Journal of Health Economics 58 (2018) 228–252

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Increasing breast-cancer screening uptake: A randomized

controlled experiment
Léontine Goldzahl a,∗ , Guillaume Hollard b , Florence Jusot a
a
PSL, Université Paris-Dauphine, Leda-Legos, France
b
CNRS and Ecole Polytechnique, France

a r t i c l e i n f o a b s t r a c t

Article history: Early screening increases the likelihood of detecting cancer, thereby improving survival rates. National
Received 31 March 2017 screening programs have been established in which eligible women receive a letter containing a voucher
Received in revised form 5 December 2017 for a free screening. Even so, mammography use is often considered as remaining too low. We test
Accepted 12 December 2017
four behavioral interventions in a large-scale randomized experiment involving 26,495 women. Our
Available online 31 January 2018
main assumption is that, due to biases in decision-making, women may be sensitive to the content and
presentation of the invitation letter they receive. None of our treatments had any significant impact on
JEL classification:
mammography use. Sub-sample analysis suggests that this lack of a significant impact holds also for
D03
I18
women invited for the first time and low-income women.
© 2017 Published by Elsevier B.V.
Keywords:
Cancer screening
Randomized controlled experiment
Behavioral interventions

1. Introduction we consider four treatments inspired by behavioral economics and

Early screening increases the likelihood of detecting cancer
when it is more localized, thereby increasing survival rates. To
prevent asymptomatic breast cancer from progressing to metas-
tasis, national-screening programs have been established in many
countries. Eligible women receive an invitation letter containing a
voucher for a free screening (namely a mammogram). Women are
free to decide whether to undertake this exam. Despite this scheme,
many policy makers and researchers consider that mammography
use remains too low (Altobelli and Lattanzi, 2014) and that there
are socioeconomic disparities in mammography use (Carrieri and
Wübker, 2013; Devaux, 2015; Jusot et al., 2012; Sicsic and Franc,
2014). Improving mammography use thus remains an important
policy objective. The present paper uses a large-scale Randomized
Controlled Trial (RCT) conducted in two French départements.1
In addition to a control group who receive the status-quo letter,
∗ Corresponding author.
E-mail addresses: leontine.goldzahl@dauphine.fr (L. Goldzahl),
guillaume.hollard@polytechnique.edu (G. Hollard).
1
In the geographical organization of France, the département is an administrative
district. There are 95 départements in metropolitan France, with an average land
area of 5,666 square kilometers (three and half times the median land area of US
Counties). In terms of population, départements in 2015 varied in size from 76,000
to 2.6 million, with an average population of 670,000.
psychology.2 Each treatment changes a different aspect of the invi-
tation letter: (1) a new logo on the envelope, (2) patient-approved
letter content, (3) a combination of the two previous treatments,
and (4) information on the number of women receiving mammo-
grams in the recipient’s area of residence.
Our behavioral interventions contrast with most previous
approaches, which either provide extra information (e.g. adding
a leaflet to the letters) or financial incentives (e.g. a cash reward for
those who take the medical test). One desirable feature of the type
of interventions analyzed here is that changing the letters is almost
costless. In addition, our behavioral interventions leave individuals
free to choose what they believe is best for themselves. It is impor-
tant to note that mammograms are not entirely without risk, and
that some women may choose not to take them for good reasons
(Pace and Keating, 2014).
Our main hypothesis is that, due to variability or biases in
decision-making, women may be sensitive to the content and pre-
sentation of the invitation letter they receive.
2
Behavioral interventions are inspired by behavioral economics. They include
nudges (Thaler and Sunstein, 2008), but also refer to the provision of particular infor-
mation using behavioral insights (for example, providing new information about the
behavior of others). See below for a more detailed discussion.

