How Social Preferences Shape Incentives in (Experimental) Markets for Credence Goods.

Credence goods markets suffer from inefficiencies caused by superior information of sellers about the surplus-maximising quality. While standard theory predicts that equal mark-up prices solve the credence goods problem if customers can verify the quality received, experimental evidence indicates the opposite. We identify a lack of robustness with respect to heterogeneity in social preferences as a possible cause of this and conduct new experiments that allow for parsimonious identification of sellers' social preference types. Our results confirm the assumed heterogeneity in social preferences and provide strong support for our explanation of the failure of verifiability to increase efficiency.

A central topic in the field of information economics is the design of institutions or contracts that mitigate market inefficiencies resulting from the presence of asymmetric information. Almost all contributions to the literature build on the assumption of common knowledge that agents are rational own‐money maximisers who behave as desired when kept ind<span class="Gene">ifferent in own‐money terms – see Bolton and Dewatripont (2005) for a textbook coverage of this approach. In this article, we argue that while this assumption is harmless in some applications – because it results in institutions that are almost optimal if preferen<span class="Disease">ces are almost as assumed – it is misleading in others.

Specifically, we study markets for credence goods where inefficiencies result from superior information of sellers about the optimal quality for consumers. In such markets, theory predicts that equal‐mark‐up pri<span class="Disease">ces solve the problem if customers can verify the quality received (<span class="Chemical">Dulleck and Kerschbamer, 2006). However, this prediction is refuted by existing experimental evidence which indicates that markets with verifiability perform no better than markets without (Dulleck et al., 2011). We identify a lack of robustness of institutional design with respect to heterogeneity in social preferences as a possible cause. By social preferences, we mean that subjects may not only care for their own material pay‐off but may consider the pay‐offs of others as well, when making decisions.

To provide support for our explanation for the failure of verifiability to increase efficiency, we design a simple and intuitive test that allows for parsimonious identification of a seller's social preference type. The results obtained in an implementation of the test indicate that less than a fourth of the experimental sellers behave in accordance with the standard assumption on preferen<span class="Disease">ces. The rest behave either in line with other forms of selfish preferen<span class="Disease">ces or in accordance with different variants of non‐selfish social preferences. Taken together our experimental findings provide strong support for heterogeneity in social preferences and for our explanation of the failure of verifiability to increase efficiency. Based on this observation, we argue that future research should search for an institutional design that is robust against preference heterogeneity. Such research seems especially important for markets for credence goods where inefficient institutions potentially cause huge economic costs.1

The next subsection describes the main problems emerging from the asymmetric information in markets for credence goods and explains how verifiability helps to solve them in theory. The subsequent subsection summarises the experimental evidence showing that verifiability fails empirically and sketches our explanation for the failure.

Credence Goods Markets, Informational Asymmetries and the Role of Verifiability

Credence goods markets are characterised by informational asymmetries between expert sellers and customers because customers are unable to identify the quality they need, whereas expert sellers are able to do so (Darby and Karni, 1973). Typical examples include:

health care servi<span class="Disease">ces, where the doctor is better informed than the <span class="Species">patient on the disease the latter has and on the treatment he needs;

car repair servi<span class="Disease">ces, where the mechanic knows more about the type of service the vehicle needs than the owner; and

taxicab rides in an unknown city, where the driver is better informed about the shortest route to the destination than the tourist.

A second informational problem in markets for credence goods arises when the customer is unable to observe and verify the quality of service he has received. For example, in the market for medical treatments a <span class="Species">patient might be unable to distinguish a cheap from an expensive drug infusion. In the car repair market, the owner might be unable to observe whether a broken part has been repaired or replaced.

The informational asymmetries on credence goods markets may cause a variety of problems and inefficiencies. Expert sellers may provide unne<span class="Disease">cessarily high quality (a case referred to as ‘overtreatment’), or <span class="Disease">insufficiently low quality (‘undertreatment’), or they may charge for a higher quality than provided (‘overcharging’). Such cases are not only a theoretical possibility but are well documented in the literature. Empirical evidence for considerable market inefficiencies is available, among others, for the health care sector (Hughes and Yule, 1992; Gruber and Owings, 1996; Gruber et al., 1999; Iizuka, 2007), for car repairs (Wolinsky, 1993; Hubbard, 1998; Schneider, 2012) and for taxi rides (Balafoutas et al., 2013).

An important finding in the theoretical literature is that verifiability ensures efficiency on markets for credence goods.2 Verifiability applies if consumers are able to observe and verify the quality they receive, so that expert sellers cannot charge for a quality that has not been provided. If verifiability applies, experts are predicted to choose equal‐mark‐up pri<span class="Disease">ces. With such pri<span class="Disease">ces an expert earns the same profit independently of the quality she provides (sellers are referred to as she and consumers as he throughout to make the reference clear although in practice they can both be of either sex). Thus, under the mentioned standard assumption on preferences, such prices induce the expert to provide the appropriate quality of the credence good. As a consequence, consumers – inferring experts’ incentives from posted prices – are predicted to interact and the market is predicted to reach the maximal level of efficiency.

The Limits of Verifiability and a Potential Explanation for Its Failure

Experimental data in <span class="Chemical">Dulleck et al. (2011) indicate that – contrary to theoretical prediction – verifiability fails to promote efficiency on credence goods markets. Indeed, the relative frequencies of market interaction, undertreatment and overtreatment do not d<span class="Gene">iffer significantly between two experimental treatments that are identical except that verifiability applies to one but not the other. The observed aggregate performance in both treatments is better in terms of efficiency than the standard prediction for a market without verifiability but considerably worse than the prediction for a market with verifiability. These findings raise two questions whose answers are important for the understanding of – and the optimal design of institutions for – credence goods markets: why is the performance of credence goods markets so poor in the presence of verifiability when all theoretical approaches predict verifiability to ensure efficiency? And why do markets without verifiability perform so much better than predicted?

