Reference-dependent preferences and loss aversion: A discrete choice experiment in the health-care sector

Judgment and Decision Making, Vol. 3, No. 2, February 2008, pp. 162–173
Reference-dependent preferences and loss aversion: A discrete
choice experiment in the health-care sector
Einat Neuman
Shoshana Neuman?
Department of Economics and Business Administration
Department of Economics
University Center of Ariel
Bar-Ilan University
Abstract
This study employs a Discrete Choice Experiment (DCE) in the health-care sector to test the loss aversion theory that
is derived from reference-dependent preferences: The absolute subjective value of a deviation from a reference point
is generally greater when the deviation represents a loss than when the same-sized change is perceived as a gain. As
far as is known, this paper is the ?rst to use a DCE to test the loss aversion theory. A DCE is a highly suitable tool
for such testing because it estimates the marginal valuations of attributes, based on deviations from a reference point
(a constant scenario). Moreover, loss aversion can be examined for each attribute separately. Another advantage of a
DCE is that is can be applied to non-traded goods with non-tangible attributes. A health-care event is used for empirical
illustration: The loss aversion theory is tested within the context of preference structures for maternity-ward attributes,
estimated using data gathered from 3850 observations made by a sample of 542 women who had recently given birth.
Seven hypotheses are presented and tested. Overall, signi?cant support for behavioral loss aversion theories was found.
Keywords: preferences, attributes, loss aversion, reference dependence, discrete choice experiment, maternity-wards.
1 Introduction
extended model of reference-dependent preferences and
loss aversion, which is claimed to be more generally ap-
A person’s valuation of the bene?t from an outcome of
plicable. Numerous studies present evidence supporting
a choice is often determined by the intrinsic “consump-
the loss aversion hypothesis. They include: Hartman et
tion utility” of the outcome itself, combined with its con-
al. (1991); Hardie et al. (1993); Andreoni (1995); Be-
trast with a reference point. The most noteworthy mani-
nartzi and Thaler (1995); Camerer et al. (1997); Myagkov
festation of such reference-dependent preferences is loss
and Plott (1998); Bowman et al. (1999); Jullien and
aversion: the absolute subjective value of a change in an
Salanie (2000); Genesove and Mayer (2001).There are
endowment is generally greater when the deviation from
also several studies that are looking at loss aversion in the
the reference point represents a loss than when the same-
medical domain. They include: Stalmeier and Bezem-
sized change is perceived as a gain.
binder (1999); Robinson et al. (2001); Bleichrodt and
The most systematic general theory of this kind is
Pinto (2002); van Osh et al. (2004).
probably Tversky and Kahneman’s (1991) reference-
dependence model, which builds on Kahneman and Tver-
What is the reference point that is used by the individ-
sky’s (1979) Prospect Theory. The signi?cance of loss
ual to evaluate gains (positive deviations) versus losses
aversion is highlighted in Camerer’s (2000) review of the
(negative deviations)? The majority of the empirical stud-
practical implications of prospect theory: seven out of the
ies examined traded goods, and the reference point was
ten examples are derived from the loss aversion hypoth-
the endowment of the commodity under consideration.
esis. Recently, Koszegi and Rabin (2006) presented an
Expectations were mentioned by other researchers as can-
didates for the reference point: Shalev (2000) used ex-
?We would like to thank the editor and two referees for very helpful
pectations in his game-theoretic model; and Koszegi and
comments and suggestions. Part of this study was completed at the time
Rabin (2006) assumed that a person’s reference point is
one of the authors (Shoshana) was staying at IZA (summer 2006 and
2007). I would like to thank IZA for their hospitality and excellent re-
her rational expectations held in the recent past about out-
search facilities. In particular, thanks are due to Margard Ody, the IZA
comes. They speci?ed a rule for the endogenous deriva-
information manager, whose incredible pro?ciency and promptness
tion of this point, within the framework of an equilibrium
brought to my table (computer) all requested publications. Shoshona
utility-maximizing model. van Osch et al. (2006) argued
Neuman is also af?liated with CEPR, London. Address: Shoshona
Neuman, Department of Economics, Bar-Ilan University, Ramat-Gan,
that goals (aspirations) in?uence the reference point in
Israel. Email, neumans@mail.biu.ac.il
the health domain. Combining qualitative and quantita-
162

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
163
tive data they provided evidence of the reference point in
was conducted in maternity-wards in three large public
life-year certainty equivalent (CE) gambles and explored
hospitals located in the Greater Tel-Aviv area in Israel.
the psychology behind the reference point.
Women who had given birth were approached by inter-
This empirical study takes a new and different ap-
viewers and requested to ?ll out a questionnaire. In the
proach to the determination of the reference point and
questionnaire, DCE was used to present individuals with
the testing of the loss aversion hypothesis.
