deltie.net archive

Extracted abstract

Cooperation without third-party enforcement is particularly puzzling in illicit online markets given the anonymity of online exchanges in the "dark web" and the asymmetry of information between buyers and sellers. Most of the literature investigates the effects of reputation systems on sales. Less is known about the role of (semi)institutionalized solutions to trust problems, such as the escrow service, which deposits payments for online purchases with the market platform and releases them only upon confirmation of the item delivery by a customer. We study the effect of such a trust intermediary on sales in a cryptomarket for illegal drugs. Using a large dataset of illegal online transactions, we estimate two sets of fixed effects models predicting the sellers' choice to offer the trust intermediary and examine the effects of such a choice on sales.

Our results indicate that the trust intermediary reduces online drug sales. We explain this finding by showing suggestive evidence that escrow may crowd out traders' trust and reciprocity. Our findings have implications for theories of the role of institutions in online markets and offer policy recommendations for law enforcement agencies.

Introduction

Third-party enforcement is considered to be one of the most effective solution to the problem of social order (Nozick, 1974; Fehr and Fischbacher, 2004; Baldassarri and Grossman, 2011) . One important modern manifestation of the problem of social order is the trust problem that occurs among exchanging partners in anonymous settings, such as online marketplaces. 1 To address this problem, online marketplaces have developed various institutional arrangements to facilitate and protect anonymous exchange, even in the absence of third-party enforcement. Most platforms offer a reputation system that disincentivizes dishonest behavior by allowing customers to look into the history of transactions involving particular vendors (Diekmann et al., 2014) .

Other institutional forms of assurance in online exchange-embodied in the market platform-oversee the behavior of the exchanging partners by acting as an intermediary in trust. One of the most common trust intermediaries in online exchange is the escrow service. Using this assurance device, the trading platform acts as a guarantor that holds payments for online purchases until the buyers have confirmed the receipt of a purchased product. In the absence of such confirmation, the platform adjudicates disputes. Such a trust intermediary is meant to solve trust problems that are inherent in online exchangenamely, the fact that the buyer has to pay in advance for products of uncertain quality and that s/he faces the risk that the products are not sent or delivered.

However, there is a crucial difference between the role of the market platform as trust intermediary and the classical guarantor described by Coleman (1990) . A classical guarantor has a trust relationship with the final trustee and incurs financial risks associated with the behavior of that trustee. By contrast, offering the escrow service, the market platform simply reverses the time asymmetry in the exchange between buyer and seller, thereby protecting the buyer from a possible misconduct of the seller, but does not incur financial risks. How does such a guarantor affect exchange?

It is widely believed that semi-formal trust intermediary institutions such as escrow solve the trust problems in peer-to-peer online exchanges (Hu et al., 2004; Odabaş, Holt, and Breiger, 2017 ). Yet, some scholars have questioned these beneficial effects of escrow (Holt, Smirnova and Hutchings, 2016) . Theoretical predictions are likewise mixed. The classical studies of assurance devices find that trust intermediaries facilitate exchange.

Others, however, argue that the costs of escrow tend to be higher than the expected benefits, undermining the efficiency of this trust intermediary (Antony, Lin, and Hu, 2006) .

In the present study, we investigate how the escrow service affects online trading in a very large cryptomarket (Alphabay). Cryptomarkets are online marketplaces located on the dark web, trading in illegal goods and services (more details about Alphabay in Online Appendix A.1). We focus on transactions involving illegal drugs.

Cryptomarkets for illegal drugs offer an interesting setting to study the effects of trust intermediaries. On the one hand, cryptomarkets are the kind of marketplace where escrow should arguably work best. First, the illegal nature of the trade implies that buyers and sellers are unable to rely on other institutional legitimate third-party enforcement services, namely, contract laws or legal oversight (Beckert and Dewey, 2017; Bakken, Moeller, and Sandberg, 2018) . Second, theoretically, the adoption of a trust intermediary is strategically rational only if the probability of encountering a dishonest exchange partner is not too small (Hu et al., 2004) . Third, the anonymity of cryptomarket exchanges makes the development of stable dyadic trading outside the platform-based on reciprocal acquaintance-difficult (for an exception, see Childs et al., 2020) . All these factors should reinforce the demand for semi-institutional forms of assurance in cryptomarkets.