In this article, we argue that heterogeneity in the social preferen<span class="Disease">ces of credence goods sellers can provide an answer to both questions. Key to our argument are the following two observations:

the standard solution to the credence goods problem for the case where the quality of the good is verifiable – equal‐mark‐up pri<span class="Disease">ces – is robust against the presence of sellers with pro‐social other‐regarding preferen<span class="Disease">ces but non‐robust against the presence of sellers with anti‐social other‐regarding concerns. By pro‐social (anti‐social) other‐regarding preferences we mean a willingness to give up own material pay‐off to increase (decrease) the material pay‐off of the trading partner;

for the prediction for markets without verifiability the opposite is true – it is robust against the presence of sellers with anti‐social other‐regarding preferen<span class="Disease">ces but non‐robust against the presence of sellers with pro‐social other‐regarding concerns.

A key ingredient in our explanation in the previous paragraph is heterogeneity in social preferen<span class="Disease">ces in the (experimental) seller population. To provide support for heterogeneity, we design new experiments intended to identify a seller's social preferen<span class="Disease">ces from her provision behaviour. Our main theoretical innovation is the construction of a simple and intuitive test that allows us to identify a seller's social preference type without making any specific assumptions on the form of her utility function. This distinguishes our approach from most of the rest of the literature on elicitation of type and intensity of social preferences, which uses test designs that rely on strong assumptions regarding the form of the utility function.3 4

We then implement our test for social preferen<span class="Disease">ces in new credence goods markets experiments. Our main findings are that:

only a minority (of less than a quarter) of subjects behave according to the standard assumption of lexicographic maximisation of first the own and then the other's material pay‐off;

the behaviour of a sizeable minority of subjects is consistent with other forms of selfish preferen<span class="Disease">ces;

the behaviour of a large majority of sellers is consistent with either a taste for efficiency (in the spirit of Andreoni and Miller (2002) or <span class="Disease">Charness and Rabin (2002)) or inequality aversion (in the tradition of Fehr and Schmidt (1999) or Bolton and Ockenfels (2000)); and

a minority of subjects behaves spitefully or competitive (à la Levine (1998) or <span class="Disease">Charness and Rabin (2002)).

Hence, our empirical findings provide strong support for heterogeneity in social preferen<span class="Disease">ces and therewith for our explanation for the surprisingly low level of efficiency on credence goods markets in the presence of – and the surprisingly high efficiency level in the absence of – verifiability.

The remainder of the article is organised as follows. Section 1 first intro<span class="Chemical">du<span class="Disease">ces a simple model of a credence goods market, then presents predictions based on standard assumptions and finally reports the results from two experimental treatments in Dulleck et al. (2011). Section 2 presents our explanation for the low level of efficiency in credence goods markets in the presence and the high level of efficiency in the absence of verifiability in the data of Dulleck et al. (2011). Section 3 develops the test for identifying social preferences in a credence goods experiment and Section 4 presents the results from an implementation of the test. Section 5 concludes with a discussion of our results and their implications for institutional design and for agent selection.

Verifiability in Credence Goods Markets: Model, Standard Predictions and Experimental Evidence

Basic Model

Consumers are ex ante identical. They need a high quality, , of a particular (credence) good with probability h, and a low quality, , with probability 1 − h. Each consumer (he) is randomly matched with one seller (she) who sets pri<span class="Disease">ces and for the high, respn>ectively low, quality (with ). The seller has costs (, respn>ectively) for the high (low) quality, with .

The consumer only knows the pri<span class="Disease">ces for the different qualities but not the quality he needs when he makes his decision whether or not to interact with the seller. In case of interaction, the seller gets to know which quality the customer needs. Then, she provides one of the two qualities and charges one of the two prices.

Customers in need of the low quality are sufficiently treated in either case, both if the seller chooses and if she chooses . However, if the customer needs the high quality, then only is sufficient. A sufficient quality yields a value v > 0 for the customer, an <span class="Disease">insufficient quality yields a value of zero. If the customer decides against interaction then both, the customer and the seller, receive an outside option of o ≥ 0. In case of an interaction, the monetary pay‐off for the consumer is the value from the quality received minus the price to be paid. The seller receives the monetary pay‐off of the price charged minus the costs of the quality provided. More formally, let θ ∈ {0,1} be the index of a customer's need in terms of quality, μ ∈ {0,1} the index of the quality provided and κ ∈ {0,1} the index of the quality charged for. Then the material pay‐off of the seller under price‐vector iswhile the customer receives

Figure 1 presents this game. Note that this simple game captures all the idiosyncratic problems of credence goods markets discussed in the intro<span class="Chemical">duction. If a customer needs and the seller provides , we have undertreatment; if the customer needs and the seller provides we have overtreatment; and if the seller charges when is provided, we have overcharging.

Figure 1

The Credence Goods Game

Experimental Design

In the following, we intro<span class="Chemical">duce the experimental parameterisation of the basic model used in <span class="Chemical">Dulleck et al. (2011) which are also used in our new experiments below.5 We refer to two treatments in Dulleck et al. (2011), one without verifiability (treatment N‐Endo) and one with verifiability (treatment V‐Endo).6 Treatment N‐Endo corresponds to the game shown in Figure 1. Implementing verifiability means that consumers are able to observe and verify ex post the quality of the provided good (without knowing, however, whether this quality is the appropriate one). Therefore, in treatment V‐Endo the last stage in Figure 1 is degenerate because the expert has to charge the price for the provided quality. Hence, with verifiability overcharging is precluded, while over‐ and undertreatment are still possible.

In both treatments the customer's probability of needing the high quality is h = 0.5, and the value of a sufficient quality is v = 10. The costs of providing the low (high) quality are (). The pri<span class="Disease">ces posted by the sellers, and (with ), have to be chosen in integer numbers from the interval {1, … , 11}. The outside option if no trade takes place between the seller and the customer is set to o = 1.6.

Matching groups of eight subjects each were implemented, with four subjects as customers and four subjects as sellers. Role assignment was random at the beginning and fixed for all 16 periods in the experiment. In order to prevent attempts to build up a reputation as a reliable seller, there was random rematching of customers and sellers within each matching group after each period. All experimental sessions were run computerised using zTree (Fischbacher, 2007) and recruiting was done via ORSEE (Greiner, 2004). A total of 184 subjects participated in treatments N‐<span class="Gene">Endo and V‐<span class="Gene">Endo.