Discrete
a series of pairs of hypothetical scenarios (maternity-
Choice Experiments (DCEs) are used for the estimation
wards), which were described in terms of some relevant
of a preference structure for a multi-dimensional con-
attributes with different levels in the various scenarios.
sumption good or service, by establishing the relative
For each pair of scenarios the subjects were asked to
importance of different attributes in the provision of the
choose which they prefer. It is assumed that subjects will
good/service under discussion, vis-à-vis a constant refer-
choose the alternative that provides the higher level of
ence bundle. It follows that the employment of a DCE
utility. A DCE setup is appropriate for the analysis of
also facilitates the testing of the loss aversion hypothe-
public health-care services such as a delivery where rele-
sis for each attribute separately. The decomposition into
vant revealed-preference data are unavailable.2 It is also
attribute-speci?c components adds richness and insight:
especially appropriate for the assessment of utilities of in-
It is not obvious that people are loss averse regarding all
tangible characteristics, such as, information transferred
kinds of attributes and there are most probably different
from supplier to purchaser or attitude of supplier/staff.
degrees of loss aversion, which can be compared across
The attributes, their levels and the wording used in the
attributes. As far as is known, this is the ?rst published
questionnaire to describe the attributes and levels, were
study that employs DCEs to test attribute-speci?c loss
determined during three preliminary stages: (i) A lit-
aversion that is derived from reference-dependent pref-
erature survey (the most relevant studies are: McGuirk
erences.
and Porell, 1984; Rahtz and Moore, 1988; Bronstein and
The empirical illustration presented in this paper re-
Morrisey, 1991; Phibbs et al., 1993; Brown and Lumley,
lates to a DCE that was conducted among 542 women
1994; Wilcock et al., 1997; Janssen et al., 2000; Sadler
who had recently given birth. Their preferences for ?ve
et al., 2001); (ii) in-depth face-to-face interviews with
maternity-ward attributes (number of beds in hospital
ten women who had recently given birth; and (iii) a pi-
room; attitude of staff toward the patient; medical staff’s
lot study involving 48 women.
professionalism; information transfer from staff to pa-
The following attributes (levels) were identi?ed: (a)
tients; and travel time from residence to hospital) were
number of beds in hospital room (three beds; two beds;
estimated and loss aversion was tested for each of the at-
or a private room), (b) attitude of staff towards the patient
tributes separately.
(reasonable; very good), (c) medical staff’s professional-
The main results of the study were that the loss aver-
ism (good; very good), (d) information transfer from staff
sion hypothesis was con?rmed for four of the ?ve hospital
to patient (basic; extensive), and (e) travel time from res-
attributes investigated. The results were less conclusive
idence to hospital (45; 30; or 15 minutes).
for “travel time from residence to hospital.”
The levels of the ?rst attribute - number of beds in
The following section describes the DCE method em-
hospital room - relate to the current facilities in most Is-
ployed for estimation. The third section presents the
raeli maternity-wards, where a standard room has two or
econometric model, followed by the hypotheses derived
three hospitalization beds. There are few private rooms
from the loss aversion theory. The preference structures
and some hospitals have also rooms with more than three
used to test the outlined hypotheses are presented in sec-
beds. A private room as one of the options could also
tion 5. The last section concludes and poses questions
have policy implications: if it will be found that a be-
that merit further research.
ing hospitalized in a private room leads to a signi?cant
increase in utility, hospitals might consider favorably the
costly transformation of multi-bed hospitalization rooms
2 Method
into one-bed rooms.
The three qualitative attributes — attitude of staff,
The statistical tool used to elicit preferences and detect
professionalism of staff, and transfer of information —
loss aversion was a Discrete Choice Experiment,1 which
have two similar levels each. The levels are similar but
1
not identical — the gap between “reasonable” and “very
DCEs were ?rst introduced in Mathematical Psychology (Luce and
Tukey, 1964; Green et al., 1972) and then adopted by economists for
good” is larger than between “good” and “very good”, be-
use in the ?elds of transportation (e.g., Wardman, 1988), environment
(e.g., Opaluch et al., 1993), marketing (e.g., Cattin and Wittink, 1982
2002).
who survey the marketing literature), and recently in health (e.g., Ryan
2Israel has a public health-care sector. Hospitalization data have
and Hughes, 1997; Bryan et al., 1998; Ryan et al, 1998a, 1998b; Vick
been used to represent revealed preferences, but they suffer from basic
and Scott, 1998; Salkeld et al., 2000; San Miguel et al., 2002; Scott,
statistical and methodological problems.

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
164
Table 1: Attributes, levels and coding in Type 1 and Type 2 questionnaires
Maternity-ward
Ward A1
Ward A2
Ward B (all
Coding of
Xi Type 1
Xi Type 2
attributes
(Type 1
(Type 2
aternatives)
all alter-
(Value of
(Value of
constant)
constant)
natives
independent
independent
variable =
variable =
B?A1)
B?A2)
Number of
2 beds
3 beds
3 beds
3
+1 (loss)
0
beds in room
2 beds
2
0
?1 (gain)
private room
1
?1 (gain)
?2 (gain)
Attitude of staff
reasonable
reasonable
reasonable
0
0
0
(towards the patients)
very good
1
+1 (gain)
+1 (gain)
Professionalism of staff
good
very good
good
0
0
?1 (loss)
very good
1
+1 (gain)
0
Transfer of information
extensive
extensive
basic
0
?1 (loss)
?1 (loss)
from staff to patient
extensive
1
0
0
Travel time
30 minutes
45 minutes
45 minutes
2
+1 (loss)
0
to hospital
30 minutes
1
0
?1 (gain)
15 minutes
0
?1 (gain)
?2 (gain)
Note: “Number of beds” was de?ned by two dummy variables: “3 beds versus 2 beds” and “private room versus 2
beds” in Type 1 regressions; and by “2 beds versus 3 beds” and “private room versus 3 beds” in Type 2 regressions.
