Psychonomic Bulletin & Review

https://doi.org/10.3758/s13423-018-1531-9

BRIEF REPORT

The comparison process as an account of variation in the attraction,

compromise, and similarity effects
Andrea M. Cataldo 1 & Andrew L. Cohen 1

# Psychonomic Society, Inc. 2018

Abstract
Context effects are changes in preference that occur when alternatives are added to a choice set. Models that account for context
effects typically assume a within-dimension comparison process; however, the presentation format of a choice set can influence
comparison strategies. The present study jointly tests the influence of presentation format on the attraction, compromise, and
similarity effects in a within-subjects design. Participants completed a series of choices designed to elicit each of the three context
effects, with either a by-alternative or by-dimension format. Whereas the by-alternative format elicited a standard similarity
effect, but null attraction and reverse compromise effects, the by-dimension format elicited standard attraction and compromise
effects, but a reverse similarity effect. These novel results are supported by a re-analysis of the eye-tracking data collected by
Noguchi and Stewart (Cognition, 132(1), 44–56, 2014) and demonstrate that flexibility in the comparison process should be
incorporated into theories of preferential choice.

Keywords Context effects . Attraction effect . Compromise effect . Similarity effect . Information presentation format

Introduction increases preference for Apartment X relative to Apartment

Y. The compromise effect (Simonson, 1989) is when the avail-
Preferential choice scenarios, such as choosing an apartment, ability of Apartment CX increases preference for Apartment
restaurant, or laptop, are ubiquitous. An important collection X, which now has intermediate values on both dimensions,
of results demonstrates that adding an alternative to a choice relative to Apartment Y. The similarity effect (Tversky, 1972)
set can change preferences among the original alternatives. is the finding that the availability of Apartment SX, which is
The attraction, compromise, and similarity effects are three similar to, but not dominated by, Apartment Y, increases pref-
well-studied examples of these Bcontext effects.^ To demon- erence for Apartment X relative to Apartment Y.
strate, consider the scenario of choosing between apartments These effects serve as central examples of how the decision
that vary in their rated size and location (Fig. 1). Assuming process deviates from rationality (e.g., Roe, Busemeyer, &
both dimensions are equally important, a choice between Townsend, 2001) and are therefore often used as theoretical
Apartments X and Y would be difficult – whereas benchmarks. Indeed, many computational models of prefer-
Apartment X rates well on location, but poorly on size, the ential choice strive to account for all three effects simulta-
reverse is true for Apartment Y. The attraction effect (Huber, neously (Bhatia, 2013; Noguchi & Stewart, 2018; Roe et al.,
Payne, & Puto, 1982) is when the availability of Apartment 2001; Trueblood, Brown, & Heathcote, 2014; Usher &
AX, which is similar to, but dominated by, Apartment X, McClelland, 2004). Behavioral data, however, suggest that
this constraint may be premature (Trueblood, Brown, &
Electronic supplementary material The online version of this article Heathcote, 2015). For example, a number of studies have
(https://doi.org/10.3758/s13423-018-1531-9) contains supplementary found the attraction and compromise effects to be positively
material, which is available to authorized users.
correlated with each other but negatively correlated with the
similarity effect (Berkowitsch, Scheibehenne, & Rieskamp,
* Andrea M. Cataldo
[email protected] 2014; Liew, Howe, & Little, 2016; Trueblood et al., 2015).
Further, it is not simply the case that particular effects do or do
1
not occur in different scenarios (e.g., Frederick, Lee, &
Department of Psychological and Brain Sciences, University of
Massachusetts, Amherst, MA, USA
Baskin, 2014; Huber, Payne, & Puto, 2014); rather, the
 Psychon Bull Rev