On the other hand, in cryptomarkets, the trust intermediary intended to protect the exchanging partners from fraud is offered by the market platform, which is by definition legally unaccountable and thus less credible -as a guarantor -than a legal platform (Moeller, Munksgaard, and Demant, 2017) . In addition, law enforcement authorities can disrupt the service and seize cryptocurrencies deposited in escrow at any time (Buskirk et al., 2017; Ladegaard, 2020) . For all these reasons, traders may be skeptical about the benefits of using escrow in cryptomarkets. Therefore, whether the availability of an escrow service affects sales positively or negatively in a cryptomarket remains an empirical question. The present study addresses this question by investigating 1) determinants of the sellers' decision to offer escrow payment and 2) the effects of the payment method chosen by the sellers (i.e., escrow vs advance payment) on the volume of sales.

Theory & Hypotheses

The literature on the use of escrow and other semi-formal assurance mostly points to the likely positive effects of escrow on sales (Hu et al., 2004) . However, these predictions are typically derived from theoretical models that are rarely tested in empirical analyses (for exceptions, see Holt, Smirnova and Hutchings, 2016; Munksgaard, 2022) . Evidence from lab experiments, by contrast, suggests more sobering conclusions (Bracht and Feltovic 2007) . Moreover, conclusions drawn from legal online marketplaces may not apply to illicit online trade, where all trading activities are anonymous and the platform administrators who set the rules are largely unaccountable.

A useful framework to analyze this ambivalent relationship between escrow and illicit online exchange comes from game-theoretical models of trust games with commitment devices (e.g., Raub, 2004 ). 2 The escrow system serves the same purpose as the commitment devices postulated in these theoretical models: it is meant to facilitate transactions in a risky environment. Raub's (2004) model makes three assumptions that are likely to hold in the cases of escrow. First, the model assumes that both honest and dishonest types of sellers exist, and the buyers do not know the seller's type before entering the interaction. Second, even if the seller is honest, the outcome of the transaction is uncertain, i.e., there is a small probability that something goes wrong. The item can be lost, and the seller may appear to have behaved dishonestly even though she did not, for example, because the police discovered and impounded the illegal item. Third, the model also assumes that the seller has the option to post a hostage, offering some form of guarantee that protects the buyer against the transaction risks. Such an action is equivalent of providing security for customers by signaling one's honesty and thus increasing competitiveness. Hu et al. (2004) present a more complex game-theoretical model -specifically designed to analyze the viability of the escrow services as trust intermediaries in online (legal) markets -that also assumes the existence of honest and dishonest types. In this model, the feasibility of the escrow service from the traders' point of view depends crucially on its cost and on the prevalence of dishonest types in the market. In cryptomarkets, the prevalence of dishonest types is certainly higher than in legal markets.

Thus, the traders could be willing to pay a higher fee for the escrow service (Hu et al., 2004) . However, as stated above, in cryptomarkets more sources of uncertainty are present than in legal online markets, including possible mistrust towards the platform administrators and fear that law enforcement authorities overtake the platform and freeze all the assets, including cryptocurrencies deposited in the escrow.

In sum, in cryptomarkets, such as Alphabay, sellers decide whether to offer escrow or not. Buyers cannot select the payment method, but they can choose between alternative offers of the same item-sometimes even from the same seller-that come with different payment methods. The choices of sellers and buyers are thus sequential and closely interrelated, mirroring the structure found in formal game-theoretical models (e.g., Hu et al. 2004 , Raub 2004) . Who thus uses semi-formal trust-intermediaries, such as the escrow services, while trading illegal drugs online? And, how does the use of these trustintermediaries affect sales in cryptomarkets?

Determinants of escrow

From the seller's point of view, the escrow alters the decision sequence in a way that introduces an element of risk (Afilipoaie and Shortis, 2018; Ladegaard, 2019) . Under advance payment, the seller only ships the goods upon receiving payment. Therefore, the seller faces no risk of incurring losses. By contrast, under escrow payment, the seller receives the payment only after the buyer has received the item and confirmed it to the platform. Therefore, selling under escrow implies that the payment to the seller is delayed and somewhat less certain: cryptocurrencies are highly volatile, and sums deposited in an escrow wallet -pending confirmation from the buyer -are subjects both to the risk of devaluation and to the risk of law enforcement seizure of the platform's assets (Ladegaard, 2020) . Sellers may also fear platforms' exit scam, in which markets close and their administrator steal all cryptocurrencies deposited in escrow (Van Buskirk et al., 2016) . Moreover, to some extent, sellers are also uncertain about the true quality of their product (Reuter and Caulkins, 2004; Lakhdar et al., 2013) . 3 By offering escrow payment, vendors thus incur the additional risk that clients will be dissatisfied with the product and refuse to release the payment. These factors arguably make the escrow payment a less preferred option from the seller's perspective. At the same time, escrow is a default payment method in many cryptomarket, including Alphabay (see Online Appendix A.1).

