MSc Marketing Intelligence Master Thesis

MSc Marketing Intelligence

Master Thesis

Candidate: Giovanni Riefolo – s2746697 Supervisor: Prof. Dr. L.M Sloot

“Understanding the Effect of Decoy Removal

Through Arbitrary Coherence”

Faculty of Economics & Business, Department of Marketing

Abstract

Introduction ………... 2

Literature Review ……….. 3

Choice Theories ..………... 3

Shelf Management .………. 8

1. Visualization Effects …………... 2. The Decoy Effect ……… 8 11 Conceptual Framework ………. 14 The Experiment ………. 18 Methodology ………. 19 The Data ………….……….. 22

Analysis & Results ...….……… 1. Main Analyses …….. ……….. 2. Robustness Check ...……… 24 24 39 Hypotheses Testing & Discussion ……… 44

Conclusions & Limitations ..………. 46

INTRODUCTION

The present study seeks to gain the first evidence concerning the ‘decoy-removal effect’ through an unconscious choice mechanism known as arbitrary coherence, through which will be possible to replicate the real-life decision-making context in which the effect occurs. Moreover, the experiment will be designed in order to also test for discrepancies between the decoy effectiveness and the associated perceived satisfaction of the product assortment.

The study will be structured as follows: first, the literature section will review all the different points of view from which choice mechanisms have been analyzed. These include the large body of choice theories, to which will follow the one concerning shelf management: this includes a section on the most studied visualization effects plus the stream focused on the decoy effect. Subsequently, the conceptual framework and the hypotheses will be presented, together with an explanation of the experiment. Following, the methodology and the data sections, the analysis, the results and a general discussion.

LITERATURE REVIEW Choice Theories

or be indifferent between two of them. An agent would therefore have preference ordering over A which satisfies the conditions of completeness and transitivity, which jointly define the so-called weak ordering over all prospects (Sen, 1979). The first condition assumes that the agent will always order his preferences, that is, either Xi Xj or Xj Xi for all Xi, Xj A. The condition of transitivity instead assumes that the ordering of prospect must be consistent, in the sense that if Xi Xj and Xj Xz, then Xi Xz for all Xi, Xj, Xz A. Other assumptions concerning rational choice theory state that (1) each rational agent has a coherent set of probabilistic beliefs, meaning that beliefs can be represented as probability distributions; and (2) the independence condition, meaning that if two prospects are mixed with a third one, the ordering preference remains the same, as obtained when the two prospects are presented independently from the third one.

The issue with rational choice theory, however, is that its conditions might be violated in real-choice contexts. One example of violation that is particularly relevant for the present research is the Allais’ paradox (Allais, 1953): suppose we have the following scenarios, 1 and 2, in which we are asked to choose respectively either A or B, or C or D

It is glaring now that alternative A is identical to C*, and B* is identical to D. The paradox is that according to the independence condition, these specular outcomes should be disregarded in the deliberation, meaning that the original preferences of the first scenario should be replicated also in the second, without regard (that is, independently) to the addition or subtraction of alternatives in the choice sets. Indeed, if in the second choice set an agent chooses A (and consequently C*) in the first choice set should also choose A and C, vice-versa if he chooses B* (and consequently D in the second choice set) then he should also choose B and D in the first choice set. Surprisingly, Kahneman and Tversky (1979) found out that more than 82% of the respondents in a controlled experiment chose B, and 83% subsequently chose C. This proves that the independence condition is clearly violated, as the agent’s preferences vary depending on the mixture between Xi, Xj and a third element Xz.

Another well known example that shows how the transitivity assumption of rational choice theory can be violated is presented by Ariely (2008): suppose you would like to purchase the subscription to a magazine, and you are confronted with the following set of offers:

- A: Online Subscription for 12 months, Price: 59$ - B: Print Subscription for 12 months, Price: 125$

- C: Print + Online Subscription for 12 months, Price 125$

The second option, the print subscription, is usually referred to as the decoy in the marketing literature (Ariely, 2008; Geyskens et al., 2010). It is glaring to notice how this alternative in particular is absolutely irrelevant: why should I spend 125$ for the print subscription if I can get, for the exact same price, both the print and online subscriptions? Of course, there is no valid reason to do that. So when a sample of respondents was asked to make a selection, 84% of the interviewees chose option C, 16% opted for option A while no respondent chose option B, as expected. Now suppose you are asked to chose an option from this new set of offers:

- A: Online Subscription for 12 months, Price: 59$

- B: Print + Online Subscription for 12 months, Price 125$

depends on the composition and disposition of the items in the choice set, which do not have to be necessarily relevant offers (ie. offers designed to compete with the other alternative offers).

In order to account for this fallacies in the postulates of rational choice theory, the concept of utility, defined as the assignation of numerical rankings to each possible choice (Levin and Milgrom, 2004), has been embedded in the classic rational choice theory. Considering the four conditions of rational choice theory, namely completeness, transitivity, the coherence of the set of probabilistic beliefs and the independence conditions defined before, it is possible to express the set of preferences over the prospects as a utility function:

u: Χ ⇒ R, Xi Xj

⇔

Σ

_≥

Σ

This means that, within rational choice theory, a rational agent is the one that chooses the alternative in the set of choices that maximizes his utility. An ordinal utility function is a function that represents the preferences from a choice set of an agent on an ordinal scale, meaning that the actual magnitude of the difference between two alternatives is not taken into account (Pareto, 1906). A cardinal utility functions by contrast, keeps into account the existence of many different levels of satisfaction, and makes it possible to compare the magnitude of the marginal utilities associated with each preference (Ellsberg, 1954). Roughly speaking, differently from the ordinal utility function, the second allows for the concavity or the convexity of the utility function.

The limits of rational choice theory, however, are not overcome with the introduction of the concept of utility, even if it solves many of the problems associated with its idea of ‘universal rationality’. In fact, the main contribution of utility is that it takes into account an underlying subjectivity, explaining why, from a theoretical perspective, two rational agents might behave differently from one another in the exact same situation.

Despite this, many scholars have argued that rational choice theory presents some limitations that, for the way in which the theory is postulated, refrain it from being applicable to a broad range of situations, in particular to real-life scenarios:

of the decision-maker, the addition of seemingly extraneous items to the choice set, and a host of other environmental factors appear to influence choice behavior. The existence of the marketing industry is testament to this, and many other examples are possible” (Levin and Milgrom, 2004).

For this reason, a number of other theories governing choice behaviors have been developed over the past: the weighted utility theory was developed by Chew (1982) and solves the problems arisen with the Allais’ paradox. In fact, this theory is based on the exact same axioms of rational choice theory with the only difference that it allows for the so-called weak (instead of strong) independence (Chew and Waller, 1986): it states that preferences over a choice set may actually vary when a new item is mixed with the existing offer. Said it roughly, differently from rational choice theory it allows for a greater flexibility in the preferences of an individual.