Trueblood et al., 2015) and individual differences (Liew et al.,

2016) by differentially highlighting dimension-level (e.g.,
extremeness and dominance; Simonson, 1989) and
alternative-level (e.g., dimension dispersion; Chernev, 2004,
2005) stimulus characteristics.
The present research targets a flexible comparison process
as a critical mechanism driving diversity in context-dependent
choice. To date, the effect of format on the attraction effect has
not been tested. Furthermore, the compromise (Chang & Liu,
2008) and similarity (Cataldo & Cohen, 2018) effects were
tested in two different paradigms with between-subject format
manipulations. We extend previous work by jointly testing the
influence of presentation format on all three context effects in
an entirely within-subjects design. These methodological
changes allow a unified view of the role of the comparison
process in context-dependent choice. The current data also
allow for response-time analysis, which has not been consid-
ered in this framework.
Following previous work (Cataldo & Cohen, 2018; Chang
Fig. 1 Each label represents the dimension values of an apartment.
Subscripts denote the apartment from the base pair X and Y targeted by & Liu, 2008), we predicted that a by-dimension format would
the decoy. The presented ratings are for the 3 EV condition. The precise produce strong compromise and reverse similarity effects, but
(Size, Location) stimulus values were as follows: X = (2.5, 3.5), Y = (3.5, that a by-alternative format would produce reverse compro-
2.5), AX = (2.25, 3.25), AY = (3.25, 2.25), CX = (1.5, 4.5), CY = (4.5, 1.5), mise and strong similarity effects. Given previously observed
SX = (3.75, 2.25), SY = (2.25, 3.75). For the 2 EV condition, all ratings
were one unit lower correlations between context effects (e.g., Berkowitsch et al.,
2014), we predicted that a by-dimension format would pro-
duce a strong attraction effect, but a by-alternative format
occurrence of reversed effects suggests a more diverse land- would produce a weakened attraction effect. The data support
scape of context-dependent choice (Cataldo & Cohen, 2018; these choice predictions. Furthermore, longer response times
Spektor, Kellen, & Hotaling, 2018). Thus, it is important for in the by-dimension condition suggest a process that requires
theories of preferential choice to consider how context effects greater deliberation. We conclude with a re-analysis of the
co-vary and what factors influence their presence, absence, or eye-tracking data collected by Noguchi and Stewart (2014)
reversal. demonstrating that, while the compromise effect is driven by
The processes underlying context effects have been a cen- a greater number of transitions within dimensions, the simi-
tral point of debate between theories of preferential choice; larity effect is driven by a greater number of transitions within
however, there is evidence to suggest that the information- alternatives.
acquisition process may play a primary role. Most recent
models assume that choice information is compared between
alternatives within a single dimension (Bhatia, 2013; Noguchi
& Stewart, 2018; Roe et al., 2001; Trueblood et al., 2014; Method
Usher & McClelland, 2004; Wollschläger & Diederich,
2012). Work by Chang and Liu (2008) and Cataldo and Participants
Cohen (2018), however, suggests the need to consider a more
flexible process. These studies show that differences in pre- Four-hundred and seventy undergraduates from UMass
sentation format (Bettman & Kakkar, 1977; Biehal & Amherst participated for course credit. The sample size was
Chakravarti, 1982) can modulate context effects, implicating selected to be larger than that used in past work eliciting si-
the information-acquisition process. Specifically, whereas the multaneous context effects (Berkowitsch et al., 2014; Liew et
compromise effect is facilitated by a format encouraging al., 2016; Noguchi & Stewart, 2014). Following previous
within-dimension comparisons and impeded by a format en- work (Cataldo & Cohen, 2018), participants were excluded
couraging within-alternative comparisons (Chang & Liu, for failing more than four of 24 catch trials. Thirty-four par-
2008), the opposite is found for the similarity effect (Cataldo ticipants were excluded, leaving 436 in the analyses. Separate
& Cohen, 2018). Together, these studies suggest that a flexible analyses including all participants and excluding 198 who
comparison process may be a key mechanism underlying the failed any catch trial are provided in the Supplementary
correlations between context effects (Berkowitsch et al., 2014; Materials, but yield the same qualitative results.
Psychon Bull Rev