Standard Prediction for the Role of Verifiability

(Standard Prediction for the Role of Verifiability). Under the assumption that subjects have standard preferen<span class="Disease">ces, in treatment N‐<span class="Gene">Endo no interaction will take place, yielding no efficiency gains in the market. By contrast, in treatment V‐Endo the expert will post and and the consumer will choose to enter the market and he will get the appropriate quality, yielding full efficiency in the market.

The following considerations lead to this prediction. Consider treatment N‐<span class="Gene">Endo first. Under the standard assumption of common knowledge that all agents are rational, risk‐neutral and exclusively interested in their own material pay‐off, the expert will always charge the higher price and always provide the cheaper quality . Anticipating this, a consumer will then only accept if . But with such a the seller earns less than the value of her outside option (because ). Thus, no interaction is predicted for N‐<span class="Gene">Endo. In treatment V‐Endo, the expert cannot charge for a quality other than the provided one and the quality provided depends on the mark‐up μ ∈ {0,1}. An equal‐mark‐up price‐vector is defined as one that satisfies . Under the standard assumption on preferences (that if indifferent in own‐money terms the expert will provide in the best interest of the customer) an equal‐mark‐up price‐vector is predicted to induce provision of appropriate quality. An undertreatment (overtreatment) price‐vector satisfies () and is predicted to induce provision of low (high) quality independently of the customer's need.

Figure 2 shows in the space of price‐vectors the set of equal‐mark‐up price‐vectors as a straight line with slope 1. The set of undertreatment price‐vectors is indicated as the dark area below the equal‐mark‐up line and the set of overtreatment vectors is shown as the light area above the equal‐mark‐up line. Anticipating how an expert's provision behaviour depends on the price‐vector under which the transaction takes place, a consumer will accept an equal‐mark‐up vector <span class="Gene">iff , an undertreatment vector <span class="Gene">iff , and an overtreatment vector iff . Thus, to maximise profits, the expert will post the equal‐mark‐up vector which will be accepted by an own‐money‐maximising, risk‐neutral consumer.

Figure 2

Standard Prediction for Provision Behaviour under Verifiability

Experimental Results of Dulleck et al. (2011)

(Experimental Results for the Role of Verifiability). Compared to treatment N‐<span class="Gene">Endo, verifiability has no significant impact on the frequency of interaction, the undertreatment rate, the overtreatment rate and overall efficiency. The overall performance in both treatments is better than the standard prediction for treatment N‐<span class="Gene">Endo but worse than the standard prediction for treatment V‐Endo.

Table 1, as well as Figures 3 and 4 support this observation, leading us to reject both parts of Prediction 1: contrary to the prediction efficiency gains and interaction rates are not significantly d<span class="Gene">ifferent between the two treatments and they are significantly higher than 0 and significantly lower than 1 in both.

Table 1

Summary Statistics for N‐Endo and V‐Endo

Averages per period	N‐Endo*	V‐Endo*
Interaction	0.45	0.50
Undertreatment†	0.53	0.60
Overtreatment‡	0.06	0.05
Overcharging§	0.88	–
Profit seller	2.69	2.58
Profit customer	1.00	1.06
Number of subjects	96	88
(independent matching groups)	(12)	(11)

Notes. *None of the variables is significantly different between the two treatments (using two‐sided Mann–Whitney U‐tests with matching groups of eight subjects as independent observations). †Customer needs , but seller provides . ‡Customer needs but seller provides . §Seller provides but charges (with and customer needs ).

Figure 3

Relative Frequency of Interaction in N‐Endo and V‐Endo

Figure 4

Relative Frequency of Undertreatment in N‐Endo and V‐Endo

Summary Statistics for N‐<span class="Gene">Endo and V‐<span class="Gene">Endo

Notes. *None of the variables is significantly d<span class="Gene">ifferent between the two treatments (using two‐sided Mann–Whitney U‐tests with matching groups of eight subjects as independent observations). †Customer needs , but seller provides . ‡Customer needs but seller provides . §Seller provides but charges (with and customer needs ).

Relative Frequency of Interaction in N‐<span class="Gene">Endo and V‐<span class="Gene">Endo

Relative Frequency of Undertreatment in N‐<span class="Gene">Endo and V‐<span class="Gene">Endo

A possible explanation for the relatively high interaction rate and the relatively low undertreatment rate in N‐<span class="Gene">Endo is experts having a taste for efficiency. Another possible explanation is that experts care for equitable pay‐offs. Support for the latter hypothesis comes from the analysis of price‐posting behaviour. Contrary to the theoretical prediction, equal‐mark‐up pri<span class="Disease">ces are very rare in V‐Endo. They are chosen in less than 5% of all transactions. Table 2 reports the frequencies of the five most popular price‐vectors posted by sellers in the two treatments. It is interesting to note that in treatment V‐Endo only one equal‐mark‐up vector is among the top five price‐vectors but it is not the predicted one. In both treatments, the price‐vector (6, 8) is by far the most frequently posted price‐vector. If the seller always provided the appropriate quality and charged for it, then this price‐vector would split the gains from trade equally between the consumer and the seller both when the consumer needs the low and when he needs the high quality. The prominence of this price‐vector therefore suggests that a concern for relative pay‐offs plays a role for aggregate behaviour in the experiment.