https://doi.org/10.1016/j.jhealeco.2017.12.004
0167-6296/© 2017 Published by Elsevier B.V.
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Fig. 1. Organized screening uptake rate since the implementation of the national program.
Note: The Eure and Seine-Maritime are the two départements in which we conducted the RCT.
Despite having a large sample of N = 26,495 women, none of
our four behavioral interventions was found to have any signifi-
cant impact on mammography use, where we know that our design
does not allow us to detect effects under 2.7%. We then ran com-
plementary analyses to better understand this negative result: even
women who received their invitation letter for the first time do not
seem to be sensitive to these treatment effects.
Although we find no evidence of an average effect, it is possible
that uptake rates change in particular subgroups and for particular
outcomes (e.g. across social groups and screening inside or outside
of the national program). However, we found no evidence of this.
2. Literature review
In the late 1990s and early 2000s, national programs were intro-
duced in many developed countries to increase mammography
use. The general principle in most countries is to send a letter
every two years containing an invitation for a free mammogram.
In France, the organized screening program is carried out at the
département level.3 These organized screening programs increased
breast-cancer screening by 17% in Europe (Carrieri and Wuebker,
2016). However, uptake rates often seemed to reach a plateau. As
shown in Fig. 1, the French rate has been stable at 52%4 for the last
eight years (as of 2016), as is the rate in the two départements in
which the RCT was carried out. The current experiment took place
in a context in which a number of attempts have been made to
increase take-up rates, but with only limited success. To the best of
our knowledge, our work here is the first attempt to test behavioral
interventions on cancer screening in France.
2.1. Nudges and behavioral interventions: definitions
Policy makers often want to change behavior. Traditional pub-
lic policies rely on two main channels to do so: information and
incentives. By providing easy access to relevant information and
3
The national screening program is financed at the national level, but the pro-
gram is managed at the département level. The managing structure is typically a
non-profit organization, which is in charge of sending out the invitation letters and
double-checking the results via a centralized département-level second reading of
the mammograms.
4
This figure is the uptake rate in the national screening program. The French
health authority (Haute Autorité de Santé) evaluated the uptake rate outside of the
national program to be 10% of the eligible population in 2008.
229
reducing the cost (or increasing the benefits) of adopting the rec-
ommended behavior, these aim to make individuals reconsider the
trade-off between a decision’s costs and benefits. A third channel
has been popularized by Thaler and Sunstein (2008), under the
name of nudges. A nudge intervention takes advantage of cognitive
or psychological biases (status quo and hyperbolic discounting, for
example) by changing the architecture of choice in order to influ-
ence behavior. A simple nudge example is to put healthy foods in
school cafeterias at eye level, and less-healthy junk food in harder-
to-reach places. Individuals are then not prevented from eating
whatever they want, but the re-arrangement of food choices leads
to less junk food and more healthy food being chosen. As can be
seen in this example, a typical nudge intervention is an inexpensive
tool to change decision-making, while respecting the individual’s
freedom to choose what they consider best for themselves. Policy
makers have devoted particular attention to nudges, as attested by
the creation of the Behavioral Insights Team in the UK Government
and the US “Nudge Unit”.
A pure nudge would change decisions without adding any new
information. Nudges here have an effect as they change the way in
which individuals process information. Individuals can, for exam-
ple, change from fast and automatic decisions to slow and explicit
ones, as suggested by Kahneman (2003), or stick to the status quo. It
is however difficult in practice to draw a clear line between nudges
and information provision based on behavioral insights. Provid-
ing information on what others do will add new information but
also affects decision-making. As will be explained below, our treat-
ments are certainly inspired by behavioral economics but whether
they qualify as nudges is open to debate, depending on the defini-
tion used. We consequently prefer to refer to them as behavioral
interventions, in a broad sense, rather than nudges.
2.2. Why use behavioral interventions (and not incentives or
information)?
We can classify the attempts made to increase mammography
use according to whether they provided financial incentives or
more information.5 As will be explained, incentives and informa-
tion did not provide the expected increase in uptake rates, opening
the road for behavioral interventions.
5
For more-extensive reviews see Galizzi (2012, 2014).
230 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Incentives. A number of contributions have targeted finan-
cial incentives rewarding cancer-screening uptake. Merrick et al.
(2015) analyze breast-cancer screening and find no difference in
mammography use for fixed and random payments compared to
the control group. With a very large sample (N = 145,467), and
focusing on low-income women, Slater et al. (2005) found a small
effect of only 0.75% for a $10 reward.6 In general, the results for
financial incentives and cancer screening are mixed, suggesting
at best a small effect at a rather high cost. Another intervention
specifically addressing transportation costs to go to see radiolo-
gists involves mobile mammography-screening programs (a truck
with mammogram materials, called “Mammobiles”, which goes to
remote areas). Vallée (2016) reports that Mammobiles are expen-
sive, have technical limitations and do not seem to increase the
uptake rate. Transaction costs do not then seem to hinder uptake.
Information. The information channel was explored in two
large-scale studies in England and France (Trials 1 and 2 of Wardle
et al. (2016) and Bourmaud et al. (2016)), for bowel and breast can-
cer respectively. The intervention consisted in adding a leaflet to the
existing invitation letter that summarized key information about
cancer. Despite the large sample size (N = 163,525 and N = 149,871
in Wardle et al. (2016), and N = 15,844 in Bourmaud et al. (2016)),
the provision of more information did not increase screening rates.7
This result led us first to concentrate our efforts on the design of
invitation letters (rather than adding an extra leaflet) and, second,
to avoid any large increase in the total quantity of written material.
2.3. Designing invitation letters: interest and limitations
One key issue with interventions on letters is that we cannot
be sure that the individuals to whom they are addressed actually
open them. We only know that the women who were screened
within the program used the voucher contained in the envelope.
This suggests that treatments signaling that the envelope contains
important information may be effective. Our first treatment then
added the logo of the three main National Health Insurance funds
to the envelope containing the invitation letter, to signal the impor-
tance of its contents. In behavioral terms, these signals are designed
to attract attention. We also expect a “messenger” effect, i.e. the fact
that a message that comes from a socially-recognized body will be
perceived as signal of the reliability of the information (Webb and
Sheeran, 2006).
Another possibility is to use letters to reduce “psychological
frictions”, such as minor procedural hassle and confusion. We in
particular want to avoid conveying any confusion with the descrip-
tion of the mammogram procedure. This is very similar to Bhargava
and Manoli (2015), who compare different letters that aim to
increase the take up of social benefits in California. We thus ran a
targeted experiment to come up with a letter explaining, as clearly
as possible, the benefits and procedures regarding mammograms.
The behavioral effects we target here are complexity and limited
attention.
6
This note reviews evidence on bowel-cancer screening, as this also involves
an organized screening program, although the screening procedure and financial-
incentive interventions differ from those for breast cancer. Gupta et al. (2016) test
the relative effects of $5 and $10 Walmart vouchers and no voucher on fecal blood
screening. They find no substantial differences in a dataset with 1000 individuals per
treatment. The same lack of effect on bowel-cancer screening is found by Kullgren
et al. (2014) for various amounts ($5, $10 and $20, with N = 713). However, a lottery
incentive (i.e. a random payment) increased test completion by 19% (p-value <0.01)
as compared to the control group in their study.
7
Bourmaud et al. (2016) find a 2% decrease in uptake rate but this decrease is
driven by a 50% drop in screening rate in a specific geographical area. We interpret
this result as a null result.
Our next treatment relates to social psychology and community-
based approaches.8 Peer behavior influences the individual’s own
behavior by making social norms more salient. The effect is sum-
marized by the following citation “If everyone is doing it, it must be
a sensible thing to do” from Keizer et al. (2008). Social-norm inter-
ventions have been tested in various domains9 and found to be
more effective when individuals identify with the reference group,
i.e. when they consider the group members as peers.10 Recent field
experiments have however highlighted that social-norm inter-
ventions may produce the opposite effect to that intended, a
so-called “boomerang” effect (Schultz et al., 2007; Beshears et al.,
2015). Boomerang effects may come about due to negative belief
update: individuals learn that the activity being promoted is less
widespread than they previously thought. As current screening
rates are around 50%, we tried to avoid boomerang effects by refer-
ring instead to the number of women who were screened in the
past year.
2.4. Other possible behavioral interventions
Other interventions have been attempted, although either the
literature has not reached a consensus on their efficacy or they were
not feasible in our context. Some work has changed the default
option of the appointment system of letting patients call and set
the time of their appointment. Individual status-quo bias implies
that individuals will stick to the proposed default option (date and
time for their appointment are already set). However, evidence
on the effect of changing the default option in invitation letters
is mixed: Narula et al. (2014) find a fall in bowel-cancer atten-
dance and Segnan et al. (1997) a rise in cervical- and breast-cancer
screenings. Another intervention is to have tailored letters, which
use information on the receiver (such as any perceived barriers to
attending screening) and adjust the letter content accordingly.11
However, there is no strong empirical evidence for an effect of tai-
lored letters. Another possibility is to add the signature of the GP, or
some other authority or celebrity, to the invitation letter.12 While
only the GP’s signature seems to increase screening uptake, this
intervention requires the sharing of information between the GP
and the managing structure in charge of the screening program.
This is not currently possible in France.
Last, other biases in human decision can be addressed by behav-
ioral interventions. For example, procrastination and information
aversion are two well-documented biases that may explain why
some individuals prefer to postpone screening or even not to
know the result of their screening tests (Ganguly and Tasoff, 2017;
Golman et al., 2017). These biases were not considered in our work
here for practical reasons.
8
For instance, the managing structure can set up a network of informed women
who had already been screened to spread information about breast-cancer screening
in their community.
9
Among others: restaurant menu choices (Cai et al., 2009), taste in music
(Salganik et al., 2006), pro-environmental behavior (Griskevicius et al., 2008; Ayres
et al., 2013), contributions to retirement plans (Beshears et al., 2015) and voting
participation (Gerber and Rogers, 2009).
10
For example, the normative messages in hotel bathrooms increased towel reuse
of occupants by 4% when the reference group changed from “most guests at the
hotel reuse their towel” to “most previous occupants of the room reused their
towel” (Cialdini, 2003). In the same vein, and closer to our context here, Bronchetti
et al. (2015) used a controlled field experiment to test interventions through an e-
mail aiming to increase flu vaccination among American students. The peer-effect
intervention comes from a peer endorsing the flu vaccine: the peer was a refer-
ence individual on campus (tutor or athlete) who sent the e-mail directly to the
participants.
11
See Vernon et al. (2008), Bodurtha et al. (2009), Lipkus et al. (2005) and Myers
et al. (2008).
12
SeeWardle et al. (2016), Segnan et al. (1997), and Senore et al. (2015). The
signature can also come from various Health Authorities (Stein et al., 2005).
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Fig. 2. Timeline of the experiment and data collection.
3. Experimental design
After providing some background information on cancer screen-
ing in France, we present our four treatments.
3.1. Contextual features
The experiment was carried out simultaneously in 2015 in two
départements, Seine-Maritime and Eure, in the Haute-Normandie
region of France. There are 1.85 million inhabitants and 275,652
women aged 50–74 years old (and therefore eligible for the national
screening program) in these two départements. The choice of these
two départements was both based on demographics (one is more
urban and educated, the other more rural) and practical. The man-
aging structures in charge of the program in both départements
were willing to undertake a RCT Randomized Controlled T. Without
full support from the local teams, our experiment would have been
impossible. As shown on the timeline in Fig. 2, the intervention took
place in April 2015, and the data was collected the following year.
We also collected data from the previous time the women were
invited or screened, which was two years before our intervention
(i.e. 2013), as the national program has a two-year invitation cycle.
An employee of each managing structure carried out the individual
randomization (it was conducted on Structured Query Language
(SQL) using a unique identifier for each woman). The treatments
appeared in the invitation letter that was sent to eligible women,
which contained a voucher for a free mammogram. The managing
structure sent out the letters and collected the information each
time a women used her voucher. However, women can screen out-
side the program independently of this invitation letter: we call this
opportunistic screening. The medical test in this case is exactly the
same (mammograms can only be carried out at clinics that accept
vouchers). There are however two differences in the procedure.
Women who screen opportunistically have to pay a fee (a minimum
of 66 Euros) and require a prescription.13 Second, every mammo-
gram undertaken within the national program is double-checked
via a second reading at the managing structure (while those from
opportunistic screening are not). We collect data on opportunis-
tic screenings from the Health Insurance funds, as the managing
structure is not informed about these.14
The invitation letter had to follow the guidelines defined by the
National Institute of Cancer (i.e. it has to include information about
13
In France, 70% of the regulated fees are reimbursed by the Health Insurance fund
and the last 30% is covered by complementary health insurance, if the women has
this insurance. This applies to both opportunistic screening and the consultation to
obtain the prescription.
14
The National Health Insurance funds can identify each item of health-care con-
sumption via the medical acts classification code (Code de Classification des Actes
Médicaux). This coding system allows us to know if women screened for breast
cancer in the 12 months after the invitation was sent or if they screened during
the 12 months before the invitations were sent. This classification code distin-
guishes between organized screening (coded QEQK004), opportunistic screening
(coded QEQK001) and follow-up mammograms (coded QEQK005).
231
mammography and the program, as well as a list of radiologists par-
ticipating in the program in their département). These guidelines
leave sufficient leeway for substantial modification. We designed
four treatments jointly with the head of each managing structure
and a representative of each National Health Insurance fund in
Haute-Normandie. Each treated invitation was randomly sent to
women who were due to be invited in April or May 2015.
3.2. Treatments
Women who were due to receive an invitation letter in April or
May 2015 were randomly divided into five groups, each receiving
one of the treatments described below.
3.2.1. Control group
Women in the control group received the usual invitation from
each managing structure. In both départements, the letter had not
been changed since the national program was introduced in 2004.
Even though they have the same default option (i.e. the existing
letter), these differed enough across the two départements that
the treatment effects will be presented separately by département.
Both letters can be found in Figs. A.7 and A.8 in Appendix B.
3.2.2. Logo treatment
By default, the logo of the managing structure of the breast-
cancer prevention program already appeared on the envelope.
However, this logo is not very well known.15
Our “logo” treatment added the official logos of the three
National Health Insurance funds16 : these three funds are trustwor-
thy well-known health institutions. Adding these logos provided
a clear signal that the letters came from an official institution. An
image of the logo-treated envelope appears in Figs. A.9 and A.10 in
Appendix B.
Based on the saliency effect of the logos, we expect this treat-
ment to raise the probability of the envelope being opened. We
also expect them to increase the trust placed in the information
contained in the letters via a messenger effect (Webb and Sheeran,
2006).
On the contrary, there may also be adverse effects. Unknown
logos may trigger more opening out of curiosity. Alternatively, if
the Health Insurance funds are believed to be untrustworthy, their
logos may reduce envelope opening.
3.2.3. Chosen treatment
This treatment aimed to produce a letter that was clearer and
easier to understand. It is well known that perceived complexity
can lead individuals to stop collecting and processing information
(Bertr et al., 2006). This bias can be addressed by changing the way
in which information is presented. Most research in this area does
not make explicit how the written content sent to recipients was
selected. We here used a sample of eligible women to evaluate the
invitation letters in a separate experiment carried out in both the
Paris area and Seine-Maritime a few months before our RCT. The
128 women who participated filled out a questionnaire and then
evaluated a set of five invitation letters: the two default or control
invitation letters in each département (labeled “Eure” and “Seine-
Maritime” in the paper), the initial letter suggested to the managing
15
There are almost 90 managing structures in France, each with a different logo,
whose only purpose is to manage the breast- and bowel-cancer programs for people
aged 50–74.
16
These are the CNAMTS (National Health Insurance Fund for Employees), the RSI
(National Insurance Fund for the Self-Employed) and the MSA (Agricultural Social
Mutual Fund). These three insurance agencies together comprise the National Health
Insurance funds (Assurance Maladie).
232 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Table 1
The content of letters and the selection process.
Information on riska Number of lines
Seine-Maritime None 16
Test 1 Vague 13
Test 2 Vague 11
Eure None 17
Nationalb Precise 15
No. Obs.
Notes: The respondent’s geographic origin did not turn out to be significant, so pooled results are presented. We excluded respondents from the sample who did not complete
either the questionnaire or the task (24 women out of 128 (18.7%)).
a
The information on the risk of developing breast cancer: vague when the letter only mentions breast cancer as frequent; precise when the probability of developing
cancer over the lifetime is mentioned.
b
This letter is called “Chosen” hereafter. Additional selection process analysis appears in Table A.1 in Appendix A.
structures by the National Institute of Cancer (labeled “National”)
and two additional letters that we wrote in collaboration with the
managing structures (labeled “Test 1” and “Test 2”). The full invita-
tion letter set appears in Appendix A. The five letters were displayed
in a neutral format, so that only the textual content differed. The
initial order of the letters was fixed. Respondents had to physically
rank the five letters, putting their preferred choice on top. They
were asked to “Rank the letters by putting on top the one you prefer
(clearer and providing more incentives, for instance) and last the one
you liked the least”.
We are here faced with a social-choice problem, i.e. we need
to aggregate individual letter preferences. It is well known that a
social ordering can be obtained in a number of (potentially conflict-
ing) ways. However, the individual rankings aggregated nicely into
an unambiguous social preference. In particular, the “National” let-
ter was obviously found to be better, i.e. ranked first according to
all criteria. It was, for instance, preferred to its closest competitor
(“Eure”) by 61.5% of the respondents. The robustness of the ranking
according to various criteria can be assessed from Table 1.
We wanted to select the best letter in order to increase uptake
rates. We assumed that the preferred letter would also be the most
appropriate way to convey information. We designed Test 1 and Test
2 in collaboration with the individuals who are usually in charge of
writing these letters. We brought together a large set of letters in
use in other départements and tried to come up with a mix of all of
their desirable features. However, the eligible women did not select
the Test 1 and Test 2 letters, underlining the need to test letters with
actual recipients. The best letter, for our sample, was that suggested
by the National Cancer Institute to the managing structures.
The exact reasons why this letter won are not entirely clear.
However, it is worth noting that it is the only one that provides
precise information regarding the magnitude of the risk. Its very
first sentence reads “Over the course of her life, nearly one in eight
women will face this cancer”. Other letters (Test 1 and Test 2), one the
contrary, include vaguer formulations such as “Breast cancer is com-
mon, but discovered early, it is easier to treat”, while Seine-Maritime
and Eure include no reference to risk at all. The other noticeable
difference relates to the length of the letters: the chosen letter is
shorter than the Seine-Maritime and Eure letters by respectively
1 and 2 lines. The differences among letters (in terms of reference
to risk and length) are summarized in Table 1. The chosen letter
was also considered the clearest in the sense that all of its content
was easily understood. From anecdotal evidence provided by the
respondents, no part of it was obscure as it contained no adminis-
trative references. Other letters often contain statements that are
confusing.17
17
For example, the first paragraph of the Eure control letter refers to the Regional
Health Authority. This authority is unknown to recipients as its only goal is to allo-
cate resources between hospitals and define, finance and evaluate health-promotion
Number of words Average ranking Ranked 1st (prop.)
183 2.95 12 (11.5%)
132 2.97 17 (16.3%)
96 4.06 18 (17.4%)
176 2.76 22 (21.2%)
155 2.26 44 (42.3%)
104 104 (100%)
3.2.4. Mixed treatment (L + C)
The mixed treatment is a combination of the logo and chosen
treatments, aiming to ensure that the envelope would be opened
and that the letters could be easily understood.
3.2.5. Social-norm treatment
As explained above, social-norm interventions consist in provid-
ing information about what others are doing. In particular, rather
than anonymous others, social norms work best when reference is
made to peers (i.e. similar individuals). We thus chose our refer-
ence group to be women living in the same département. The peer
information provided in this treatment is the number of women in
the département who screened in the program the previous year
(which is a large number). Using a number instead of the uptake rate
(around 55%) avoids telling recipients that screening avoidance is
more pervasive than they previously believed, and so negatively
affecting beliefs. We believe that using the number screened will
avoid this effect, although we have no direct evidence on women’s
beliefs in this respect. The exact wording is: “In 2014, 17,682 women
screened in the national program in Eure. Why not you?”. This sen-
tence was introduced between the 1st and 2nd paragraphs of the
invitation letter. It constitutes a paragraph on its own, as shown in
Figs. A.11 and A.12 in Appendix B.
4. Data and empirical strategy
A total of 27,137 invitation letters were sent out. A small fraction
of these, 2.37%, were returned to the managing structures marked
“not living at this address”. Our final sample is hence N = 26,495,
with NEure = 10,411 (40%) and NSeineMaritime = 16,084 (60%). The main
outcome is Total screening, obtained by merging data from the man-
aging structure and the National Health Insurance funds. It is almost
impossible for someone to have obtained a mammogram without
appearing in our dataset. Total screening is 1 if a participant had a
mammogram, either as part of the program or outside of it, within
one year of being invited. The other available individual charac-
teristics in our dataset are described in Appendix C. The random
assignment of treated letters means that we expect women in the
control and various treatment groups to have similar background
characteristics. As such, they would have had similar average out-
comes in the absence of any change to the letters. The comparison of
outcomes between the treated groups and the control group reveals
the effect of each treated letter.
If the multiplicity of tests is not addressed, then the probability
of a true null hypothesis being rejected increases with the num-
policies. The Seine-Maritime control letter refers to the possibility for the recipient
to consult her gynecologist or GP to obtain a voucher. This could be confusing as there
is no need to have a consultation with a doctor prior to obtaining an appointment
with the radiologist in the national screening program.
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 233

