
16
,
to
Ja
nu
ar
y
8,
20
17
(t
hi
s
is
th
e
pr
e-
ex
p
er
im
en
t
p
er
io
d
fo
r
w
hi
ch
w
e
ha
ve
at
te
nd
an
ce
da
ta
fo
r
al
l
in
di
vi
du
al
s)
.
8
Baseline
email
July 1
2016
Dec 1 Jan 9
2017
Apr 9 June 30
Sample inflow period
Pre-experiment
observation period
Experiment period
Post-experiment
observation period
Figure 2: Timeline of the experiment
2.3 Experiment details
The intervention consisted of a series of email reminders that were dispatched to the treatment
group between January 9, 2017, and April 9, 2017 (see Table 10 in Appendix B for the dispatch
dates). The reminders were sent from the gym’s email account once per week, and they
contained a short message (typically no longer than three sentences) encouraging the recipient
to attend the gym. In addition, they contained a picture related to the gym and a link to
the club’s website (for an example, see Figure 9 in Appendix B). The text and the picture
differed every week in order to attract the recipients’ interest and make them more likely to
open and read the reminders. Most emails referred to the various exercise options available
at the club or reminded the reader about the potential benefits of attending the gym.10 An
example of a reminder is the following: Did you know that we offer more than 200 classes per
week? Drop in any time to work out, or sign up for a class on www.fysiken.nu/en/book/. We
hope to see you soon! The complete list of reminders can be found in Appendix B.
Each reminder email contained a line at the bottom explaining that the readers had
received this email because of their gym membership and that they could unsubscribe by
clicking a link. When members unsubscribed, they were removed from our email list and did
not receive any further email communication related to the study.
Receiving emails from the gym is not unusual for members. The gym has a policy of send-
ing out an automated welcome email when a new member signs up. Furthermore, automated
emails are sent when a new member has not attended the gym for 90 days (though these
emails were suspended during our experiment). A few times per year, all members receive a
newsletter (unless they have unsubscribed from this service). Both the control and the treat-
ment groups received such a newsletter once during the experiment. The starting date and
the end date of the email reminders were unknown to participants. Finally, the days on which
the reminders were dispatched were randomized between Sundays, Mondays and Tuesdays in
order to make sure that members in the treatment group did not expect to receive a reminder
10We opted for generic reminders, as several papers have shown that adding other contents to general
reminders or applying gain versus loss frames does not change their effectiveness (Apesteguia et al., 2013;
Altmann and Traxler, 2014; Karlan et al., 2016a). Only in the domain of electricity consumption does the
inclusion of social comparisons or social norms seem to significantly alter the effects of messages (Schultz et al.,
2007; Nolan et al., 2008).
9
on a particular weekday.
The effectiveness of email reminders obviously depends on whether the recipient receives
(and notices) the reminders. An email might not be received because, for example, members
have misspelled their email address or registered an address that they fail to use, or the
emails have ended up in their spam folders. Even if the email appears in the user’s inbox, it
may be deleted without being opened (although in that case, its impact could already have
been achieved if it brought the gym to the top of the receiver’s mind). Since email reading
behavior is likely to depend on individual characteristics, selecting a sample of recipients who
have responded to an email, as in Calzolari and Nardotto (2017), may imply that the sample
is also more responsive to the following reminders than the average gym population.
Analyzing the extent to which such selection issues drive the results is difficult because
whether someone opens and reads a particular email is not typically observed. We address
this issue in two ways. First, we observe a measure of whether each email has been opened
using so-called ‘web beacons’, which record whether the picture included in the email has
been downloaded. This measure of email opening is not perfect because not every email client
downloads the picture by default.11 An email can thus be read without downloading the
picture, implying that our obtained measure provides a lower bound for the opening rate of
emails.
Second, we sent out a ‘baseline email’ to our entire sample about five weeks prior to
the start of the study. In this email, we briefly described the widely known challenge of
achieving one’s exercising goals and asked the recipients whether they believed that regular
email reminders could help them improve in this regard. If they thought that regular reminders
could indeed be helpful, they were asked to click a button in the email.12 The baseline email
provides a measure of which members – in both the control and treatment groups – open
emails. In addition, when someone clicks on the button, we know with certainty that the
email was read. We will make use of the web bugs and the baseline email in the empirical
analysis to identify a subgroup of individuals who are likely to open and read emails so that
we can estimate heterogeneous effects for this group (see Section 3.3).
2.4 Opening of emails and unsubscribing
Before we analyze the impact of our reminders, we show some statistics on the email opening
rate for our reminders and on members’ unsubscription from our email list.
Weekly unsubscribing is measured after each reminder email. It is important to note that
11Gmail, for example, automatically downloads the picture once the email is opened, and downloading
statistics are thus a perfect proxy for opening statistics. For clients such as Outlook, however, the user can
choose whether to download the picture or not.
12The email explained that the gym was considering offering such a service in the future and would like to
collect its members’ general opinion on whether they believed this practice would be useful. We made it very
clear in the email that indicating interest did not imply signing up to any service or future reminders.
10
Table 2: Regression of the email opening rate on individual char-
acteristics
(1)
Share of reminders that is opened
Male -0.066***
(0.020)
Age <20 0.17***
(0.048)
Age 21-25 0.29***
(0.037)
Age 26-30 0.29***
(0.038)
Age 31-40 0.25***
(0.037)
Age >40 0.36***
(0.038)
Student 0.028
(0.022)
Recurring member 0.024
(0.020)
Opened the initial email 0.11***
(0.018)
Clicked link initial email 0.100**
(0.046)
Observations 1111
R2 0.166
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors
clustered by individual in parentheses. In addition to the
variables above, six email client dummies are also included
in the regression (Gmail, Hotmail, Live, Outlook, Univer-
sity, Other). The sample contains only the treatment group
and excludes 120 individuals for which the gender variable is
missing.
11
Figure 3: Number of emails sent and opened
0
25
0
50
0
75
0
10
00
12
50
15
00
Ind
ivi
du
als
1 2 3 4 5 6 7 8 9 10 11 12
Email reminder week
Emails sent Emails opened
unsubscription is not equivalent to attrition in our study. We observe gym attendance (or non-
attendance) for all individuals in the study. Unsubscription implies only that fewer emails
are received. In Figure 3, we present the weekly number of emails that were dispatched.
While the number of email addresses on our list declined steadily, 91 % of all individuals
in the treatment group remained on the mailing list at the end of the intervention. Thus,
weekly unsubscription is approximately 1 %. In the same figure, we show the number of
delivered emails that we know with certainty were opened, i.e., those for which the picture
was downloaded. Considering that this percentage is a conservative estimate, the opening
rate is quite substantial, beginning at 46 % in the first week and decreasing to 36 % in the
final week.
To assess whether certain types of individuals are more likely to open the email, we
regress the share of reminders that are opened on several individual characteristics. The
results are shown in Table 2 (obviously only the treatment group is included). We find that
men are significantly less likely to open emails (-7 %-points). The same holds true for gym
members who are less than 20 years old (with the difference between this group and most other
age groups being more than 10 %-points). We find no differences for students or recurring
members. Furthermore, we find that responding to the baseline email is, unsurprisingly, a
good predictor of opening the reminder emails. Opening the baseline email is associated
with an 11 %-points increase in the share of reminders that are opened during the study.
Clicking the link in the baseline email adds another 10 %-points increase. Even though much
unexplained variation remains in the opening rates, the included set of variables explains a
12
non-trivial share of the variation (R2 = 0.16).13
3 Experimental results
In this section, we discuss the effect of our intervention on gym attendance for both the short
term (during the intervention) and the long term (after the intervention). We also consider
the effect on class bookings and the cancellation of bookings. Furthermore, we estimate
heterogeneous effects and consider individuals’ contract renewal decisions.
3.1 Impact on gym attendance
We show the average weekly number of gym visits in the control and treatment groups in
panel (a) of Figure 4. On average, individuals visit the gym slightly more than once per
week during the whole study period (with the exception of a large dip around the Christmas
break). The time trend is negative and can have three possible explanations. First, because
our experiment begins shortly after the new year, it is in line with a ‘fresh start effect’ at
the beginning of the year and a subsequent decline in attendance.14 Second, our sample
contains members with new contracts, and attendance typically declines over the course of
the membership (e.g., DellaVigna and Malmendier, 2006). Finally, higher temperatures and
longer hours of sunshine might drive gym members from indoor activities toward outdoor
activities in spring and summer.
The first vertical line in Figure 4 indicates the start of the experiment period in which
the email reminders were sent. As mentioned earlier, the control group has a slightly higher
gym attendance than the treatment group in the weeks before the experiment begins. Almost
directly after the reminders begin, this pattern reverses and the treatment group has a slightly
higher attendance than the control group (panel a). We can split gym visits into free training
and class training visits. In panel (b), we show the average number of free training sessions
per week, which is just below one visit. In panel (c), we find that the impact of the reminders
on class training is particularly pronounced. Class attendance is higher for the treatment
group in almost all experiment period weeks, while the two groups have almost identical pre-
experiment class attendance. Finally, in panel (d), the difference between the treatment and
control groups for each of the three variables is depicted. The diagram clearly shows that
13We will use the estimated coefficients in a robustness analysis to obtain predicted email opening rates
not only for the treatment group but also for the control group. Since these predicted values are based only
on baseline (pre-experiment) variables, we can use them to classify individuals as ‘likely to open emails’ and
‘unlikely to open emails’ and subsequently estimate treatment effects separately for these two groups. This
process allows us to assess whether opening the email is essential for the reminder to affect behavior. We return
to this issue in Section 3.3.
14This phenomenon is found by Dai et al., 2014, who describe how salient temporal landmarks can increase
aspirational behaviors such as gym visits.
13
Figure 4: Weekly gym attendance
.
4
.
7
1
1.3
Tota
l vis
its p
er w
eek
2016w49 2017w1 2017w5 2017w9 2017w13 2017w17 2017w21 2017w25
Control group Treatment group
(a) Total visits
0
.
1
.
2
.
3
Clas
s tra
ining
 per
 wee
k
2016w49 2017w1 2017w5 2017w9 2017w13 2017w17 2017w21 2017w25
Control group Treatment group
(b) Class training
.
25
.
5
.
75
1
1.25
Free
 trai
ning
 per
 wee
k
2016w49 2017w1 2017w5 2017w9 2017w13 2017w17 2017w21 2017w25
Control group Treatment group
(c) Free training
−
.
2
−
.
1
0
.
1
.
2
Diffe
renc
e tre
atm
ent−
cont
rol
2016w49 2017w1 2017w5 2017w9 2017w13 2017w17 2017w21 2017w25
Total attendance Free trainingClass training
(d) Treatment-control difference (weekly)
14
the lower attendance of the treatment group in the pre-experiment period reverses to higher
attendance after the email reminders begin.
The second vertical line indicates the end of the intervention, after which no further
reminders are sent. We observe attendance until June 30. After that, the vacation period
begins, during which gym attendance is traditionally very low and the number of offered
classes is greatly reduced. Furthermore, extending the post-experiment beyond this day would
imply that we would lose participants in our sample since some of them will not prolong their
contracts. The negative trend in weekly attendance clearly continues after the experiment
ends. While average weekly attendance is 1.25 in January, it is only 0.5 at the end of June.