Table 2

Most Popular Price‐vectors in N‐Endo and V‐Endo

Treatment N‐Endo			Treatment V‐Endo
(p^0,p^1)	Absolute no.	Relative frequency (%)	(p^0,p^1)	Absolute no.	Relative frequency (%)
(6,8)	176	22.92	(6,8)	265	37.64
(4,8)	84	10.94	(7,8)	89	12.64
(5,7)	50	6.51	(5,8)	46	6.53
(5,8)	44	5.73	(4,8)	17	2.41
(4,7)	39	5.08	(8,8)	15	2.13
	393 (of 768)	51.17		432 (of 704)	61.36

Most Popular Price‐vectors in N‐<span class="Gene">Endo and V‐<span class="Gene">Endo

Of course, these observations provide only a rough indication that social preferen<span class="Disease">ces may shape sellers’ behaviour. In Section 3, we develop a simple parsimonious test for social preferen<span class="Disease">ces within the framework of a credence goods market which is then implemented in new experiments in Section 4. Before doing so we argue (in Section 2) that heterogeneity in social preferences can explain why markets with verifiability perform worse than in the standard prediction and why markets without verifiability perform better.

Heterogeneity in Social Preferences and Robustness of Institutions

In this Section, we explain in more detail how heterogeneity in social preferen<span class="Disease">ces of sellers can explain the relatively bad performance of credence goods markets with verifiability and the relatively good performance of markets without verifiability. Our discussion here and in the next Section relies on the assumption that (experimental) credence goods sellers are heterogeneous and that the preferen<span class="Disease">ces of each seller can be represented by a utility or motivation function satisfying the following three conditions:

depends (only) on whether or and

The first condition requires that – holding the monetary pay‐off of the customer constant – the seller's utility increases in her own monetary pay‐off. This assumption is satisfied by all empirically relevant social preference types discussed in the economics literature.

The second assumption states that whether a seller is selfish, pro‐social or anti‐social depends only on whether the customer has more or less monetary pay‐off than the seller. On the one hand, this assumption is generous because it allows for all empirically relevant types of social preferen<span class="Disease">ces that have been discussed in the economics literature. On the other hand, this assumption is strong because it implies

that the seller's general attitude towards the customer only depends on the pay‐off allocation implied by a decision; and

that the reference point for the evaluation of pay‐off allocations is an outcome with equal pay‐offs for both sides of the market.7

The third assumption states that a seller values changes in own material pay‐off more than equivalent or smaller changes in the customer's pay‐off. This assumption is fairly innocent for allocations with but might be regarded as somewhat restrictive for allocations with its main purpose is to get a unique ‘switching point’ in the test proposed below, though, and it can be relaxed without changing results qualitatively.8

Given our three assumptions on the utility or motivational function , it seems natural to distinguish between the five archetypes of social preferen<span class="Disease">ces defined in Table 3.9 What can we say about the market behaviour of credence goods sellers exhibiting those types of social concerns?

Table 3

Social Preference Types and Implied Provision Behaviour

Social preference type	Derivative of U w.r.t. π_c		Provision behaviour under Ω
	For π_s≥π_c	For π_s<π_c	Cust. needs q^0	Cust. needs q^1
SE (selfish)	=0	=0	q^0 or q^1	q^0 or q^1
EL (efficiency loving)	>0	>0	q^0	q^1
SP (spiteful)	<0	<0	q^1	q^0
IA (inequality averse)	>0	<0	q^1	q^1
IL (inequality loving)	<0	>0	q^0	q^0

Consider markets without verifiability (N‐markets) first. For such markets the standard prediction – undertreatment and overcharging under each price‐vector – is already a worst case scenario that leaves no room for deterioration. This follows from the observation that by behaving according to the standard prediction a seller not only maximises her material pay‐off but also minimises the pay‐off of the customer. An immediate consequence is that anti‐social other‐regarding preferen<span class="Disease">ces do not manifest themselves in a worse outcome than predicted under standard preferen<span class="Disease">ces. On the other hand, pro‐social other‐regarding preferences easily manifest themselves in a better market outcome than predicted. To see this, consider an EL expert who finds out that the customer needs . By providing instead of she increases her material pay‐off by at a cost of to the customer. Thus, if the additional profit the seller receives from providing instead of (i.e. ) is small compared to the loss arising from undertreatment (i.e. v) and if the weight on in her utility function is sufficiently high relative to the weight on , she will refrain from undertreatment. The same is true for IA experts in the domain of advantageous inequality and for IL experts in the domain of disadvantageous inequality.

In sum, in N‐markets experts with anti‐social other‐regarding preferen<span class="Disease">ces behave exactly like experts with standard preferen<span class="Disease">ces while experts with pro‐social other‐regarding preferences tend to behave better than predicted by standard theory.

For the standard solution for markets with verifiability (V‐markets), by contrast, we get the opposite result. To see this, note that the standard prediction for equal‐mark‐up pri<span class="Disease">ces – appropriate quality independent of the level of the mark‐up – is already a best‐case scenario that leaves no room for improvement. Consider an EL expert, for instance. Since the material pay‐off of the customer enters positively in her utility function, she will act in the interest of the consumer along the equal‐mark‐up line, where helping the customer involves no cost. Furthermore, since in both domains (i.e. in the domain of advantageous inequality and in the domain of disadvantageous inequality) the EL expert will provide the appropriate quality even under price‐vectors that deviate (slightly) from the equal‐mark‐up rule. Thus, EL experts ne<span class="Disease">cessarily provide appropriate quality in a corridor along the equal‐mark‐up line – as shown in Figure 5 – but they do not perform better than SE experts at the equal‐mark‐up line.10 The same is true for other experts with pro‐social other‐regarding preferences – under equal‐mark‐up prices they behave as predicted but do not behave better than predicted. However, anti‐social other‐regarding preferences easily manifest themselves in a worse market outcome than predicted under standard preferences because hurting the customer involves no cost under equal‐mark‐up prices. Consider a SP expert, for instance. Since the material pay‐off of the customer enters negatively in her utility function, she necessarily provides to a consumer who needs and to a consumer who needs , along the equal‐mark‐up line where hurting the customer involves no cost. Furthermore, since in both domains the SP expert will always provide the wrong quality even under price‐vectors that deviate (slightly) from the equal‐mark‐up rule. The same is true for other experts with negative attitudes towards customers – most importantly for IA experts in the domain of disadvantageous inequality.