Table 2
Treatment effects on total screening.

Total sample Seine-Maritime Eure

(1) (2) (3) (4) (5) (6) (7) (8) (9)

N % screening T-C N % screening T-C N % screening T-C

Control 5277 47.58 3186 48.4 2091 46.4

Logo 5296 46.51 −1.07 3227 48.1 −0.3 2069 44.0 −2.4
(0.970) (1.25) (1.54)
Chosen 5315 46.83 −0.75 3214 47.3 −1.1 2101 46.2 −0.2
(0.970) (1.25) (1.54)
Norm 5307 47.09 −0.49 3234 48.5 0.1 2073 44.9 −1.5
(0.970) (1.25) (1.54)
L+C 5300 46.85 −0.73 3223 47.3 −1.1 2077 46.1 −0.3
(0.971) (1.25) (1.55)

Total 26,495 46.97 16,084 47.9 10,411 45.5

Standard errors in parentheses. T-C is the difference between % screening in the treatment group and in the control group. Differences would be in bold if they were
statistically-significant adjested differences.
*p < 0.1.
**p < 0.05.
***p < 0.01.

ber of tests. We thus here use the Holm–Bonferroni (Holm, 1979)

procedure that controls for the family-wise error rate.18 The family
of hypotheses here corresponds to the null hypothesis that each
of the four treatments has the same effect on mammography use
(total, organized or opportunistic screenings) as the control group
for the full sample, as well as in each sub-population of interest.
This method does not assume any dependent structure between
the tests, and therefore provides fairly conservative results for tests
that may be dependent. In our case, we can assume that, while the
treatments are independent, the three outcomes are not.
To control for potential imbalances between groups, we also
regress total screening on the treatment groups while controlling
for individual characteristics using logit models.

5. Results
We first check that the experimental design was actually imple-
mented as planned. Table A.2 in Appendix D presents the summary
statistics on experiment implementation in the two départements
separately. We test for differences between the treatment and con-
trol groups, while correcting for the family-wise error rate using the
Holm–Bonferroni method. Differences between the treatment and
control groups are small in size, and only 2 out of 57 adjusted esti-
mated p-values are below 0.1, suggesting that the randomization
was effective in creating balanced groups. This is confirmed by the
regression analysis in Table A.3 in Appendix D.
Despite having a large sample, we are only able to detect effects
above 2.7 percentage points.19
5.1. Treatment effects on total sample
Fig. 3 depicts the percentage of women receiving a mammogram
for each treatment and control group in each département. There
appears to be no differences between treatment and control mam-
mography use. Table 2 lists the percentages of women receiving a
Fig. 3. Treatment effects on total screening. Notes: The vertical lines are 95% con-
fidence intervals. Each bar depicts the percentage of mammograms undertaken
within a year after the invitation among women who were invited in each group. We
consider both mammograms undertaken in and outside of the national screening
program.
mammogram and the differences between the treatment and con-
trol groups for the total sample and each département separately.
Bold figures would reflect statistically-significant adjusted p-values
using the Holm Bonferoni method. Bonferroni adjusted p-None of
the differences are statistically significant.
Table 3 provides further evidence of this lack of effect by
showing the logit regression results. Columns 1, 3 and 5 include
a département fixed effect to account for control-group differ-
ences, and columns 2, 4 and 6 include the individual characteristics
described in Appendix C.
The RCT results in Tables 2 and 3 are clear and unambigu-
ous: none of our four treatments caused any significant change in
screening behavior.

5.2. Treatment effects on the newly eligible

18
The multiple hypothesis-testing problem has received increasing attention, with
many authors re-analysing RCT data to produce new results with adjustment for The newly eligible are women aged 50 who are invited to under-
multiple hypotheses (for instance the RCTs on education and health programs in Lee
take a mammogram for the first time. As they have never been
and Shaikh (2014) and Carneiro et al. (2011)), as well as a charity field experiment
(List et al., 2016).
invited before, there may be a greater chance that they will pay
19
Assuming ˛ = 0.05, ˇ = 0.20, the variance of the outcome  being 0.25 and each attention to the letter. Women who have already been invited a
group sample size  being close to 5300, the minimum detectable effect is MDE = number of times may just take the voucher and screen (or not) as
(1.96 + 0.85) ∗ 2 2 /n = 0.027. usual.
234 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Table 3
Treatment effects on total screening.

Total sample Seine-Maritime Eure

(1) (2) (3) (4) (5) (6)

Logo −0.011 −0.004 −0.003 0.003 −0.024 −0.019

(0.010) (0.008) (0.012) (0.010) (0.015) (0.012)
Chosen −0.008 −0.004 −0.011 −0.009 −0.002 0.000
(0.010) (0.008) (0.012) (0.010) (0.015) (0.012)
Norm −0.005 −0.001 0.001 0.007 −0.015 −0.016
(0.010) (0.008) (0.012) (0.010) (0.015) (0.012)
L+C −0.007 −0.004 −0.011 −0.004 −0.003 −0.007
(0.010) (0.008) (0.012) (0.010) (0.015) (0.012)
p-value of joint hyp. test 0.8468 0.9712 0.7842 0.6067 0.4843 0.3844

Covariates No Yes No Yes No Yes

Observations 26,495 26,495 16,084 16,084 10,411 10,411

These are the average marginal effects from logit regressions. Standard errors in parentheses. There is no correction for the number of hypotheses tested.
*p < 0.1.
**p < 0.05.
***p < 0.01.

Table 4
Treatment effects on total screening among the newly eligible (age 50).

Total sample Seine-Maritime Eure

N % screening T-C N % screening T-C N % screening T-C

Control 456 53.07 296 49.0 160 60.6

Logo 433 48.27 −4.8 265 49.8 0.8 168 45.8 −14.8***
(3.35) (4.24) (5.51)
Chosen 469 50.11 −2.96 329 51.1 2.1 140 47.9 −12.8**
(3.33) (4.01) (5.73)
Norm 444 46.17 −6.9** 294 47.3 −1.7 150 44.0 −16.6***
(3.33) (4.12) (5.61)
L+C 412 46.36 −6.7** 268 45.5 −3.5 144 47.9 −12.7**
(3.40) (4.21) (5.7)

Total 2214 48.87 1452 48.6 762 49.3

Standard errors in parentheses. T-C is the difference between % screening in the treatment group and in the control group. Differences would be in bold if they were
statistically-significant adjested differences.
*p < 0.1.
**
p < 0.05.
***
p < 0.01.