The difference between the control and treatment groups is smaller than it is during the
experiment but still suggests a slightly higher average attendance for the treatment group
(especially when taking the pre-experiment difference into account). As shown in panel (d),
for each of the three lines, the difference is either (slightly) positive or close to zero during
the post-experiment period.15
To obtain estimates of the impact of the reminders, we compute the average weekly at-
tendance in the pre-experiment period and experiment period for the control and treatment
groups.16 The simple difference-in-differences (DiD) estimate for the total number of gym
visits is 0.12 additional visits per week as a result of the email reminders, which is equivalent
to an 11 % increase. The additional visits consist of 0.03 additional classes (15 % increase)
and 0.09 additional free training sessions (11 % increase). Over the 12 weeks of the experi-
ment, this effect aggregates into 1.5 extra visits on average for each individual who received
the reminders. Given that marginal costs of the intervention are close to zero, this shift can
be considered a substantial effect.
We now turn to a regression framework and estimate a DiD model. The impact of the
reminders is identified from the assumption that in the absence of the intervention, the control
and treatment groups would have had the same time trend in weekly gym attendance. Weekly
gym attendance (Ait) of individual i in week t depends on week fixed effects αt, individual
fixed effects γi and treatment assignment Tit, which is a dummy equal to 1 if individual i is
in the treatment group and week t is in the experiment period. Furthermore, we control for
membership duration in a flexible way by including a set of dummies for different membership
duration intervals, which is indicated by pid. We include membership duration because of the
abovementioned negative time trend in attendance and because individuals in our sample
15There is a particularly huge spike in attendance of the treatment group in week 22, mostly as a result of
an increase in free training. After checking with the health club, we found that several emails were sent by the
gym on May 29, May 30 and May 31, with the latter being sent to all individuals who had a valid contract
and, of course, registered a valid email address. We later perform several robustness checks with respect to
this particular week.
16To take into account the correlation between an individual’s multiple observations, we first compute in-
dividual averages before estimating the group mean and standard error of the mean. The resulting numbers,
including standard errors, are shown in Figure 10 in the Appendix.
15
entered their current membership at different dates.
Ait = αt + β1T
during
it + β2T
post
it + pid + γi + it (1)
The parameters of interest are β1 and β2, which measure the immediate effect of the email
reminders and the long-term effect of the reminders, respectively. In panels (c) and (d) of
Figure 8 in Appendix A, it is clear that weekly gym attendance is a count variable with many
zeros. Thus, a Poisson regression is more suitable than a linear regression. We report the
results for a linear model as a robustness check (see Section 3.2) and find that the results are
very similar. We estimate standard errors clustered by individual.
The regression results are presented in Table 3, where we report exp(β) − 1, which is
the percentage effect. We first discuss the immediate impact of the reminders (β1), which is
shown in the first row. We find that total visits increase by 13 %, which is significant at the
1 % level (column 1). We split these visits into class training and free training in columns (2)
and (3). In column (2), we find that attendance in class exercise training increases by 17 %,
while free training visits increase by 12 % (column 3). These percentage effects suggest that
the reminders led predominantly to more class training visits. However, the much lower base
level for class training visits implies that the impact in terms of additional visits is greater for
free training (and the standard errors imply that the difference between the two estimates is
not statistically significant).17
As a robustness check, we also present regressions in which we exclude the pre-experiment
period and thus do not include individual fixed effects (columns 4-6 in Table 3). Instead, we
include a number of time-invariant individual covariates (a gender dummy, a student dummy,
seven age category dummies and a dummy for being a recurring member of the gym).18
The results are similarly positive, although somewhat smaller in magnitude. Combined with
somewhat larger standard errors, only the effect on the total number of visits is statistically
significant. This finding was expected, as the slightly lower attendance level of the treatment
group in the pre-experiment period is ignored in these regressions. Note that while the
control and treatment groups are balanced with respect to most observed characteristics
(see Table 1), the share of recurring members is 5 %-points higher in the control group.
Recurring members typically exercise more frequently, which may explain the slightly higher
pre-experiment number of visits within the control group. Thus, we prefer the DiD model
that corrects for these (and potentially other) differences.
So far, we have looked at the immediate effect of email reminders on gym attendance.
17There could be positive spillovers from the treatment group to the control group via peer effects. Treated
individuals could inspire untreated ones to meet up and exercise with them more frequently. If that is true,
our study underestimates the positive effects of reminders, although we expect that this bias is rather small.
18For 230 individuals (9 % of the sample), some of these covariates are missing, such that these individuals
are excluded from the regressions with covariates.
16
Table 3: Regressions on weekly gym attendance (Poisson models)
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.13*** 0.17** 0.12*** 0.072* 0.12 0.059
(0.035) (0.090) (0.038) (0.040) (0.097) (0.044)
Experiment (post) 0.13*** 0.10 0.13*** 0.059 0.032 0.064
(0.046) (0.11) (0.050) (0.053) (0.12) (0.058)
Recurring member 0.21*** 0.15 0.22***
(0.050) (0.11) (0.057)
Male 0.21*** -0.66*** 0.61***
(0.048) (0.034) (0.075)
Student 0.0020 0.060 -0.0040
(0.048) (0.13) (0.052)
Daytime card -0.31*** -0.69*** -0.21**
(0.074) (0.13) (0.087)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
N 66900 28320 65250 55825 55825 55825
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. We
report exp(β)− 1, which is the percentage effect. Seven age dummies and six gym location dummies are
included but not reported in columns (4)-(6). Note that the number of observations in columns (1)-(3)
varies because individuals with zero attendance in all time periods drop out in Poisson regressions with
fixed effects.
One might expect that once the reminders cease, treated individuals will stop attending the
gym at an increased frequency. Habit formation might, however, explain persistent effects of
a short-term intervention. Habits are acquired when ‘repetition of a behavior in a consistent
context progressively increases the automaticity with which the behavior is performed when
the situation is encountered’ (Lally et al., 2010, p. 998).19 In our context, this phenomenon
means that the email reminders would provide a stimulus upon which individuals decide to
go to the gym, and the increased frequency of going makes it easier for individuals to attend.
Only weak evidence of habituation has been found by studies in the field of health behavior
(see, e.g., Charness and Gneezy, 2009; Acland and Levy, 2015; Calzolari and Nardotto, 2017),
although this might be due partly to a lack of statistical power.
We present estimates of the impact of the reminders during the post-experiment period
(β2) in the second row of Table 3. We find the surprising result that the post-experiment
impact on total visits is of a similar magnitude as the immediate impact (a 13 % increase).
The same result holds for the impact on free training visits (column 3). The coefficient for
19For early work on the relationship between repetition and habit strength, see Hull (1943, 1951); more
recent contributions on habits and habit formation include Verplanken (2006) and Wood and Neal (2007).
17
class training in column (2) is slightly smaller, which, combined with less precision due to a
lower base rate, leads to the impact not being statistically significant. However, the magnitude
of the class training coefficient is very similar to that of the free training coefficient. A simple
difference model (columns 4-6) again leads to slightly smaller coefficients that are positive but
not statistically significant.
3.2 Robustness checks
We perform a series of robustness checks to investigate whether these results are sensitive to
certain choices in the empirical model. First, we estimate the baseline models with simple
OLS to assess the sensitivity of the results to the Poisson model. The results are presented
in Table 11 in Appendix C. The results are very robust. The statistically significant increase
of 0.095 weekly visits translates into a 10 % increase, which is very similar to the Poisson
regression estimate (13 %). Additionally, the effects on class training (16 %) and on free
training (8 %) are very similar to the Poisson estimates. When using a simple difference
model (columns 4-6), the coefficients are slightly smaller, and only the effect on total visits is
statistically significant (at the 10 % level).
Second, we show that removing the Christmas holidays from the pre-experiment period
hardly affects the estimates. During the last two weeks of 2016 (weeks 51 and 52) and the
first week of 2017 (week 1), gym attendance was substantially below its long-run level (see
Figure 4). In particular, in week 52, attendance was extremely low. Since these weeks are part
of the pre-experiment period, they may affect the DiD estimates. In Table 12 in Appendix
C, we exclude either week 52 or weeks 51, 52 and 1 from the analysis. Note that the total
pre-experiment period consists of weeks 49, 50, 51, 52 and 1, and thus, we can estimate the
DiD model even when excluding the Christmas holidays. Excluding either one or three weeks
hardly affects our results. In all columns, we find statistically significant positive effects on
gym visits, which are very close to our baseline estimates.
Third, we consider the length of the pre-experiment period. In all DiD estimations, the
pre-experiment period consists of five weeks. This period is restricted by the fact that our
sample consists of gym members who purchased their contracts between July 1, 2016, and
December 1, 2016. As a robustness check, we select a subsample of individuals who purchased
their contracts earlier (July 1 - November 1) and extend the pre-experiment period by four
weeks. The results are presented in Appendix C in Table 13 and are similar, although the
magnitude and significance of the estimates are smaller in the DiD model (but not in the
simple difference model).
Fourth, we consider using daily data on gym visits (instead of aggregated weekly data). We
use a specification that is similar to the one used for weekly visits, but we include additional
weekday fixed effects. The results are presented in Table 16 and Figure 11 in Appendix C.
18
We find that the impact of reminders is very similar to our baseline model in terms of both
magnitude and significance. Note that in the simple difference models, we can also include
two weather dummies (‘Sunny day’ and ‘Rainy day’) and find that while sunny days lead to
lower gym attendance, rain has no impact.
Finally, we perform two robustness checks with respect to the spike in attendance in the
treatment group in week 22. We show that the positive post-experiment effects are robust to
(i) excluding week 22 from the post-treatment period (see Table 14 in Appendix C) and (ii)
restricting the post-treatment period to all weeks up to (but not including) week 22 (Table
15 in the Appendix). Thus, the positive long-term effects are not (solely) driven by this
spike. In week 22, the gym sent an email to all its members (including both the treatment
and control groups). We hypothesize that the spike suggests that treated gym members were
more responsive to the email than control group members. This phenomenon could mean
that the reminder emails sensitized individuals to respond to a similar stimulus in the same
way.
To conclude, we find clear evidence of a persistent positive effect of reminders on atten-
dance. Interestingly, the percentage effects of the reminders are quite similar for the exper-
iment period and the post-experiment period. We interpret this finding as a sign of habit
formation. Note that the average number of weekly visits is lower in the post-experiment
period, and thus, the same percentage effects imply a lower increase in terms of the number
of visits in the post-experiment period.
3.3 Heterogeneous effects
The impact of email reminders may vary, depending on an individual’s characteristics. Given
our large sample size, we can investigate different sources of heterogeneous effects, and we
focus on gender, baseline gym attendance, previous gym membership (at the same gym),
student status and a pre-experiment measure of the likelihood of opening emails. We simply
estimate the same model for each particular subsample. Identification follows from a similar
standard DiD assumption. For example, in the case of heterogeneous effects by gender, we
assume that in the absence of the intervention, men in the treatment group would have had
the same trend in attendance as men in the control group (the same holds for women). We
report results only for the total number of weekly gym visits (rather than distinguishing
between classes and free training), and for simplicity, we consider only the short-run effect of
the reminders (and thus include only observations until the end of the intervention period).
We first consider heterogeneous effects by gender. While men and women have similar
average numbers of weekly visits (1.16 for men and 1.00 for women), they exercise in different
ways. One-third of gym visits by women are class training, while the corresponding number
is less than 10 % for men. In Table 4, we present estimation results for men in column (1) and
19
Table 4: DiD regressions on weekly gym attendance (Poisson models): heteroge-
neous effects
(1) (2) (3) (4) (5)
Men Women
High
attenders
Low
attenders
Recurring
members
Experiment (during) 0.13*** 0.11* 0.084*** 0.17** 0.13***
(0.046) (0.059) (0.034) (0.078) (0.047)
N 23104 16416 22211 19665 19418
(6) (7) (8) (9) (10)
New
members Student
No
student
Opened
baseline
Not opened
baseline
Experiment (during) 0.13*** 0.13*** 0.13*** 0.12*** 0.15***
(0.053) (0.050) (0.049) (0.045) (0.057)
N 22458 22135 19741 24244 17632
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses.