Materials conditions (DX and DY) by two expected values (2 and 3) by

three contexts (attraction, compromise, and similarity) by two
Each choice set consisted of three apartments varying on size presentation formats (by-alternative and by-dimension) by six
and location. Every test trial included Apartments X and Y. alternative orderings.
The unweighted mean of X and Y were always equivalent; To identify participants who were not sufficiently engaged,
that is, assuming no dimension bias, X and Y were of equal each participant also completed 24 catch trials that included a
overall value. For generalizability, the unweighted mean (EV) dominating apartment, resulting in a total of 168 choices.
of X and Y was varied (either 2 or 3). The size and location
dimension values in the 3 EV condition are described in Fig.
Procedure
1, and the 2 EV dimension values are one unit less.1
The third apartment was either AX or AY for the attraction
The trials were blocked by presentation format, with block
effect, CX or CY for the compromise effect, and SX or SY for
order randomized across participants. Participants were told
the similarity effect. These stimuli will be collectively referred
that the different formats represented different rental maga-
to as decoy stimuli (D). The subscripts index the target alter-
zines. All other factor levels were randomly ordered within
native, i.e., whether adding the decoy is expected to increase
each block.2 Participants were given detailed instructions at
choices for Apartment X or Y. The non-target, non-decoy
the beginning of each block, including the meaning of the
alternative will be referred to as the competitor. The distances
dimensions, a description of the rating scale, and a preview
from A and S to the nearby alternative were equal and
of the presentation format. Participants completed three prac-
subjectively selected to be as small as possible while remain-
tice trials before the test trials. Each response involved a key
ing perceptually distinct.
press (1, 2, or 3) corresponding to the desired alternative.
Following Tversky (1972, Task B) and Cataldo and Cohen
(2018), the dimension values were depicted as filled, horizon-
tal bars (see Fig. 2, center panels). The values were goodness-
of-fit ratings, from Bworst for me^ (unfilled) to Bbest for me^ Results
(completely filled). This scale standardizes the dimensions
and minimizes concerns about differences in preferred dimen- Choice
sion directions, e.g., preferring small versus large apartments.
The horizontal length of the bar was determined by multiply- Following Wedell (1991), we measure each context effect as a
ing the constant, vertical height of the bar (50 px) by the comparison between two three-choice scenarios targeting X or
dimension rating. Y. A context effect is obtained if both ΔPX = P(X | X, Y, DX) -
The choice sets were displayed in one of two presentation P(X | X, Y, DY) and ΔPY = P(Y | X, Y, DY) - P(Y | X, Y, DX)
formats – by-alternative or by-dimension. Consider the center are positive. Negative values indicate a reverse effect. The left
panels of Fig. 2. The top and bottom rows show sample stim- panels of Fig. 2 show ΔP for X and Y for each context effect
uli in the by-alternative and by-dimension conditions, respec- and presentation format.
tively. In both cases, the ratings are presented as horizontal Participants in the by-alternative condition (top row) dis-
bars in a matrix, strongly encouraging comparisons within play a classic similarity effect, but a weak or null attraction
columns rather than within rows. In the by-alternative condi- effect and a reversed compromise effect. That is, SX and SY
tion, the columns of the matrix denote alternatives and the increase preference for the target, but AX and AY have little
rows denote dimensions, encouraging within-alternative com- effect on choice proportions, and CX and CY actually decrease
parisons. In the by-dimension condition, the columns denote preference for the target.
dimensions and the rows denote alternatives, encouraging In perfect contrast, participants in the by-dimension condi-
within-dimension comparisons. The bar lengths were constant tion (bottom row) display classic attraction and compromise
across presentation format conditions. The size dimension effects, but a reversed similarity effect. That is, whereas AX,
was always presented on the top in the by-alternative condi- AY, CX, and CY increase preference for the target, SX and SY
tion and on the left in the by-dimension condition. For each decrease preference for the target.
choice set, participants saw all six possible alternative order- A hierarchical Bayesian conditional logistic regression
ings, randomly ordered. model was used to test for differences in choice proportions
The factors described above were manipulated within-sub- across target, EV, context, and presentation format conditions.
ject, resulting in 144 test trials for each participant: two target Details of the model are provided in the Supplemental
2
Participants who viewed the by-dimension format second exhibited a weak-
1
Due to a coding error, 227 participants were dropped from the 3 EV condi- ened effect. The Supplementary Materials include further discussion and anal-
tion. A separate analysis excluding these participants is provided in the ysis of this result. These analyses yield the same critical interactions as the full
Supplementary Materials, and yields the same qualitative results. data set.
 Psychon Bull Rev