Table 5
Estimated treatment effects on total screening among the newly eligible.

Total sample Seine-Maritime Eure

(1) (2) (3) (4) (5) (6)

Logo −0.048 −0.046 0.008 0.009 −0.148*** −0.146***

(0.034) (0.033) (0.042) (0.042) (0.054) (0.053)
Chosen −0.029 −0.031 0.021 0.017 −0.128** −0.115**
(0.033) (0.032) (0.040) (0.039) (0.057) (0.056)
Norm −0.069** −0.081** −0.017 −0.026 −0.166*** −0.186***
(0.033) (0.033) (0.041) (0.040) (0.056) (0.055)
L+C −0.067** −0.071** −0.035 −0.040 −0.127** −0.136**
p-value of joint hyp. test 0.2062 0.0951 0.7005 0.5775 0.0292 0.0125

Covariates No Yes No Yes No Yes

Observations 2214 2214 1452 1452 762 762

These are marginal effects from logit models. Standard errors in parentheses. No correction for the number of hypotheses tested.
*p < 0.1.
**
p < 0.05.
***
p < 0.01.

From Tables 4 and 5, it seems that the social-norm and mixed
treatments reduced the probability of screening by the newly eligi-
ble. Breaking down the analysis by département suggests a slightly
different story. The treatment and control letters led to similar
screening rates in Seine-Maritime, while in the Eure all treatments
reduced the screening rate. There are three possible interpretations
of the sharp fall in screening rates among the newly eligible in the
Eure (Fig. 4).
A first possibility would be that the randomization failed for
the particular sub-group of newly eligible women. In particular,
we double-checked whether there are any imbalances between
the groups, but concluded that this was not the case (as can be
seen in Table A.4 in Appendix D). Furthermore, these results are
robust to the introduction of covariates, as shown in Table 5. A sec-
ond possibility is that the sharp fall in screening rates reflects a
higher screening rate in the control group rather than a fall in the
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Fig. 4. Treatment effects on total screening among the newly eligible. The vertical
lines are 95% confidence intervals. Each bar depicts the percentage of mammograms
undertaken within a year after the invitation among women who were invited for
the first time in each group. We consider both mammograms undertaken in and
outside of the national screening program.
treatment groups. In fact, the screening rate in the control group
in the Eure is 60%, while that in the treatment group is close to the
total sample average. We thus interpret this as a Type-I error, i.e.
the null was falsely rejected. Given the number of subgroups, it is
likely that this kind of Type-I error would arise for one subgroup.
To further investigate, we collected additional data from the man-
aging structure, allowing us to compare the organized screening
rate during our experiment with that found several months before
and after our experiment. These organized screening rates appear
in Table A.5 in Appendix E. We have no information on opportunis-
tic screening for other time periods. Assuming that opportunistic
screening is fairly stable over time, the comparison of the take-
up rates in the organized program provides support for the Type-I
error interpretation. The newly eligible in our experiment have an
uptake rate of 60%, including 53.8% who screen within the program.
In sharp contrast, the comparable figure for other time periods is
never over 50.9%, with an average of 46.9%, supporting the view that
Fig. 5. The change in screening probability across social groups.
Note: Each dot represents the coefficient (and its 95% confidence interval) of the linear combination of two coefficients estimated using the interaction effect model in
Table A.7 in Appendix F. These indicate the size and significance of the change in screening probability between each treatment and the control group for each deprivation
quintile. For example, the effect of the mixed treatment on the 5th quintile corresponds to the following linear combination of coefficients from Table A.7: (L + C × 5th
quintile) + (L + C) = −0.007 + (−0.007) = −0.014.
235
we observed abnormally high rates in our experiment. In addition,
the screening-rate difference between the control and treatment
groups in the Eure does not attain the usual significance bench-
marks when controlling for multiple hypothesis testing. A third
possibility is that treatments did have a real depressing effect on
newly-eligible women in the Eure. We cannot rule out this possi-
bility.
Overall, the negative differences for the newly eligible are driven
by the Eure and especially by the surprisingly high screening rate
in the control group. We should be cautious in interpreting this
as a negative effect of our behavioral interventions on the newly
eligible.
What do we then learn from the newly eligible? We believed
that treatment effects would be more likely in a group that had
a greater chance of reading the invitation letters. For instance, if
learning occurs through trial and error, these women may change
the information they consider to be relevant and thus their behav-
ior. However, our analysis provides suggestive evidence that they
turned out to behave in the same way as those women who had
already received a number of invitations.
6. Evaluation with respect to alternative policy objectives
We have shown that none of our treatments led to positive
effects for a policy maker who wishes to increase mammography
use. However, even with no effect on average, uptake rates may
have moved for some particular sub-groups. Two sub-group analy-
ses are particularly important from a policy perspective. First, it
is well-known that there is a social gradient in mammography
use: Policy makers may thus want to reduce the gap in uptake
across social groups, even if average uptake does not change. To
see whether our treatments had any impact on social inequality,
we approximate socioeconomic status by a deprivation index at the
local level (2000 inhabitants per area). The higher is this index, the
more deprived the area is and the more likely that individuals liv-
ing there are low social status. We run a linear-probability model
with an interaction between deprivation-index quintiles and the
treatments. There were no significant treatment effects on mam-
236 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Table 6
Treatment effects on organized and opportunistic screenings in the total sample.

Organized screening

N % screening T-C

Control 5277 45.27

Logo 5296 43.67 −1.6*
(0.967)
Chosen 5315 44.4 −0.87
(0.967)
Norm 5307 44.58 −0.69
(0.967)
L+C 5300 43.85 −1.42
(0.967)