“Travel time to hospital” included the two dummy variables of “15 minutes versus 30 minutes” and “45 minutes versus
30 minutes” in Type 1 regressions; and “30 minutes versus 45 minutes” and “15 minutes versus 45 minutes” in Type
2 regressions.
cause hospitals are believed to be more diverse in terms of
vide a fractional factorial orthogonal design.3 The proce-
attitude of staff than in terms of professionalism. Diver-
dure’s application gave rise to 16 different scenarios, each
sity is prevalent also in terms of information transfer and
representing a hypothetical maternity-ward. If all 16 op-
therefore the two levels are “basic” and “extensive.” The
tions were pair-wise compared, a large number of pos-
pilot survey and the pilot interviews indicated that unifor-
sible discrete choices would emerge. To overcome this
mity simpli?es the task of choice between scenarios and
dif?culty, one scenario was randomly chosen to be con-
that respondents were fully aware of the gap between the
stant throughout the questionnaire (scenario A1) and each
two levels that were assigned to each of the attributes.
of the remaining 15 scenarios was compared to this cho-
sen scenario, concluding in 15 pair-wise combinations.
Travel time has the levels of 15, 30 and 45 minutes to
One pair of scenarios had a dominant option (one alterna-
re?ect the fact that actual travel time is relatively short
tive had superior or identical levels for all attributes) and
because several hospitals are located in the center of the
was used to test for internal consistency.4 Realizing that
country (the average actual travel time of the respondents
it is dif?cult for women who recently delivered to cope
from residence to the maternity-ward was 19 minutes,
with 15 complex pair-wise choices, the 15 paired combi-
with minor differences between the three hospitals: 23,
nations were split into two subsets. The dominant option
20 and 16 minutes to each of the hospitals, respectively).
was included in each subset, thus giving rise to two ques-
A gap of 15 minutes between two successive levels seems
tionnaires each with eight choices. The two subsets were
therefore reasonable.
randomly distributed among the respondents.
As will be elaborated below, the constant scenario
A full factorial design that will use all possible
serves as the respondent’s reference point. One of the
combinations of attributes gives rise to 72 scenarios
(3*3*2*2*2 = 72; 2 attributes have 3 possible levels each,
3
and each of the other 3 attributes has 2 alternative levels).
The choice of scenarios that yield an orthogonal design means that
the statistical analysis excludes interactions between attributes.
In order to reduce the number of scenarios to a manage-
4A few women (6), who failed the test by preferring the inferior
able size, the SPSS Orthoplan procedure was used to pro-
alternative, were excluded from the sample.

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
165
Table 2: Discrete choice questions, Type 1 questionnaire.
You can choose for a delivery, either Ward A1 or Ward B. They differ with respect to a number of attributes.
• Assume that all other attributes (on top of the 5 listed ones) are identical in the two wards.
• In each question, Ward A1 is the same and ward B is different.
• Which ward would you prefer? (Please tick box below).
• Please answer all questions.
Question 1
Attributes
Ward A1 (constant)
Ward B
Number of beds
2 beds
private room (1 bed)
Attitude of staff (towards you)
reasonable
reasonable
Professionalism of staff
good
good
Information
extensive
extensive
Travel time to hospital
30 minutes
45 minutes
Prefer Ward B
Prefer Ward A
Question 2
Attributes
Ward A1 (constant)
Ward B
Number of beds
2 beds
3 beds
Attitude of staff (towards you)
reasonable
reasonable
Professionalism of staff
good
good
Information
extensive
basic
Travel time to hospital
30 minutes
15 minutes
Prefer Ward B
Prefer Ward A
devices that will be used for the testing of the loss aver-
two sets (A1 and A2) were different. Table 1 presents the
sion theory is a comparison of preference structure that
two alternative constant scenarios.
are based on different reference points (constant scenar-
Not all attributes changed levels in the two constant hy-
ios). For this purpose, we also experimented with an al-
pothetical scenarios: The attributes of “attitude” and “in-
ternative constant scenario (A2) that was also randomly
formation” displayed the same level in A1 and A2, while
chosen from the full set of 16 orthogonal scenarios. Four
“professionalism” appeared at the lower level (good) in
dominant options were detected within the 15 paired
Questionnaire Type 1 and at the higher level (very good)
combinations. Three were excluded and one left for the
in Questionnaire Type 2. “Number of beds” and “travel
internal consistency test. The remaining pairs of sce-
time” was found on the middle level (out of 3 options) in
narios were also split into two subsets with six or seven
Questionnaire Type 1 and yet on the least desirable level
choices, respectively (the dominant option was included
in Questionnaire Type 2. Table 2 contains an example
in the two subsets) and distributed randomly among the
of two of the pair-wise combinations presented to the re-
interviewed women. The questionnaires based on each of
spondents who ?lled out Questionnaire Type 1.
the two constant wards, are Type 1 and Type 2 question-
In hypothetical Ward B, one or more of the attributes
naires. It follows that all the respondents who received
were “better” than in the constant hypothetical Ward
questionnaire Type 1 had the same constant reference set
A1 (representing a “gain”), one or more attributes were
(A1). All the women who answered questionnaire Type 2
“worse” (representing a “loss”) and the rest identical. In
also had an identical constant reference set (A2), but the
Question 1: Moving from A1 to B results in one “gain”

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
166
and one “loss’, whereas in Question 2 the move leads to
4 The loss-aversion hypotheses
two “losses” and one “gain.” The respondent is in effect
referring to scenario A1 as a reference point while con-
A central statement of the loss aversion theory is that
sidering deviations from A1 to B. She therefore must con-
utility is reference dependent and that individuals value
sider the gains versus the losses and trade-offs between
losses (vis-a-vis the attribute’s reference level), signif-
the attributes when making her quite complex choices.