Fig. 2 Left: Mean differences in overall choice proportions for alternatives (top) and by dimensions (bottom). In this example,
Alternatives X and Y in attraction, compromise, and similarity choice Apartments 1, 2, and 3 correspond to Apartments X, Y, and AX in Fig.
sets, when choice information is grouped by alternatives (top) and by 1, respectively. Right: Mean response times broken down by target (X or
dimensions (bottom). Error bars are between-subject standard errors. Y) in attraction, compromise, and similarity choice sets, when choice
Middle: Sample stimuli depicting an apartment choice set grouped by information is grouped by alternatives (top) and by dimensions (bottom)

Material. Inferences are made by calculating the 95% highest The classic attraction and compromise effects and reverse
density interval (HDI) around the mean of the posterior esti- similarity effect observed in the by-dimension condition are
mated choice proportions for a given condition. A difference both statistically supported. The HDIs for the attraction and
between conditions is indicated by non-overlapping HDIs. compromise effects are greater than zero, demonstrating a
Consider the posterior estimates and HDIs for choice pro- change in preference towards the target. The HDIs for the
portions provided in Table 1. First, as previewed above, the similarity effect are less than zero, indicating a change in pref-
attraction decoy was selected least often. Second, there was a erence towards the competitor.
preference for X over Y, suggesting a bias for location. Third, It is important to note that the HDIs for ΔPX and ΔPY are
there was no effect of expected value; thus, we collapse across greater in the by-dimension condition for the attraction and
expected value in all following analyses. compromise effects but greater in the by-alternative condition
Next, we use this model to address which context effects for the similarity effect. This result supports the hypothesis
were present. The estimated choice proportions and HDIs for that the attraction and compromise effects are facilitated by a
ΔPX and ΔPY are provided in Table 2, broken down by con- presentation format that encourages within-dimension com-
text and format. The null attraction effect, reverse compromise parisons but the similarity effect is facilitated by a presentation
effect, and classic similarity effect observed in the by- format that encourages within-alternative comparisons.
alternative condition are all statistically supported. The ΔPX To ensure that the results are not due to averaging, it is
and ΔPY HDIs for the attraction effect comfortably include informative to examine the results at the individual participant
zero. The HDIs for the compromise effect are less than zero, level (Berkowitsch et al., 2014; Cataldo & Cohen, 2018; Liew
demonstrating a change in preference towards the competitor. et al., 2016; Trueblood et al., 2015). Figure 3 summarizes the
The HDIs for the similarity effect are greater than zero, dem- co-occurrence of the attraction (A), compromise (C), and sim-
onstrating a change in preference towards the target. ilarity (S) effects within subjects. For each participant, each

Table 1 Posterior estimates and 95% HDIs for choice proportions by target, expected value (EV), context effect, and presentation format conditions

P(X) P(Y) P(D)

Factor Level M HDI Low HDI High M HDI Low HDI High M HDI Low HDI High
Target X 0.526 0.523 0.531 0.356 0.352 0.359 0.118 0.115 0.120
Y 0.500 0.496 0.504 0.378 0.373 0.381 0.122 0.119 0.125
EV 2 0.513 0.510 0.517 0.370 0.367 0.374 0.117 0.114 0.119
3 0.513 0.509 0.517 0.363 0.360 0.367 0.123 0.120 0.126
Context effect A 0.539 0.534 0.543 0.381 0.377 0.386 0.080 0.077 0.083
C 0.521 0.516 0.525 0.364 0.360 0.369 0.115 0.112 0.118
S 0.480 0.475 0.486 0.355 0.349 0.361 0.165 0.160 0.170
Format By-Alt 0.538 0.534 0.542 0.348 0.344 0.352 0.114 0.111 0.117
By-Dim 0.489 0.484 0.493 0.385 0.381 0.390 0.126 0.123 0.129

Notes: A, C, and S are the attraction, compromise, and similarity effects, respectively
Psychon Bull Rev

Table 2 Posterior estimates and 95% HDIs for ΔPX and ΔPY and their mean, broken down by context and presentation format