Prospect theory, instead, takes into account what is known as bounded rationality of individuals, when making choices over alternatives that involve risks: this concept refers to the limited human ability to behave in a fully rational fashion when dealing with choice, in which individuals make rational evaluations and choices relative to their knowledge and beliefs (Simon, 1957). From this assumption, experimental psychologist Kahneman and Tversky (1979) developed prospect theory, which states that individuals evaluate lotteries mainly thinking in terms of losses and gains, and that this evaluation is based on the so-called available heuristics: individuals are not precise in their evaluations, as these would often require a great deal of effort in order to collect, analyze and use objective information. In order to overcome this issue, individuals often base their decisions based on the available heuristics, which are simple rules that often involve focusing on one single aspect of a complex problem (Lewis, 2008). As a consequence, in many circumstances this oversimplification result in systematic deviations from logic, violating the postulates of rational choice theory, ultimately resulting in what are known as cognitive biases (Kahneman, 2011).

an important contribution as opposed to Luce’s IIA theory, stating that the probability for an individual to choose a particular item in a choice set is strictly dependent on the alternatives in the same set. The choice is hence based on the contextual elements, based on which individuals can make an evaluation of each item relative to the others. This is in line with the theory of available heuristics in the sense that individuals, in the absence of absolute references for an effective comparison, recall the information they have concerning that product. When these informations are not available, they proceed by evaluating the focal item as compared to the others (in terms of price, perceived quality or any other information that is, indeed, available). However, it has been showed in the present literature review how the Luce’s theory does not hold in many circumstances, such as the non-independence of irrelevant alternatives and the violation of the proportionality effect (seen in the previous section of this study).

A theory which will be of particular relevance for this study is the one which refers to the so-called arbitrary coherence (Ariely et al., 2003): it can be defined as the ability of an arbitrary anchor to affect subjective evaluation (Scott & Lizieri, 2011). The effect encompasses two distinct ideas, (i) that a value in a subject’s mind can easily be established arbitrarily by anchoring; and (ii) that once established the anchor will shape decision making through a process where every decision is shaped and built on previous decisions taken. The term anchor refers to one of the “purest form of behavioral phenomena” and is a form of heuristics, the mental shortcuts our brain use to simplify complex evaluations when in lack of informations (Kahneman et al., 1999). In the present research, the manipulation of the effect of arbitrary coherence will be implied to simulate the removal of a decoy item within a set of alternatives. A more detailed explanation can be found in the experiment section.

Shelf Management

1. Visualization Effects

compromise effect states that a product, within a choice set, can benefit from a relatively larger utility and a disproportionately larger probability of being selected when it becomes a compromise in the assortment after the introduction of a new product. Said it roughly, a product has much higher chances of being chosen with respect to the others when said product is better than other in the assortment, but not the best. From a purely behavioral perspective, this same effect has also been referred to as

theory of partially dominant choice (Gerasimou, 2015): the selection of the compromise item is

reduced to a rational choice when preferences are complete in two ways that are made precise. Geyskens et al. (2010) identify a number of reasons to justify the occurrence of this phenomenon: first of all, there is a broadly accepted belief in marketing research that consumers in general want to be positively evaluated by others (Simonson, 1989). This means that consumers tend to select those products that, based on the (usually) limited informations available when they have to make a selection, seem more justifiable to others observing their choices (Huber and Puto, 1983). That is, ‘I

bought this item because is better than others, and will probably give me a satisfaction (utility) similar to the best product, whose price was higher’, which seems to be a perfectly justifiable explanation of

one’s own selection process. This phenomenon in turns also affects the perceptions related to the expected quality of the items in the choice set: since consumers are often uncertain about their preferred quality level, they rely on choices made by others (recalling that most consumers would select the compromise product in the set) in order to get an estimate of their preferred quality level (Wernerfelt, 1995). Moreover, the perceived difference between the mid-quality item and the top-quality item tends to decrease when an extreme low-top-quality item is added to the choice set (Parducci, 1974; Bultez and Guerra, 2005). This means that even if the difference between the low and mid-quality options is less than the one between the mid and top-mid-quality items, the mid-mid-quality product would be chosen following the reasoning that ‘it is certainly much better than the low-quality item,

but not that worse compared to the top-quality item’.

selected) from each of the items already in the set, in proportion to their original shares (Huber and Puto, 1983; Luce, 1959). In fact, the similarity effect proposed by Tversky accounts for a generally held view that the actual disposition or composition of items in the choice set will affect judgments of individual members (Payne, 1982), and it does so following a simple principle: similar products that shares certain aspects (or attributes) can be intuitively be perceived as dividing the loyalty associated with the original, most similar item. Despite the efforts of the academic community to strengthen the understanding and the acceptance of the similarity effect by proposing many variations to the original model of Tversky (McFadden, 1980; Daganzo, 1979; Batsell, 1980), the postulate still presents some problematic implications since there are too many parameters the theory should account for, in order to be useful for predictions.

A further explanation in support of the attraction effect was given by Simonson (1989), who stated that when consumers expect others to evaluate their choices, they will try to predict and anticipate what is likely to affect others’ preferences. Moreover, the salience of the dominance relationship (that is, the extent to which one of the similar products is clearly superior to the other) might lead consumers to believe that that particular attribute will guide the judgment of others evaluating the same choice set (Taylor and Fiske, 1978).

2. The Decoy Effect

The decoy effect is the phenomenon in which the introduction of an irrelevant alternative (i.e. the decoy) in a choice set will significantly affect and change the probability shares of being selected of the items in the original choice set (Huber, Payne and Puto, 1982; Hedgcock et al., 2009). The decoy itself can essentially be of two distinct types: either viable, meaning that the newly introduced product in the choice set is not an irrelevant alternative but rather a less convenient one concerning certain attributes, therefore still potentially selectable by the respondent, or to use the words of Hedgcock et al (2009) “force the decision maker to evaluate a trade-off between the decoy and other competing alternatives”; or it could be an asymmetrically dominated decoy, meaning that the new item is clearly dominated by another item in the choice set (which is referred to as the target item) but at the same time dominates other items either completely or only concerning the level of certain attributes. The main difference hence lies on the fact that viable decoys can still be potentially evaluated in terms of being the preferred alternative in the choice set, while dominated ones cannot. A large body of academic literature has focused on the study of dominated decoys, and on the mechanisms that allow an irrelevant alternative to alter the probability shares of items in a choice set with its simple presence. In fact, if one one hand it is nowadays broadly accepted the way in which dominated decoys challenge Luce’s IIA or the principle of regularity (or proportionality), on the other hand, still many mechanisms need to be uncover concerning the magnitude of the decoy effect in various different contexts.

robustness, holding without respects of the type of product or service offered (Sellers-Rubio & Nicolau-Gonzalbez, 2015). More in detail Heat & Chatterjee (1995) analyzed the decoy effect of both low- and high-quality brands: the results of their meta-analysis and own experiment revealed how the decoy effect is more effective versus low-quality brands (in the sense that its presence mostly reduces the original probability shares of low-quality brands) rather than high-quality ones. Also, whereas Luce’s IIA is violated in every case (meaning that the introduction of an irrelevant alternative is always expected to affect the choice probabilities) the principle of regularity, or proportionality effect, is only violated when high-quality brands are targeted with the aid of a decoy. Furthermore, their study reveals how the decoy effect is mostly effective in increasing the choice probabilities of high-quality brands, while it rarely applies also to low-high-quality ones.