icantly more (in absolute terms) than they value same-
The same applies to interviewees who completed Ques-
sized gains8: e.g., adding one bed to the number of beds
tionnaire Type 2: they refer to the constant hypothetical
in the reference constant scenario (loss) leads to a larger
scenario A2 as their reference point.
decrease in utility than the increase in utility that is as-
sociated with the removal of one bad (gain). The very
rich data set, generated by the discrete choice experimen-
3 The econometric model
tal design, is used to test several hypotheses, all derived
from the reference-dependence assertion.
Assuming a linear utility function, the marginal change
Two complementary approaches were employed for
in utility when moving from A5 to B is given by
testing hypotheses: The ?rst was a comparison of the
marginal valuation (utility scores) of positive versus neg-
n
ative same-sized deviations of attributes from the con-
?UA?B =
?iXi + u + ?
(1)
stant scenario. Such comparisons can be performed for
i=1
attributes that exhibit at least three possible levels, with
The observed value of the dependent variable of the
the constant scenario including the middle level. The util-
estimated preference equation (?U ) is dichotomous and
ities of same-sized positive and negative deviations can
takes the value of 1 if maternity-ward B is chosen and the
then be compared. If individuals are loss averse, the disu-
value of 0 if maternity-ward A is preferred.
tility of negative deviations will be larger than the utility
The independent variables are theX
of positive deviations;
is, where Xi is the
difference in the level of attribute i between B and A.
The second approach was the use of two questionnaire
They express changes from the reference level of the con-
types, based on different constant scenarios (A1 and A2):
stant scenario and are outlined in Table 1. Each of the
If the respondent was referring to the constant maternity-
two three-level quantitative attributes was de?ned using
ward as her reference point, then different constant sce-
two dummy variables (see note to Table 1).
narios should lead to different estimated utilities for the
?
same attribute. A comparison of the preference structures
i are the parameters of the model that represent
marginal utility scores (relative importance) of the at-
estimated using data generated by the two questionnaires
tributes; u is the error term that represents differences be-
facilitates statistical testing of loss aversion theory. For
tween the various choices of the same respondent (each
instance, assume that a two-level attribute x has different
respondent provides 6–8 discrete choice observations);
levels in the two alternative constant scenarios: In the ?rst
and ? is the error term representing differences between
questionnaire it exhibits the less favorable level and in the
respondents. Interactions between independent variables
second it exhibits the more desirable one. It is expected
are not included because an orthogonal fractional design
that estimates of the marginal utility score of x based on
was used.
data generated by the ?rst questionnaire will be smaller
The data6 compiled from the completed questionnaires
than the respective estimates when based on data from
of the two types were used to estimate two sets of main-
the second questionnaire. The justi?cation for this con-
effects regressions. To account for the fact that each re-
clusion rests on the fact that in the ?rst case a deviation
spondent makes several choices, a Random-Effects Probit
represents a “gain” vis-à-vis the reference point whereas
was used for estimation.7
in the second case, a deviation represents a “loss.” The
DCE is therefore an ef?cient instrument for soliciting
5The constant hypothetical scenario A is denoted by A1 in Ques-
preferences in general and loss aversion components in
tionnaire Type 1 and by A2 in Questionnaire Type 2.
particular.
6Each observation had the value of the dependent variable (either 1
if scenario B was chosen or 0 if A was the preferred scenario) and the
Table 3 illustrates the attribute levels used in the test of
deviations between the levels of scenario B and scenario A. See Table
loss aversion. For the sake of clarity, the attribute levels of
1 for more detail. Each respondent made several choices and therefore
the constant scenarios in the two types of questionnaires
contributed several observations.
7
(A1 and A1), as well as the de?nition of all possible at-
Each respondent had an identi?cation number (ID) that was used
for all the responses of this subject (ID=1 for the 6 choices of woman
tribute levels, are repeated. In the Type 1 Questionnaire
#1; ID=2 for the 6 responses of woman #2 etc.). The statistical Stata
program that was used for estimation was “informed” that ID is the
8This is the de?nition proposed by Kahneman and Tversky (1979).
subject identi?er when applying the random-effects (re) estimation (the
Other plausible de?nitions include those suggested by Wakker and
Stata command is: re, i (ID), where i refers to the subject ).
Tversky (1993) and Kobberling and Wakker (2005).

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
167
Table 3: Loss aversion hypotheses: levels of the constant scenarios that are used in the two types of questionnaires, as
well as all possible attribute levels.
Attributes
Constant ward A1
Constant ward A2
All possible levels
(Type 1 Questionnaire)
(Type 2 Questionnaire)
Number of beds
2 beds
3 beds
1, 2, 3 beds
Attitude of staff
reasonable
reasonable
reasonable, very good
Professionalism of staff
good
very good
good, very good
Information
extensive
extensive
basic, extensive
Travel time to hospital
30 minutes
45 minutes
15, 30, 45 minutes
the attributes “number of beds” and “travel time” exhibit
based on Type 1 Questionnaires it represents the valua-
the middle level in the constant ward A1, facilitating the
tion of a gain (moving from “good” to “very good’, rep-
following hypotheses that are based on the coef?cients of
resented by a difference of +1).