We report exp(β) − 1, which is the percentage effect. All regressions include individual fixed
effects and membership duration dummies.
those for women in column (2). We find that the estimates for the impact of the reminders
are very similar, and we conclude that reminders are equally effective for both genders.
Second, we consider heterogeneous effects by baseline gym attendance. We split the sample
at the median weekly attendance in the pre-experiment period, which is 0.6 visits per week.
The estimation results are presented in column (3) (high attenders) and column (4) (low
attenders) and suggest a greater impact of email reminders on members who attended less
than the median. This finding is in line with the results from Calzolari and Nardotto (2017)
and are not particularly surprising since for high attenders, there is a lower scope to increase
the amount of exercising beyond the current level (e.g., due to time constraints). Note,
however, that a formal test again cannot reject that the coefficients for high and low attenders
are equal.
Third, we test whether recurring members (those who had a contract at the same gym prior
to the current contract) are more or less affected by the reminders. Since recurring members
extended their previous contract, they are likely to be more committed to exercising, and
thus, we expect them to benefit less from email reminders. However, we do not find evidence
supporting this hypothesis, with the coefficients for the two groups being almost identical
(columns 5 and 6). This finding suggests that the group of recurring members is perhaps less
selective than one might expect.20
Fourth, we consider student status. As shown in Table 1, approximately half of our
20A potential explanation is that many new members of this particular health club may have previously
attended a different gym and are thus equally experienced and committed.
20
sample consists of students.21 While we have no strong hypothesis for differing impacts
of the reminders for students and non-students (except for maybe differently binding time
constraints), we still believe it is important with respect to external validity. For example,
previous findings on the effectiveness of reminders by Calzolari and Nardotto (2017) were
based only on a student population, and thus, we are interested in whether the impact extends
to a more general population. The results are presented in columns (7) and (8), and we find
that the estimates are basically identical for students and non-students. Thus, the positive
effect is present even for a population that is more likely to be (full-time) employed and that
has (perhaps) more stringent time restrictions for exercising.
Finally, we consider heterogeneous effects according to whether one opens the baseline
email (1,405 out of the 2,463 individuals in the sample opened this email; see Section 2.3).
Ideally, one would measure whether those who actually opened the reminders were more likely
to visit the gym. However, since those members are a selective subset of the treatment group
and because we do not observe opening rates for the control group, we have no valid control
for those who opened the reminders. Thus, we use opening the baseline email as a proxy for
opening gym emails in general. The results for those who opened the email are presented
in column (9) and for those who did not open the email in column (10). We find the rather
surprising result that there is no significant difference between the two groups. The coefficient
for those who did not open the baseline email is even slightly larger than the coefficient for
those who opened the email.22 Thus, even for those who are (relatively) less likely to open
and/or read emails, we find a significantly positive effect. While this finding is somewhat
puzzling, one explanation could be that opening the email is not a necessary condition for the
message to be received. Seeing the email appear in one’s inbox could be sufficient to bring
the gym to one’s attention and thus overcome limited attention. However, an alternative
explanation is that the statistics on downloading web bugs severely underestimate the actual
opening rate, meaning that our results are (to some extent) diluted by measurement error.
In summary, we do not find strong evidence for heterogeneous effects of email reminders.
The impact does not appear to vary by gender, recurring membership, student status or the
likelihood of opening emails. We find only some suggestive evidence that low attenders are
more affected than high attenders, but the difference is not statistically significant. This
finding suggests that nearly all groups benefit from reminders in terms of overcoming their
limited attention, which strengthens the potential for reminder policies.
21Student status is recorded in the data only if one (voluntarily) reports being a student at the time he or she
purchases the contract. Thus, this measure might not be a perfect measure of student status. However, given
the sizable discount that students receive (around 30 %), most students will likely disclose this information.
22We can extend this approach by predicting who opens emails, based on a range of individual characteristics
(including whether the baseline email was opened). To do so, we regress (for the treatment group) the share of
the 12 email reminders that were opened on individual (pre-experiment) characteristics. Splitting the sample at
the median predicted opening rate, we again find that there is no significant difference between the effectiveness
of the reminders for those who are likely to open the emails and those who are less likely to open the emails.
21
3.4 Booking and canceling
In the previous sections, we established that gym members who received email reminders
attended gym classes more frequently both during the treatment period and in the three
months after the reminders had ended (although the long-run effect is not significant). As
we have detailed data on class bookings and booking cancellations, we now consider whether
reminders lead to more bookings, fewer cancellations or both. The following analysis sheds
light on whether reminders lead only to an (almost) immediate response of visiting the gym
or whether they also stimulate the recipient to plan a future visit by booking a class. Before
we proceed with the empirical analysis, we briefly describe the features of the booking and
cancellation system in place at the health club. Note that we postpone a more thorough
discussion of booking and canceling behavior to Section 4.
The classes offered at the gym have a maximum capacity. Members can book a spot in a
class online or through a mobile app up to seven days prior to the class.23 Once all spots are
booked, new bookings are placed on a waiting list and result in a spot when earlier bookings
are canceled. A booking can be canceled online or through the mobile app up to one hour prior
to the class. In the last 60 minutes before the class, members without a booking can claim a
spot with the gym’s reception staff if a spot is still available (‘drop-in attendance’). Finally,
those with a prior booking must show up at least 10 minutes prior to the class; otherwise,
the booking is canceled and the spot becomes available for drop-in attendance. If a booking
is not canceled in time and the individual does not show up, his or her card is blocked from
further bookings until a fine of SEK25 (approximately $3) is paid. We refer to these cases
as ‘no-shows’. In the analysis in this section, we label no-shows simply as cancellations for
simplicity, but we discuss the difference in detail in the next section.
It is not uncommon for all spots in a class to be taken, often several days in advance. For
example, of all 9,513 classes for which we have data, 30.7 % had a waiting list (at some point).
This figure does not mean that all spots were eventually taken for one-third of the classes
because canceling is also very common (for more details, see the next section). However, it
does illustrate that booking classes beforehand is often necessary and very common. To keep
the analysis simple, we use only the first and the last observable actions, i.e., the very first
booking and the observation of whether the person attended or canceled, and we ignore all
actions in between.24
We estimate a DiD model with individual fixed effects, similar to equation (1), but with
different outcome variables. First, we estimate the impact of the reminders on the number
of weekly bookings using a linear model (column 1 in Table 5). We find a positive impact of
23A member is limited to five active bookings at a time.
24For the majority of observations, this makes no difference, but a non-trivial share of observations are
somewhat more complicated. Many alternative scenarios are possible, for example, a member may book a
class, then cancel it, and then attend through a last-minute drop-in.
22
Table 5: Regressions on weekly class bookings and cancellations of bookings
(1) (2) (3) (4) (5)
Number of
bookings
(linear)
Number of
bookings
(Poisson)
Positive #
of bookings
(LPM)
Number of
cancellations
(linear)
Share of
bookings
canceled
Experiment (during) 0.056* 0.11 0.014* 0.026 -0.023
(0.031) (0.076) (0.0080) (0.018) (0.026)
Experiment (post) 0.016 0.049 0.0013 0.0028 -0.038
(0.031) (0.092) (0.0083) (0.018) (0.028)
Mean value dep. var. 0.42 0.42 0.16 0.18 0.37
Percent effect (during) 13 9 14 -6
N 73890 33270 73890 73890 11653
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. In
column (2), we report exp(β)− 1, which is the percentage effect. Note that the number of observations is
lower in columns (3)-(4) because individuals with zero attendance in all time periods drop out in Poisson
regressions with fixed effects. Columns (5)-(6) are linear probability models in which the outcome has
been set to 1 in the case of a positive number of bookings/cancellations.
0.056 bookings, which is significant at the 10 % level and corresponds to a 13 % increase in
the number of bookings per week. There is no impact after the experiment ends. Since the
number of bookings is a count variable, we also estimate a Poisson regression in column (2).
The result is again positive and very similar in magnitude (11 % increase), although it is not
statistically significant. Given the relatively small share of gym members who attend classes,
we also consider the extensive margin of making any positive number of bookings in column
(3). We find that reminders lead to significantly more members making at least one booking
(a 9 % increase).
We now consider cancellations. In column (4), we show that the number of weekly cancel-
lations increases during the experiment; however, the coefficient is small and not statistically
significant. Cancellations are obviously possible only ifa booking was made earlier. Thus, we
consider in column (5) the share of bookings that are canceled, and as a result, this regres-
sion includes only observations for weeks in which the individual made at least one booking.
We find a negative coefficient, although it is again small in magnitude and not statistically
significant. We conclude that the increase in class visits as a result of the reminders is due to
(1) an increase in the number of bookings and (2) a lack of an increase (and perhaps even a
decrease) in the share of bookings that are canceled.
The effects of the reminders for the post-experiment period all share the sign of the effect
during the experiment. However, none of the effects is statistically significant. This finding is
in line with the estimated impact on class attendance for the post-experiment period, which
is smaller in magnitude and not statistically significant (see Table 3).
23
Table 6: Contract renewal after initial 12-month period
Control Treatment p-value test
group group equal proportions
Full sample:
Active after 13 months 0.57 0.59 0.55
Active after 14 months 0.56 0.55 0.92
Observations 1232 1231
Upfront (12-month) paid contract:
Active after 13 months 0.49 0.48 0.77
Active after 14 months 0.51 0.48 0.39
Observations 579 554
Monthly direct debit payments:
Active after 13 months 0.65 0.67 0.38
Active after 14 months 0.60 0.61 0.59
Observations 653 677
3.5 Impact on contract duration and renewal
We have seen that there are persistent effects of the intervention in terms of increased at-
tendance in the post-experiment period. We now investigate whether these increases carry
over to the decision on contract renewal once the 12-month contract period ends. One would
expect that individuals who were induced by the intervention to exercise more regularly are
more likely to remain members of the gym. Rational decision makers base this decision on
projections of future gym attendance (though in practice, many may underestimate their fu-
ture self-control, suffer from overconfidence, etc.; see, e.g., DellaVigna and Malmendier, 2006).
The reminders might increase projections of future attendance if past experiences are used as
a basis for predictions about the future. It is doubtful though whether the magnitude of the
increase in attendance suffices to affect the (larger) decision of buying another contract.
We obtained data on whether individuals buy a new contract (or extend the first contract)
and define as the outcome measure an indicator for being a member 13 or 14 months after
purchasing the initial contract.25 Table 6 shows the share of individuals in the control and
treatment groups that were active members 13 and 14 months after the initial contract began.
In the upper panel, we observe that slightly more than half of our sample had active contracts
both 13 and 14 months after the initial contract began, while the share decreased slightly from
month 13 to 14. The share of active members is very similar in the treatment and control
groups, and the difference is not significant (with a p-value of 0.55 after 13 months and 0.92
after 14 months).
The type of payment of the initial contract (upfront or monthly) is likely to affect the
25We include both measures (13 and 14 months) to take potential (short) gaps between the initial contract
and the new contract into account.
24
likelihood of contract renewal, and thus, we report the same statistics, split by the type of
payment, in the lower panels of Table 6. Again, we find no significant differences between
the control and treatment groups for both payment methods after 13 months or after 14
months. Interestingly, the number of active contracts is much higher among those who had
initially chosen direct debit payments compared to those who had paid upfront. This finding
suggests that there is some kind of cancellation delay for contracts with direct debit payment,
as also reported by DellaVigna and Malmendier (2006).26 This delay can also be explained
by limited attention, because for these contracts, an active choice (the choice to terminate the
contract) must be made, and members might frequently forget to make this decision. Overall,
we do not find any impact of our reminders on contract duration and renewal, which suggests
that persistent effects of an intervention fade once a decision at a margin different from that
targeted in the intervention is made.