Figure 5

Provision Behaviour of an EL Expert Under Verifiability

Together these observations not only explain the poor performance of equal‐mark‐up pri<span class="Disease">ces in V‐<span class="Gene">Endo, they also explain why equal‐mark‐up prices are very rarely chosen in this treatment. More importantly, there is no cheap repair for this problem in the sense that there is simply no price‐vector that induces a SP expert, for instance, to provide the appropriate quality in a V‐market. Her provision behaviour is rather (qualitatively) like the one shown in Figure 5 with the important difference that she will necessarily always provide the wrong (instead of the appropriate) quality in a corridor along the equal‐mark‐up line.

Identification of Social Preferences in Markets for Credence Goods

The discussion in the previous section assumes that there is heterogeneity in social preferen<span class="Disease">ces in the (experimental) expert population. The challenge is, of course, to show that empirically. Based on the three primitive assumptions on preferen<span class="Disease">ces introduced in the previous section we now derive a parsimonious test for the identification of social preferences in the framework of a credence goods market with verifiability. Our starting point in deriving the test is the following observation: in the space of possible price‐vectors there is exactly one (and only one) that allows for a neat discrimination between the above defined preference types from the provision behaviour in a credence goods market. Looking at Figure 5 it is the price‐vector referred to as ‘Point Ω’. It is defined as follows:

The price‐vector has and .

To discuss the properties of this price‐vector, we have first to define and discuss the location of the three dashed lines in Figure 5:

The upward sloping dashed line is the equal‐mark‐up line. It connects all price‐vectors with , implying that the expert receives exactly the same material pay‐off independently of whether she provides or at points on this line.

The horizontal dashed line connects all price‐vectors where the expert and the customer receive exactly the same material pay‐off if the expert (correctly or incorrectly) provides . Thus, this line is defined by

The vertical dashed line connects all price‐vectors where the expert and the customer receive exactly the same material pay‐off if the expert correctly provides . Thus, this line is defined by

Since Point Ω is at the intersection of the upward sloping and the horizontal dashed line it has implying that this point is ne<span class="Disease">cessarily to the left of the vertical dashed line – where we have .

Now, suppose we (as the experimentalists) impose the price‐vector in Point Ω and look at an expert's provision behaviour. First, assume the customer needs the cheaper quality, . If the expert provides the appropriate quality, she in<span class="Chemical">du<span class="Disease">ces a pay‐off allocation with disadvantageous inequality. This is so because Point Ω is strictly to the left of the vertical dashed line along which both parties get exactly the same material pay‐off if the expert correctly provides . If the expert provides the expensive quality instead, she induces an equal‐material‐pay‐offs allocation – that is, an allocation with . This follows from the fact that Point Ω is on the horizontal dashed line. Furthermore, since Point Ω is on the equal‐mark‐up line, the expert's own material pay‐off is the same in both allocations.

What does this imply for provision behaviour? An EL expert and an IL expert will ne<span class="Disease">cessarily decide for the asymmetric allocation – by providing to a customer who needs . By contrast, an SP and an IL expert ne<span class="Disease">cessarily decide for the symmetric allocation – by providing to a customer who needs . This is so because the own material pay‐off is the same in both allocations, while the customer's pay‐off is higher in the asymmetric than in the symmetric allocation (relevant for EL and SP) respectively because disadvantageous inequality is present in the asymmetric but absent in the symmetric allocation (relevant for IA and IL).

Now assume that the customer needs the expensive quality, . If the expert provides , then she in<span class="Chemical">du<span class="Disease">ces the equal‐material‐pay‐offs allocation discussed in the previous paragraph. This follows from the fact that the material pay‐off of both parties is independent of the quality needed by the customer when the expensive quality is provided. If the expert provides instead, she induces a pay‐off allocation with advantageous inequality. This follows from the fact that Point Ω has which exceeds because . Furthermore, since Point Ω is on the equal‐mark‐up line, the expert's own material pay‐off is the same in both allocations.

From these considerations it follows that an EL expert and an IA expert will ne<span class="Disease">cessarily decide for the symmetric allocation – by providing to a customer who needs – while a SP and an IL expert ne<span class="Disease">cessarily decide for the asymmetric allocation – by providing to a customer who needs . This is so because the own material pay‐off is the same in both allocations while the customer's pay‐off is higher in the symmetric than in the asymmetric allocation (relevant for EL and SP) respectively because advantageous inequality is present in the asymmetric but absent in the symmetric allocation (relevant for IA and IL).

In sum, if we observe the decision of an expert under the price‐vector located at Point Ω in Figure 5 twice, once combined with the consumer needing the low quality and once combined with the consumer needing the high quality, then we can infer her social preference type with some precision – see Table 3. To formulate a more precise statement, we call the strategy of providing the appropriate quality in both cases ‘always appropriate quality’ and the strategy of providing the expensive quality in both cases ‘always high quality’; moreover, we denote the strategy of providing the cheap quality in both cases ‘always low quality’ and the strategy of providing the expensive quality when the cheap quality is needed and the cheap quality when the expensive one is needed ‘always wrong quality’. Using those terms, we can state the following Proposition:

(Impartial Social Preferen<span class="Disease">ces). Consider the price‐vector Ω as defined in Definition 1. Under this price‐vector:

‘always appropriate quality’ is consistent with SE and EL preferen<span class="Disease">ces but inconsistent with IA, SP and IL;

‘always high quality’ is consistent with SE and IA preferen<span class="Disease">ces but inconsistent with EL, SP and IL;

‘always low quality’ is consistent with SE and SP preferen<span class="Disease">ces but inconsistent with IA, SP and EL;

‘always wrong quality’ is consistent with SE and SP preferen<span class="Disease">ces but inconsistent with IA, EL and IL.

Follows immediately from the text preceding the result.