Total 26,495 44.36

Fig. 6. Percentage mammography use across the deprivation index.
Notes: This graph shows the proportion of mammograms undertaken within a year Opportunistic screening
after the invitation among the women invited during the experiment. We brake
down the total sample by département. The X-axis represents the five quintile of N % screening T-C
the deprivation index. Q5 represents the highest level of deprivation. Control 5277 2.31
Logo 5296 2.83 0.52*
(0.31)
mography use in any deprivation-index quintile,20
as shown in Chosen 5315 2.43 0.12
Fig. 5. (0.3)
As a result, we continue to find a social gradient in mammog- Norm 5307 2.51 0.2
(0.3)
raphy use, as shown in Fig. 6, with a 10–15 percentage-point
L+C 5300 3 0.69**
screening difference between women in the lowest and highest (0.31)
deprivation-index quintiles.
Total 26,495 2.62
Another policy objective may be to increase uptake rates in
the national program and thus reduce opportunistic screening (i.e. Standard errors in parentheses. Differences in bold indicate statistically-significant
adjusted differences.
mammograms received without using the voucher). This objective *
p < 0.1.
reflects that national-program screening provides an additional **
p < 0.05.
safeguard to reduce false negatives (with a second reading of ***p < 0.01.
every normal mammogram by an independent radiologist) and
leads to more regular mammogram use than opportunistic screen-
ing, as shown in Goldzahl and Jusot (2017). A higher uptake rate received: we sent the letters out randomly, but cannot know if
could also increase the cost-effectiveness ratio of the national pro- these letters were actually received, opened, read and their con-
gram. tents taken into account. At the extreme, almost all women may
Again, even if average uptake remains constant, our treatments have already made up their mind before receiving the letters, so
may affect the type of mammogram. It is important to bear in mind that no behavioral intervention would have much effect. In the
here that opportunistic screening is only limited (between 2 and same vein, we have little control over contagion effects that may
3% in both départements). The detailed results can be found in blur the distinctions between treatments. Contagion means that it
Table 6. The regression analysis confirms this finding (see Table A.9 is more difficult to identify any treatment effect econometrically:
in Appendix F) both overall and in both départements. There may if the treatment works, then both the treated and the non-treated
be a slight tendency for the treatments to reduce the uptake will screen more. However, we find no evidence that our treat-
rate in the national program by about 1% and increase them by ments here have increased the average screening rate. These two
about the same amount outside of the national program. How- limitations (ITT and contagion) reinforce our main conclusion that
ever, none of our treatments had any significant effect once we there is little hope, under the current system and our experimental
control for the number of hypothesis tested. Overall, even the sec- design, that behavioral interventions will have an effect. In terms of
ondary policy objectives are not achieved by our four behavioral comparing one treatment to another, we do not believe that there is
interventions. a problem as ITT and contagion will affect the treatments proposed
here in the same way.
Statistical power is another possible limitation of our work.
7. Robustness and limitations
Despite having a large sample, we only detect effects that are over
2.7% (i.e. which produce a change of at least 2.7% in mammography
We first list the potential problems that may affect our results,
use). A study using a larger population (and/or fewer treatments),
and then explore the external validity of our results, i.e. the extent
would be a welcome addition.
to which our results may hold in different contexts.
Last, there is an intrinsic limit due to reminders. Women who
do not screen within one year receive a reminder. These reminders
7.1. Limitations and internal validity have a considerable effect, as over 10% screen using the reminder-
letter voucher (which incidentally explains why the uptake rates
We here adopt an intention-to-treat (ITT) approach, based on are different in Fig. 1 and Table 3), although they should have a
the initial allocation of treatment rather than the actual treatment similar effect across treatment groups. In our work here, we only
measure treatment effects over the first twelve months. A more
complete study would also include interventions on the reminders
20
The detailed regression results appear in Table A.7 in Appendix F for the full before drawing conclusions about the effect of behavioral inter-
sample and the Eure and Seine-Maritime samples. Note that only one interaction ventions. We cannot exclude that behavioral interventions have a
term is statistically significant at the 10% level in the full sample. This difference
is no longer statistically significant when we control for the number of hypotheses
larger effect in the reminder letters (although we do not think that
tested. this is likely).
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
7.2. Scaling-up and external validity
While the two départements border each other and belong to
the same administrative region, they differ in some ways. For exam-
ple, the Eure is more rural than the Seine-Maritime. We thus find it
reassuring that the same results are found in both départements
(except for the newly eligible), as policy recommendations will
only be made for treatment effects that are robust enough to con-
textual features. Furthermore, the two départements in which the
experiment was carried out are representative of France in terms
of age distribution, the proportion of women with a chronic dis-
ease and health-insurance status. The national figures21 appear in
Table A.2. To further explore external validity, we can regress our
available covariates on the decision to screen. The existing litera-
ture provides ample evidence that low income, being sick, limited
health-insurance coverage and low health-care consumption are all
associated with a lower screening probability (Carrieri and Wübker,
2013; Goldzahl and Jusot, 2017; Sicsic and Franc, 2014; Devaux,
2015; Wübker, 2014; Jusot et al., 2012; Hsia et al., 2000; Trivedi
et al., 2008). We also find these correlations in our sample (Table A.6
in Appendix E.3). The behavior of our sample women is very much
in line with the existing evidence, so that our results may well be
valid for women in similar conditions.
One potential limit to scaling-up the experiment would be the
national program itself over the last decade. It may be that the cur-
rent programs have reached a limit, since uptake rates are no longer
rising. Behavioral interventions in situations with lower uptake
rates, e.g. before or at the introduction of a national program, may
be able to increase uptake rates.
8. Discussion
Behavioral economics provides useful empirical insights by
underlining the most common flaws affecting individual decision-
making. Applying these insights in the field is a major challenge
to the design of better policies. We here considered screening
behavior in this context, asking whether we can use behavioral
interventions to increase mammography use. Our behavioral inter-
ventions were based on the previous literature and feasibility.
Among these interventions in the national program, we selected
those that seemed the most promising, based on a literature review,
in terms of the effect size. However, none of our treatments had a
significant impact.
Two hypotheses can explain this negative result. The first is that
many women may already have all the information they think that
they need and have taken their decision before receiving the letter.
Non-screening results from a calculated decision where the costs
outweigh the benefits of screening. Behavioral interventions will
then not be effective. On the contrary, a second hypothesis is that,
despite the intervention, the gap to an informed decision is still too
large for many women, who may not open the letters or not under-
stand them if they do. In short, they remain confused or unaware
and our behavioral interventions are not able to help. These two
hypotheses are not mutually exclusive, and may apply to different
population sub-groups. The main lesson here is probably that the
subset of women for whom our interventions affected behavior is,
at best, small.
Is there then no role for behavioral interventions? Not necessar-
ily. A number of biases were not addressed here for feasibility rea-
sons: procrastination, for example, is likely to play a role, as some
women may want to screen but keep on putting off the decision.
21
Our sample of women aged 50–74 is representative of the French population.
The data is from the 2010 and 2012 waves of the Enquete sur la Santé et la Protection
Sociale (ESPS) survey.
237
In terms of policy implications, it is perhaps time to rethink
national screening programs, at least for breast cancer. To the best
of our knowledge, these programs typically use mass mailings to
inform individuals of their eligibility. In addition to letters, a num-
ber of attempts have been made to improve screening uptake based
on incentives and information. We here contribute to this litera-
ture by testing some behavioral interventions. None of these three
approaches to mammography uptake (incentives, information and
behavioral interventions) has proved effective. Indeed, under the
two hypotheses discussed above, marginal changes to the current
system are unlikely to trigger substantial behavioral change. Some
women have already obtained the information they require else-
where and stick to their screening decisions (80% of women who
screened in a given year screen again two years later). On the
contrary, some women may remain confused or undecided about
participating, and marginal changes to the content and presen-
tation of their letters does not suffice to change their behavior.
Addressing this confusion is a challenge. New technologies offer
additional tools (text messages or e-mail invitations for example).
However, the existing evidence suggests that incomplete take-up
can persist, even for individuals who have clear incentives (for
instance in terms of claiming a sum of money, as in Bhargava and
Manoli (2015)). Last, there may be alternative behavioral inter-
ventions to improve screening rates (e.g. commitment devices to
address procrastination), although these are difficult to imple-
ment within the current screening system. Overall, the increase
of screening rates above their current level appears to be difficult
with the national programs as they were initially designed.
We finish by speculating about the type of changes that could
be implemented. For instance, when the programs were intro-
duced it was considered natural to invite a subset of women every
month based on their date of birth (at least in France, but also in all
European countries in which we have been able to collect this infor-
mation). Women who are born in the same month are then invited
simultaneously. Geolocalizing women is now feasible at reasonable
cost, and it is possible to invite all women from the same neigh-
borhood at the same time. Changing the way in which invitations
are decided is a detail that matters. There is indeed evidence that
neighborhood effects have an impact on health (Kling et al., 2007).
If women in the same neighborhood are all contacted at the same
time, it may well be possible to reinforce the message with the
appropriate posting of public information.
Acknowledgments
The authors gratefully acknowledge the support of the Ligue
Contre le Cancer for providing funding to conduct this research
project. We also acknowledge the support of the Health Chair –
a joint initiative by PSL, Université Paris-Dauphine, ENSAE, MGEN
and ISTYA under the aegis of the Fondation du Risque (FDR). The
first author’s research has been supported by a post-doctoral grant
co-financed by the Health Chair at the Fondation du risque (FDR),
the project ANR 11-LABX-0019 and the grant “Prévention Primaire”
number 2014-100) from the french National Institute of Cancer
(INCA). People working in the managing structures deserves a
special thanks: Guillaume Fiton, Sandrine Flechais, Sylvie Guiller-
Devillers, Ahmed Benhammouda, Catherine Daunou, Catherine
Verleye, Louise Denéchère and Laurent Verzaux. Sigolène Duver
help us with administrative data. Audrey Bertinet greatly help us
in getting approval from CNIL (accreditation number DR-2015-
183). The authors would like to thank Pierre-Emmanuel Couralet
for his help with the data analysis and Andrew Clark for editing.
The audience in various seminars provided very helpful comments
in the 2016 ESA world meeting in Jerusalem, 2016 EUHEA Phd
Conference in Barcelona, 2016 Social Capital and Health Work-
shop in Jerusalem, 2016 EUHEA Conference in Hamburg, 2016
238 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

JESF in Lyon, 2017 AFSE Conference in Nice, 2017 ASFE Confer- reports from anonymous referees and critical guidance from the
ence in Rennes, 2017 Health Chair Conference in Paris, 2017 SHARE editor.
Meeting in Nice, and participants to seminars at Université Paris
Dauphine, Université Lille 2, Université Paris-Descartes and in Appendix A.
Paris School of Economics. Last, we benefited from three superb
Letters from the selection process

Fig. A.1. Instruction.

L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 239

Fig. A.2. Eure.

240 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Fig. A.3. Test 2.

L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 241

Fig. A.4. Seine-Maritime.

242 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Fig. A.5. Test 1.

L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 243

Fig. A.6. National.

244 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Appendix B.

Treatments

Fig. A.7. Seine-Maritime control-group letter.

Fig. A.8. Eure control-group letter.

 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 245

Fig. A.9. Seine-Maritime envelope with the health-insurance fund logos.

Fig. A.10. Eure envelope with the health insurance-fund logos.

246 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Fig. A.11. Eure letter with the social-norm treatment (2nd paragraph).

Fig. A.12. Seine-Maritime letter with the social-norm treatment (2nd paragraph).
Table A.1
Additional selection-process analysis.

Average ranking Ranked 1st (prop.) Ranked 1st or 2nd Majority ranking

National 2.26 44 (42.3%) 63 (60.6%) 1

Eure 2.76 22 (21.2%) 47 (45.2%) 2
Seine-Maritime 2.95 12 (11.5%) 41 (39.4%) 3
Test 1 2.97 17 (16.3%) 39(37.4%) 4
Test 2 4.06 9 (8.7%) 18 (17.4%) 5

No. Obs 104 104 (100%) 208 (200%)

Majority rankings are obtained by comparing each pair of options. An option that is preferred by a majority to all other options is ranked first, and so on. Note that majority
rankings may not exist as the aggregation of individual preferences can lead to cycles.
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 247

Fig. A.13. Chosen treatment in the Seine-Maritime and Eure.

Appendix C.

Randomization-check tables

Table A.2
The sample.

Seine-Maritime Eure France

Control Logo Chosen Social Norm L+C Total S-M Control Logo Chosen Social Norm L+C Total Eure

Full sample 19.81 20.06 19.98 20.11 20.04 20 20.08 19.87 20.18 19.91 19.95 20
Past screening
Screened 39.42 37.99 38.08 38.1 38.5 38.42 37.11 36.25 37.6 37.14 37.79 37.18
Not screened 46.61 49.05 46.83 48.3 48.84 47.93 44.91 45.82 45.98 44.72 44.58 45.2
Refused 1.63 1.98 2.02 1.73 1.52 1.78 6.98 7.39 6.95 7.86 8.28 7.49
Newly eligible 9.29 8.21 10.24 9.09 8.32 9.03 7.65 8.12 6.66 7.24 6.93 7.32 5.6
Newcomer 3.04 2.76 2.83 2.78 2.82 2.85 3.35 2.42 2.81 3.04 2.41 2.8
Deprivation index
1st quintile 18.3 17.97 18.01 18.37 18.68 18.27 23.43 21.02 21.09 22.05 22.82 22.08
2nd quintile 12.77 14.29 13.94 12.18 13.84 13.4 18.84 18.27 17.99 17.37 16.9 17.88
3rd quintile 10.42 11.59 11.14 10.27 10.58 10.8 11.19 12.76 13.9 11.96 12.33 12.43
4th quintile 20.97 20.05 20.72 21.61 19.7 20.61 19.18 20.2 19.75 21.76 19.74 20.12
5th quintile 37.54 36.1 36.19 37.57 37.2 36.92 27.36 27.74 27.27 26.87 28.21 27.49
Compl. HI
Private HI 70.34 71.15 70.38 70.22 69.16 70.25 68.24 64.52 65.06 63.15* 65.48 65.3
CMUC-ACS 7.6 6.79 7.81 8.01 8.28 7.7 3.92 4.49 5.14 4.87 4.43 4.57 5.77
None 22.07 22.06 21.81 21.77 22.56 22.05 27.83 30.98 29.8 31.98 30.09 30.13
Age groups
[50–54] 29.28 28.91 30.52 28.7 29.88 29.46 25.16 24.12 23.89 24.31 24.46 24.39 27.22
[55–59] 26.71 25.19 25.05 25.82 26.75 25.9 24.39 23.2 24.89 25.66 24.55 24.54 25.12
[60–64] 17.86 19.58 18.17 18.95 17.84 18.48 20.95 21.41 21.13 22.19 20.56 21.25 21.67
[65–69] 15.57 15.74 15.87 15.77 16.01 15.79 17.69 18.37 17.66 16.16 18.54 17.68 15.42
[70–75] 10.58 10.57 10.39 10.76 9.53 10.36 11.81 12.9 12.42 11.67 11.89 12.14 10.57
Health insurance funds
Regime Generale 90.14 90.18 90.54 90.14 89.98 90.2 85.03 82.65 83.34 81.86* 83.29 83.24 78.81
MSA 1.76 1.95 1.59 2.01 2.2 1.9 3.68 3.38 3.24 4.1 3.08 3.5 6.11
RSI 0.31 0.34 0.44 0.34 0.31 0.35 4.54 4.88 4.71 5.45 5.3 4.98 5.37
MGEN 4.65 4.77 4.73 4.73 4.62 4.7 3.49 4.93 4.71 5.26 4.33 4.54 9.71
Other 3.14 2.76 2.71 2.78 2.89 2.85 3.25 4.16 4 3.33 4 3.75
Has a chronic disease 28.71 28.65 27.08 29.14 28.65 28.45 28.15 26.7 26.57 27.39 27.09 27.18 23.98
GP visits
0 13.97 13.02 13.5 13.64 13.4 13.5 23.39 23.1 23.04 24.6 23.5 23.52
[1–3] 27.02 26.22 25.95 24.86 27.65 26.34 27.64 28.37 27.61 28.27 27.35 27.85
248 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Table A.2 (Continued)