ΔPX ΔPY Mean

Presentation Context M HDI Low HDI High M HDI Low HDI High M HDI Low HDI High
Format
By-Alternative A 0.008 -0.003 0.019 0.011 -0.000 0.021 0.009 -0.001 0.019
C -0.017 -0.029 -0.005 -0.015 -0.026 -0.004 -0.016 -0.027 -0.005
S 0.049 0.035 0.062 0.031 0.018 0.043 0.040 0.028 0.052
By-Dimension A 0.111 0.099 0.124 0.095 0.083 0.106 0.103 0.091 0.114
C 0.046 0.034 0.059 0.049 0.036 0.060 0.047 0.035 0.059
S -0.038 -0.053 -0.023 -0.040 -0.054 -0.027 -0.039 -0.052 -0.026

Notes: ΔPX = P(X | X, Y, DX)-P(X | X, Y, DY), ΔPY = P(Y | X, Y, DY)-P(Y | X, Y, DX). A, C, and S are the attraction, compromise, and similarity effects,
respectively

effect was marked present (+) if the average of ΔPX and ΔPY That is, participants were much more likely to display classic
was greater than 0.04, reversed (-) if the average was less than attraction and compromise effects with a reversed similarity
0.04, and null (0) otherwise. effect (A+, C+, S-). Far less likely was a null attraction effect,
The two most frequent response patterns for each presen- reversed compromise effect, and classic similarity effect (A0,
tation format are indicated by darker colors. The individual C-, S+).
subject results are consistent with the group results. In the by-
alternative condition, the modal response patterns included Response times
null or reversed attraction and compromise effects with a clas-
sic similarity effect (A-, C-, S+ or A0, C-, S+). In the by- The right panels of Fig. 2 present mean response times corre-
dimension condition, there was an especially clear result. sponding to the conditions presented in the left panels of Fig.

Fig. 3 Co-occurrence of reverse (-), null (0), and classic (+) attraction, compromise, and similarity effects within subjects, when choice information is
grouped by alternatives (top) and by dimensions (bottom). Darker bars indicate the two most frequent categories within each format condition
 Psychon Bull Rev