4 Booking and canceling behavior
In this section, we go beyond the analysis of the experimental data pertaining to the reminders.
We first document a number of observations related to behavioral patterns in the data on gym
class bookings, booking cancellations and class attendance. A standard classical choice model
assuming a rational individual with perfect foresight and the mental capacity to keep track
of all decisions would never predict that the individual will book a class and subsequently
cancel the booking. Therefore, we develop a theoretical model that deviates from this simple
assumption and incorporates several factors and mechanisms that influence an individual’s
intertemporal choices (limited attention, time-inconsistent preferences, i.e., present bias, and
uncertainty about future attendance costs). Finally, we discuss which observations can be
explained by which factors.
4.1 Descriptive statistics and behavioral patterns in the data
For the following analysis, we use an observational dataset that is much larger than the
one for the reminder experiment. This dataset contains observations for class bookings,
cancellations and no-shows among all gym members for the period from July 1, 2016, to June
30, 2017 (for more details on the booking system, see Section 3.4). We observe approximately
190,000 bookings and 28,000 drop-in attendances by over 7,000 gym members. Our first two
observations are related to the booking of classes and cancellation of bookings.
Observation 1 Most gym class visits are booked prior to the class (78 %) and are not made
through drop-in.
26Note that it cannot be excluded that (part of) the difference is due to self-selection into the two payment
types.
25
Figure 5: Distribution of bookings and cancellations
0
5
10
15
20
25
Per
cen
t
−8 −6 −4 −2 0Days
(a) Bookings in the week prior to class
0
2
4
6
8
10
Per
cen
t
−25 −20 −15 −10 −5 0Hours
(b) Bookings in the 24 hours prior to class
(25.4 % of all bookings)
0
20
40
60
80
Per
cen
t
−8 −6 −4 −2 0Days
(c) Cancellations in the week prior to class
0
5
10
15
20
25
Per
cen
t
−25 −20 −15 −10 −5 0Hours
(d) Cancellations in the 24 hours prior to class
(83.1 % of all cancellations)
In the upper panel of Figure 5, we show the distribution of class bookings during the week
before a class (note that bookings can be made beginning seven days prior to a class). In
panel (a), we find that the number of daily bookings is approximately uniform across the first
six days prior to a class (roughly 10 % per day). Around 15 % of all bookings are made on
the day before a class and another 25 % on the day of the class (‘day zero’). Most of the
bookings on day zero are made in the very last hours before the class (panel b).27 This finding
shows that there is considerable interest in booking classes quite early and that an increasing
number of people book a class in the final days and even in the last hours before the class.
Observation 2 A significant share of bookings are canceled (41 %), and most of these can-
cellations occur on the final day before the class (83%).
Thus, canceling is very common, as can be seen from the first column of Table 7. In Figure
27The low level of bookings between the final 10 and 18 hours before the class are likely due to this time
often being nighttime, as a majority of classes are held in the late afternoon or evening hours.
26
Table 7: Bookings, cancellations and drop-in attendance by class popularity
All Unpopular Fairly popular Popular
classes classes classes classes
(< 0.5% (< 5% (> 5%
waiting list) waiting list) waiting list)
Number of bookings 190454 24912 40543 124999
Of those (in %):
Attendance (%) 0.53 0.57 0.53 0.52
Cancellation (%) 0.41 0.38 0.41 0.41
No-shows (%) 0.057 0.057 0.057 0.058
On final waiting list (%) 0.008 0.000 0.001 0.012
Number of drop-in attendance 27828 10024 7281 10523
As % of all attendance 0.22 0.42 0.25 0.14
Number of classes 238 98 44 96
The categorization of classes by popularity is made based on the share of all bookings that goes to
the waiting list (which is, due to cancellations, much larger than the share that eventually remains
on the list when the class occurs). The ‘Cancellation’ share includes cancellations by people on the
waiting list. Being ‘on the final waiting list’ means that a person is still on the list when the class
occurs (and thus is not allowed to participate).
5 (c), we show how cancellations are distributed by booking day (no-shows are included in
the figure as canceling zero minutes prior to the class). We see that the vast majority (83.1
%) of all cancellations occur on the day of the class, and 62.8 % of all cancellations occur
within the final six hours (panel d). Thus, we find that most people who change their minds
and cancel a booking do so at the very last moment.
Our dataset contains large variation across classes in terms of their popularity. While some
classes are never fully booked and thus never have a waiting list, others are (almost) always
fully booked. Arguably, the popularity of a class, which is exogenous from an individual’s
point of view, must have an impact on booking and cancellation decisions. For example, if a
particular class is always fully booked, there is a need to book in advance because the class
cannot be attended otherwise. If, by contrast, a class is never fully booked, there is no obvious
need to book this class (unless the class’s unpopularity is unknown to the individual).
To define the popularity of a class, we take the following approach. First, we define a
class relatively broadly by the type of activity (yoga, cycling, etc.) and its scheduled time
(day of the week plus ‘daytime’ or ‘evening’). This approach leads to 238 different types of
classes (see Appendix C.3 for more details). In the next step, we categorize these classes into
‘unpopular’, ‘fairly popular’ and ‘popular’ based on the share of bookings that end up on the
waiting list. For a popular class, more than 5 % of all bookings appear on the waiting list.
Given a typical class size of 20, there will thus always be a waiting list for this type of class
on average (with at least one person on this list). For a fairly popular class, between 0.5 %
27
and 5 % of the bookings end up on the waiting list; for an unpopular class, less than 0.5 % of
bookings end up on the waiting list. Applying this procedure implies that the likelihood of a
waiting list appearing for an unpopular class is close to zero. Overall, 40 % of all classes are
categorized as popular, and half of these even have a waiting list share of more than 20 %.
The next two observations are related to bookings and booking cancellations for unpopular
classes.
Observation 3 Even unpopular classes, for which waiting lists are extremely rare, are booked
well in advance (58 % of all visits are booked in advance).
In columns (2)-(4) of Table 7, we present statistics on the number of bookings, cancellations
and drop-in attendances for the three popularity categories. We find that, as expected,
attendance through drop-in occurs more often in unpopular classes than in the other two
categories. However, even in unpopular classes, where waiting lists virtually never appear,
more than half of all class participants make a booking. Not only do many attendees of
unpopular classes make a booking, but many do so very early. For example, approximately
30 % of the bookings are made at least three days in advance (see Figure 12 in the Appendix).
Observation 4 For unpopular classes, a significant share of bookings are canceled (38 %).
In Table 7, we show that of all bookings, only between 52 and 57 % are actually followed by
class attendance (with unpopular classes having a slightly larger share than popular classes).
The remaining share of all bookings is not followed up on (thus resulting in a fine), as the
next observation states.28
Observation 5 There is a substantial share of no-shows (6 % of all bookings) across all
popularity types.
Again, this can be seen from Table 7. As a final observation, we consider the evidence
from the experiment that class participation increases after receiving reminders:
Observation 6 Gym members who receive regular reminders from the gym (i) make more
bookings, (ii) are less likely to cancel their bookings and (iii) attend more classes.
The first finding is based on the result shown in Table 5 (columns 1 and 2) that the weekly
number of bookings increases during the experiment. The second finding is based on the
same table (column 5) and suggests that the share of bookings that are canceled is lower
when reminders are received. We acknowledge, however, that this result is far from statis-
tically significant and suggests only weakly that canceling becomes less likely as a result of
the reminder emails. The third finding, that class attendance increases when reminders are
received, is most robust (Table 3, column 2).
28A small share of all bookings for popular classes remain on the waiting list (1 %). Note that this small
share does not necessarily mean that excess demand is rare. Members might cancel their bookings while on
the waiting list or decide not to book in the first place if a waiting list exists.
28
4.2 A model of booking, canceling and attendance
We now proceed by laying out a simple two-period discrete choice model that describes an
individual’s decisions on booking, canceling and attending a class. The goal is to investigate
the relevance of several behavioral biases, and the model provides a framework to think about
them. Although the model is highly stylized, it contains the key elements that play a role in
the decision-making process. In particular, we assume that (i) there is uncertainty about the
future attractiveness of attending a class (and corresponding beliefs might be overoptimistic),
(ii) there is uncertainty about whether the class will be fully booked, (iii) there is a risk of
forgetting about a class that one would actually like to attend and for which one might have
even made a booking, and (iv) preferences can be time-inconsistent (and individuals are either
na¨ıve or sophisticated regarding their time inconsistency).
Consider an individual who can attend a class that takes place in period t = 2. In period
t = 1, she decides whether or not to book a spot in the class, and in period two, she decides
whether or not to attend the class.29 Her second-period utility, U2, depends on two (binary)
state variables: whether she made a booking in the first period (B, equal to one in case of a
booking and zero otherwise) and whether the class is fully booked (F , equal to one if the class
is fully booked and zero otherwise). As a result, there are four possible states of the world in
the second period, where the individual decides to attend (A = 1) or not to attend the class
(A = 0). Only if she made no booking (B = 0) and the class is fully booked (F = 1) is there
no possibility to attend and thus no decision to be made.
Attending the class is associated with time and effort cost cA. In period two, these costs are
assumed to be known (they are not known in period one; see below). The returns to attending
the gym are a future stream of health benefits u3, . . . , uT , which are assumed to materialize in
period three (one period after attending the class) and the subsequent periods. The period-
two utility from attending the class is given by −cA+β
∑T
t=3 δ
t−2ut. For simplicity, we denote
the present value of these future (accumulated) health benefits by U¯ ≡
∑T
t=3 δ
t−2ut. We apply
the β− δ (quasi-) hyperbolic discounting framework (Phelps and Pollak, 1968; Laibson, 1997)
to allow for time-inconsistent preferences. The per-period discount factor is δ ≤ 1, while
payoffs between the present and all future periods are additionally discounted with factor β.
For β < 1, discounting is quasi-hyperbolic and preferences are time-inconsistent, while for
β = 1, discounting is exponential and preferences do not change over time. We will discuss
the individual’s knowledge about β later. Depending on the two state variables, optimal
29Note that period t = 1 represents all days prior to the class date. The model can be extended to seven
periods, with each representing one day and the last representing the day of the class or, more specifically, the
time the class takes place. In each period, a booking can be made or canceled. While complicating the model
significantly, this extension will not enrich the key mechanisms in any way, and thus, we opt for the simplest
model that still contains all relevant mechanisms.
29
period-two utility is given by
U2(B,F ) =



max
A=0,1
[
A · [βU¯ − cA]
]
if B = 0, F = 0
0 if B = 0, F = 1
max
A=0,1
[
A · [βU¯ − cA] + (1−A) · [−cT ]
]
if B = 1
(2)
If the individual did not book in period one and the class is not full, she attends if the
discounted benefits (βU¯) outweigh the costs of attending (cA). If she did not book in period
one and the class is full, she cannot attend, there is no choice to make and she obtains her
outside option, which is normalized to zero. If she booked in period one, it is irrelevant
whether the class is full. Again, she attends if the benefits outweigh the costs; however, not
attending requires canceling the booking through the website or app, which is assumed to be
associated with a transaction cost of cT (this cost also applies to making a booking in period
one, see below). If the individual does not show up to a booked class, she has to pay a fine
cF . We assume that cF > cT , so it is always optimal to cancel the booking when one does
not want to attend the class.30
Now, we move on to the period-one decision on whether or not to book a spot in the class.