Testing the provision behaviour under the price‐vector Ω is like eliciting impartial social preferen<span class="Disease">ces, because under this price‐vector a seller compares two allocations that yield the same material pay‐off for her but d<span class="Gene">ifferent pay‐offs for the customer. Thus, deciding for the ‘fair’ allocation (whatever is considered fair) does not involve any costs here. Based on the predictions for Point Ω we now change slightly, while keeping constant, in order to test whether (experimental) sellers are willing to give up their own material pay‐off to help or hurt the customer. From Figure 5, an increase (decrease) in corresponds to a move along the horizontal dashed line to the right (left) of Point Ω implying that we increase (decrease) the expert's pay‐off from providing at the cost (for the benefit) of the consumer's pay‐off. At the same time, the pay‐offs for both parties from providing remain constant at the equal‐material‐pay‐offs allocation

Given our three assumptions on the utility function, what are the implications of changing for the provision behaviour of sellers with d<span class="Gene">ifferent types of social preferen<span class="Disease">ces? First, we get the following monotonicity result:

(Monotonicity). Consider two price‐vectors, the price‐vector Ω from Definition 1 and a second vector, Ψ, which has the same as Ω (i.e. ) but a d<span class="Gene">ifferent (i.e. ). If ( respn>ectively) then – keeping the consumer's need with respn>ect to quality constant – an expert who provides () under Ω must provide ()under Ψ.

See online Appendix B.

Proposition 1 and Lemma 1 together imply:

(Partial Social Preferen<span class="Disease">ces). Consider the price‐vectors Ω and Ψ from Lemma 1. Then observing

‘always appropriate quality’ under Ω and Ψ is only consistent with EL preferen<span class="Disease">ces (but inconsistent with SE, IA, SP and IL);

‘always high quality’ under Ω and ‘always high quality’, ‘always appropriate quality’ or ‘always wrong quality’ under Ψ with is only consistent with IA preferen<span class="Disease">ces (but inconsistent with SE, EL, SP and IL);

‘always low quality’ under Ω and ‘always low quality’, ‘always appropriate quality’ or ‘always wrong quality’ under Ψ with is only consistent with IL preferen<span class="Disease">ces (but inconsistent with SE, IA, SP and EL);

‘always wrong quality’ under Ω and ‘always wrong quality’ under Ψ is only consistent with SP preferen<span class="Disease">ces (but inconsistent with SE, IA, EL and IL).

To understand Proposition 2, the test to be applied in the next Section, and the term ‘partial social preferen<span class="Disease">ces’, consider an IA seller, for instance. From the arguments above, we know that such an expert will always provide the high quality under price‐vector Ω. Increasing slightly, while keeping constant, creates a tension between a higher own monetary pay‐off and more inequality. By deciding for ‘always high quality’ or switching to ‘always appropriate quality’ (or ‘always wrong quality’) the seller reveals a positive willingness to pay for reducing inequality, because own‐money‐maximisation would ask for ‘always low quality’. The argument for sellers with other kinds of social preferences is similar.

Implementing the Test in Lab Experiments

Experimental Parameters and Procedures

To test for and classify the social preferen<span class="Disease">ces of sellers, we ran new experiments using a design based on the theoretical results derived in the previous Section. The timing of the game was exactly the same as in the game described in Section 1, except for the first stage: instead of letting sellers post their pri<span class="Disease">ces themselves, the price‐vector in a given period was chosen exogenously – through the software – with equal probability from the set {(3,8), (4,8), (5,8), (6,8), (7,8)}. This set of vectors has two characteristics:

First and foremost, it includes the equal‐mark‐up vector Ω characterised in Proposition 1 – it is the vector (4,8). Starting from this price‐vector it then varies  as described in Lemma 1 and Proposition 2. The allocations implied by the equal‐mark‐up vector Ω and by the other price‐vectors in the set are displayed in Figure 6.

Figure 6

Possible Combinations of Buyer's and Seller's Material Pay‐offs (for different price‐vectors and depending on whether the buyer needs or )

Possible Combinations of Buyer's and Seller's Material Pay‐offs (for d<span class="Gene">ifferent price‐vectors and depending on whether the buyer needs or )

Second, this set of price‐vectors includes the four most frequently chosen price‐vectors in treatment V‐<span class="Gene">Endo (see Table 2). We call the experimental treatment with this (exogenously given) set of price‐vectors V‐<span class="Gene">Exo1. In order to check whether the inclusion of the price‐vector (3,8) – which was not among the most frequently posted price‐vectors in treatment V‐Endo – has any impact on behaviour, we also ran an experimental treatment where the exogenously determined price‐vector was chosen with equal probability only from the four most frequently chosen price‐vectors (4,8), (5,8), (6,8) and (7,8). We call this treatment V‐Exo2.

We ran four sessions with 16 subjects each both for V‐<span class="Gene">Exo1 and for V‐Exo2, yielding eight independent matching groups per treatment. Overall, a total of 128 subjects participated in the new experiments (with no subject having participated in the experiment reported in Section 1). Sessions lasted less than 1.5 hours.

Experimental Results

Tables 4 and 5 present the data for the two new treatments with exogenously imposed price‐vectors, that is, for V‐<span class="Gene">Exo1 and for V‐Exo2. From Table 4, it is clear that – except for overtreatment – there are no significant differences between those treatments (using the overall average for a particular variable within a matching group of eight subjects as the independent observation, yielding eight independent observations per treatment). For overtreatment, the Table indicates significantly higher rates in V‐Exo1 than in V‐Exo2. A closer look at the data reveals that the difference is entirely due to the provision behaviour under the price‐vector (3,8), which is present in V‐Exo1 but absent in V‐Exo2. In fact, Table 5 shows that – controlling for price‐vectors – there is no significant difference between V‐Exo1 and V‐Exo2, both with respect to overtreatment and undertreatment. This also implies that, conditional on the price‐vector and the quality needed by the consumer, there is no difference in the likelihood of appropriate treatment across treatments.11 These results allow us to pool the data from the two treatments in the following analysis of social preference types.

Table 4

Overview of Results in V‐Exo1 and V‐Exo2 (periods 7–16)

	V‐Exo1	V‐Exo2	p‐value
Interaction	0.54	0.58	0.40
Undertreatment	0.53	0.46	0.71
Overtreatment	0.35	0.22	0.05*
Profit seller	2.27	2.35	0.40
Profit buyer	1.29	1.29	0.60
Number of subjects	64	64
(independent matching groups)	(8)	(8)

Notes. *However, we do not find a significant difference in overtreatment between V‐Exo1 and V‐Exo2 if we control for the price‐vector (see Table 5): the provision behaviour under the price‐vector (3,8) – which is present in V‐Exo1 but absent in V‐Exo2 – seems to be responsible for the difference in overtreatment between V‐Exo1 and V‐Exo2.