Type 1 regressions (and on the assumption that the degree
of diminishing sensitivity/utility is comparable for gains
Hypothesis 6: A larger absolute coef?cient for “3
and for losses):
beds” (a loss of one bed compared to the “2 bed” ref-
erence level) in the regression based on Type 1 Question-
Hypothesis 1: The coef?cient of the dummy variable
naires will be obtained, in comparison to the coef?cient
for “3 beds’, that represents a loss, will be signi?cantly
of “two beds” (a same-sized gain vis-à-vis the “3 bed”
larger (in absolute value) than the coef?cient of the
reference level) in the regression using the Type 2 Ques-
dummy variable “1 bed” that represents a same-sized
tionnaire data.
gain.
Hypothesis 7: The coef?cient that relates to “travel
time of 45 minutes” in Type 1 regression (moving from
Hypothesis 2: The coef?cient of the dummy variable
“30 minutes’ in the reference level to “45 minutes”, that
for “travel time of 45 minutes” that relates to a loss, will
relates to a loss of 15 minutes) is expected to be signif-
be signi?cantly larger (in absolute terms) than the coef?-
icantly larger (in absolute terms) than the coef?cient of
cient of the “travel time of 15 minutes” dummy variable
“travel time of 30 minutes” in Type 2 regression (same-
that relates to the same-sized gain.
sized gain, moving from the reference level of “45 min-
Turning to a comparison of regression coef?cients
utes” to “30 minutes”).
based on data generated by the two questionnaires, the
The hypotheses testing will be based on the prefer-
following hypotheses can be derived:
ence structure for maternity-ward attributes that were es-
timated using the data generated by the two types of ques-
tionnaires.
Hypothesis 3: The coef?cients of “attitude of staff”
will not statistically differ by questionnaire type, as in the
two questionnaire types they represent the valuation of a
5 Results
gain in attitude.
Five hundred and forty-two (542) women who had given
birth in three large public hospitals located in the Greater
Hypothesis 4: The coef?cients of “information” will
Tel-Aviv area in Israel9 comprised the primary study
not statistically differ by questionnaire type, as in the two
sample. They were surveyed while still in the hospi-
questionnaire types they represent the valuation of a loss
tal maternity-wards, by interviewers who provided ex-
in information.
planations and instructions10. The overall response rate
was about 50% and 542 questionnaires have been fully
Hypothesis 5: The coef?cient of “professionalism of
9The Rabin medical Center (in Petach Tikva), Sheba (in Ramat-Gan)
staff” will be signi?cantly larger in the regressions based
and Meir (in Kfar Saba).
10
on Type 2 Questionnaires where it relates to the valua-
It is recognized in the literature that interviews are the most effec-
tive and appropriate means for conducting DCEs, even though they are
tion of a loss (moving from “very good” to “good’, repre-
rarely used, due to their high costs. Postal questionnaires are regularly
sented by a difference of ?1), whereas in the regressions
used instead.

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
168
Table 4: Main-effects regressions with two different constant scenarios: Women who gave birth — Israel, 2003.
Explanatory variables
Coef?cients
Explanatory variables
Coef?cients
Type 1 Questionnaires
Type 1 Questionnaires
Type 2 Questionnaires
Type 2 Questionnaires
Number of beds
Number of beds
(reference: 2 beds)
(reference: 3 beds)
Three beds (1 more)
?0.7115 (8.32)
Two beds (1 less)
0.0524 (0.67)
Private room (1 less)
0.2137 (2.23)
Private room (2 less)
0.1098 (1.38)
Attitude (reasonable=0;
1.1691 (14.69)
Attitude (reasonable=0;
1.4567 (17.11)
very good=1)
very good=1)
Professionalism of staff
1.1655 (15.76)
Professionalism of staff
1.8153 (20.35)
(good=0; very good=1))
(very good=0; good=-1)
Information (extensive=0;
0.5882 (8.48)
Information (extensive=0;
0.8402 (11.61)
basic=-1)
basic=-1)
Travel time to hospital
Travel time to hospital
(reference: 30 minutes)
(reference: 45 minutes)
45 minutes (15 more)
?0.3270 (4.13)
30 minutes (15 less)
0.4794 (5.69)
15 minutes (15 less)
0.5077 (4.89)
15 minutes (30 less)
0.5093 (5.25)
Sample size
1751
Sample size
2099
Number of women
219
Number of women
323
Log Likelihood
?800.25
Log Likelihood
?919.81
?
0.1816
?
0.4460
?2 to test ? = 0
32.36 (p = 0.00)
?2 to test ? = 0
172.02 (p = 0.00)
Note: The coef?cients of the following pairs of attributes
Note: The coef?cients of the following pairs of attributes
are not signi?cantly different (at a signi?cance level of
are not signi?cantly different (at a signi?cance level of
0.05): Attitude and Professionalism; Time of 15 minutes
0.05): Two beds and Private room; Time of 15 minutes
more and of 15 minutes less (in absolute values); Infor-
less and Time of 30 minutes less.
mation and Time of 15 minutes less; A private room and
Time of 15 minutes more (in absolute value); Three beds
and Information (in absolute value).
Notes: 1. Z statistics are in parentheses.