The process by which the class becomes fully booked (or not) is exogenous to the individual,
and thus, we assume that there is simply a probability (1− γ) that between periods one and
two, all available spots are booked. Furthermore, we need to specify what information the
individual has in period one. We assume that the costs of attending the class in period two
are unknown to the individual in period one. This assumption is reasonable, as uncertainty
exists about many factors that influence the cost of attending (e.g., the weather, one’s plans,
one’s energy and motivation level), and it is common in the literature (e.g., DellaVigna and
Malmendier, 2006). Rather than knowing the precise cost, the individual obtains a signal c∗A,
which is correlated with the true value cA.31 Given her prior belief about the distribution of
the cost and the observed signal, the individual uses Bayesian updating to the distribution of
the cost of attending the class.32
Finally, we include limited attention in the model. In particular, we assume that in period
two, there is a probability (1 − pi) that the individual forgets about the class, regardless of
whether a booking has been made. If a booking was made, this individual will be a no-show;
if no booking was made, she will simply not attend. This interpretation of limited attention
30Given that it takes only a minute to cancel the booking and the fine for not attending a booked class is $3,
individuals would need to have a very high value of time in order for canceling to be more costly than paying
the fine. This situation seems rather implausible, especially for the student population.
31Depending on the joint distribution of c∗A and cA, the signal might be unbiased. Alternatively, it might
reflect systematical over- or underestimation of future attendance costs (and thereby the likelihood of atten-
dance). We discuss overconfidence with respect to future attendance costs in more detail below.
32Individuals who have been gym members for a while are likely to have a good sense of their true distribution
of the attendance cost and use this distribution as the prior.
30
is similar to that of, e.g., Taubinsky (2014). Given her expectations about the future cost of
attending, the likelihood of the class being fully booked and the risk of forgetting, she decides
whether to book a spot during period one. Her optimal utility in period one is given by
U1 = max
B=0,1
[
B·
[
piβδE1[U2(B = 1, ·)]− (1− pi)βδ · cF ]− cT
]
+
(1−B)· piγβδE1[U2(B = 0, F = 0)]
]
(3)
If she decides to book (B = 1) and does not forget (with probability pi), she obtains the
expectation of the value U2(B = 1, ·), discounted with factor βδ. If she does forget (with
probability 1 − pi), she will pay cF in period two for not showing up to the booked class,
discounted with βδ. In either case, she immediately incurs the transaction cost cT (for making
the booking). The expectation E1 is taken conditional on period-one information. If she does
not book, she obtains with probability piγ the period-one expectation of U2(B = 0, F = 0),
again discounted with βδ. If she does not book and the class is fully booked (with probability
1 − γ) or if she forgets to attend in period two without a prior booking (with probability
1 − pi), she obtains a payoff of zero in the second period, and thus, these terms are omitted
from the above equation. The decision to book in period one simply depends on whether the
additional option value that results from booking outweighs the costs in terms of booking
costs, potential canceling costs and the no-show fine.
The model allows for three distinct types of time preferences. First, for β = 1, discounting
is exponential and choices are consistent across periods. Second, if β < 1 and the agent
is aware of the value of β, the model describes a sophisticated time-inconsistent discounter.
Such an agent is present-biased and knows this. As a result, she may have a demand for
commitment devices that restrict future choices or that make some choices more unlikely.
Third, if β < 1 but the agent erroneously assumes that β = 1 when making decisions, the
model describes a na¨ıve time-inconsistent discounter. Such a na¨ıve individual falsely believes
that her future self will maximize her current preferences, so in period one, she would compute
the period-two terms as if β were equal to one (see, e.g., Frederick et al., 2002, for an in-depth
discussion of these concepts).33
33The model describes the single (two-period) decision regarding attendance in one particular class. Our
setting is one in which this decision is repeatedly made, for example, one could imagine this situation occurring
once every week. While we do not discuss what may change when modeling the repeated decision, there are
some obvious possibilities. First, the prior distribution should become more accurate each time the individual
observes the attendance cost in period two. As a result, overconfidence will decrease. Second, the belief about
the value of γ would become more accurate. Third, one might learn about commitment problems, thereby
becoming less na¨ıve and more sophisticated about a potential present bias.
31
4.3 Comparing the model to observed behavior
We now elaborate on each of the mechanisms included in the model and discuss which model
features can explain the six observations that we made in Section 4.1. In Table 8, the obser-
vations from the data are listed across columns, from Observations (1) to (6).
Baseline. As a baseline, we consider a model where the individual has time-consistent
preferences and does not suffer from limited attention and where there is no uncertainty about
future attendance costs (see the first row of Table 8). This model predicts that bookings may
be made for popular classes, which might be fully booked in period two, that is, for which
γ < 1, in order to secure a spot (column 1). However, such a booking would never be canceled
(column 2) because the optimality of the booking does not change over time. Furthermore,
making a booking for an unpopular class would never be optimal (columns 3 and 4), as
unnecessary booking costs would be incurred. Finally, we would not expect this individual
to not show up to a booked class (column 5) or behave differently after receiving reminders
(column 6).
Uncertainty about attendance cost. As a first step, we enrich the baseline model with
period-one uncertainty about the attendance cost in period two (see the second row in Table
8). An individual will now make a booking in period one if the option value of having a
guaranteed spot in the class outweighs the cost of making the booking. Note that we still
assume that preferences are time-consistent and that there is no limited attention, i.e., β = 1
and pi = 1.
Booking in period one is optimal if it yields a higher expected period-two payoff than not
booking, i.e., if the following condition holds:
δEmax(U¯ − cA,−cT )− cT > γδEmax(U¯ − cA, 0) (4)
The value of booking (the left-hand side of the inequality) is the discounted expected value of
the maximum of either attending (U¯ − cA) or canceling (−cT ), while an immediate booking
cost cT is also incurred in either case. The expectation is taken over the distribution of cA
(conditional on the observed signal in period one). The value of not booking is again the
discounted expected value of the maximum of attending or not attending; however, this value
is obtained only if the class is not fully booked (which happens with probability γ). From
condition (4), it is clear that for sufficiently low values of γ, booking is an optimal decision.
Furthermore, given a booking, a sufficiently large realization of cA will make canceling optimal
in period two. Thus, this model can explain both booking and canceling classes that are
sufficiently popular. However, for classes that are unpopular and never fully booked (γ = 1),
condition (4) shows that booking is never optimal. Thus, this model cannot explain booking
32
Table 8: How can the observed behavior of gym members be explained?
Popular Unpopular No-shows Impact of
classes classes reminders
(γ < 1) (γ = 1)
Book Cancel Book Cancel
Observations: (1) (2) (3) (4) (5) (6)
Mechanisms:
(1) No uncertainty, TI or LA yes no no no no no
(2) Uncertainty yes yes no no no no
(3) Na¨ıve TI yes yes no no no no
(4) Sophisticated TI yes no yes no no no
(5) Uncertainty + Na¨ıve TI yes yes no no no no
(6) Uncertainty + Soph. TI yes yes yes yes no no
(7) Limited attention no no no no yes yes
Uncertainty refers to period-one uncertainty about period-two attendance cost (cA). TI is short for time
inconsistency. LA is short for limited attention.
(or canceling) unpopular classes.34
As a sidestep, uncertainty also predicts that as γ decreases (as a class becomes more
popular), cancellations become more likely because if a class is sufficiently popular, booking
is optimal even for unfavorable signals about attendance costs. Since those who receive
unfavorable signals are more likely to cancel their bookings, we should observe a larger share
of cancellations for popular classes. This is indeed the case, as can be seen in Table 7.
Na¨ıve time inconsistency. We now consider how time-inconsistent preferences affect the
model predictions. To this end, we eliminate uncertainty about attendance cost and assume
that the agent knows the exact value of cA in period one. Additionally, we begin by assuming
that time inconsistency is na¨ıve (Laibson, 1997). Thus, while β < 1, the individual erroneously
assumes β = 1 for future periods. In this case, the booking trade-off is between the value of
having a guaranteed spot and the immediate booking cost. She will book if the attendance
cost cA is such that
βδ(U¯ − cA)− cT > γβδ(U¯ − cA) ⇐⇒ (1− γ)βδ(U¯ − cA) > cT (5)
34It should be noted that the booking and canceling of popular classes can also be explained by uncertainty
in combination with overconfidence. Individuals might systematically underestimate attendance costs and
thus be systematically overoptimistic when they are confronted with the decision to make a booking. In the
model, overconfidence results if the belief about the attendance cost distribution does not resemble the true
distribution. Such biased beliefs can give rise to overestimation of the probability of attending in period two
and thus increase the likelihood of booking (and canceling). Such overconfidence has, e.g., been identified by
DellaVigna and Malmendier (2006) as a likely explanation for the observation that many people sign up at
gyms but do not attend as often as they intend to. The overconfidence channel would reinforce the uncertainty
channel but would not change the main conclusions.
33
In period two, she will change her mind and cancel her booking if the discounted returns
minus the attendance costs are small enough that she prefers to incur the cancellation costs:
−cT > βU¯ − cA (6)
If we assume that δ is close to unity for such a short period as a week, we can ignore it and
combine conditions (5) and (6) into
cT
(1− γ)(U¯ − cA)
< β <
cA − cT
U¯
(7)
If this condition holds, the individual will book in period one and cancel in period two,
even in the absence of any uncertainty. The interpretation of this condition is that a low β
makes booking unattractive because it puts relatively more weight on the directly incurred
transaction cost, while a large β (close to unity) implies less discounting of future health
benefits in period two, making canceling in period two less attractive. There exists an interval
for β, dependent on cT , cA, γ and U¯ , for which time-inconsistent preferences lead to booking
in period one and subsequent canceling in period two, and thus, na¨ıve time inconsistency can
explain Observations 1 and 2.35 However, it cannot explain such behavior when classes are
unpopular (γ = 1) since booking such classes would again never be optimal (straightforward
to see from condition 5). This conclusion is summarized in row 3 of Table 8.
Sophisticated time inconsistency. If individuals are sophisticated and thus are aware
of their inconsistent behavior, they understand that if they book in period one, they might
cancel that booking in period two because they will discount the future benefits U¯ relatively
more once period two arrives. Thus, they wish to commit themselves to future attendance.
The simple act of booking a class may function as a commitment device and increase the
likelihood of attending the class. While a booking ultimately does not commit one to attend
the class, it does place a cost on not attending (in terms of the canceling cost or in terms of
the fine that has to be paid in case of failing to attend a booked class). Thus, it enlarges the
set of values of cA for which attending is the optimal period-two choice.36
In period one, a sophisticated individual would like to attend the gym in period two if
cA < U¯ , i.e., if attendance costs are sufficiently small. However, she knows that she will
attend in period two only if cA < βU¯ . By booking the class in period one, she changes the
period-two condition for attending to cA < βU¯ + cT , which implies that it becomes relatively
35Note that the bounds of the interval become wider if cT is smaller. For small transaction costs, time-
inconsistent preferences are thus more likely to explain booking and subsequent canceling.
36Furthermore, it has been found that the simple act of planning to take an action may make its implemen-
tation more likely (Beshears et al., 2016), as one feels more guilty if an explicitly set plan is not realized. While
this mechanism is not present in the model, it is a reason why the cost of canceling might be larger than what
would be implied by the (small amount of) effort and the time spent canceling a booking.
34
Table 9: Booking and canceling by class popularity: experienced attenders that have attended
at least 5 times before
Unpopular Fairly popular Popular
classes classes classes
(< 0.5% waiting (< 5% waiting (> 5% waiting
list) list) list)
Number of bookings 6251 10877 47442
Of those (in %):
Attendance (%) 0.66 0.61 0.59
Cancellation (%) 0.29 0.35 0.36
No-shows (%) 0.04 0.04 0.04
On final waiting list (%) 0.00 0.00 0.01
Number of drop-in attendance 3864 1869 2798
As % of all attendance 0.48 0.22 0.09
Number of classes 51 40 83
more attractive to attend in period two. Of course, the commitment device comes at the cost
of incurring the booking cost in period one and provides the additional benefit of having the
option to attend in period two even if the class is fully booked. The commitment device is
valuable for both popular and unpopular classes. Thus, as opposed to na¨ıve time inconsistency,
sophisticated time inconsistency predicts the booking of unpopular classes that will never
be fully booked and can explain Observation 3. Cancellations, on the other hand, cannot
be explained by sophisticated time inconsistency because a sophisticated person perfectly
anticipates her choices in the second period and will never change her mind in the absence of
uncertainty (see row 4 of Table 8).