Seine-Maritime Eure France

Control Logo Chosen Social Norm L+C Total S-M Control Logo Chosen Social Norm L+C Total Eure

[4–6] 27.5 29.63 29.15 29.44 28.3 28.81 25.39 27.98 27.46 25.04 27.3 26.64
≥7 31.51 31.14 31.39 32.07 30.65 31.35 23.58 20.54 21.89 22.09 21.86 22
Gynecologist visit
≥1 9.07 8.46 8.09 8.69 8.87 8.64 10.57 10.58 9.85 11 10.54 10.51

This table shows the individual characteristics of the treatment and control group for each département. Figures in bold indicate statistically-significant adjusted differences.
*
p < 0.1.
**p < 0.05.
***p < 0.01.

Table A.3
Randomization check.

Logo Chosen Social Norm L+C

Seine-Maritime 0.006 0.001 0.006 0.004

(0.009) (0.009) (0.009) (0.009)
2nd quintile 0.012 0.007 −0.010 −0.000
(0.008) (0.008) (0.008) (0.008)
3rd quintile 0.007 0.011 0.007 0.001
(0.008) (0.008) (0.008) (0.008)
4th quintile −0.000 0.002 0.008 −0.004
(0.008) (0.008) (0.008) (0.008)
5th quintile −0.010 −0.012 −0.001 0.001
(0.008) (0.008) (0.008) (0.008)
Gynecologist visit −0.004 −0.009 −0.001 −0.001
(0.006) (0.006) (0.006) (0.006)
Compl. HI 0.002 −0.006 −0.006 −0.006
(0.005) (0.005) (0.005) (0.005)
Chronic disease 0.004 0.017 −0.006 0.002
(0.009) (0.009) (0.009) (0.009)
N.C. −0.006 −0.015 −0.002 −0.005
(0.008) (0.008) (0.008) (0.008)
Not screened 0.019 0.006 0.010 0.012
(0.010) (0.010) (0.010) (0.010)
Refused 0.003 0.002 0.004 0.004
(0.004) (0.004) (0.004) (0.004)
Newly eigible −0.005 0.002 −0.003 −0.009
(0.005) (0.005) (0.005) (0.005)
New comers −0.005 −0.003 −0.003 −0.005
(0.003) (0.003) (0.003) (0.003)
[55–59] −0.014 −0.008 −0.000 0.001
(0.008) (0.008) (0.008) (0.008)
[60–64] 0.012 0.003 0.011 −0.002
(0.008) (0.008) (0.008) (0.008)
[65–69] 0.004 0.002 −0.005 0.006
(0.007) (0.007) (0.007) (0.007)
[70–75] 0.004 0.001 0.001 −0.006
(0.006) (0.006) (0.006) (0.006)
GP visits [1–3] −0.002 −0.007 −0.011 0.003
(0.009) (0.009) (0.009) (0.009)
GP visits [4–6] 0.023* 0.018** 0.011 0.012
(0.009) (0.009) (0.009) (0.009)
GP visits ≥7 −0.014 −0.007 −0.002 −0.012
(0.009) (0.009) (0.009) (0.009)
MSA 0.000 −0.003 0.003 0.000
(0.003) (0.003) (0.003) (0.003)
RSI 0.001 0.001 0.003 0.003
(0.003) (0.003) (0.003) (0.003)
MGEN 0.006 0.005 0.007 0.003
(0.004) (0.004) (0.004) (0.004)
Other 0.001 0.000 −0.002 0.001
(0.003) (0.003) (0.003) (0.003)

p-value: Joint Test (Cross Equation) 0.1535 0.4094 0.3651 0.704

These are coefficients from seemingly unrelated regressions. Standard errors appear in parentheses. Each row presents the results of a regression of observable characteristics
on all the listed treatments (rows correspond to a single regression).
*
p < 0.1.
**
p < 0.05.
***p < 0.01.
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 249

Appendix D. Table A.5 (Continued)

Organized Number of
The newly eligible in the Eure screening rate women invited

Social Norm 40.7

L+C 43.8
Table A.4 Average without the control group 42
Randomization check for the newly eligible in the Eure.
January 2015 is the farthest back the managing structure can go in time. The missing
Logo Chosen Social Norm L+C months are due to the fact that the managing structure prefers not to invite women
during the holidays (July, August and February) or in May (several long weekends
2nd quintile 0.007 −0.054 −0.005 0.010
with Bank Holidays). Invitations are instead sent in the surrounding months.
(0.045) (0.047) (0.046) (0.047)
3rd quintile −0.010 0.007 0.007 −0.047
(0.043) (0.045) (0.045) (0.045)
4th quintile −0.008 −0.004 −0.028 −0.001 Appendix E.
(0.041) (0.043) (0.042) (0.042)
5th quintile −0.016 0.045 −0.079* 0.024
Individual characteristics and their effect on screening outcomes
(0.044) (0.046) (0.046) (0.046)
Gynecologist visit −0.045 −0.021 0.060 0.057
(0.044) (0.047) (0.046) (0.046) E.1 Individual characteristics
Compl. HI 0.009 −0.017 −0.018 0.034 Breast-cancer screening profiles. The data provided by each
(0.027) (0.029) (0.028) (0.029) managing structure allows us to distinguish between five breast-
Chronic disease −0.025 −0.032 −0.068 0.017
(0.047) (0.050) (0.049) (0.049)
cancer profiles. The first three are determined by women’s behavior
N.C −0.002 0.038 0.031 −0.058 the last time they were invited to screen in 201322 and the last
(0.041) (0.043) (0.042) (0.042) two by their eligibility in 2015 when we invited them. Women
GP visits [1–3] 0.001 −0.021 −0.050 −0.017 who were invited as part of our experiment were women who (a)
(0.052) (0.054) (0.053) (0.054)
screened 22 months ago,23 (b) were invited 22 months ago but
GP visits [4–6] 0.032 0.042 0.020 0.028
(0.045) (0.047) (0.047) (0.047) have not screened in the program since then, (c) refused to par-
GP visits ≥ 7 −0.047 −0.002 0.036 −0.071 ticipate after being invited 22 months ago, (d) are newly eligible
(0.046) (0.048) (0.047) (0.047) for the program as they have just turned 50, and (e) are newcom-
MSA 0.029 0.002 0.028 0.022 ers as they have just moved to the département and received our
(0.018) (0.019) (0.019) (0.019)
invitation. Note that the women in category (c) are a very particular
RSI 0.010 0.013 0.016 0.019
(0.025) (0.026) (0.026) (0.026) group in the sense that they wrote to or telephoned the managing
MGEN 0.004 −0.003 −0.011 −0.010 structure to say that they refused to screen in the program.
(0.018) (0.019) (0.018) (0.019) Socioeconomic characteristics. We proxy socioeconomic
Others −0.001 0.009 −0.013 0.008
characteristics by two variables: a deprivation index and comple-
(0.012) (0.013) (0.013) (0.013)
mentary health-insurance status.
p-value: Joint Test 0.962 0.991 0.356 0.601 The deprivation index
(Cross Equation)
To offset the lack of socioeconomic data in National Health Insur-
These are coefficients from seemingly unrelated regressions. Standard errors appear ance fund datasets, every woman invited to the experiment had her
in parentheses. Each row presents the results of a regression of observable char-
postal address geocoded by experts from the National Institute of
acteristics on all listed treatments (rows correspond to a single regression). The
département, age categories and screening profiles were not introduced as the sam- Geographic and Forest Information (IGN). From these geographical
ple here is restricted to women aged 50 who are newly eligible and who live in the coordinates, we were able to match each individual to a specific
Eure. area with 2000 inhabitants, for which we have aggregate socioe-
*
p < 0.1. conomic information from census data collected by the French
**p < 0.05.
National Institute for Statistics and Economic Research (INSEE).
***p < 0.01.
From this census data, we were able to construct an ecologi-
cal deprivation index reflecting the socioeconomic situation with
respect to the surrounding population. The selection of variables
Table A.5
included in the deprivation indices is usually carried out using a
Organized screening rates among the newly eligible in the Eure before and after our
experiment.
number of methods such as principal-component analysis, factor
analysis, and the opinions of health experts. The indices are used as
Organized Number of
a proxy of individual socioeconomic deprivation but do not refer to
screening rate women invited
the individual’s own deprivation experience. We use the index in
Months before our experiment Pornet et al. (2012) and Guillaume et al. (2016), called the French
January 2015 47.3 498
version of the European Deprivation Index (EDI), as it selects census
March 2015 50.9 163
Months after our experiment variables that closely correspond to the individual’s perceptions of
June 2015 45.7 766 deprivation.
July 2015 45.3 371 Complementary health-insurance coverage
September 2015 47 628
Coverage by a private or public complementary health-
November 2015 46 657
January 2016 45.8 698
insurance scheme provides information on individual income.
February 2016 46.8 344
Average 46.9
Our experiment
22
April 2015 44.5 762 We unfortunately do not have information on what happened prior to 2013.
23
Control 53.8 This differs between départements. In the Eure, women were invited every 22
Logo 42.3 months, while in the Seine-Maritime they are invited every 20 months.
Chosen 41.4
250 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252