2. Stata 9 was used for estimation (Random-Effect Probit, with no constant).
3. The constant set in Type 1 questionnaires has the following attributes: Number of beds, 2; Attitude, reasonable;
Professionalism of staff, good; Information, extensive; Travel time, 30 minutes. The constant set for Type 2 ques-
tionnaires has the following attributes: Number of beds, 3; Attitude, reasonable; Professionalism of staff, very good;
Information, extensive; Travel time, 45 minutes. The levels of Attitude and Information are therefore the same. Levels
of all other attributes are different.
4. The signi?cance of the differences between the main-effects of two groups (not reported) are derived from a ?2test
for equality of coef?cients.
completed. Questionnaire Type 1 has been ?lled out by
their ?rst delivery; the rest, 69%, had had two or more de-
219 women (109 and 110 women completed the two sub-
liveries.11 Over one quarter (28%) had undergone high-
versions, respectively), with a total of 1751 observations.
Questionnaire Type 2 has been completed by 323 women
11Another interesting issue is the effect of experience on preferences.
(161 women completed the ?rst version and 162 the sec-
Neuman and Neuman (2007) used the same data set to examine whether
ond one), with 2099 observations.
experience changes preferences. They compared preferences of three
sub-groups of women: (i) women in pre-natal classes who were preg-
The age range of the participants in the sample was 18–
nant with their ?rst child (no experience); (ii) women who gave birth
for the ?rst time (single experience); and (iii) women who had more
47: average age was 31. Four percent were aged over 40.
than one delivery (multiple experience). It was found that there were no
Thirty-one percent of the interviewees were experiencing
signi?cant differences in preferences of the latter two sub-groups, but

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
169
Table 5: Loss aversion hypotheses results: The following table summarizes the relevant regression results and the
conclusions of seven hypotheses concerning loss aversion.
Hypotheses
Regression Results
Conclusion:
Hypothesis . . .
1. in Type 1 regression: absolute coef?cient of “3 beds” (loss) >
0.711 > 0.214 signi?cant
accepted
coef?cient of “private room” (gain)
difference
(p = 0.001)
2. in Type 1 regression: absolute coef?cient of “45 minutes” (loss) >
not signi?cantly different
not accepted
coef?cient of “15 minutes” (gain)
(?s of 0.327 and 0.508)
(p = 0.238)
3. coef?cient of “attitude” in Type 1 regression (gain) = coef?cient
not signi?cantly different
accepted
of “attitude” in Type 2 regression (gain)
(?s of 1.169 and 1.457)
(p = 0.290)
4. coef?cient of “information” in Type 1 regression (loss) = coef?-
not signi?cantly different
accepted
cient of “information” in Type 2 (loss)
(?s of 0.588 and 0.840)
(p = 0.123)
5. coef?cient of “professionalism” in Type 1 (gain) < coef?cient of
1.165 < 1.815 signi?cant
accepted
“professionalism” in Type 2 (loss)
difference
(p = 0.000)
6. absolute coef?cient of “3 beds” in Type 1 (loss) > coef?cient of “2
0.711 > 0.052 (the latter
accepted
beds” in Type 2 (gain)
not signi?cant)
(p = 0.000)
7. absolute coef?cient of “45 minutes” in Type 1 (loss of 15 minutes)
not signi?cantly different
not accepted
> coef?cient of “30 minutes” in Type 2 (same-sized gain)
(?s of 0.327 and 0.479)
(p = 0.366)
The p value is derived from the test for the equality of the respective coef?cients.
risk pregnancies. The socio-economic characteristics of
= 0.50; and even to a private room; p = 0.17). But, in
the sample were representative of the general Israeli pop-
Type 1 regressions a change in the number of beds has
ulation for the relevant age group (see Appendix 1for de-
a signi?cant effect on utility (or disutility): moving from
tails).
a two-bed room to a three-bed room results in a signi?-
cant drop of 0.711 in utility score (Z = 8.32; p = 0.00),
5.1 Main-effects preference structures
and improving the room conditions from a two-bed room
to a private room leads to a signi?cant increase in utility
Table 4 presents the main-effects preference structure of
(coef?cient of 0.214; Z = 2.23; p = 0.025).
maternity-ward attributes.
Before examining if our respondents are loss averse, by
5.2 Testing the loss aversion hypotheses
testing the seven hypotheses that were formulated above,
it is instructive to investigate their preferences for these
Table 5 summarizes the tests and conclusions of the seven
attributes. However, these two issues are intertwined: if
hypotheses outlined above. The signi?cance of the differ-
loss aversion is in evidence, then the estimated preference
ence between respective coef?cients of the two regression
structure will depend on the reference point (constant sce-
equations (Table 4) are derived from a ?2 test for equality
nario) and data generated by experiments that are based
of coef?cients.
on different reference points, will lead to different prefer-
Discussion of the results follows:
ence equations.