How thus do sellers get away with not offering the escrow service?

Transactions in a cryptomarket differ from the interaction assumed in the gametheoretical models of trust with commitments (Raub 2004) in some important respects.

These models generally assume that the interaction between seller and buyer is a one-shot game and the buyer has no information about the seller, besides what can be inferred from the seller's choice. For example, in signaling games, the choice to offer a commitment is a costly signal that reveals that the seller is an honest type (Spence 1973, Przepiorka and Diekmann, 2013) . By contrast, in terms of web service design, a cryptomarket is similar to any other online market platform. In exchange for a commission, customers can buy from a wide selection of sellers, they can buy repeatedly from the same seller, and there is a feedback mechanism allowing buyers to publicly rate every transaction and write a short review. Consequently, sellers can build a reputation.

In the literature on trust in embedded settings there is robust evidence that trust problems are easily solved when interactions are repeated (e.g. Camerer and Weigelt, 1988 , see Barrera, 2008 for a review). Actors display higher trust in others with whom they had successfully exchanged in the past (Barrera, 2007) . Once an exchange relationship is established, they tend to remain committed to their exchange partners even when presented with more attractive offers (Kollock, 1994) . Commitment to a specific dealer is likely to be even stronger in a cryptomarket for illegal drugs, where customers are especially likely to return to the same vendors and buy more of the same product, provided that their trust was not abused (Décary-Hétu and Quessy-Doré, 2017; Duxbury and Haynie, 2018; Norbutas, Ruiter, and Corten, 2020) .

Thus, once vendors have established relationship with customers who are likely to return to buy products from them, irrespective of the payment method, they have incentives to gradually opt out of escrow in favor of advance payment. This allows them to avoid the costs and risks associated to the escrow without losing clients.

Hypothesis 1. The higher the number of recurrent customers a seller has, the less s/he is likely to prefer escrow to advance payment.

As argued by Yamagishi (2011) , problems of trust are more salient when the social uncertainty inherent in an interaction is higher. Accordingly, every piece of information that reduces the uncertainty has the effect of moderating the problem of trust. As we stated earlier, the escrow system serves the purpose of reducing risk (social uncertainty in Yamagishi's terms) for the buyer. However, the reputation scores that are available in online markets, including cryptomarkets, serve the same purpose. The positive effects of reputation scores both on the number of sales per offer and on the price are consistent with various rational choice models describing embedded settings (Buskens and Raub, 2013) . While it is still debated whether reputation increases prices (see Holt, Chua and Smirnova, 2013 vs. Hardy and Norgaard, 2016; Munksgaard and Tzanetakis, 2022) , there is some evidence linking reputation to higher sales in another cryptomarket (Przepiorka, Norbutas, and Corten, 2017) . Therefore, sellers with a good reputation should be more likely to opt out of escrow. 4

Hypothesis 2: The higher the reputation score of a seller, the less s/he is likely to prefer escrow over advance payment.

Effects of escrow

Turning to the effects of escrow on the buyers' decision, theoretical models of trust games incorporating commitment devices generally predict that the use of commitments, such as the escrow, facilitates cooperation through various mechanisms. First, a commitment device, such as the escrow, can bind the seller to a cooperative behavior, thereby significantly reducing the risk that trust is abused. Second, escrow can compensate the buyer for the loss that occurs when trust is abused or when the item is lost due to exogenous contingencies. Third, a commitment device can signal the seller's type, thereby allowing the buyer to avoid exchanges with dishonest types (Raub 2004, Przepiorka and Diekmann, 2013) . All these mechanisms predict a higher rate of successful transactions when commitment devices are available and used.

Hypothesis 3: Offers with escrow generate a higher number of successful transactions than comparable offers requiring advance payment.

In contrast to this line of reasoning, some scholars studying cryptomarkets highlighted possible counterproductive effects of escrow. Sellers are likely to raise the price for offers with escrow payment, in order to compensate for the risk that they are taking (Holt, Chua, and Smirnova, 2013; Munksgaard and Tzanetakis, 2022) . Higher prices, in turn, can reduce the appeal of escrow as an assurance device, and discourage buyers from purchasing products through escrow.

Moreover, the escrow would not have the same properties of a commitment device-as postulated in theoretical models-if the platform was not perceived to be as reliable as the contract law in a legal market. We can hardly rule out such a possibility.

Note that, although the cryptomarket administrators have the power to ban users at their discretion, buyers do not have the possibility to hold the administrators accountable for their behavior (Horton-Eddison and Di Cristofaro, 2017; Van Buskirk et al., 2017) . If buyers are indeed suspicious of escrow's reliability, the system is likely to backfire.