Kim et al. (2006) found positive evidence supporting the hypothesis that the decoy effect is negatively moderated by the strength of the brand: the theoretical foundation of such view resides into two distinct theories identified by the authors, namely the averaging process view (Anderson, 1971) and the category-based process view (Meyers-Levy & Tybout, 1989). The first theory refers to the fact that attribute’s importance is adjusted according to the weight of the other attributes being considered, hence averaging the attribute’s strength. The second one instead refers to how a brand name can signal the category label and hence lead to category-based evaluations. In their study, the authors did not record a decoy effect in the case in which one or some of the brands in the product category were well known by the decision makers. The results were hence in line with what predicted by the category-based process view, while no evidences where found in support of the averaging process view. Moreover, Kim et al. (2006) also recorded an opposite tendency of the decoy effect in those cases where the category under evaluation was particularly popular and well known by consumers: indeed, in such cases the inclusion of a dominated decoy of the same brand as the target product yielded to a decrease in the choice shares of the target product (compared to the choice shares recorded without the presence of the decoy). The authors explain this phenomenon as the result of the so-called persuasion knowledge (Friestad & Wright, 1994) according to which customers are sometimes equipped with enough product and persuasion knowledge to respond negatively to similar kinds of manipulations, therefore inferring a negative motive about that brand.

• Weight Shift. The first class of explanations focuses the attention on the relative importance and strength on the attributes. Accordingly, focusing the attention of the decision maker on a particular attribute may alter the associated perceptions and thus leading to over-emphasize the relative importance of that attribute. In turn, this poses the attention on those attribute ‘stimulated’ by the introduction of the decoy, which by definition underlines a particular strength of the target product. This explanation is also consistent with the averaging process view taken into account by Kim et al. (2006). Moreover, this view is also supported by other authors: Ariely & Wallsten (1995) concluded that the evaluation process over a set of alternatives differing on one particular dimension will drive decision makers to select the option that dominates on that attribute. Mellers & Biagini (1994) argue that “the similarity along one attribute magnifies differences with others”, and accordingly the item which dominates on that attribute will have a greater choice share. • Value Shift. The second class of explanations follows the idea of a change in the perceived value

of an object when a new item is introduced in the choice set. The introduction of a decoy has two possible outcomes according to this view: either increases the range in one dimension (thus potentially focusing the attention on that dimension and increasing its relative strength; see weight shift in the previous paragraph), or it might change the stimulus rank, by changing the frequency of options on a particular dimension (Pettibone & Wedell, 2000; Hedgcock et al., 2009).

• Loss Aversion. Pettibone & Wedell (2000) also identified loss aversion as a mechanisms accounting for shift values associated with the options: In fact, if the decoy becomes a reference point for evaluation, the target will always be the preferred option as it dominates the decoy entirely, while competing products which only partially dominate the decoy will suffer from a reduced probability share. That is, because of loss aversion, the target appears more attractive than the competing products. Also in the case in which the reference point results from an averaging process of all attributes, the introduction of a decoy may induce a shift in the reference point ultimately increasing the attractiveness of the target. The fact that decisions are influenced by the way in which decision makers assign reference points is also consistent with other studies, which proved how consumers often use attributes levels of commonly chosen alternatives as a reference point for current and future items evaluations (Heat et al., 2000; Hardie et al., 1993) • Emergent Value. This last class of explanations is instead somehow different from the others, as

(Hedgcock et al., 2009). In fact, as it as been already shown in previous sections, the importance of attribute-based heuristics is at the core of many decision processes and one of the main pillars of prospect theory (Huber, Payne and Puto, 1982; Kahneman and Tversky, 1979; Kahneman, 2011).

CONCEPTUAL FRAMEWORK

As mentioned in the introduction, the main goal of this research is to deepen the understanding of the decoy effect, in particular concerning the way in which probability shares assigned to the items in the choice set change after the removal of the decoy: in fact, if on one hand there exist many studies which analyze the decoy effect from a multitude of points of view like the ones reviewed in the literature sections, on the other it has been never analyzed how the removal mechanism of the decoy affects the decoy-influenced probability shares.

In particular, this is the first study attempting to understand how the removal of a decoy affect the original and the decoy-influenced probability shares, and it is the first one which tries to simulate this effect through the manipulation of the anchor that serves as a foundation of the arbitrary coherence of individual decision makers.

moving the choice shares up (down) if the target is a premium (discount) item, or producing an opposite reaction resulting in an overall reduction of the target’s choice share.

H1: The probability shares after the removal of the decoy (Group 1) will be significantly different

from the ones obtained from the choice set without the decoy (Group 3) and the one with the decoy included (Group 2).

Premium vs Discount. This set of hypotheses regards the difference effect that decoy removal will have on PLs and NBs or more broadly on premium versus discount products. Beside the almost total absence of studies investigating the moderating role that premium/discount or PL/NB have on the removal of the decoy. However, some of the studies we have already seen in the literature section concerning the decoy effect can offer some hints about the effect that we could expect. Recall that Kim et al. (2006) found positive evidence supporting the fact that the strength of the brand negatively moderates the decoy effect: their analysis show how a decoy item related to a premium target item has more power in ‘stealing’ choice shares from lower quality items than a decoy related to a discount item would do to a premium one. Heat and Chatterjee (1995) found evidence supporting the same thesis, that high quality products benefit much more from decoy effect with respect to their lower quality competitors. It can therefore be stated the following:

H2: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

premium target products (olive oil category) than for discount target products (coffee category).

H3a: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

Premium target products if the respondent is a traditional subject (quality oriented), rather than a non-traditional respondent (cost oriented).

H3b: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

Discount target products if the respondent is a non-traditional respondent (cost oriented), rather than a traditional respondent (quality oriented).

Moreover, based on the reasoning above and recalling the findings of Kim et al. (2006) and Friestad & Wright (1994), it can also be expected that the inclusion of a decoy, beyond its effectiveness in enhancing the choice probabilities of the target item, will have the indirect effect of decreasing the satisfaction associated with the assortment. In particular, this effect will be more visible in those customers that might be equipped with the previously mentioned persuasion knowledge, that in turn should be the ones which are more focused on a thorough evaluation of the assortment because of budget constraints, hence non-traditional respondents:

H3c: Non-traditional subjects (cost oriented) will be more likely to express a lower assortment

satisfaction in the presence of a decoy, than traditional subjects.

Control Variables. Beyond the ones described above, control variables will also be collected during the survey. These variables are: gender, age, household income, level of education reached and the number of times the respondents buy groceries in a week. For the first two variables no hypotheses will be developed, as the choice mechanisms under investigation are not expected to vary in accordance to a respondent’s age and gender.