One might worry that the above conclusions hinge on the agent being aware of the risk of
a particular class becoming fully booked.37 In particular, she would need to know that γ is
(very close to) one for some classes. If she is uncertain about this risk, she might still book to
be sure that she will be able to attend rather than to commit herself. We argue that this is
unlikely to drive our results for two reasons. First, we have categorized a class rather broadly,
which makes it much more likely that a gym member is aware of the popularity of the class.
In our classification, a 45-minute spin class on Thursday at 7 pm is classified as the same type
of class as a 60-minute spin class on Thursday at 8 pm. Because of this broad classification,
gym members are likely to quickly learn which classes are potentially popular. Additionally,
a class type is called ‘unpopular’ only if all versions of this class are unpopular. Second, we
can focus on a subset of gym members who attend a particular type of class regularly. Those
37Another concern might be that those who participate in unpopular classes systematically differ from those
who attend popular classes. We show in Table 17 in the Appendix that there are some differences between the
groups, but those are small compared to the variation in the sample.
35
individuals certainly know the popularity of the class. We restrict the sample to observations
for gym members who book a certain type of class that they have attended at least five times
before. In Table 9, we reproduce Table 7 for this specific sample of experienced attenders. As
expected, among more experienced gym members, the share of drop-in attendance is larger
for unpopular classes. However, even for very experienced attenders, we still find that a large
majority book the class in advance rather than drop in. Thus, bookings indeed seem to be used
to commit to attend the class. Note that the statistics in Table 9 reveal another interesting
finding, which is that as the members become more experienced, they are less likely to cancel
their booking (this finding holds for all popularity categories). This finding can be explained
by sequential learning about the attendance cost distribution, which enables members to more
precisely predict whether they will actually attend the class. Thus, potential overconfidence
will decrease over time.
Since only sophisticated time inconsistency can explain Observation 3 (that unpopular
classes are commonly booked), we conclude that bookings are used as commitment devices to
cope with time-inconsistent preferences. This finding is complementary to results from recent
field studies (e.g., Goldhaber-Fiebert et al., 2010; Bhattacharya et al., 2015; Royer et al.,
2015), which show that a significant share of people are willing to sign a contract that puts
a penalty on failing to achieve a specified number of workouts or gym visits over a particular
time period. Significant demand for commitment devices has also been shown to exist for
savings products (Ashraf et al., 2006), restricted lunch menu options (Toussaert, 2016) or
self-imposed deadlines for tasks (Bisin and Hyndman, 2014). Our setting differs from these,
as we consider a repeated decision on whether or not to obtain a commitment device (rather
than a one-off decision). It is not obvious that the value of commitment devices in single
decisions, as shown in the literature, carries over to situations where a decision is repeatedly
made. Thus, we contribute by showing that even for the gym attendance decision, which is
made repeatedly on a weekly basis, commitment devices play a relevant role.
Uncertainty plus time inconsistency. We have shown that booking and canceling pop-
ular classes can be explained by either uncertainty about period-two attendance cost or by
na¨ıve time inconsistency. Explaining that an unpopular class is booked requires sophisticated
time inconsistency. However, such preferences cannot explain by themselves why a booking
would be canceled. Thus, jointly explaining Observations (1)-(4) requires a combination of
factors. Combining uncertainty with na¨ıve time inconsistency (see row 5 in Table 8) still
cannot explain the booking and canceling of unpopular classes. Thus, only a combination
of uncertainty and sophisticated time inconsistency predicts each of these four observations
in our data. In such a model, popular classes are booked to secure a spot (Observation 1),
unpopular classes are booked as a commitment device (Observation 3), and each booking
is canceled if the attendance cost becomes sufficiently unfavorable (Observations 2 and 4).
36
Of course, not all gym members need to have sophisticated time-inconsistent preferences, al-
though the large number of unpopular class bookings suggests that at least a sizable share
do. As is clear from Table 8, we cannot conclude that some share of gym members are also
na¨ıve about their time inconsistency. Nevertheless, it seems likely that this is the case, given
the large share of cancellations (41 %), which otherwise would have to be fully explained by
uncertainty about the future.
Limited attention. Observations 5 and 6 point more directly towards limited attention.
First, we observe that 13 % of all cancellations (or 6 % of all bookings) in our data are no-
shows. In the model, there is a risk 1 − pi < 1 of forgetting about the class in period two,
which seems to be a natural explanation for no-shows. In the absence of limited attention, one
would never book a class and subsequently not attend (without canceling) because, as argued
earlier, the cancellation costs, cT , are low while the costs of not showing up to a booked class,
cF , are non-trivial (around $3). Thus, we can safely assume that cT < cF , and no-shows
therefore must be the result of limited attention.38 Interestingly, the share of no-shows is the
same across all types of classes. This finding suggests that limited attention is a common
phenomenon and does not depend on the type of class.
Furthermore, we find in the experiment that regular reminders affect decisions regarding
class attendance (Observation 6). In particular, we observe that reminders increase bookings,
decrease the share of cancellations and increase class attendance (where the impact on cancel-
lations is weak and conclusive). A natural approach to incorporating reminders in the model
is to assume that an individual who receives them will no longer suffer from limited attention,
that is, pi = 1. Compared to the situation where pi < 1, this situation leads to predictions
in line with Observations 6. First, if individuals anticipate the reminders (they might do so
after receiving repeated emails), they realize that they will not forget in period two, which
makes booking more attractive and, thus, depending on parameter values, could increase the
number of bookings (as observed). Second, reminders may eliminate no-shows, decreasing the
share of bookings that are not followed by class attendance (as observed). Third, reminders
would increase the total number of class attendances due to both an elimination of no-shows
and to a higher number of drop-in attendances. Thus, each element of Observation 6 can be
explained by limited attention and its elimination through reminders.
38Note that in reality (though not in our model), no-shows can also be caused by extreme present bias.
Imagine that someone books a class and, one hour prior to the class, believes that she will attend. However,
half an hour later, she weighs the direct effort and time costs much more highly relative to the returns and
decides not to attend. Since she can no longer cancel at this point, she will not show up. It is, however,
difficult to imagine that the change in discounting during the final hour is so large that it can explain the 13
% of no-shows out of all cancellations.
37
5 Conclusion
It is often suggested that one of the main reasons for inefficiently low frequencies of exercise
are behavioral explanations, such as limited attention and time-inconsistent preferences. Our
analysis of exercising behavior and planning by a large sample of gym members shows that
both are indeed relevant.
Using a large-scale randomized experiment with email reminders, we find that short and
simple weekly reminders lead to a significant 13 % increase in the number of weekly gym visits.
The fact that reminders are so effective at increasing exercising frequencies suggests that
limited attention is an important hurdle for reaching individual exercise goals. In addition, we
find that gym members suffer from not only limited attention but also present bias. Analyzing
a large dataset on the booking, cancellation and attendance of gym classes, we show that many
members book classes simply to commit themselves to attending the class in the future. This
analysis allows us to conclude that commitment problems, which stem from (sophisticated)
time-inconsistent preferences, are a common phenomenon among gym members.
While our results are based on a sample of gym members (thus, individuals who decided to
purchase a gym membership), we believe that the external validity of our results is relatively
strong. Our large sample contains a diverse set of individuals with both students and employed
members and a wide range of age groups. Since the gym offers many different sports activities,
our results are not based solely on individuals interested in weightlifting but apply to all types
of people interested in regular exercise and sports. Finally, average attendance in our sample
(approximately once per week) is similar to the commonly found attendance rates but well
below the general advice from a health perspective.
Our findings point toward the importance of policies that guide those who are interested
in attaining a higher level of regular exercise. Simple reminders, which essentially have zero
marginal cost, have the potential to stimulate individuals to exercise more by ensuring that
the option to exercise is not forgotten during one’s busy day-to-day life. Our experimental
results show that the impact might even persist after reminders end, potentially due to habit
formation. Additionally, offering help to commit oneself to one’s exercising goals can be very
effective and, as our results imply, would be welcomed by many. For example, one could
consider disallowing cancellation of an exercise class booking in the 24 hours before the class
in order to strengthen the commitment value of such a booking. Since many seem to be aware
of their commitment problems, one option would be to offer such policies on a voluntary basis,
ensuring that they are taken up specifically by those individuals who benefit most from them.
38
References
Acland, D. and Levy, M. R. (2015). Naivete´, projection bias, and habit formation in gym
attendance. Management Science, 61(1):146–160.
Allcott, H. and Rogers, T. (2014). The short-run and long-run effects of behavioral inter-
ventions: Experimental evidence from energy conservation. American Economic Review,
104(10):3003–3037.
Altmann, S. and Traxler, C. (2014). Nudges at the dentist. European Economic Review,
72:19–38.
Apesteguia, J., Funk, P., and Iriberri, N. (2013). Promoting rule compliance in daily-life:
Evidence from a randomized field experiment in the public libraries of Barcelona. European
Economic Review, 64:266–284.
Ashraf, N., Karlan, D., and Yin, W. (2006). Tying odysseus to the mast: Evidence from
a commitment savings product in the philippines. The Quarterly Journal of Economics,
121(2):635–672.
Battigalli, P. and Dufwenberg, M. (2009). Dynamic psychological games. Journal of Economic
Theory, 144(1):1–35.
Beshears, J., Milkman, K. L., and Schwartzstein, J. (2016). Beyond beta-delta: The emerging
economics of personal plans. American Economic Review, 106(5):430–434.
Bhargava, S. and Manoli, D. (2015). Psychological frictions and the incomplete take-up
of social benefits: Evidence from an IRS field experiment. American Economic Review,
105(11):3489–3529.
Bhattacharya, J., Garber, A. M., and Goldhaber-Fiebert, J. D. (2015). Nudges in exercise
commitment contracts: A randomized trial. NBER Working Paper, No. 21406.
Bisin, A. and Hyndman, K. (2014). Present-bias, procrastination and deadlines in a field
experiment. NBER Working Paper, No. 19874.
Bornstein, B., Tomkins, A., Neeley, E., Herian, M., and Hamm, J. (2012). Reducing courts’
failure-to-appear rate by written reminders. Psychology, Public Policy, and Law, 19(1):70–
80.
Calzolari, G. and Nardotto, M. (2017). Effective reminders. Management Science, 63(9):2915–
2932.
39
Castleman, B. L. and Page, L. C. (2015). Summer nudging: Can personalized text messages
and peer mentor outreach increase college going among low-income high school graduates?
Journal of Economic Behavior & Organization, 115:144–160.
Charness, G. and Gneezy, U. (2009). Incentives to exercise. Econometrica, 77(3):909–931.
Chen, Y., Lu, F., and Zhang, J. (2017). Social comparisons, status and driving behavior.
Journal of Public Economics, 155:11 – 20.
Dai, H., Milkman, K. L., and Riis, J. (2014). The fresh start effect: Temporal landmarks
motivate aspirational behavior. Management Science, 60(10):2563–2582.
Damgaard, M. T. and Gravert, C. (2018). The hidden costs of nudging: Experimental evidence
from reminders in fundraising. Journal of Public Economics, 157:15–26.