Table 3 Posterior estimates of mean response times from the Bayesian Furthermore, participants were biased towards making
model, separately broken down by target, expected value (EV), context,
pairwise comparisons within a dimension; however, that bias
and presentation format conditions
was small, suggesting variability across trials.
Factor Level M HDI Low HDI High Our re-analysis partitions the trials by the relative number
of within-dimension and within-alternative transitions.
Target X 3.162 3.141 3.184
Additional details are included in the Supplementary
Y 3.180 3.155 3.203
Materials. If the comparison process affects choice behavior
EV 2 3.227 3.203 3.251
as predicted, as the relative number of within-dimension tran-
3 3.115 3.086 3.141
sitions increases, the attraction and compromise effects should
Context effect A 3.167 3.143 3.194
increase and the similarity effect should decrease. Conversely,
C 3.135 3.105 3.164 as the relative number of within-alternative transitions in-
S 3.212 3.183 3.240 creases, the similarity effect should increase and the attraction
Format By-Alt 2.843 2.820 2.866 and compromise effects should decrease.
By-Dim 3.499 3.470 3.528 Consistent with the original analysis, there was a small bias
Notes: A, C, and S are the attraction, compromise, and similarity effects,
towards within-dimension comparisons. The first row of Fig.
respectively 4 shows the distribution of within-alternative minus within-
dimension transitions for all trials for each context. Positive
2. There is a marked effect of presentation format, such that and negative values correspond to trials including more
responses are much slower in the by-dimension format. There within-alternative and within-dimension eye movements, re-
is also an observable effect of context, such that response spectively. For attraction, compromise, and similarity trials,
times are slowest for similarity choice sets. An individual- there was an average of 1.40, 1.50, and 0.90 more within-
subject analysis of the response-time results is provided in dimension transitions.
the Supplemental Material. These distributions were partitioned into five regions. Each
A hierarchical Bayesian regression model was used to test region is associated with a color ranging from blue (Region 1)
for differences in response times across target, EV, context, to red (Region 5). Region 1 includes trials strongly favoring
and presentation format conditions. Details of the model are within-dimension comparisons and Region 5 includes trials
provided in the supplemental material. Inferences are again strongly favoring within-alternative comparisons. The region
based on the 95% HDIs of a response time in a given condi- criteria (-5, -2, 2, 5) were fixed for all contexts and were
tion. The estimated choice proportions and 95% HDIs for selected to separate the distributions into qualitatively mean-
response times are provided in Table 3. The group differences ingful regions, be symmetric around 0, and keep enough data
described above are statistically supported. That is, the 95% in each region for analysis. Different values were tried with
HDI for mean response times is greater in the by-dimension similar qualitative results. Because we are interested in how
condition than in the by-alternative condition, and greater for the comparison process on a particular trial influences behav-
similarity choice sets than for the attraction and compromise ior, these analyses are at the trial level.
choice sets. The second row of Fig. 4 shows how each context effect
changes as the comparison process changes. The measure of
interest is the relative proportion of target and competitor
Eye tracking: A re-analysis of Noguchi and Stewart choices in each region. Positive values indicate a classic con-
(2014) text effect and negative values indicate a reverse effect.
Although the attraction effect is relatively unchanged, as pre-
The previous analyses provide strong evidence that presenta- dicted, the compromise effect decreases and the similarity
tion format influences context effects. Although these differ- effect increases, even reversing direction, as the proportion
ences imply a switch from a within-dimension to a within- of within-alternative comparisons increases.
alternative comparison process, the evidence is indirect be-
cause the comparison process was not directly measured.
Here, we re-analyze the results of Noguchi and Stewart Fig. 4 Reanalysis of Noguchi and Stewart (2014). Top row: Distribution„
of within-alternative – within-dimension comparisons for each context
(2014) to provide this direct evidence. effect broken into Regions 1–5. Second row: Proportion of target
Similar to the current experiment, participants in Noguchi choices minus proportion of competitor choices within each region for
and Stewart (2014) were asked to select one of three alterna- each context effect; statistics (details are provided in the supplemental
tives from a set designed to elicit an attraction, compromise, or material) are a comparison to Region 1; + p<.10, * p<.05, ** p<.01,
*** p<.001. Bottom three rows: Choice proportions for all data and
similarity effect. Importantly, eye movements were tracked Regions 1 and 5 for the target, competitor, and decoy for each context
during the decision process. The authors found strong attrac- effect. Numbers above points are sample sizes. Each region is associated
tion and compromise effects and a weak similarity effect. with a color ranging from blue (Region 1) to red (Region 5)
Psychon Bull Rev
 Psychon Bull Rev