Concerning the household income, we can expect, based on what has been said concerning traditional (quality oriented) and non-traditional (price oriented) subjects, that as the income grows the respondents are increasingly driven to purchase more expensive, high quality products rather than budget items. Therefore, similarly to traditional subjects, it can be expected that respondents disposing of higher incomes will be more likely to purchase the target product. Hence, the following hypothesis can be formulated:

H4a: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

Regarding the other two control variables, their effect can be expected to be specular: recalling what stated by Friestad & Wright (1994), decision makers equipped with either (1) higher level of educations or that (2) perform enough trip to the store to become, over time, acquainted with retailer’s tools and techniques, and therefore acquire a certain degree of persuasion knowledge. It can hence be expected that such respondents will favor less the target products, while at the same time will associate lower levels of satisfaction to the assortment:

H4b: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

target products if the respondent exhibits a lower level of education.

H4c: Respondents with lower levels of education will be more likely to express a lower assortment

satisfaction in the presence of a decoy, than traditional subjects.

H4d: The probability shares after the removal of the decoy (Group 3) will be significantly larger for

target products if the respondent performs a lower number of trips to the store per week.

H4e: Respondents performing more trips to the store each week will be more likely to express a lower

assortment satisfaction in the presence of a decoy, than traditional subjects.

Below, the graphical representation of the conceptual framework of this study, with the dependent variable on the right blue box, and the independent and control variables in the white boxes:

THE EXPERIMENT

As we have seen in the literature section, many different studies focused their attention on understanding the consequences of items introduction and from a variety of points of view. But when it comes to understand the effect of item removal prior to this introduction, the task is much more complicated to be replicated in a controlled experiment: many of the unconscious flaws we have seen so far are sometimes an outcome of the precise way in which these get portrayed and introduced in the set, rather than the effect that that particular product’s introduction yield in a choice set. The main problem is that understanding the decoy effect in a retail setting requires a longitudinal data set of real retailers in which certain categories of products have been altered with the addition of a decoy for a certain period of time. But given the absence of such a dataset, and given the (necessarily) limited scope of this master thesis, how is it possible to understand how choice shares in a given set of alternatives change after the removal of a product? One idea could be to imply and manipulate those same cognitive flaws that alter our supposedly rational behavior when dealing with choices. Therefore, the experiment will be structured in the following way: three distinct groups of respondents will be surveyed; the first group will first have to look at one product assortment (with the decoy) and then a second time (this time without decoy that will be removed) and will be asked to (i) assign probability shares to each item in the assortment, and (ii) to evaluate the overall assortment in terms of perceived quality, depth, width and completeness. It has to be noticed that between the first time in which respondents will see the image and the second, an intermediate section has been developed in order to collect general informations about the respondents. This section has the twofold aim of gathering variables that will be used in the analysis (such as traditional/non-traditional nature of the subject and control variables) and to buffer the attention of the respondents, to ensure a certain degree of ‘distraction’ between the two times the images are shown. The goal, is to stimulate an unconscious recall of information from the first time the decision maker sees the shelves, in order to bias the final product selection.

The second group will undertake the same procedure with the only difference that in this case the decoy will be included in both the first and the second time the images will be shown. The third group, instead, will never be exposed to the decoy, neither the first nor the second time in which the choice sets will be shown.

will seek to answer the question concerning the items and the whole assortment keeping in mind the original choice set with the decoy included, hence reproducing the effect of a decoy removal. Concerning the decoy, it has to be noticed that the differences in prices among products have been standardized, meaning that in both categories the prices always increase by 10%, and that the decoys are target product offering 50% of amount less of the product, at a price that is 5% lower than the target.

Below, a representation of the experiment’s design:

Figure 2 – The Experiment of this Study

METHODOLOGY

Another FA will be performed on the six items involving the evaluation of the assortments of the two categories (three items per each category), from which other two factors will be extracted, in order to identify the degree to which a single respondent can be classified as ‘satisfied’ from the assortment or not. The FA will be performed using principal component analysis, with varimax rotation and pairwise deletion of missing data. Also in this case, since the aim of this FA was to identify two factors (one for Coffee and one for Olive oil) the FA was designed to extract 2 factors.

Other two FAs will be run on the eight items (four per each category) through which the level of knowledge of the brands in each choice set can be assessed for each respondent. The two FAs will be performed using principal component analysis, with varimax rotation and pairwise deletion of missing data. In this case, these FAs were designed to extract factors with Eigenvalues > 1.

For all of the four FAs explained above, the factor scores will be saved and recorded as variables in order to be used in the main analyses. After the initial preparation of the dataset (each step is extensively explained in the ‘Extended Walkthrough Analysis’ in Appendix A) the main analyses will be performed.

First, an ANOVA (Analysis of Variances) test will be performed on the different groups of respondents in order to test whether there is a statistically significant difference in the selection of the target item in the choice sets (Hair et al., 2010). In particular, the test is needed to identify differences in the choice shares of the target items in the two product categories, olive oil and coffee.

Afterwards, in order to finally test most of the hypotheses two different types of regressions will be performed: first of all, when the DV variable is a categorical variable (for example, ‘choose the preferred product in the choice set’, with 5 levels indicating the 4 different products plus the non-choice, or ‘general assortment satisfaction’ with 5 levels ranging from strongly disagree to strongly agree) a Multinomial Logistic Regression will be performed (Hair et al., 2010). This type of analysis basically automatically performs a set of binary logistic regressions for each label in response to the reference level, therefore allowing for categorical variables to be treated as DV. In particular, three different MLR will be run, two related to the relative product categories (one for the olive oil category and one for the coffee category), while the third one will have as a DV the general assortments satisfaction in which the two assortments are evaluated together, expressed by the respondent on a 1 to 5 Likert scale. The formula expressions for the three MLRs are reported below:

Prob[Choose_Oil = i] = β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5 (Choice_Set) + β6

(H/U_OliveOil) + β7 (NoticingD) + β8 (Factor_COST) + β9 (Factor_QUALITY) + β10 (Factor_KNOWLEDGEoil) + β11

(Frequency_OliveOil) + εI

Where i N, [1 ; 5]

2) MLR, DV: Choose_Coffee (Segafredo is the target product in the choice set)

Prob[Choose_Coffee = i] = β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5 (Choice_Set) + β6

(H/U_Coffee) + β7 (NoticingD) + β8 (Factor_COST) + β9 (Factor_QUALITY) + β10 (Factor_KNOWLEDGEcoffee) + β11

(Frequency_ Coffee) + εi

Where i N, [1 ; 5]

3) MLR, DV: General_Assortment_Satisfaction

Prob[General_Assortment_Satisfaction = i] = β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5

(Choice_Set) + β6 (H/U_Coffee) + β7 (H/U_Oil) + β8 (NoticingD) + β9 (Factor_COST) + β10 (Factor_QUALITY) + β11

(Factor_KNOWLEDGEoil) + β12 (Factor_KNOWLEDGEcoffee) + β13 (Frequency_ Oil) + β14 (Frequency_ Coffee) +

β15 (ChooseToil) + β16 (ChooseTcoffee) + εi

Where i N, [1 ; 5]

Robustness Check. In order to counter verify and strengthen the findings of the MLRs, a second kind of analysis will be performed. A logit regression will be run for each MLR, this time having as a DV the Choose_Target variable, which takes on value 1 when the target product in the choice set is selected, 0 in every other case. A logit regression is used to calculate the regression coefficients in the case of a linear probability model. A LPM is simply a regression model in which the dependent variable is binary. Therefore, a Logit regression will be used as it relies on the standard cumulative probability distribution function (CDF), hence avoiding non-sense outcomes (Hair et al., 2010). A logit is, as has been pointed out before, the basic analysis that constitute the building blocks for the MLR, in which a set of binary logistic regressions is run independently vis-à-vis the reference level. For each of the three MLR presented above, a logit regression will be run afterwards, made exception for the assortment satisfaction, for which two logit regressions will be performed, one having as DV the olive oil assortment satisfaction and one with the coffee assortment satisfaction (both variables taking on value 1 for med-high to high values of satisfaction expressed, 0 for lower values).