Last, there may be better options for ensuring trust in cryptomarkets than escrow.

Scholars have shown that online drug traders' communities exhibit high levels of trust, operating according to strong reciprocity norms supported by reputation systems (Przepiorka, Norbutas, and Corten, 2017; Masson and Bancroft, 2018; Munksgaard et al., 2022 ; see also, Moeller and Sandberg, 2019) . The introduction of the escrow payment may be incompatible with these forms of informal social control, thus crowding out trust between traders. In general, crowding out can occur when the introduction of institutions designed to promote cooperation undermines the intrinsic motivation to act cooperatively on the basis of trust (see Frey and Jegen 2002, Bohnet and Baytelman 2007) . As a result, in our case buyers may be inclined to pass on the escrow offers and select into transactions that rely on informal social control. Based on these insights, we propose an alternative hypothesis regarding the effects of escrow on sales.

Hypothesis 4: Offers with escrow generate a lower number of successful transactions than comparable offers requiring advance payment.

How can we reconcile these contradictory predictions under Hypotheses 3 and 4? Paying a higher price for a safer offer may be more appealing for buyers with no or little information on the trustworthiness of vendors in the cryptomarket. Once a buyer has identified a seller with a history of positively rated transactions, however, things could change. A seller with a good reputation faces large potential damage from defrauding customers. Thus, once trust in a specific dealer has been established due to their reputation, a buyer will be inclined to pass on the escrow offer and opt for advance payment, especially if the latter comes at a lower price. Accordingly, we expect sellers to complete less sales with escrow, when they have a good reputation.

Hypothesis 5: The higher the reputation of a seller, the more the number of successful transactions per offer using escrow will decrease.

Until now we have discussed the risk associated with completing a transaction in a cryptomarket focusing on the buyer's perceived probability that things can go wrong and the buyer loses her money. However, even if that probability was perceived to be very small, the trust problem can be severe if the amount at stake is very high. In terms of the overall value, transactions on cryptomarkets are very heterogeneous, not only because the price per gram varies considerably depending on the substance sold, the country in which it is sold, and the country of origin, but also because the quantity exchanged in a single offer range between a few grams and several kilograms. High-stake transactions incentivize untrustworthy behavior of the sellers. For very high-value transactions, the long-term costs of a single defection (negative feedback, possible closure of account, or loss of platform deposit of 300 USD) are much lower than potential short-term gains from dishonest behavior (see Décary-Hétu and Leppänen, 2013; Holt, Smirnova, and Hutchings, 2016) . Ceteris paribus, we thus expect that actors are more cautious when the stakes are higher. An escrow offer should thus be attractive especially to buyers intending to purchase large quantities.

Hypothesis 6: The higher the transaction value, the more the number of successful transactions per offer using escrow will increase.

Design

Data

We use a dataset of auction-listings and transactions from the Alphabay cryptomarket between March 2015 and January 2017 (collected by McKenna and Goode, 2017) . Our data includes the following information on listings: product description, number of sales per listing, origin and destinations of the listed goods, method of payment, and transaction feedback (see Figure A1 in the Online Appendix). We also have information on sellers' nicknames and their lifespan on the platform. Following Przepiorka, Norbutas, and Corten (2017) , we restrict our data to transactions concerning a subset of illegal drugs, that is, buds and flowers, cocaine, hashish, heroin, ketamine, MDA, MDMA, and methamphetamine. We do so to avoid bias due to unobserved item heterogeneity (Diekmann et al., 2014) . Additionally, by focusing on illegal drugs we are able to use some common metrics to control for observable heterogeneity (e.g. quantity and price per gram). (Note that similar metrics do not exists for forged documents or illegal weapon, for instance.) The final dataset includes 466,714 transactions linked to 30,459 listings posted by 2,566 sellers.

Measurement

We measure escrow payment with a dummy variable (Escrow) taking value 1 if the offer uses escrow and 0 if it requires advance payment (also known as "finalized early" or "FE" in the cryptomarket jargon).

Our first explanatory variable in the models testing Hypotheses 1 and 2 is the number of recurrent customers the seller has had by the time their item is first listed online (Recurrent clients). We identify recurrent customers analyzing the text of feedback comments. We focus on keywords pointing to the evidence of repeated transactions between specific buyers and the seller (details in Online Appendix A3). Our second set of explanatory variables in the models testing Hypotheses 1 and 2 is the number of positive and negative feedback the seller has received by the time their item is first listed online (Positive reviews and Negative reviews, respectively). We follow Przepiorka, Norbutas, and Corten (2017) and interpret the latter two variables as measures of reputation. Since most of Alphabay feedback is positive (see Table 1 ), we also use the length of seller's history of trading on the platform (in days) as a complementary measure of positive reputation (Days selling). By doing so, we follow Reichelt, Sievert, and Jacob (2014) who show that seniority is a signal of trustworthiness in online environments. We use natural logs + 1 of the above explanatory variables.