Indeed, Table 4 indicates that the two respective pref-
Hypothesis 1: In Type 1 regressions: The absolute co-
erence structures differ, not only in size of attribute coef-
ef?cient of the dummy variable for “3 beds” is more than
?cients but also in the ranking of attributes’ utility: The
three times larger compared to the coef?cient of a “pri-
most striking difference is related to the room facilities.
vate room” (respective coef?cients of -0.711 and 0.214),
In Type 2 regressions the two dummy variables that re-
indicating that a loss (one bed more) is much more ap-
late to the number of beds are not signi?cant, indicating
preciated than a same-sized gain (one bed less) and thus
that the interviewed women do not gain utility from a de-
supporting Hypothesis 1. Moreover, the results based on
crease in the number of beds (from 3 beds to 2 beds; p
the Type 2 Questionnaire indicate that gains are not sig-
ni?cantly valued at all. A two-bed hospital room or even
those who had no experience at all exhibited different preference pat-
a private room are not valued more than a three-bed room
terns than those with any experience. In the study reported in this paper
we excluded the ?rst sub-group and combined the other two that have
(the A2 scenario reference level), as is demonstrated by
the same preference pattern.
the insigni?cant coef?cients of the two dummy variables.

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
170
Hypothesis 2: Hypothesis 2 that traveling 15 minutes
questionnaires the deviation from the reference scenario
more (a loss in relation to the reference level of 30 min-
is associated with a loss (a change from “very good” to
utes) is more negatively valued than traveling 15 minutes
“good”).
less (a gain), is not supported. As Type 1 regressions in-
dicate, the difference between the coef?cients of the two
Hypothesis 6: The marginal utility score of “3 beds”
dummy variables in not signi?cant (p = 0.238, for the test
in responses to Type 1 Questionnaire (absolute value of
of equality of the two coef?cients), indicating that the in-
0.711) is 14 times (!) larger compared to that of “2 beds”
terviewed women have a similar valuation of a loss and a
in the responses to the Type 2 Questionnaires (insignif-
same-sized gain.
icant coef?cient of 0.052). The former represents a loss
compared to the “2 bed” reference level, while the latter
Hypothesis 3: We ?nd support for Hypothesis 3 that ar-
manifests a gain versus the “3 bed” reference level. It
gues that the coef?cients of “attitude of staff” are not sig-
appears that a gain, in terms of fewer beds in the room,
ni?cantly different in the two regressions, based on Type
is not appreciated by the women in our sample while a
1 and Type 2 Questionnaires (p = 0.290, for the test of
loss (more beds) is painful and highly (negatively) val-
equality of coef?cients).12. The two coef?cients repre-
ued. This is a very distinct recon?rmation of the asym-
sent a gain compared to the identical reference point in
metry between gains and same-sized losses.
the two questionnaires that is “reasonable attitude” and
therefore give an estimate of the marginal utility score of
a gain in attitude.
Hypothesis 7: The difference between a 15-minutes
gain in travel time (a decrease from 45 minutes to 30
minutes, Type 2 regression) and a same-sized loss (an
Hypothesis 4: Hypothesis 4 is also supported as the
increase from 30 to 45, type 1 regression) is not statis-
difference between the corresponding coef?cients of the
tically signi?cant (p = 0.238). Hypothesis 7 is therefore
“information” attribute in Type 1 and Type 2 regressions
not supported.
is not signi?cant (p = 0.123). In the two questionnaires
This result is consistent with the rejection of Hypoth-
this attribute is assigned the same level of “extensive in-
esis 2 and also with the observation that in the Type 2
formation’, indicating that the estimated coef?cients re-
regression, insigni?cant differences were found between
late to the marginal valuation of a loss of information,
the valuations of 15 and 30 minutes less travel time, i.e.
from “extensive” to “basic.”
the two different gains in time have a similar marginal
Moreover, combining the two pieces of evidence,
utility score.
namely that the marginal utility score of the “attitude”
To conclude, ?ve of the seven hypotheses have been
attribute relates to a gain in attitude, while the marginal
supported by the regression results, indicating that loss
utility score of “information” is associated with a loss in
aversion is relevant for the attributes of “professional-
information (that is more highly valued than a gain), we
ism,” “attitude,” “information” and in particular “num-
can speculate that using the difference between the cor-
ber of beds in hospital room.” The rest two hypotheses,
responding coef?cients in order to evaluate the difference
which relate to the “travel time” attribute, did not obtain
in the valuations of “attitude” and “information” is an un-
support but have not been reversed either. Could be that
derestimation and gives a lower-limit for the difference.
travel time is a minor factor within the maternity ward
Had “information” also been associated with a gain, we
preference structure because the participants had experi-
would have arrived at a larger difference, i.e. at the con-
enced only two short episodes of travel (to the maternity
clusion that the attribute of “attitude” ranks much higher
ward and back home), leading to neutrality between the
than the attribute of “information” (with a larger differ-
(absolute) valuation of a loss and a gain.
ence than is indicated by our preference equations).
Hypothesis 5: The coef?cient of “professionalism” is
6 Summary and discussion
indeed signi?cantly larger in Type 1 regression, thus sup-
porting Hypothesis 5: there is a positive signi?cant dif-
In this study, DCEs were used to estimate preference
ference of 0.65 (p = 0.000). These results are explained
structures for maternity ward attributes, which were later
by the fact that in the data of Type 1 questionnaires a
used to test seven hypotheses derived from the loss aver-
gain versus the reference level is in evidence (a change
sion theory. The data used for the estimation and testing
from “good” to “very good”), while is the data of Type 2
reported here were based on two experiments conducted
on large samples of 219 and 323 Israeli women, respec-
12Based on a x2 Test for signi?cance between coef?cients of two re-
gressions. Not reported in the table.