1) Logit, DV: Choose_Target_OliveOil (Il. Kalamata is the target product in the choice set) Prob[Choose_T_OliveOil = 1] = Φ [β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5 (Choice_Set) +

β6 (H/U_Oil) + β7 (NoticingD) + β8 (Factor_COST) + β9 (Factor_QUALITY) + β10 (Factor_KNOWLEDGEoil) + β11

(Frequency_ Oil)] + εi

2) Logit, DV: Choose_Coffee (Segafredo is the target product in the choice set)

Prob [Choose_T_Coffee = 1] = Φ [β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5 (Choice_Set) + β6

(H/U_Coffee) + β7 (NoticingD) + β8 (Factor_COST) + β9 (Factor_QUALITY) + β10 (Factor_KNOWLEDGEcoffee) + β11

(Frequency_ Coffee)] + εi

3) Logit, DV: OliveOil_Assortment_Satisfaction & Coffee_Assortment_Satisfaction

Prob [Category_Assortment_Coffee = 1] = Φ [= β0 + β1 (Age) + β2 (Gender) + β3 (Income) + β4 (Education) + β5

(Choice_Set) + β6 (H/U_Coffee) + β7 (H/U_Oil) + β8 (NoticingD) + β9 (Factor_COST) + β10 (Factor_QUALITY) + β11

(Factor_KNOWLEDGEoil) + β12 (Factor_KNOWLEDGEcoffee) + β13 (Frequency_ Oil) + β14 (Frequency_ Coffee) +

β15 (ChooseToil) + β16 (ChooseTcoffee)] + εi

The Greek term Φ indicates the standard cumulative distribution function, while the betas refer exactly to the same variables as in the MLR analysis. Differently from the MLR is the interpretation of the parameters: while the MLR reports the changes associated in terms of log-odds, and can then be transformed in probabilities, the logit regression output allows directly to compare changes in probability through the computation of the CDF, Φ.

Moreover, a complete and detailed list of all the variables and their values/labels can be found in the ‘Extended Analysis Walkthrough’ in Appendix A.

THE DATA

The survey responses were collected from 477 individuals around the world, using the Author’s personal network of relatives and acquaintances. Initially, the different nationalities of the respondents have been recorded, but said variable has been discarded at the beginning of the computations since the vast majority of respondents came from the US, therefore making insignificant the amount of other nationalities collected. This

is not necessarily an issue however, given the structure of the research and its main target: recalling that the overall objective is to understand the effect of decoy removal through a controlled experiment involving individuals’ arbitrary coherence through the selection of existing items from fictitious choice set, the fact that US respondents were confronted with choice set composed of popular products sold by the

largest Dutch retailer (Albert Heijn) made it possible to buffer the effect of brands and marketing activities of the producers, given the low levels of brands’ knowledge (see figure 3). In fact, since I could not control for such product features that are often found to be crucial in influencing grocery choice such as commercial and marketing activities, this required respondents to pay a greater amount of attention to the items in the survey.

Below, an infographic of the socio-demographic characteristics of the 477 respondents who took part in the study: -100,000 200,000 300,000 400,000 500,000 0 100 200 300 400 Average Household Income 37% 63% Gender Male Female 0% 20% 40% 60% 80%

Level of Brands' Knowledge

not at all very little indifferent quite enough very well

Figure 4 – Infographic of Socio-Demographic Characteristics

In general, the socio-demographic characteristics of the respondents portrait an heterogeneous sample: made exception for the gender variable, through which is possible to notice that the majority of respondents were females, the age and income variables seems to be normally distributed. For the level of education, it can be noticed how the majority of the respondents exhibits a level that is higher than the high school diploma, which is indicative of an averagely highly educated population. The reason might simply lie behind the fact that the ‘promoter’ of this survey (people who were asked to send or to post the survey link on social networks and to lists of acquaintances) were mainly fellow MSc and specialization students.

ANALYSIS & RESULTS

21% 12% 13% 12% 42%

Choice Shares (Decoy

Removal)

The 3 images above represent the distribution of the product shares, therefore the answer to the question ‘How likely are you to purchase the items?’. As it is possible to notice, the target product, Iliada Kalamata (in the charts detached from the other products) did experience an increase, even if slight, in the associated choice shares of the target product in the moment in which the decoy was included in the choice set. What is also interesting, is to notice how, when the decoy is present, the target product records an increase in preference shares but all the other products loose shares in favor of the non-choice. Also, in the group in which the decoy has been removed (on the right) it can be noticed how the shares of the target are exactly the same as when the decoy was not present, but the shares of the competing products, in particular the lowest-priced items in the choice set (namely Monini) experienced a big loss in favor of the most similar product in terms of price, namely Carbonell. In general, it seems that Luce’s IIA theory is violated.

Below, the same output, this time relative to the ground coffee category:

Figure 5 – Choice Shares for the Ground Coffee Category in the Three Groups of Respondents

In this second set of charts the distribution of choice shares across the groups is different from the olive oil category: first, it is easy to notice how at a first glance (recalling this charts are simply computed through the weighted average of the raw responses) there is no striking difference between group 1 (decoy removal) and 3 (no decoy), where the shares of not only the target, but also of the other items including the non-choice, are roughly identical. Another thing that is easily noticeable regards the shares of the non-choice, first because its value is pretty much constant across groups, and secondly because this value is much larger than the values associated to the non-choice in the olive oil category. The strangest thing, however, regards the shares of the target item, namely Segafredo: differently from the other category in this case the presence of the decoy reduces the choice shares of the target item. At a first sight this seems in line with the findings of Kim et al. (2006) who found that the decoy is counterproductive in those categories that are popular and well known by the decision makers. This phenomenon could have also been given by the similarity effect (Tversky, 1972) according to which the two least similar products to the target in the choice set received a disproportionate amount of attention: both Lavazza and Illy are of higher price and because of the available heuristics, probably, are perceived of higher quality (recalling the reference price effect, Ariely, 2008). This however is in contrast to the pure theory of the similarity effect as it predicts that the is the most similar product, and not the target, which should suffer a decrease in choice shares in favor of the least similar ones. In our case, the most similar products in terms of price, namely Jacob, has exactly the same amount of shares between groups 2 (decoy) and 3 (no decoy), and the only product whose shares are lower is the target, Segafredo. An effect that would better fit this situation is instead the attraction effect (Huber, Payne and Puto, 1982) where the most similar, superior products receive a disproportionate greater amount of attention. In any case, also in this category, at a first sight, it seems that Luce’s IIA theory is violated, as predicted.