The final three rows of Fig. 4 show the raw choice propor- proportion of within-alternative transitions increased. The
tions for all data (compare to Fig. 5 from Noguchi & Stewart, lack of a significant decrease for the attraction effect differs
2014) and Regions 1 and 5 only. The black circles and white from the present work, although we note it is historically a
squares are trials on which A (X from Fig. 1) and B (Y from very robust effect (e.g., Berkowitsch et al., 2014).
Fig. 1) are targets, respectively. An effect is supported if the The increased response times for similarity choice sets sup-
black circle is above the white square for A and below the ports the idea that a different process is associated with this
white square for B. Again, more within-alternative compari- effect. Indeed, models of preferential choice accounting for all
sons greatly weaken or reverse the compromise effect and three effects often assume that the similarity effect is generated
more within-dimension comparisons reverse the similarity by a distinct mechanism (Bhatia, 2013; Noguchi & Stewart,
effect. 2018; Roe et al., 2001; Trueblood et al., 2014; Usher &
McClelland, 2004; Wollschläger & Diederich, 2012). The da-
ta suggest that this assumption is appropriate. Although spec-
Discussion ulative, it is even possible that a deliberative process underlies
the attraction and compromise effects, whereas a more heuris-
The attraction, compromise, and similarity effects are critical tic process underlies the similarity effect. In the present data,
phenomena in preferential choice that serve as key examples of although similarity choice sets are associated with increased
how the decision process deviates from rationality (Huber et al., response times, a by-alternative format promoting the similar-
1982; Simonson, 1989; Tversky, 1972). Several studies have ity effect is associated with reduced response times. This idea
found that the attraction and compromise effects tend to be pos- is consistent with previous work finding that while the simi-
itively correlated with each other but negatively correlated with larity effect can be accounted for by a heuristic model
the similarity effect (Berkowitsch et al., 2014; Liew et al., 2016; (Tversky, 1972), the attraction and compromise effects are
Trueblood et al., 2015). Work by Chang and Liu (2008) and inhibited by the use of a lexicographic heuristic (Simonson,
Cataldo and Cohen (2018) suggest that a flexible comparison 1989). Further research directly testing this account is needed.
process may be a key mechanism underlying these correlations. Previous researchers have noted that models of preferential
Specifically, whereas the compromise effect is facilitated by a choice can account for the correlation between context effects
presentation format encouraging within-dimension comparisons seen both here and in prior work (Berkowitsch et al., 2014;
and impeded by a format encouraging within-alternative compar- Liew et al., 2016). For example, Roe et al. (2001) state that
isons (Chang & Liu, 2008), the opposite is found for the similar- Multi-alternative Decision Field Theory (MDFT) predicts a
ity effect (Cataldo & Cohen, 2018). negative correlation between the attraction and similarity ef-
The present research extends previous work by jointly testing fects, but a positive correlation between the attraction and
the influence of presentation format on the compromise, similar- compromise effects. With few exceptions, such modeling
ity, and attraction effects with an entirely within-subjects design. work has almost exclusively considered the presence or ab-
Previous work has only studied the similarity and compromise sence of context effects (e.g., Trueblood et al., 2015), but not
effects separately (Cataldo & Cohen, 2018; Chang & Liu, 2008), whether they can be reversed. Consistent with the current
and the attraction effect has not been tested at all in this paradigm. results, however, Cataldo and Cohen (2018) showed that var-
As predicted, a by-alternative presentation format elicited the iations of both the MDFT and Multi-attribute Linear Ballistic
standard similarity effect, but weak and reversed attraction and Accumulator (MLBA; Trueblood et al., 2014) predict that a
compromise effects, respectively, whereas a by-dimension pre- reverse similarity effect is more likely than a standard similar-
sentation format produced standard attraction and compromise ity effect when the attraction and compromise effects are pres-
effects but a reverse similarity effect. These results not only rep- ent. To account for the reverse similarity effect, the MDFT
licate effects found in past work but demonstrate their relation- relies on increased forgetting and increased inhibition between
ship at the subject level. alternatives and the MLBA relies on increased attention to
The current data also allow for the incorporation of negative differences of attributes. Although the models can
response-time analyses, which have rarely been considered potentially predict the present results, more work is needed
in context effects and never in this framework. Interestingly, to determine whether the mechanisms used to do so are psy-
longer response times were produced in the by-dimension chologically supported. The current research suggests a poten-
condition and the similarity choice sets. The effect of format tial alternative account in which these effects are driven, in
likely results from a difference in the number of within- part, by changes in the comparison process.3 More broadly, it
alternative and within-dimension comparisons. Though not is important for models of preferential choice to take the
directly assessed in the current data, a re-analysis of eye-
tracking data collected by Noguchi and Stewart (2014) sup- 3
A step in that direction is provided by the Associations and Accumulation
ports this interpretation. In their study, the similarity effect Model (Bhatia, 2013), which assumes that attention to attributes varies based
increased and the compromise effect decreased as the on the available alternatives.
Psychon Bull Rev