Dean, M., Kıbrıs, O¨., and Masatlioglu, Y. (2017). Limited attention and status quo bias.
Journal of Economic Theory, 169:93–127.
DellaVigna, S. and Malmendier, U. (2006). Paying not to go to the gym. The American
Economic Review, 96(3):694–719.
Ericson, K. M. (2017). On the interaction of memory and procrastination: Implications
for reminders, deadlines, and empirical estimation. Journal of the European Economic
Association, 15.
Fellner, G., Sausgruber, R., and Traxler, C. (2013). Testing enforcement strategies in the
field: Threat, moral appeal and social information. Journal of the European Economic
Association, 11(3):634–660.
Frederick, S., Loewenstein, G., and O’Donoghue, T. (2002). Time discounting and time
preference: A critical review. Journal of Economic Literature, 40(2):351–401.
Goldhaber-Fiebert, J. D., Blumenkranz, E., and Garber, A. M. (2010). Committing to exer-
cise: Contract design for virtuous habit formation. NBER Working Paper, No. 16624.
Guthold, R., Stevens, G. A., Riley, L. M., and Bull, F. C. (2018). Worldwide trends in
insufficient physical activity from 2001 to 2016: a pooled analysis of 358 population-based
surveys with 19 million participants. The Lancet Global Health.
Himmler, O., Ja¨ckle, R., and Weinschenk, P. (forthcoming). Soft commitments, reminders
and academic performance. American Economic Journal: Applied Economics.
Hull, C. L. (1943). Principles of Behavior: An Introduction to Behavior Theory. D. Appleton-
Century Company, Incorporated, New York, London.
40
Hull, C. L. (1951). Essentials of Behavior. Published for the Institute of Human Relations
by Yale University Press, New Haven.
Karlan, D., McConnell, M., Mullainathan, S., and Zinman, J. (2016a). Getting to the top of
mind: How reminders increase saving. Management Science, 62(12):3393–3411.
Karlan, D., Morten, M., and Zinman, J. (2016b). A personal touch in text messaging can
improve microloan repayment. Behavioral Science & Policy, 1(2):25–31.
Kool, J., de Bie, R., Oesch, P., Knusel, O., van den Brandt, P., and Bachman, S. (2004).
Exercise reduces sick leave in patients with non-acute non-specific low back pain: A meta-
analysis. Journal of Rehabilitation Medicine, 36(2):49–62.
Laibson, D. (1997). Golden eggs and hyperbolic discounting. The Quarterly Journal of
Economics, 112(2):443–478.
Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., and Wardle, J. (2010). How are habits
formed: Modelling habit formation in the real world. European Journal of Social Psychology,
40(6):998–1009.
Nolan, J. M., Schultz, P. W., Cialdini, R. B., Goldstein, N. J., and Griskevicius, V. (2008).
Normative social influence is underdetected. Personality and Social Psychology Bulletin,
34(7):913–923.
O’Donoghue, T. and Rabin, M. (1999a). Doing it now or later. American Economic Review,
89(1):103–124.
O’Donoghue, T. and Rabin, M. (1999b). Incentives for procrastinators. The Quarterly Journal
of Economics, 114(3):769–816.
O’Donoghue, T. and Rabin, M. (2001). Choice and procrastination. The Quarterly Journal
of Economics, 116(1):121–160.
Phelps, E. and Pollak, R. A. (1968). On second-best national saving and game-equilibrium
growth. Review of Economic Studies, 35(2):185–199.
Royer, H., Stehr, M., and Sydnor, J. (2015). Incentives, commitments, and habit formation
in exercise: Evidence from a field experiment with workers at a Fortune-500 company.
American Economic Journal: Applied Economics, 7(3):51–84.
Schultz, P. W., Nolan, J. M., Cialdini, R. B., Goldstein, N. J., and Griskevicius, V. (2007). The
constructive, destructive, and reconstructive power of social norms. Psychological Science,
18(5):429–434.
41
Sonntag, A. and Zizzo, D. J. (2015). On reminder effects, drop-outs and dominance: Evidence
from an online experiment on charitable giving. PLoS ONE, 10(8):1–17.
Taubinsky, D. (2014). From intentions to actions: A model and experimental evidence of
inattentive choice. Technical report, Department of Economics, Harvard University.
Toussaert, S. (2016). Connecting commitment to self-control: a field experiment with partic-
ipants in a weight loss challenge. Mimeo.
Van Amelsvoort, L. G., Spigt, M. G., Swaen, G. M., and Kant, I. (2006). Leisure time
physical activity and sickness absenteeism; a prospective study. Occupational Medicine,
56(3):210–212.
Verplanken, B. (2006). Beyond frequency: Habit as mental construct. British Journal of
Social Psychology, 45(3):639–656.
Wood, W. and Neal, D. T. (2007). A new look at habits and the habit-goal interface. Psy-
chological Review, 114(4):843–863.
42
A Appendix: Additional descriptive statistics
Figure 6: Visits per class
0 1,000 2,000 3,000
frequency
zorboll
gymclass
christmas
mindfulness
fight
gympa
q−performance
spökboll
triggerstretch
powercage
cirkelgym
indoorwalking
volleyboll
innebandy
bodystep
basket
fys
funktionellträning
powercircuit
tabata
fotboll
vattenträning
pilates
badminton
bodybalance
bodycombat
trx
dansklass
crossfit
cxworx
yoga
krafttag
cykel
Figure 7: Distribution of dates of buying contract for control and treatment group
0
50
10
0
15
0
0
50
10
0
15
0
1 J
ul 2
016
1 A
ug 
201
6
1 S
ep 
201
6
1 O
ct 2
016
1 N
ov 
201
6
1 D
ec 
201
6
Control group
Treatment group
Fr
eq
ue
nc
y
Date of buying contract
43
Figure 8: Comparing control and treatment: pre-experiment gym attendance
0 1,000 2,000 3,000frequency
Facility 4
Facility 2
Facility 3
Facility 1
Control group Treatment group
(a) Free training per facility
0 100 200 300 400frequency
dance
ballsport
core
cycle/run/swim
body and mind
endurance and strength
Control group Treatment group
(b) Group classes
0
10
20
30
per
cen
t
0 1 2 3 4 5 6 7 8
Control group Treatment group
(c) Distribution of weekly free training visits
0
10
20
30
40
per
cen
t
1 2 3 4 5 6
Control group Treatment group
(d) Distribution of weekly class training visits, exclud-
ing zeros (90.9 % of all observations)
44
B Appendix: Email reminders
Table 10: Calendar schedule email reminders
December 2016 January 2017 February 2017 March 2017 April 2017
15 (general email) 9 (reminder) 6 (reminder) 7 (reminder) 3 (reminder)
15 (reminder) 13 (reminder) 13 (reminder) 9 (reminder)
24 (reminder) 19 (reminder) 21 (reminder)
[26 (newsletter)] 26 (reminder)
Note: Reminders were always sent at the beginning of a week. To vary the day of dispatch, we randomized
between Sundays, Mondays and Tuesdays. The newsletter on January 26 was sent to all gym members. The
lack of a reminder in the fourth week of February was due to an expected newsletter, which was, eventually,
not sent out by the gym.
Figure 9: Example of an email reminder
45
List of email reminders (in chronological order):
1. Did you know that we offer more than 200 classes per week? Drop in any time to work
out, or sign up for a class on www.fysiken.nu/en/book/. We hope to see you soon!
2. At Fysiken, we offer a huge range of classes and training facilities. You can find the
schedule and sign up for classes at www.fysiken.nu/en/book/. Get inspired! We hope to
see you soon!
3. Did you know that you can train at Fysiken 24/7? Our facility at Gibraltargatan is
now open 24 hours a day. Also our facilities Lindholmen and Kaserntorget have long
opening hours. See www.fysiken.nu/en/ for more information. See you soon!
4. Did you know that research has shown that working out makes you feel more energetic?
Go to our website and get inspired for your next energizing workout: www.fysiken.nu/en/.
We hope to see you soon.
5. Did you know that regular exerce has huge health benefits? Reap the benefits of regular
exercise and sign up for one of Fysiken’s classes at www.fysiken.nu/en/book/. We hope
to see you soon!
6. Research has shown that regular workout reduces your stress levels significantly. Try out
one of the many classes we offer. Look at www.fysiken.nu/en/ for more information.
See you soon!
7. Fysiken offers a huge variety of training activities. For example, you can try out var-
ious ball sports, climbing activities and water training. Find more information on
www.fysiken.nu/en/ and get inspired! We hope to see you soon!
8. Did you know that exercising helps reduce stress and improve concentration? Body and
mind are connected - book your yoga, Pilates or Mindfulness class at www.fysiken.nu/en/
book/. Welcome!
9. We all have problems with committing ourselves to a strict workout schedule. Did you
know that simply putting reminders in your calendar might help you to better stick to
your exercise goals? Inspired? Take a look at www.fysiken.nu/en/. We hope to see you
soon.
10. Research has shown that regular exercise improves the quality of your sleep. Improve
your life now and take a look at our wide range of gym classes at www.fysiken.nu/en/.
We hope to see you soon!
46
11. Working out together is more fun - and it helps you commit to going to the gym.
Why don’t you contact a friend and schedule a training session right away? Go to
www.fysiken.nu/en. We hope to see you soon!
12. At Fysiken, we offer a huge range of classes and training facilities. You can find the
schedule and sign up for classes at www.fysiken.nu/en/book/. Get inspired!
47
C Appendix: Extended results
C.1 Regressions on weekly gym attendance
Table 11: Regressions on weekly gym attendance (linear models)
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.12*** 0.029** 0.091*** 0.075* 0.028 0.047
(0.031) (0.014) (0.028) (0.043) (0.019) (0.039)
Experiment (post) 0.11*** 0.014 0.093*** 0.049 0.0053 0.044
(0.034) (0.013) (0.032) (0.040) (0.013) (0.037)
Recurring member 0.19*** 0.022 0.16***
(0.042) (0.016) (0.039)
Male 0.19*** -0.18*** 0.37***
(0.039) (0.017) (0.035)
Student 0.0030 0.0041 -0.0011
(0.047) (0.019) (0.043)
Daytime card -0.30*** -0.14*** -0.16**
(0.077) (0.030) (0.071)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
Mean value dep. var. 0.95 0.16 0.79 0.96 0.17 0.79
Percent effect (during) 13 18 12 8 17 6
N 73890 73890 73890 55825 55825 55825
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. Columns (4)-(6) also include
seven age dummies which are not reported.
48
Figure 10: Weekly gym attendance: group means (with 95 % confidence interval)
.
6
.
7
.
8
.
9
1
1.1
1.2
We
ekl
y a
tten
dan
ce
  
Before experiment
.
6
.
7
.
8
.
9
1
1.1
1.2
We
ekl
y a
tten
dan
ce
  
During experiment
.
6
.
7
.
8
.
9
1
1.1
1.2
We
ekl
y a
tten
dan
ce
  
After experiment
Control group Treatment group
(a) Total visits (DiD estimate during: 0.12, DiD estimate post: 0.11)
0
.
05
.
1
.
15
.
2
.
25
.
3
We
ekl
y c
las
stra
inin
g
  
Before experiment
0
.
05
.
1
.
15
.
2
.
25
.
3
We
ekl
y c
las
stra
inin
g
  
During experiment
0
.
05
.
1
.
15
.
2
.
25
.
3
We
ekl
y c
las
stra
inin
g
  
After experiment
Control group Treatment group
(b) Class training (DiD estimate during: 0.03 : DiD estimate post: 0.01)
.
6
.
7
.
8
.
9
1
1.1
We
ekl
y fr
eet
rain
ing
  
Before experiment
.
6
.
7
.
8
.