In order to assess with statistical precision whether the differences seen above are significant, the first analysis to be performed is an ANOVA test, as explained in the methodology section of this study. Below, the output of the first ANOVA test on the changes in percentage terms recorded for the targets products versus the choice sets, for the olive oil category:

ANOVA

% Iliada Kalamata Extra Virgin (0.5L / 17oz fl.) Sum of

Squares df Mean Square F Sig.

Total 17,416 391

For the olive oil category, it can be seen that the difference in percentage terms for the choice share of the target item (Il. Kalamata) in the three choice sets are not significant. This however does not help in untangling the question of how and due to which factors the change (however small it was) occurred. In particular, if we look at the graph in the output above reporting the mean values in each choice set, we see that there is indeed a small difference: the inclusion of a decoy alternative does increase the shares of the target, while its exclusion lowers again the shares of the target but while still maintaining a ‘premium’ (in terms of choice shares) for the target.

For the ground coffee category, the ANOVA output suggests a different relationship between the low-price target and the decoy:

ANOVA

% Segafredo (500g / 17,5oz.) Sum of

Squares df Mean Square F Sig.

Between Groups ,738 2 ,369 4,020 ,019

Within Groups 34,309 374 ,092

Total 35,046 376

First of all, in this case there is indeed a statistically significant difference in the mean percentage choice share value of the target item. What is particular in this case, is that the effect portrayed in the image above is exactly the opposite, if we look at the pattern of the line: the decoy presence in the choice set dramatically reduces the choice shares for the discount target item, while the decoy removal allows the choice shares to go back toward to the no-decoy level, but still, as in the previous category, keeping a larger (in this case smaller) share for the target.

This first set of ANOVAs show that the differences in responses do vary in according to the choice set, but this difference is not too strong based on these tests, as it has been shown how there is a strong statistical significance only in one of the two categories, ground coffee, in which the target item is a low-priced product in the choice set. Accordingly, H1 can only be partially accepted, based on the results so far. In fact, apart from general differences that can be captured by the ANOVA test, the next analyses give an idea of how, in practical term, the variables bear an effect on the likelihood of choosing the target, or in general at the choice shares of the entire set.

As explained previously, the first extended analysis consists of a Multinomial Logistic Regression (MLR): such method is appropriate when the dependent variable is a categorical variable with multiple levels, in our case the answer to the question ‘If you would have to choose ONE product in the choice set, which one would you buy?’, where each level represents one of the possible choices. For both categories, the reference label in the analysis is the last one, namely ‘None of the products’ so that it is possible to evaluate the various impacts on each of the items, including the target.

variable, instead of performing three different regressions, one per each group. In any case both cases have been tried, but when the variable ‘Choice Set’ is not included as an IV the overall model drops in significance. The reason is that it would have required to split the main population sample, which is not too irrelevant when aggregated, but whose numbers becomes much less coherent to the minimum ones for this type of analysis (since instead of a sample of 473 respondents the three separate regressions would have roughly one third of the respondents, therefore undermining the overall significance and strength of the model). As a result, the aggregated model has been chosen as the preferred option.

Below, the output of the MLR for the olive oil category:

Pseudo R-Square

Cox and Snell ,327

Nagelkerke ,343

McFadden ,130

Model Fitting Information

Model Model Fitting Criteria Likelihood Ratio Tests

-2 Log Likelihood Chi-Square df Sig.

Intercept Only 1063,167

Final 925,344 137,823 48 ,000

Parameter Estimates

If you would have to choose one product, which OLIVE

OIL would you buy? a B

Std.

Error df Sig.

(TARGET) Income ,001 ,004 1 ,679 ,995 1,008 Education ,123 ,124 1 ,321 ,887 1,442 FCOSTOr -,549 ,269 1 ,041 ,341 ,978 FQUALITYOr ,371 ,260 1 ,154 ,870 2,414 Freq (olive o) -,722 ,409 1 ,078 ,218 1,084 Fknowl(oil) ,635 ,260 1 ,015 1,134 3,139 Age ,008 ,016 1 ,597 ,978 1,040 Gender=0(m) ,491 ,459 1 ,284 ,665 4,017 Gender=1(f) 0b . 0 . . . [CS=1] -,701 ,662 1 ,290 ,136 1,817 [CS=2] -,741 ,551 1 ,178 ,162 1,403 [CS=3] 0b . 0 . . . [U/H OIL=0] -1,200 ,460 1 ,009 ,122 ,742 [U/H OIL=1] 0b . 0 . . . NoticingD=0 -,916 ,637 1 ,151 ,115 1,395 NoticingD=1 0b . 0 . . .

a. The reference category is: None of the products in the assortment b. This parameter is set to zero because it is redundant

Recalling that the reference category for the choice set under examination is the last value (‘None in the products in the assortment’) every alternative product is evaluated versus the non-choice. The above output is in the same line of the ANOVA test performed previously, where the differences among choice sets are not significant, for any of the products in the olive oil category.

First of all, it can be noticed how the coefficient for the ‘cost’ factor is negative and significant, indicating that the more cost-oriented the individual, the lower the attraction to the target, vis-à-vis the non-choice. The frequency of buying follows the same pattern, hence the higher the frequency of purchasing olive oil the lower the likelihood to choose the target item (as theorized through the literature review). Knowing the category, instead, increases the chances of preferring the target premium item against a non-choice. Also the knowledge effect is as forecasted. Moreover, individuals who perceive the olive oil category as ‘hedonic’ are much more likely to prefer the premium target versus the non-choice (in the above table the parameter is referred to an ‘utilitarian’ oriented shopper, whose likelihood to choose the target are severely lower with respect to a non-choice).

The MLR presents the parameter estimates in terms of log-odds. To understand for instance the impact of ‘hedonic’ versus ‘utilitarian’ in choosing the target product instead of the non-choice, we can transform the coefficient of the parameter (which currently indicates the differences in terms of log odds) in order to obtain probabilities:

Log odds = 0.612 -1.200 = -0.588

Odds = e (-0.588) = 0.555

Probability (Utilitarian) = 0.555/ (1 + 0.555) = 0.357

Probability (Hedonic) = e (0.612) = 1.844 / (1 + 1.844) = 0.648

This means that, ceteris paribus, an individual who perceives the olive oil category as ‘hedonic’ has (0.648 - 0.357) greater probability of choosing the premium product versus the non-choice than someone who perceives it as ‘utilitarian’, that is 29.1% of greater probability of choosing the premium target. For the shopping orientation variable, it is pointless to compute the probability value, as the coefficient is not significant, but it can still be noticed how (given what has been said above) a cost-oriented customer would, predictably, express a lower likelihood of choosing a premium, high priced product. Concerning the differences with respect to the choice sets, this test shows no statistically significant difference in the log-odds of selecting the premium target item.

The next set of analyses instead focuses on the second product category under investigation, namely ground coffee, in which the target product is the discount, low priced item in the choice set, Segafredo. Below the output of the MLR analysis:

Pseudo R-Square

Cox and Snell ,271

Nagelkerke ,285

McFadden ,103

Model Fitting Information

Model

Model Fitting

Criteria Likelihood Ratio Tests -2 Log

Parameter Estimates

If you would have to choose one product, which

COFFEE would you buy? a B

Std.