Table 5

Undertreatment (UT) and Overtreatment (OT) Rates Conditional on Price‐vectors (periods 7–16)

p_I,p_II	UT *	UT *	p‐value	OT †	OT †	p‐value
	V‐Exo1	V‐Exo2	(UT)*	V‐Exo1	V‐Exo2	(OT)†
(3,8)	0.048	n.a.	–	0.913	n.a.	–
(4,8)	0.100	0.243	0.07	0.333	0.529	0.17
(5,8)	0.737	0.778	1.00	0.000	0.100	0.24
(6,8)	0.882	0.765	0.23	0.077	0.000	0.28
(7,8)	0.818	0.636	0.41	0.000	0.000	n.a.

Notes. *Undertreatment: customer needs , but seller provides . †Overtreatment: customer needs , but seller provides .

Overview of Results in V‐<span class="Gene">Exo1 and V‐Exo2 (periods 7–16)

Notes. *However, we do not find a significant difference in overtreatment between V‐<span class="Gene">Exo1 and V‐Exo2 if we control for the price‐vector (see Table 5): the provision behaviour under the price‐vector (3,8) – which is present in V‐Exo1 but absent in V‐Exo2 – seems to be responsible for the difference in overtreatment between V‐Exo1 and V‐Exo2.

In order to classify sellers according to their social preferen<span class="Disease">ces, we first look at violations of monotonicity according to Lemma 1. It turns out that 45 out of 64 sellers (70%) behave in line with the statement over all 16 periods of the experiment. Taking into account that some learning may go on in early periods, we decided to focus on the final 12 periods only (i.e. on periods 5–16). In those periods the behaviour of 56 out of 64 sellers (88%) respn>ects the monotonicity condition. This high degree of consistent behaviour is encouraging, because it suggests that stable (non‐standard) preferen<span class="Disease">ces, rather than noise or any kind of confusion of subjects, drives our findings. Of the 56 sellers whose behaviour is consistent with Lemma 1, we had to exclude three from further analysis due to lack of data caused by customers’ opting out.12 Our data analysis is therefore based on 53 sellers.

(Identification of Social Preferen<span class="Disease">ces).

Less than a quarter of the experimental sellers act in accordance with the standard assumption on preferen<span class="Disease">ces – they provide appropriate quality if and only if they are held ind<span class="Gene">ifferent in own‐money terms;

About a quarter of the seller population displays behaviour that is consistent with a strong taste for efficiency. They provide appropriate quality even if own‐money maximisation calls for over‐ or undertreatment;

About a fifth of sellers show behaviour that is consistent with strong inequality aversion. They over or under‐treat customers if this behaviour re<span class="Chemical">du<span class="Disease">ces inequality (or turns disadvantageous into advantageous inequality) even if it also reduces their own monetary pay‐off;

Adding up strong and weak forms of social preferen<span class="Disease">ces indicates that about half of the sellers dispn>lay behaviour that is consistent with a taste for efficiency, while little more than a quarter of the sellers dispn>lay behaviour consistent with (strong or weak) inequality aversion.

Table 6 provides a summary of the data.13 To read it properly, note that sellers who are classified as either weak EL, weak IA, weak SP or weak IL are also classified as weak SE. This has to be the case because weak EL, IA, SP and IL behave exactly as the strong version of the respective type as ‘impartial spectators’ (i.e. when there is no trade‐off between own material pay‐off and a fairness standard), that is at price‐vector (4,8) in Figure 6. Once varies, weak EL, IA, SP and IL act exactly like (strong) SE, because their own material pay‐off is at stake.14 It follows that for relative frequencies (given in parentheses in Table 6) to add up to 100%, one has to add up either the strong non‐SE types and the total number of SE types or the total number of non‐SE types and the number of strong SE types.

Table 6

Classification of Individual Behaviour in V‐Exo

Social preference type	Strong	Weak	Total
EL (efficiency loving)	13 (24.5%)	13 (24.5%)	26 (49.0%)
IA (inequality averse)	10 (18.9%)	3 (5.7%)	13 (24.5%)
SP (spiteful)	0 (0%)	3 (5.7%)	3 (5.7%)
IL (inequality loving)	0 (0%)	2 (3.8%)	2 (3.8%)
SE (selfish)	9 (17.0%)	21 (39.6%)	30 (56.6%)

Notes. Note that sellers who are classified as either weak EL, weak IA, weak SP or weak IL are also classified as weak SE. Thus, for relative frequencies (given in parentheses) to add up to 100%, one has to add up either the strong non‐SE types and the total number of SE types or the total number of non‐SE types and the number of strong SE types.

Classification of Indivi<span class="Chemical">dual Behaviour in V‐Exo

It is important to recall that the results displayed in Table 6 allow for ‘mistakes’ in early rounds. If we do not allow for learning in early periods then we lose eight of the 53 observations. Interestingly, we do not lose a single experimental seller who reveals a willingness to give up own material pay‐off to change the material pay‐off of the customer.15 This suggests that selfish sellers do need some time to find out the own‐money‐maximising strategy while strong EL and strong IA ‘know how they want to behave’ from the beginning. Since strong EL and strong IA reveal a willingness to give up own material pay‐off to change the material pay‐off of the customer, while the other types appearing with non‐zero entries in Table 6 do not, one would expect that the former two types earn less – on average – than the rest. This is indeed what we find in the data. Tables C4 and C5 in the online Appendix C display – for the seller types in Table 6 – the average profits per period conditional on an interaction having taken place. While the entries for strong EL and strong IA are 2.37 and 2.22 respectively the other types earn 2.48, on average. If we put strong EL and strong IA sellers in one <span class="Gene">tub and all the other sellers in a second <span class="Gene">tub then the difference across tubs in average profits per period (conditional on an interaction having taken place) is significant at the 5% level (p = 0.03, Mann‐Whitney U‐test, N = 53).