9
1
1.1
We
ekl
y fr
eet
rain
ing
  
During experiment
.
6
.
7
.
8
.
9
1
1.1
We
ekl
y fr
eet
rain
ing
  
After experiment
Control group Treatment group
(c) Free training (DiD estimate during: 0.09, DiD estimate post: 0.09)
49
Table 12: Regressions on weekly gym attendance: excluding Christmas holidays (Pois-
son models)
Total visits Class training Free training
(1) (2) (3) (4) (5) (6)
Experiment (during) 0.13*** 0.11*** 0.19** 0.15* 0.11*** 0.10***
(0.034) (0.037) (0.090) (0.096) (0.036) (0.040)
Experiment (post) 0.12*** 0.11** 0.12 0.090 0.12*** 0.11**
(0.046) (0.049) (0.11) (0.12) (0.050) (0.054)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes yes yes yes
Include pre-period yes yes yes yes yes yes
Exclude weeks: 52 51,52,01 52 51,52,01 52 51,52,01
N 64641 60102 27289 25218 62988 58509
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. We
report exp(β) − 1, which is the percentage effect. Note that the number of observations varies,
because individuals with zero attendance in all time periods drop out in Poisson regressions
with fixed effects. Columns (1), (3) and (5) exclude the last week of 2016 (week 52), in which
attendance is very low. Columns (2), (4) and (6) in addition exclude week 51 (2016) and week 1
(2017) in which attendance is also below the trend. See Figure 4 for clarification.
50
Table 13: Regressions on weekly gym attendance: subsample with longer pre-experiment period (Poisson
models)
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.034 0.14* 0.0084 0.085*** 0.16*** 0.066***
(0.032) (0.080) (0.034) (0.017) (0.042) (0.019)
Experiment (post) 0.064 0.061 0.064 0.11*** 0.044 0.12***
(0.048) (0.11) (0.052) (0.025) (0.058) (0.028)
Recurring member 0.19*** 0.092*** 0.22***
(0.016) (0.035) (0.019)
Male 0.19*** -0.71*** 0.62***
(0.016) (0.010) (0.025)
Student -0.030* 0.071* -0.045***
(0.015) (0.040) (0.016)
Daytime card -0.31*** -0.69*** -0.21***
(0.025) (0.039) (0.030)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
N 63070 29190 61635 42550 42550 42550
* p < 0.10, ** p < 0.05, *** p < 0.01. Robust standard errors in parentheses. We report exp(β)−1, which is the percentage
effect. Note that the number of observations varies, because individuals with zero attendance in all time periods drop out
in Poisson regressions with fixed effects.
51
Table 14: Regressions on weekly gym attendance (Poisson models): post-experiment effect w/o week 22 2017
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.13*** 0.17** 0.12*** 0.072* 0.12 0.059
(0.035) (0.090) (0.038) (0.040) (0.097) (0.044)
Experiment (post) 0.11** 0.11 0.11** 0.047 0.040 0.048
(0.046) (0.11) (0.050) (0.052) (0.12) (0.058)
Recurring member 0.20*** 0.14 0.22***
(0.049) (0.11) (0.056)
Male 0.21*** -0.66*** 0.61***
(0.047) (0.033) (0.075)
Student 0.00055 0.058 -0.0053
(0.048) (0.13) (0.052)
Daytime card -0.31*** -0.70*** -0.20**
(0.074) (0.12) (0.087)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
N 64670 27347 63017 53592 53592 53592
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. We report exp(β)− 1, which
is the percentage effect. Seven age dummies and six gym location dummies are included but not reported in columns (4)-(6).
Note that the number of observations in columns (1)-(3) vary because individuals with zero attendance in all time periods
drop out in Poisson regressions with fixed effects.
52
Table 15: Regressions on weekly gym attendance (Poisson models): post-experiment effect until week 21
2017
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.13*** 0.17** 0.12*** 0.069* 0.12 0.056
(0.035) (0.090) (0.038) (0.040) (0.097) (0.044)
Experiment (post) 0.11** 0.14 0.11** 0.052 0.068 0.049
(0.047) (0.12) (0.051) (0.052) (0.12) (0.058)
Recurring member 0.19*** 0.13 0.20***
(0.048) (0.10) (0.055)
Male 0.18*** -0.67*** 0.60***
(0.046) (0.032) (0.073)
Student 0.016 0.072 0.010
(0.048) (0.13) (0.052)
Daytime card -0.31*** -0.71*** -0.19**
(0.074) (0.12) (0.086)
Week & duration FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
N 55550 23300 54000 44660 44660 44660
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. We report exp(β)− 1, which
is the percentage effect. Seven age dummies and six gym location dummies are included but not reported in columns (4)-(6).
Note that the number of observations in columns (1)-(3) vary because individuals with zero attendance in all time periods
drop out in Poisson regressions with fixed effects.
53
C.2 Regressions on daily gym attendance
Figure 11: Gym attendance: daily data
0
.
05
.
1
.
15
.
2
.
25
.
3
Tot
al v
isits
 pe
r da
y
01dec2016 01jan2017 01feb2017 01mar2017 01apr2017 01may2017 01jun2017 01jul2017
Control group Treatment group
(a) Total visits (daily)
0
.
01
.
02
.
03
.
04
.
05
.
06
.
07
Cla
ss t
rain
ing
 pe
r da
y
01dec2016 01jan2017 01feb2017 01mar2017 01apr2017 01may2017 01jun2017 01jul2017
Control group Treatment group
(b) Class training (daily)
0
.
05
.
1
.
15
.
2
.
25
.
3
Fre
e tr
ain
ing
 pe
r da
y
01dec2016 01jan2017 01feb2017 01mar2017 01apr2017 01may2017 01jun2017 01jul2017
Control group Treatment group
(c) Free training (daily)
54
Table 16: Regressions on daily gym attendance
DiD model Simple difference
(1) (2) (3) (4) (5) (6)
Visits Class training Free training Visits Class training Free training
Experiment (during) 0.014*** 0.0037** 0.011*** 0.010* 0.0038 0.0063
(0.0042) (0.0019) (0.0037) (0.0061) (0.0027) (0.0055)
Experiment (post) 0.012** 0.0030 0.0090* 0.0070 0.0024 0.0046
(0.0054) (0.0022) (0.0049) (0.0065) (0.0026) (0.0061)
Recurring member 0.025*** 0.0024 0.023***
(0.0063) (0.0027) (0.0058)
Male 0.024*** -0.032*** 0.056***
(0.0059) (0.0028) (0.0054)
Student 0.0017 0.00075 0.00091
(0.0072) (0.0032) (0.0066)
Daytime card -0.045*** -0.025*** -0.020*
(0.012) (0.0046) (0.011)
Sunny day -0.0030* -0.00083 -0.0022
(0.0016) (0.00069) (0.0015)
Rainy day -0.0081*** -0.0026*** -0.0054***
(0.0020) (0.00090) (0.0018)
Week & duration FE yes yes yes yes yes yes
Day of week FE yes yes yes yes yes yes
Individual FE yes yes yes no no no
Include pre-period yes yes yes no no no
Mean value dep. var. 0.15 0.026 0.12 0.15 0.028 0.12
Percent effect 10 14 9 7 14 5
N 384228 384228 384228 259028 259028 259028
* p < 0.10, ** p < 0.05, *** p < 0.01. Standard errors clustered by individual in parentheses. Seven age dummies and six gym
location dummies are included but not reported in columns (4)-(6). A sunny day is a day with at least 5 hours of sunshine, and
a rainy day is a day with at least 3 mm of rain.
C.3 Class popularity: Details of categorization and some more results
As we discuss in Section 2.1, over 200 classes are scheduled each week, with 37 main class
types (and there is further variation with respect to class length, focus, instructor, instruction
language, etc.). While some classes are very popular, with all places being regularly taken,
other classes are less popular (or are simply offered at higher capacity) and are rarely fully
booked. Below we describe how we split classes by popularity into three categories (unpopular,
fairly popular and popular).
The first step to define a class. Simply defining classes by their activity (e.g., Yoga, Pilates,
etc.) and their scheduled time (e.g. Wednesday at 8 pm) will be to narrow. For example, if
someone knows that ‘30 minutes Yoga’ on Wednesday at 7 pm is very popular, she will most
likely assume that ‘45 minutes Yoga’ on Wednesday at 8 pm is also very popular (even if she
55
never attended this particular class before). Thus, we need to define a class in a rather broad
sense, such that similar classes at similar times fall into the same category. For this purpose,
we first categorize each class into 37 main activities (these are the activity types, see footnote
7). Second, we assign each class to one of 14 time slots, defined by the day of the week
interacted with ‘daytime’ or ‘evening’. A particular class is thus defined as the activity type
and the time slot. Examples of a class would be ‘Yoga on Wednesday evening’ or ‘Football
on Saturday daytime’. This approach also solves the problem that the exact scheduled time
or name of a class might change during the course of one year. As a result, we observe 238
distinct classes, with between 5 and 5,617 bookings for each class. For 73 % of these classes,
we observe at least 100 bookings during the 12 months.
Next, we categorize the 238 classes into three categories, based on how popular they are.
We measure popularity of a class by computing the share of bookings that result in a spot on
the waiting list. A share of zero means that the class is never fully booked, which is the case
for 30 % of all classes. A share of 0.05 means that one in 20 bookings ends up on the waiting
list. If a class has 20 available spots, the class is typically fully booked. Of course, due to
cancellations there is a reasonable chance of still being admitted to the class. However, we
argue that an individual’s perception of class popularity will be affected by (initially) ending
up on the waiting list, and thus we opted for this particular measure. For 40 % of the classes
the share is at least 0.05, while 22 % even have a share of more than 0.2 (indicating long
waiting lists). We categorize the classes into unpopular (share less than 0.005), fairly popular
(share between 0.005 and 0.05) and popular (share more than 0.05).
Figure 12: Cumulative distribution of the timing of class bookings by popularity of the class
0
.
2
.
4
.
6
.
8
1
Cu
m.
 de
nsi
ty b
oo
kin
g d
ay
−8 −6 −4 −2 0
Booking day
unpopular fairly popular
popular
56
Figure 13: Cumulative distribution of the timing of canceling by popularity of the class
0
.
2
.
4
.
6
.
8
1
Cu
m.
 de
nsi
ty c
an
cel
ing
 da
y
−8 −6 −4 −2 0
Canceling day
unpopular fairly popular
popular
Table 17: Characteristics of attenders of classes by popularity
Mean value P-values
(1) (2) (3) t-testa
Unpopular Fairly Popular
classes popular classes
classes (1) vs (2) (1) vs (3) (2) vs (3)
Age 32.76 32.03 31.79 0.23 0.15 0.61
Male 0.30 0.17 0.31 0.00 0.66 0.00
Student 0.52 0.58 0.53 0.01 0.61 0.02
Recurring member 0.54 0.51 0.51 0.22 0.27 0.93
Membership type:
Multi Card 0.50 0.47 0.53 0.29 0.30 0.01
Facility 1 0.41 0.42 0.39 0.66 0.46 0.11
Facility 2 0.07 0.09 0.07 0.10 0.96 0.04
Facility 3 0.01 0.01 0.00 0.61 0.05 0.04
Facility 4 0.01 0.01 0.01 0.16 0.93 0.17
Daytime card 0.02 0.02 0.01 0.75 0.05 0.02
Observations 7713 11004 29493
We regress the individual characteristic on dummies for the popularity of the class and include all partici-
pants from only two categories of classes (for example all unpopular classes and all fairly popular classes).
By clustering the standard error on the individual level, we take into account that each individual might
appear in the sample many times as a participant.
aThe reported p-values follow from a standard t-test for statistical significance of one of the popularity
dummies relative to the other one.
57