Public complementary health insurance (Universal Medical Cov- Table A.6 (Continued)
erage or a subsidy used to buy private complementary health Screening profile (ref: screened)
insurance) depends only on income. For instance, an individual has Not screened −0.582*** (0.006)
to earn under 8645 Euros per year (for a single-person household) Refused −0.258*** (0.016)
New eligible −0.336*** (0.013)
to be eligible for Universal Medical Coverage.
Newcomers −0.411*** (0.019)
Health status and health-care consumption profiles. From Age (ref: 50–54)
the administrative data of the National Health Insurance funds, [55–59] −0.034*** (0.008)
we approximate health status by dummy variables for the indi- [60–64] 0.008 (0.008)
vidual having or having had a chronic disease the year before or [65–69] 0.006 (0.009)
[70–75] −0.050*** (0.010)
after the intervention. This appears in the dataset as the individ-
Complementary HI (ref: public)
ual is fully reimbursed by the Health Insurance funds (“Affection Private complementary HI 0.071*** (0.011)
de Longue Durée” status). We were not able to obtain the type of Gynecologist visit (ref: No visits) >1 or 0.100*** (0.009)
chronic disease for personal-data security reasons. more visits
GP visits (ref: No visits)
We measure health-care consumption profiles by the number of
[1–3] 0.085*** (0.009)
GP visits in the year before the intervention and whether the indi- [4–6] 0.110*** (0.009)
vidual consulted a gynecologist in the year before the intervention. ≥7 0.136*** (0.009)
Chronic disease (ref: no chronic disease)
E.2 Descriptive statistics Has a chronic disease −0.033*** (0.006)
No chronic disease status declared 0.018 (0.040)
Département (ref: Eure)
Table A.2 in Appendix C shows the composition of our sample Seine-Maritime 0.026*** (0.005)
in each département. The most frequent screening profile in both
Pseudo R-squared 0.2635
départements is not screening when the last invitation was sent in Observations 26,495
2013. There are more women who refused to participate in screen-
These are estimated marginal effects from a logit model. Standard errors appear in
ing in the Seine-Maritime than in the Eure. We suspect that this
parentheses.
is because Seine-Maritime women who wrote back to the manag- *p < 0.1.
ing structure as they had undertaken opportunistic screening were **
p < 0.05.
***
listed as having screened, and their next invitation was rescheduled p < 0.01.
according to the date of the opportunistic screening. On the con-
trary, women who screened opportunistically in the Eure and wrote
to the managing structure are classified as having refused. The per- Appendix F. Tables for alternative policy objectives
centage of public complementary health-insurance beneficiaries is
higher in the Seine-Maritime than in the Eure. We cannot distin-
guish between those without complementary health insurance and Table A.7
Interaction terms between treatments and deprivation-index quintiles.
those with private insurance in our date, but the national figure
suggests that the group without complementary health insurance Full sample Eure Seine-Maritime
is only a small share of the population (5%). Coef. SE Coef. SE Coef. SE
The great majority of women in the experiment are affiliated to
Deprivation index (ref: 1st quintile)
the “Régime Général” (RG) which is available for all employees. The
2nd quintile 0.002 (0.018) 0.016 (0.026) −0.008 (0.024)
MGEN is the former health-insurance fund of National Education 3rd quintile −0.004 (0.018) 0.013 (0.027) −0.018 (0.024)
system employees, which although now being open to everyone 4th quintile −0.007 (0.018) −0.026 (0.026) 0.005 (0.024)
continues to be mainly composed of National Education employ- 5th quintile −0.023 (0.018) 0.012 (0.026) −0.048** (0.023)
ees (professors, teachers etc.). Almost one-third of women in the Control group
Logo 0.015 (0.018) 0.010 (0.027) 0.016 (0.024)
sample have or have had in recent years a chronic disease that is Chosen 0.021 (0.018) 0.026 (0.026) 0.014 (0.024)
fully reimbursed by the National Health Insurance funds. 22% of Norm 0.006 (0.018) −0.021 (0.026) 0.021 (0.023)
women in the Eure consulted a GP seven or more times the year L+C −0.007 (0.018) 0.001 (0.026) −0.015 (0.023)
before the intervention, while in the Seine-Maritime this propor- 2nd quintile × Logo −0.017 (0.025) −0.044 (0.037) −0.001 (0.033)
tion is 31.3%. Roughly 10% of women in both départements visited 2nd quintile × Chosen −0.014 (0.025) −0.028 (0.037) −0.003 (0.033)
a gynecologist the year before the intervention. 2nd quintile × Norm −0.002 (0.025) 0.003 (0.037) −0.007 (0.034)
2nd quintile × L + C −0.019 (0.025) −0.037 (0.037) −0.005 (0.033)
0.000 (0.000) 0.000 (0.000) 0.000 (0.000)
E.3 Effect of the covariates on total screening
3rd quintile × Logo −0.032 (0.025) −0.036 (0.038) −0.023 (0.033)
3rd quintile × Chosen −0.033 (0.025) −0.035 (0.038) −0.027 (0.033)
Table A.6 3rd quintile × Norm −0.005 (0.025) −0.010 (0.038) 0.003 (0.033)
Estimation of the marginal effects of covariates on total screening. 3rd quintile × L + C 0.012 (0.025) −0.004 (0.038) 0.031 (0.033)

Health Insurance fund (ref: Regime General) 4th quintile × Logo −0.035 (0.025) −0.029 (0.038) −0.042 (0.033)
MSA 0.050 (0.042) 4th quintile × Chosen −0.042* (0.025) 0.001 (0.038) −0.068** (0.033)
RSI 0.048 (0.042) 4th quintile × Norm −0.018 (0.025) 0.046 (0.037) −0.059* (0.033)
MGEN 0.011 (0.012) 4th quintile × L + C 0.029 (0.025) 0.032 (0.037) 0.020 (0.033)
Other −0.052*** (0.014)
5th quintile × Logo −0.009 (0.025) −0.029 (0.037) 0.002 (0.033)
Deprivation index (ref: 1st quintile)
5th quintile × Chosen −0.036 (0.025) −0.063* (0.037) −0.017 (0.033)
2nd quintile −0.009 (0.008)
5th quintile × Norm −0.008 (0.025) −0.009 (0.037) −0.004 (0.033)
3rd quintile −0.016** (0.008)
5th quintile × L + C −0.007 (0.025) −0.031 (0.037) 0.010 (0.033)
4th quintle −0.021*** (0.008)
5th quintile −0.036*** (0.008)
 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252 251

Table A.7 (Continued) Table A.9 (Continued)

Full sample Eure Seine-Maritime Opportunistic screening

Coef. SE Coef. SE Coef. SE Total sample Seine-Maritime Eure

Covariates Yes Yes Yes (7) (8) (9) (10) (11) (12)
Observations 26,495 10,411 16,084
* ** **
R-squared 0.337 0.415 0.296 Logo 0.005 0.005 0.008 0.008 0.001 0.001
(0.003) (0.003) (0.004) (0.004) (0.005) (0.005)
These are linear-probability models. Standard errors appear in parentheses. Chosen 0.001 0.002 0.001 0.002 0.001 0.002
*
p < 0.1. (0.003) (0.003) (0.003) (0.003) (0.005) (0.005)
**
p < 0.05. Norm 0.002 0.002 0.003 0.003 0.001 0.001
***p < 0.01. (0.003) (0.003) (0.004) (0.003) (0.005) (0.005)
L+C 0.007** 0.007** 0.005 0.005 0.010* 0.009*
(0.003) (0.003) (0.004) (0.004) (0.006) (0.005)
Table A.8
p-value of joint 0.1352 0.1787 0.2373 0.274 0.3383 0.4175
Treatment effects on organized and opportunistic screenings in each département.
hyp. test
Seine-Maritime Eure
Covariates No Yes No Yes No Yes
N % screening T-C N % screening T-C Observations 26,495 26,495 16,084 16,084 10,411 10,411