tively (resulting in 1751 and 2099 discrete choice obser-
vation sets, respectively), interviewed in maternity wards

---

Judgment and Decision Making, Vol. 3, No. 2, February 2008
Reference dependence in health care
171
A DCE appears to be a highly suitable tool for loss
eralization of the results.
aversion testing because it estimates the marginal valua-
A second unresolved question is: What is the role of
tions of attributes, based on deviations from a reference
experience? Economists have claimed that loss aversion
point (a constant scenario). Moreover, loss aversion can
will erode as individuals accumulate more experience.14
be tested for each attribute separately rather than for the
In most health-care events not much experience can (for-
service as a whole. The DCE method can also be applied
tunately) be accumulated; hence the effect of repeated ob-
to a non-traded service having non-tangible attributes,
servations cannot be tested and becomes irrelevant.15 Re-
which implies its ?exibility.13
peated experiences can be observed mainly when a pa-
Loss aversion theory was con?rmed for four of the ?ve
tient had a series of treatments or a sequence of health
hospital attributes investigated. The results were less con-
diagnostic tests (e.g. pap smears, blood tests, EKG tests).
clusive for “travel time”, probably because traveling is
Experiments conducted among people with chronic con-
only a short and perhaps marginally meaningful episode
ditions can therefore be used to examine the effect of ex-
for this speci?c sample population.
perience on the possible erosion of loss aversion.
The existence of loss aversion also implies that the
Third and ?nally: Are there gender differences in loss
choice of the constant scenario used in a DCE affects the
aversion? Obviously, our sample, which was composed
estimated preference structure obtained: that is, differ-
only of women and pertained to a distinctively feminine
ent constant reference sets result in different preference
scenario — a maternity ward — cannot resolve this is-
structure estimates, if not in changes in the attributes”
sue. However, the conduct of a similar study among men
ranking. It follows that reports of preference structures
could shed some light on this interesting question.
should include descriptions of the constant reference sce-
nario in order to facilitate an accurate description of the
coef?cients and a distinction between coef?cients that
References
present valuations of gains and those that relate to val-
uations of losses.
Andreoni, J. (1995). Warm-glow versus cold-prickle:
Three unsolved questions merit further study:
The effects of positive and negative framing on cooper-
The ?rst: what affects the existence and magnitude of
ation in experiments. Quarterly Journal of Economics,
loss aversion regarding the various attributes of a com-
110, 1–21.
posite commodity? Is it possible to detect which at-
Benartzi, S., and Thaler, R. H. (1995). Myopic loss aver-
tribute will be subject to a higher level of loss aversion
sion and the equity premium puzzle. Quarterly Journal
when compared to the others? Our data indicated that
of Economics, 110, 73–92.
“travel time” is not subject to loss-averse preferences and
Bleichrodt, H., and Pinto J. L., (2002). Loss aversion and
that differences arose in the valuation of a “gain” ver-
scale compatibility in two-attribute trade-offs. Journal
sus a “loss” between the other four attributes. Further
of Mathematical Psychology, 46, 315–337.
research, involving additional case studies of varied com-
Bowman, D., Minehart, D., and Rabin, M. (1999). Loss
posite commodities or services, could lead to greater gen-
aversion in a consumption/savings model. Journal of
Economic Behavior and Organization, 38, 155–178.
13DCEs belong to a set of ?eld experiments that are employed in the
Bronstein, J., and Morrisey, M. (1991). Bypassing rural
social sciences. Field experiments represent an empirical approach that
bridges laboratory data and naturally occurring data (see List, 2006, for
hospitals for obstetrics care. Medical Care, 16, 87–
a discussion of the methodological contribution of ?eld experiments and
118.
a comprehensive overview of the different types of these experiments
Brown, S., and Lumley, J. (1994). Satisfaction with care
that are employed in the ?elds of: Markets; The Economics of Charity;
Environmental Economics; Development Economics; and Discrimina-
in labor and birth: A survey of 790 Australian women.
tion). Unlike laboratory experiments, DCEs do not involve any ?nancial
Birth, 21, 4–13.
incentives for performance and respondents are not paid for answering.
Bryan, S., Buxton, M., Sheldon, R., and Grant, A. (1998).
While economists presume that experimental subjects do not work for
Magnetic resonance imaging for the investigation of
free and they work harder, more persistently, and more effectively, if
they earn more money for better performance, the view expressed by
knee injuries: An investigation of preferences. Health
most psychologists is that intrinsic motivation is usually high enough
Economics, 7, 595–603.
to produce steady effort even in the absence of ?nancial rewards; and
Cattin, P., and Wittink, D. (1982). Commercial use of
that while more money may stimulate more effort, the effort does not
always improve performance, especially if good performance requires
conjoint analysis: A survey. Journal of Marketing, 46,
subjects to induce spontaneously a principle of rational choice or judg-
44–53.
ment. However, there is no consensus on this issue and there are studies
that have found some effect of real incentives (see Hertwig and Ort-
14Economists have referred mainly to the experience of selling and
mann, 2001 for a review). Camerer and Hogarth (1999), who reviewed
buying experience as re?ected in the market place.
74 experiments with no, low, or high performance-based incentives, ba-
15More speci?cally, women do not have many repeated experiences
sically found no effect of ?nancial rewards on mean performance (al-
with deliveries (the average number of children in Israel is 2.3. In Eu-
though variance is usually reduced by higher payment).
ropean countries even less experience is evidenced).

---