Error df Sig.

Income -,003 ,002 1 ,270 ,993 1,002 Education ,163 ,117 1 ,166 ,935 1,481 FCOSTOr -,426 ,216 1 ,048 ,427 ,997 FQUALITYOr -,385 ,214 1 ,072 ,447 1,035 Freq (coff) ,004 ,315 1 ,990 ,542 1,860 FCOFknowl ,835 ,216 1 ,000 1,508 3,523 Age -,006 ,013 1 ,620 ,970 1,019 Gender=0(m) ,325 ,381 1 ,394 ,655 2,923 Gender=1(f) 0b . 0 . . . [CS=1] -,546 ,535 1 ,307 ,203 1,652 [CS=2] -,532 ,471 1 ,259 ,233 1,479 [CS=3] 0b . 0 . . . [NoticingD=0] -,975 ,484 1 ,044 ,146 ,974 [NoticingD=1] 0b . 0 . . . [U/H COF=0] -,373 ,387 1 ,336 ,322 1,472 [U/H COF=1] 0b . 0 . . . Illy Intercept -,440 ,983 1 ,654 Income -,001 ,001 1 ,429 ,996 1,002 Education ,061 ,102 1 ,550 ,870 1,300 FCOSTOr -,540 ,201 1 ,007 ,393 ,865 FQUALITYOr -,098 ,201 1 ,626 ,612 1,344 Freq (coff) ,058 ,292 1 ,844 ,598 1,878 FCOFknowl ,690 ,207 1 ,001 1,329 2,991 Age -,012 ,012 1 ,306 ,964 1,011 Gender=0(m) -,327 ,382 1 ,392 ,341 1,525 Gender=1(f) 0b . 0 . . . [CS=1] ,403 ,518 1 ,436 ,543 4,128 [CS=2] ,412 ,469 1 ,381 ,601 3,787 [CS=3] 0b . 0 . . . [NoticingD=0] -,291 ,473 1 ,539 ,296 1,889 [NoticingD=1] 0b . 0 . . . [U/H COF=0] -,446 ,368 1 ,225 ,311 1,317 [U/H COF=1] 0b . 0 . . .

a. The reference category is: None of the products in the assortment b. This parameter is set to zero because it is redundant

In this second category, where the target product is the lowest priced in the choice set, many more

1 if the respondent is female) indicates that males are much more likely to be susceptible to the decoy, with the overall log odds that are more than double for men, if also the constant term is taken into account and no other variable is considered. To continue, the factor score related to cost-oriented decision makers is, predictably, positive and significant in predicting the choice of the discount item with respect to the non-choice. Vice versa, the factor score for quality-oriented subject is negative and significant. Concerning the frequency of purchase, the coefficient is not highly significant for the target output (while is much less significant for the other possible choices) but it still suggests that a higher number of weekly trips to the store positively impact the choice of the discount target item. Moreover, the extent to which the brands in the choice set are well known to the buyer positively impact the choice of the target vis-à-vis the non choice, and the coefficient is highly significant. Concerning the differences in choice sets, in this category the inverted decoy effect seen in the beginning of this thesis section is even more visible, also strengthen by the significance of group 2 (decoy) and of group 1 (decoy removal). The difference in this case, is that respondents that saw the decoy, and had the possibility to select it from the final choice set are dramatically less likely to select the target product, while respondents to whom the decoy was removed from the choice set are still less likely to purchase the target, but the negative effect is almost half the strength of the one where the decoy is present.

The last part of this analysis concerns the assortment satisfaction. Three hypotheses deal with levels of satisfaction, in particular with respect to shopping orientation (cost versus quality), level of education and frequency of buying (relative to the category). Also in this case, the same two kind of analysis are used: MLR in order to analyze the general assortment satisfaction, and logit regressions to analyze the category-specific declared satisfaction.

Below, the output of the MLR, with General_Satisfaction as the DV:

Pseudo R-Square

Cox and Snell ,252

Nagelkerke ,269

McFadden ,105

Model Fitting Information

Model

Model Fitting

Criteria Likelihood Ratio Tests -2 Log

Likelihood Chi-Square df Sig. Intercept Only 957,684

Parameter Estimates

Overall, I am satisfied with both the OLIVE OIL and the COFFEE assortment a

B Std. Error

agree Income -,002 ,004 1 ,684 ,990 1,007 Education -,076 ,173 1 ,661 ,660 1,302 FCOSTOr ,763 ,360 1 ,034 1,059 4,340 FQUALITYOr ,116 ,333 1 ,729 ,584 2,156 Freq (cof) -,240 ,552 1 ,663 ,266 2,321 Freq (oil) ,217 ,451 1 ,631 ,513 3,007 FCOFknowl ,777 ,424 1 ,067 ,947 4,990 FOILknowl -,296 ,408 1 ,468 ,334 1,655 Age ,022 ,022 1 ,320 ,979 1,067 Gender=0(m) -,463 ,667 1 ,488 ,170 2,327 Gender=1(f) 0b . 0 . . . [CS=1] ,434 ,897 1 ,628 ,266 8,963 [CS=2] -,238 ,786 1 ,762 ,169 3,679 [CS=3] 0b . 0 . . . NoticingD=0 -,779 ,978 1 ,426 ,068 3,121 NoticingD=1 0b . 0 . . . [U/H COF=0] -,025 ,698 1 ,971 ,248 3,827 [U/H COF=1] 0b . 0 . . . [U/H OIL=0] ,231 ,702 1 ,742 ,318 4,989 [U/H OIL=1] 0b . 0 . . . ChooseToil=0 1,463 1,131 1 ,196 ,471 39,613 ChooseToil=1 0b . 0 . . . ChooseTcof=0 -,975 ,791 1 ,218 ,080 1,778 ChooseTcof=1 0b . 0 . . .

a. The reference category is: None of the products in the assortment b. This parameter is set to zero because it is redundant

Keeping in mind that the reference category for this analysis is the first one (= Strongly disagree), as the frequency of buying olive oil increases the likelihood of being satisfied from the assortment drops dramatically, even if the coefficient is not highly statistically significant. Also, higher level of knowledge of the brands in the olive oil category are negatively associated with mid- to low-ranged values of satisfaction, meaning that a moderately minimum evaluation is expected by said individuals. Higher level of knowledge associated to the brands in the coffee category instead load positively on the highest possible value for assortment satisfaction, meaning that respondents equipped with enough brand knowledge in this category are likely to express a strongly positive satisfaction.

than quality-oriented respondents.

Robustness Check. As explained in the methodology section, a second type of analysis has been performed as a robustness check, therefore in order to strengthen the findings and the results of the other analyses so far. The first logit regression has as a DV a dummy variable which takes on value 1 when the target product (Il. Kalamata) is chosen, 0 in every other case:

Omnibus Testa

Likelihood Ratio

Chi-Square df Sig.