Organized screening These are estimated marginal effects from logit models. Standard errors appear in
Control 3186 46.5 2091 43.5 parentheses.
*
Logo 3227 45.4 −1.1 2069 41.0 −2.5 p < 0.1.
**
(4.2) (1.5) p < 0.05.
Chosen 3214 45.2 −1.2 2101 43.2 −0.3 ***p < 0.01.
(4.2) (1.5)
Norm 3234 46.3 −0.2 2073 41.9 −1.6
(4.2) (1.5)
References
L+C 3223 44.9 −1.6 2077 42.2 −1.3
(4.1) (1.5) Altobelli, E., Lattanzi, A., 2014. Breast cancer in European Union: an update of
screening programmes as of March 2014. Int. J. Oncol. 45 (5), 1785–1792
Total 16,084 45.7 10,411 42.4 (review).
Ayres, I., Raseman, S., Shih, A., 2013. Evidence from two large field experiments
Opportunistic screening that peer comparison feedback can reduce residential energy usage. J. Law
Econ. Organ. 29 (5), 992–1022.
Control 3186 1.9 2091 2.9
Bertrand, M., Mullainathan, S., Shafir, E., 2006. Behavioral economics and
Logo 3227 2.7 0.8** 2069 3.0 0.1
marketing in aid of decision making among the poor. J. Public Policy Mark. 25
(0.4) (0.5)
(1), 8–23.
Chosen 3214 2.1 0.1 2101 3 0.1 Beshears, J., Choi, J.J., Laibson, D., Madrian, B.C., Milkman, K.L., 2015. The effect of
(0.3) (0.5) providing peer information on retirement savings decisions. J. Finance 70 (3),
Norm 3234 2.2 0.3 2073 3.0 0.1 1161–1201.
(0.3) (0.5) Bhargava, S., Manoli, D., 2015. Psychological frictions and the incomplete take-up
L+C 3223 2.4 0.5 2077 3.9 1.0* of social benefits: evidence from an IRS field experiment. Am. Econ. Rev. 105
(0.4) (0.6) (11), 3489–3529.
Bodurtha, J., Quillin, J.M., Tracy, K.A., Borzelleca, J., McClish, D., Baer Wilson, D.,
Total 16,084 2.3 10,411 3.2 Jones, R.M., Quillin, J., Bowen, D., 2009. Mammography screening after
Standard errors appear in parentheses. Differences in bold indicate statistically- risk-tailored messages: the Women Improving Screening through Education
and Risk Assessment (WISER) randomized, controlled trial. J. Women’s Health
significant adjusted differences.
*
18 (1), 41–47.
p < 0.1.
**
Bourmaud, A., Soler-Michel, P., Oriol, M., Regnier, V., Tinquaut, F., Nourissat, A.,
p < 0.05. Bremond, A., Moumjid, N., Chauvin, F., 2016. Decision aid on breast cancer
***p < 0.01. screening reduces attendance rate: results of a large-scale, randomized,
controlled study by the DECIDEO group. Oncotarget 7 (11), 12885.
Bronchetti, E.T., Huffman, D.B., Magenheim, E., 2015. Attention, intentions, and
Table A.9
follow-through in preventive health behavior: field experimental evidence on
Estimated treatment effects on organized and opportunistic screenings.
flu vaccination. J. Econ. Behav. Organ. 116, 270–291.
Organized screening Cai, H., Chen, Y., Fang, H., 2009. Observational learning: evidence from a
randomized natural field experiment. Am. Econ. Rev. 99 (3), 864–882.
Total sample Seine-Maritime Eure Carneiro, P., Heckman, J.J., Vytlacil, E.J., 2011. Estimating marginal returns to
education. Am. Econ. Rev. 101 (6), 2754–2781.
(1) (2) (3) (4) (5) (6) Carrieri, V., Wübker, A., 2013. Assessing inequalities in preventive care use in
Europe. Health Policy 113 (3), 247–257.
Logo −0.016* −0.009 −0.011 −0.005 −0.025 −0.020*
Carrieri, V., Wuebker, A., 2016. Quasi-experimental evidence on the effects of
(0.010) (0.008) (0.012) (0.010) (0.015) (0.011) health information on preventive behaviour in Europe. Oxf. Bull. Econ. Stat. 78
Chosen −0.009 −0.006 −0.012 −0.011 −0.003 −0.001 (6), 765–791.
(0.010) (0.008) (0.012) (0.010) (0.015) (0.011) Cialdini, R.B., 2003. Crafting normative messages to protect the environment. Curr.
Norm −0.007 −0.002 −0.002 0.004 −0.016 −0.016 Dir. Psychol. Sci. 12 (4), 105–109.
(0.010) (0.008) (0.012) (0.010) (0.015) (0.011) Devaux, M., 2015. Income-related inequalities and inequities in health care
L+C −0.014 −0.011 −0.016 −0.009 −0.012 −0.016 services utilisation in 18 selected OECD countries. Eur. J. Health Econ. 16 (1),
(0.010) (0.008) (0.012) (0.010) (0.015) (0.011) 21–33.
Galizzi, M.M., 2012. Label, nudge or tax? A review of health policies for risky
p-value of joint 0.4737 0.6085 0.6564 0.5451 0.4928 0.2955 behaviours. J. Public Health Res. 1 (1), 14.
hyp. test Galizzi, M.M., 2014. What is really behavioral in behavioral health policy? And
Covariates No Yes No Yes No Yes does it work? Appl. Econ. Perspect. Policy 36 (1), 25–60.
Observations 26,495 26,495 16,084 16,084 10,411 10,411 Ganguly, A., Tasoff, J., 2017. Fantasy and Dread: the demand for information and
the consumption utility of the future. Manag. Sci. 63 (12), 4037–4060.
252 L. Goldzahl et al. / Journal of Health Economics 58 (2018) 228–252
Gerber, A.S., Rogers, T., 2009. Descriptive social norms and motivation to vote:
everybody’s voting and so should you. J. Polit. 71 (1), 178–191.
Goldzahl, L., Jusot, F., 2017. Determinants of breast cancer screening regularity in
France. Rev. Fr. Econ. XXXI (4).
Golman, R., Hagmann, D., Loewenstein, G., 2017. Information avoidance. J. Econ.
Lit. 55 (1), 96–135.
Griskevicius, V., Cialdini, R.B., Goldstein, N.J., 2008. Applying (and resisting) peer
influence. MIT Sloan Manag. Rev. 49 (2), 84–88.
Guillaume, E., Pornet, C., Dejardin, O., Launay, L., Lillini, R., Vercelli, M.,
Marí-Dell’Olmo, M., Fernández Fontelo, A., Borrell, C., Ribeiro, A.I., et al., 2016.
Development of a cross-cultural deprivation index in five European countries.
J. Epidemiol. Community Health 70 (5), 493–499.
Gupta, S., Miller, S., Koch, M., Berry, E., Anderson, P., Pruitt, S.L., Borton, E., Hughes,
A.E., Carter, E., Hernandez, S., et al., 2016. Financial incentives for promoting
colorectal cancer screening: a randomized, comparative effectiveness trial.
Am. J. Gastroenterol. 111 (11), 1630–1636.
Holm, S., 1979. A simple sequentially rejective multiple test procedure. Scand. J.
Stat., 65–70.
Hsia, J., Kemper, E., Kiefe, C., Zapka, J., Sofaer, S., Pettinger, M., Bowen, D., Limacher,
M., Lillington, L., Mason, E., 2000. The importance of health insurance as a
determinant of cancer screening: evidence from the Women’s Health
Initiative. Prev. Med. 31 (3), 261–270.
Jusot, F., Or, Z., Sirven, N., 2012. Variations in preventive care utilisation in Europe.
Eur. J. Ageing 9 (1), 15–25.
Kahneman, D., 2003. Maps of bounded rationality: psychology for behavioral
economics. Am. Econ. Rev. 93 (5), 1449–1475.
Keizer, K., Lindenberg, S., Steg, L., 2008. The spreading of disorder. Science 322
(5908), 1681–1685.
Kling, J.R., Liebman, J.B., Katz, L.F., 2007. Experimental analysis of neighborhood
effects. Econometrica 75 (1), 83–119.
Kullgren, J.T., Dicks, T.N., Fu, X., Richardson, D., Tzanis, G.L., Tobi, M., Marcus, S.C.,
2014. Financial incentives for completion of fecal occult blood tests among
veterans, a 2-stage, pragmatic, cluster, randomized, controlled trial. Ann.
Intern. Med. 161 (10 Supplement), S35–S43.
Lee, S., Shaikh, A.M., 2014. Multiple testing and heterogeneous treatment effects:
re-evaluating the effect of Progresa on school enrollment. J. Appl. Econom. 29
(4), 612–626.
Lipkus, I.M., Skinner, C.S., Dement, J., Pompeii, L., Moser, B., Samsa, G.P., Ransohoff,
D., 2005. Increasing colorectal cancer screening among individuals in the
carpentry trade: test of risk communication interventions. Prev. Med. 40 (5),
489–501.
List, J.A., Shaikh, A.M., Xu, Y. 2016. Multiple hypothesis testing in experimental
economics, WP 21875.
Merrick, E.L., Hodgkin, D., Horgan, C.M., Lorenz, L.S., Panas, L., Ritter, G.A., Kasuba,
P., Poskanzer, D., Nefussy, R.A., 2015. Testing novel patient financial incentives
to increase breast cancer screening. Am. J. Manag. Care 21 (11), 771–779.
Myers, R.E., Hyslop, T., Sifri, R., Bittner-Fagan, H., Katurakes, N.C., Cocroft, J.,
DiCarlo, M., Wolf, T., 2008. Tailored navigation in colorectal cancer screening.
Med. Care 46 (9), S123–S131.
Narula, T., Ramprasad, C., Ruggs, E.N., Hebl, M.R., 2014. Increasing colonoscopies? A
psychological perspective on opting in versus opting out. Health Psychol. 33
(11), 1426.
Pace, L.E., Keating, N.L., 2014. A systematic assessment of benefits and risks to
guide breast cancer screening decisions. JAMA 311 (13), 1327–1335.
Pornet, C., Delpierre, C., Dejardin, O., Grosclaude, P., Launay, L., Guittet, L., Lang, T.,
Launoy, G., 2012. Construction of an adaptable European transnational
ecological deprivation index: the French version. J. Epidemiol. Community
Health 66 (11), 982–989.
Salganik, M.J., Dodds, P.S., Watts, D.J., 2006. Experimental study of inequality and
unpredictability in an artificial cultural market. Science 311 (5762), 854–856.
Schultz, P. Wesley, Nolan, J.M., Cialdini, R.B., Goldstein, N.J., Griskevicius, V., 2007.
The constructive, destructive, and reconstructive power of social norms.
Psychol. Sci. 18 (5), 429–434.
Segnan, N., Senore, C., Giordano, L., Ponti, A., Ronco, G., 1997. Promoting
participation in a population screening program for breast and cervical cancer:
a randomized trial of different invitation strategies. Tumori 84 (3), 348–353.
Senore, C., Ederle, A., DePretis, G., Magnani, C., Canuti, D., Deandrea, S., Zorzi, M.,
Barca, A., Bestagini, P., Faitini, K., et al., 2015. Invitation strategies for colorectal
cancer screening programmes: the impact of an advance notification letter.
Prev. Med. 73, 106–111.
Sicsic, J., Franc, C., 2014. Obstacles to the uptake of breast, cervical, and colorectal
cancer screenings: what remains to be achieved by French national
programmes? BMC Health Serv. Res. 14 (1), 465.
Slater, J.S., Henly, G.A., Nim Ha, C., Malone, M.E., Nyman, J.A., Diaz, S., McGovern,
P.G., 2005. Effect of direct mail as a population-based strategy to increase
mammography use among low-income underinsured women ages 40 to 64
years. Cancer Epidemiol. Prev. Biomark. 14 (10), 2346–2352.
Stein, K., Lewendon, G., Jenkins, R., Davis, C., 2005. Improving uptake of cervical
cancer screening in women with prolonged history of non-attendance for
screening: a randomized trial of enhanced invitation methods. J. Med. Screen.
12 (4), 185–189.
Thaler, R.H., Sunstein, C.R., 2008. Nudge: Improving Decisions About Health,
Wealth, and Happiness. Yale University Press.
Trivedi, A.N., Rakowski, W., Ayanian, J.Z., 2008. Effect of cost sharing on screening
mammography in Medicare health plans. N. Engl. J. Med. 358 (4), 375–383.
Vallée, A., 2016. Le «Mammobile»: une méthode pertinente en France? Santé
Publique 28 (5), 599–602.
Vernon, S.W., del Junco, D.J., Tiro, J.A., Coan, S.P., Perz, C.A., Bastian, L.A., Rakowski,
W., Chan, W., Lairson, D.R., McQueen, A., et al., 2008. Promoting regular
mammography screening II. Results from a randomized controlled trial in US
women veterans. J. Natl. Cancer Inst. 100 (5), 347–358.
Wardle, J., von Wagner, C., Kralj-Hans, I., Halloran, S.P., Smith, S.G., McGregor, L.M.,
Vart, G., Howe, R., Snowball, J., Handley, G., et al., 2016. Effects of
evidence-based strategies to reduce the socioeconomic gradient of uptake in
the English NHS Bowel Cancer Screening Programme (ASCEND): four
cluster-randomised controlled trials. Lancet 387 (10020), 751–759.
Webb, T.L., Sheeran, P., 2006. Does changing behavioral intentions engender
behavior change? A meta-analysis of the experimental evidence. Psychol. Bull.
132 (2), 249.
Wübker, A., 2014. Explaining variations in breast cancer screening across European
countries. Eur. J. Health Econ. 15 (5), 497–514.