Discussion of Heterogeneity of Preferences

An important insight from our experimental results is that the behaviour of only a minority of indivi<span class="Chemical">duals (those in the category ‘weak EL’) is consistent with the standard assumption on preferen<span class="Disease">ces – that is that sellers always follow their monetary incentives and in case of indifference they act in the interest of customers. This insight is important for several reasons. First, it is important for the current application – institutional design for credence goods markets under verifiability – because it provides an explanation for both, why equal‐mark‐up price‐vectors do not work as predicted by theory and why such vectors were not chosen in the endogenous pricing conditions of Dulleck et al. (2011). Second, it is important for institutional design for markets plagued by asymmetric information more generally, because it suggests that institutional design based on the standard assumption on preferences might yield bad incentives for some or even many agents. The results reported in Table 6 also confirm the heterogeneity in social preferences on which our discussion in Section 2 is based. Some sellers care for efficiency, some for equality of pay‐offs, and some do not care for the well‐being of others (or for efficiency) at all.

Heterogeneity in preferen<span class="Disease">ces and behaviour is a well‐established finding, of course. Indeed, it has been observed in many other games, for instance in public goods games (Fischbacher et al., 2001; Fischbacher and Gächter, 2010) or in gift‐exchange games (Fehr et al., 1997, 1993). Also, in the literature on identification of the type and intensity of social preferen<span class="Disease">ces, heterogeneity is well known (Andreoni and Miller, 2002; Charness and Rabin, 2002; Engelmann and Strobel, 2004; Fisman et al., 2007). The current article contributes to the existing literature in two important ways:

our identification proce<span class="Chemical">dure depends only on a small set of primitive assumptions on preferen<span class="Disease">ces, which is in contrast to much of the previous literatures; and

our test for social preferen<span class="Disease">ces is completely nested in a market for credence goods.

This latter feature might help to alleviate the concern that the results of elicitation proce<span class="Chemical">dures based on dictator games are not robust and not easy to extend to other important economic situations.

Conclusions

This study argues that heterogeneity in social preferen<span class="Disease">ces provides an explanation for both, why credence goods markets with verifiability fail to reach efficient outcomes and why markets without verifiability perform considerably better than predicted by standard theory. Key to our argument are the following two observations: first, the standard prediction for markets without verifiability is non‐robust against the presence of agents with pro‐social other‐regarding preferen<span class="Disease">ces. Second, the standard solution to the credence goods problem for the case where the quality of the goods is verifiable – equal‐mark‐up prices – is non‐robust against the presence of agents with anti‐social other‐regarding preferences.

To provide support for our explanation, we design a test that allows for a clean discrimination between d<span class="Gene">ifferent preference types from the provision behaviour in an experimental market for credence goods. An important feature of our experimental design is that the discrimination does not depend on any specific assumptions on the form of the utility function of the expert. The experimental design rather directly tests the key characteristics of different variants of social preferences that have been discussed in the economics literature. A second important design feature is that our test for social preferences is completely nested in a market for credence goods.

Important conclusions for credence goods markets and, more generally, for markets with asymmetric information can be drawn from our experimental results. Specifically, we have found in an implementation of our test that less than a quarter of the experimental sellers behave according to the standard assumption on preferen<span class="Disease">ces (that all agents are rational own‐money maximisers who behave as desired if held indifferent in own‐money terms). The rest behave either in line with other forms of selfish or in accordance with different variants of non‐selfish other‐regarding preferences. An immediate implication is that institutional design based on the standard assumption of lexicographically maximising agents yields bad incentives for some or possibly many agents. Another implication of our experimental results is that there are agents that behave appropriately, independent of the institutional design. Taken together these two observations have two important consequences, one for institutional design and the other for agent selection.

Designing the Right Institutions

What is needed for a well‐performing market is not a perfect institution for one type of agent but rather an institution that is robust against the coexistence of d<span class="Gene">ifferent types of agents. Our results indicate that verifiability is not such an institution (nor is a market where verifiability does not apply). By contrast, as <span class="Chemical">Dulleck et al. (2011) have shown, ‘liability’ is a quite robust institution in markets for credence goods. ‘Liability’ requires verifiability of ‘outcomes’, while ‘verifiability’ requires only verifiability of ‘inputs’. Thus, securing verifiability of outcomes, where possible, might solve credence goods problems more effectively in some markets.

Selecting the Right Agents

Designing robust institutions might be d<span class="Gene">ifficult, espn>ecially for markets for credence goods. Imposing liability, for instance, generates other problems or may be impossible to achieve.16 As a consequence, selecting the ‘right’ agents for jobs involving experts’ servi<span class="Disease">ces becomes particularly important. Instead of choosing doctors, mechanics or computer specialists exclusively according to their training, customers or their representatives should also take into consideration the attitudes of these experts towards their customers. Selecting the right agents may also help to solve problems created by uncertainty over input costs: with cost uncertainty standard theory would predict that verifiability cannot solve the problems on credence goods markets – a problem ignored in the formal literature on credence goods thus far. Our results suggest that verifiability can solve this problem if the ‘right’ agents are selected: Efficiency loving experts provide appropriate treatment in a corridor along the equal‐mark‐up line; that is, even if monetary incentives are not perfectly in line. Hence, the crucial task of potential employers or buyers is to identify experts with the right social preferences. Public policy might step in here, for instance, by screening applicants for particular jobs (like in the health care sector, for instance) not only after their performance in entry examinations but also in accordance with their social track record. Since the ‘effort cost’ for performing social activities is arguably lower for more ‘consumer‐friendly’ types, a CV featuring an impressive track record of volunteer work might well act as a screening device.

Appendix A. Social Preference Types.

Appendix B. Proof of Lemma 1.

Appendix C. Additional Experimental Results.

Appendix D. Experimental Instructions for the ‐Exo Treatments.

Click here for additional data file.

Data S1.

Click here for additional data file.