13,384 12 ,342

Dependent Variable: ChooseToil

a. Compares the fitted model against the intercept-only model.

Parameter Estimates

Parameter B Std. Error 95% Wald Confidence Interval

Hypothesis Test

Lower Upper df Sig.

(Intercept) 2,438 ,8904 ,693 4,183 1 ,006 [NoticingD=0] -,150 ,4342 -1,001 ,701 1 ,730 [NoticingD=1] 0a . . . . . [CS=1] -,278 ,4222 -1,106 ,549 1 ,510 [CS=2] ,059 ,3996 -,725 ,842 1 ,883 [CS=3] 0a . . . . . Gender=0(m) -,256 ,3256 -,894 ,382 1 ,432 Gender=1(f) 0a . . . . . [U/H OIL=0] -,137 ,3192 -,763 ,489 1 ,668 [U/H OIL=1] 0a . . . . . Age ,019 ,0112 -,003 ,040 1 ,096 Income 1,573E-5 ,0011 -,002 ,002 1 ,989 Education -,147 ,0939 -,331 ,037 1 ,118 FactOILknowl ,088 ,1709 -,247 ,423 1 ,605 Freq (oil) -,373 ,1967 -,759 ,012 1 ,058 FCOSTOr -,174 ,1688 -,505 ,157 1 ,303 FQUALITYOr ,010 ,1667 -,317 ,336 1 ,954

Dependent Variable: ChooseToil

a. Set to zero because this parameter is redundant.

The logit analysis does not scores particularly better compared to the MLR, but it is possible to notice how the parameters for age (positively) and frequency (negatively) are statistically significant. The age coefficient is hard to interpret: in principle, one could have expected that as the decision maker becomes older and, therefore, more acquainted with the grocery shopping experience, should exhibit a larger amount of ‘persuasion knowledge’, and hence be less likely to choose the decoy-enhanced item. One explanation, in this case where the target item is a premium, highly-priced item, is that younger individuals have generally a more constrained budgets at their disposal, and would then choose a cheaper alternative. The frequency coefficient on the other hand simply replicates the previous results of the MLR analysis, where the more frequent the category purchase, the less the likelihood of choosing the premium target, as a result of a greater amount of ‘persuasion knowledge’. One last variable that deserves attention is Education: although (slightly) non significant, the coefficient exhibits a negative sign, which is in the same line as the reasoning behind the related hypothesis. What the coefficient suggests, is that higher levels of education are associated with a lower probability that the target product will be chosen.

For the ground coffee category the same analysis is replicated, this time as a DV the dummy variable takes on value 1 if the respondents decided to select the target Segafredo, 0 in every other case:

Omnibus Testa

Likelihood Ratio

Chi-Square df Sig.

40,892 12 ,000

Dependent Variable: ChooseTcoffee

a. Compares the fitted model against the intercept-only model.

Parameter Estimates

Parameter B Std. Error

95% Wald Confidence

Interval Hypothesis Test

Lower Upper df Sig.

Age ,001 ,0014 -,001 ,004 1 ,317 Income -2,960E-5 ,0002 ,000 ,000 1 ,849 Education -,016 ,0118 -,040 ,007 1 ,160 FactCOFknowl ,001 ,0236 -,045 ,048 1 ,953 Freq (cof) ,047 ,0317 -,015 ,109 1 ,136 FCOSTOr ,075 ,0228 ,031 ,120 1 ,001 FQUALITYOr -,054 ,0222 -,097 -,010 1 ,016

Dependent Variable: ChooseTcoffee

a. Set to zero because this parameter is redundant.

As it is possible to notice, the exact same pattern outlined by the MLR is replicated, made exception for a few parameters that lose significance: cost-oriented customers and males are much more likely to select the low-price target item, and respondents who have been exposed to the decoy also in the moment of choosing (group 2) are drastically less likely to purchase the target. In the same way, respondents that have seen the decoy before its removal are still less likely to buy the target, but in a much smaller magnitude. In this case the parameter for group 1 (decoy removal) can only be taken as a suggestion, since the parameter is not statistically significant (even though the level of significance is high enough to suggest that there is indeed an effect). Also in this category, it is possible to notice that the coefficient of Education is negative, even if not strongly significant, which again suggests that highly educated individuals are less likely to be subject to the decoy mechanism. The last two logit regressions instead focus on the category-specific assortment satisfaction. Below, the output of the logit regression having as DV a dummy variable which takes on value 1 when there are mid-high to high levels of satisfaction associated with the olive oil assortment:

Omnibus Testa

Likelihood Ratio

Chi-Square df Sig.

38,494 13 ,000

Dependent Variable: OILsatisf

a. Compares the fitted model against the intercept-only model.

Parameter Estimates

Parameter B Std. Error

95% Wald Confidence

Interval Hypothesis Test

Lower Upper df Sig.

Output 7 – Logit Regression for Choose_Target_Coffee

(Intercept) ,497 ,1561 ,191 ,803 1 ,001 [NoticingD=0] -,172 ,0699 -,309 -,035 1 ,014 [NoticingD=1] 0a . . . . . [CS=1] ,035 ,0715 -,105 ,176 1 ,620 [CS=2] -,009 ,0636 -,133 ,116 1 ,892 [CS=3] 0a . . . . . Gender=0(m) -,088 ,0532 -,193 ,016 1 ,097 Gender=1(f) 0a . . . . . [U/H OIL=0] ,130 ,0521 ,028 ,232 1 ,012 [U/H OIL=1] 0a . . . . . ChooseToil=0 -,072 ,0733 -,216 ,071 1 ,323 ChooseToil=1 0a . . . . . Age ,003 ,0018 -,001 ,006 1 ,115 Income ,000 ,0002 -8,543E-5 ,001 1 ,129 Education ,002 ,0146 -,026 ,031 1 ,873 FCOSTOr ,031 ,0283 -,024 ,087 1 ,265 FQUALITYOr -,007 ,0276 -,061 ,047 1 ,801 FOILknowledge ,114 ,0279 ,060 ,169 1 ,000 Freq (oil) -,022 ,0374 -,095 ,051 1 ,559

Dependent Variable: OILsatisf

a. Set to zero because this parameter is redundant.

According to the logit output reported above, the more an individual is acquainted with the olive oil brands, the higher the probability that he/she will be satisfied with the assortment, as it can be expected. Another thing that is visible in this category is that individuals who perceive olive oil as Utilitarian express a higher likelihood to be satisfied with the assortment. Moreover, even if the relationship can not be regarded as completely precise because of the lack of significance, we can still notice how the coefficients for the two Choice sets with the decoy (one removed and one included in the choice set) are negative, when the decoy is always present (group 2) and positive, when the decoy is removed (group 1). Another way to evaluate the impact of the decoy is through the NoticingD variable, which takes on value 1 if the respondents noticed the decoy, 0 if they either did not notice the decoy, or noticed it but failed to identify the correct package sizes when asked: surprisingly, individuals who did not notice the decoy are less likely to be satisfied with the assortment, with respect to the ones that did notice the decoy in the olive oil category.

Below the output of the same logit regression this time with respect to the coffee category:

Omnibus Testa

Likelihood Ratio