pattern of reversals of and correlations between context effects Chang, C.-C., & Liu, H.-H. (2008). Which is the compromise option?
Information format and task format as determinants. Journal of
more seriously, and allowing for diverse comparison styles
Behavioral Decision Making, 21(1), 59–75.
may be an effective approach to doing so. Chernev, A. (2004). Extremeness Aversion and Attribute-Balance Effects
This research provides strong evidence that the comparison in Choice. Journal of Consumer Research, 31(2), 249–263.
process should be treated with more nuance in theories of Chernev, A. (2005). Context Effects without a Context: Attribute Balance
as a Reason for Choice. Journal of Consumer Research, 32(2), 213–
preferential choice. Determining relative values is intuitively
223.
an important mechanism underlying context effects, but evi- Frederick, S., Lee, L., & Baskin, E. (2014). The Limits of Attraction.
dence suggests that this mechanism may not occur strictly via Journal of Marketing Research, 51(4), 487–507.
within-dimension comparisons. Differential patterns of con- Huber, J., Payne, J. W., & Puto, C. (1982). Adding asymmetrically dom-
inated alternatives: Violations of regularity and the similarity hy-
text effects are produced by modulating the ability to compare
pothesis. Journal of Consumer Research, 9(1), 90–98.
choice information within alternatives or within dimensions. Huber, J., Payne, J. W., & Puto, C. P. (2014). Let’s Be Honest About the
The theoretical implication from these data is that the compar- Attraction Effect. Journal of Marketing Research, 51(4), 520–525.
ison process may serve to promote context-dependent choice Liew, S. X., Howe, P. D. L., & Little, D. R. (2016). The appropriacy of
averaging in the study of context effects. Psychonomic Bulletin &
as well as the apparent diversity within it. From a practical
Review, 23(5), 1639–1646.
standpoint, the data suggest that requiring models to simulta- Noguchi, T., & Stewart, N. (2014). In the attraction, compromise, and
neously account for all three effects is an unrealistic and over- similarity effects, alternatives are repeatedly compared in pairs on
ly simplified constraint. single dimensions. Cognition, 132(1), 44–56.
Noguchi, T., & Stewart, N. (2018). Multialternative decision by sam-
pling: A model of decision making constrained by process data.
Author Note Both authors contributed to all aspects of this Psychological Review, 125(4), 512–544.
work. The data have been made publically available at: Roe, R. M., Busemeyer, J. R., & Townsend, J. T. (2001). Multialternative
https://osf.io/736db/. The authors are very grateful to Drs. decision field theory: A dynamic connectionst model of decision
Noguchi and Stewart for making the data from Noguchi and making. Psychological Review, 108(2), 370–392.
Simonson, I. (1989). Choice based on reasons: The case of attraction and
Stewart (2014) available. compromise effects. Journal of Consumer Research, 16(2), 158–
174.
Spektor, M. S., Kellen, D., & Hotaling, J. M. (2018). When the Good
Looks Bad: An Experimental Exploration of the Repulsion Effect.
Psychological Science, 095679761877904.
References Trueblood, J. S., Brown, S. D., & Heathcote, A. (2014). The
multiattribute linear ballistic accumulator model of context effects
Berkowitsch, N. A. J., Scheibehenne, B., & Rieskamp, J. (2014). in multialternative choice. Psychological Review, 121(2), 179–205.
Rigorously testing multialternative decision field theory against ran- Trueblood, J. S., Brown, S. D., & Heathcote, A. (2015). The fragile nature
dom utility models. Journal of Experimental Psychology: General, of contextual preference reversals: Reply to Tsetsos, Chater, and
143(3), 1331–1348. Usher. Psychological Review, 122(4), 848–853.
Bettman, J. R., & Kakkar, P. (1977). Effects of Information Presentation Tversky, A. (1972). Elimination by aspects: A theory of choice.
Format on Consumer Information Acquisition Strategies. Journal of Psychological Review, 79(4), 281–299.
Consumer Research, 3(4), 233. Usher, M., & McClelland, J. L. (2004). Loss Aversion and Inhibition in
Bhatia, S. (2013). Associations and the accumulation of preference. Dynamical Models of Multialternative Choice. Psychological
Psychological Review, 120(3), 522–543. Review, 111(3), 757–769.
Biehal, G., & Chakravarti, D. (1982). Information-Presentation Format Wedell, D. H. (1991). Distinguishing among models of contextually in-
and Learning Goals as Determinants of Consumers’ Memory duced preference reversals. Journal of Experimental Psychology:
Retrieval and Choice Processes. Journal of Consumer Research, Learning, Memory, and Cognition, 17(4), 767–778.
8(4), 431. Wollschläger, L. M., & Diederich, A. (2012). The 2N-ary Choice Tree
Cataldo, A. M., & Cohen, A. L. (2018). Reversing the similarity effect: Model for N-Alternative Preferential Choice. Frontiers in
The effect of presentation format. Cognition, 175. Psychology, 3.