Importantly, we construct longitudinal measures of recurrent clientele and reputation by aggregating transactions that have occurred prior to the time the item is listed. We do so thanks to information about the date of each transaction. For example, vendor X sold 10g of heroine to buyer Y on January 31 st , 2016. To approximate vendor X's reputation at the time of listing the product Y later bought, we look at X's previous transactions between January 31 st and the moment when s/he started selling on Alphabay-distinguishing between transactions that led to positive and negative reviews. 5 We use analogous procedure to calculate the number of recurrent customers at each point in time by aggregating transactions marked as "recurrent" by a given date.

Following Przepiorka, Norbutas, and Corten (2017) , our main dependent variable in tests of Hypotheses 3 to 6 is the log of the number of sales per listing + 1 (Sales). To test Hypotheses 5 and 6, we interact the escrow dummy with the vendor's reputation (measured as the number of positive and negative reviews) and the value of a given transaction (total price in USD for an offered quantity of drugs). The descriptive statistics of all the variables are presented in Table 1 . Note that the unit of most forthcoming analyses is a listing. Yet, for some analyses, it is transaction. Therefore, the number of observations in different models varies, as explained in table notes.

Table 1: Descriptive statistics **** Table 1 about here **** Notes: The table shows the mean and the standard deviation of the indicated variables.

Main Results

Determinants of Escrow

Is the escrow payment more frequently chosen by vendors who do not have many recurrent customers (Hypothesis 1), and who have not (yet) established a good reputation on the platform (Hypothesis 2)? To test these hypotheses, we regress the escrow payment on our measures of recurrent clientele and reputation.

We estimate two types of models. First, we estimate a multi-level model with random intercepts. Our units of analysis are listings nested within sellers. The intraclass correlation coefficient is 0.73, thus roughly 73% of the variance is attributable to the vendor-level variables. The model includes the likely correlates of using escrow at the listing level, namely the quantity of grams per sale and destination market (dummy variable for international sales). We also use a series of fixed effects for 1) the type of substance (eight categories of drugs mentioned above), 2) substance's country of origin (56 countries), and 3) month-year of the posting (to capture periodic macro fluctuations observed in cryptomarkets). Second, we estimate an analogous listing-level longitudinal linear probability model with seller fixed effects and standard errors clustered at the seller level.

Table 2 shows the results of our analysis. First, we find that sellers with higher numbers of recurrent clients-i.e. clients who repeatedly buy from the same individuals-use the escrow service less often. The model estimates a 1 percentage-point reduction in the probability of using escrow linked to having 305 more transactions from recurrent clients. Such transactions, however, constitute between 15 and 34% of the trade at Alphabay, according to our estimates. The result is in line with Hypothesis 1. Second, we do not find consistent evidence that sellers who have established reputation on the market use escrow less often (Hypothesis 2). The coefficients for the number of positive and negative reviews have expected signs (negative and positive, respectively; see column 1) but they are not statistically significant. When we use the vendor's lifespan on the platform as an alternative measure or reputation (the log of Days selling), the correlation becomes statistically significant (columns 1 and 2 of Table A1 in the Online Appendix). The latter model suggests that 100 additional days of selling on Alphabay reduces the probability of using escrow by a 1 percentage point. 6 We probe robustness of this finding in two ways. First, we address a possible source of bias in our measurement of recurrent clientele, coming from the fact that recurrent customers may stop leaving comments after multiple successful transactions with a given seller. If this were the case, our measure would underestimate the amount of recurrent clientele. We explore how this underestimation might affect our results by adopting the lower and upper bound approach. We consider the original measure of recurrent clients as a lower bound estimate. To produce an upper bound estimate, we code every transaction without a feedback as recurrent. Table A2 in the Online Appendix shows the results of our models using the upper bound estimates. The results are substantively the same.

Second, we exclude offers that did not generate a single purchase (Table A3 in the Online Appendix). In these cases, drug dealers may have used the Alphabay platform to advertise their products online, but sold them offline to local clientele. Encouragingly, the results remain unchanged. The table shows point estimates and standard errors of regression of the indicated outcomes on the indicated variables. Robust standard errors clustered at the date (monthyear) and vendor levels. Note that the above analyses rely on a counterfactual logic. Yet, there are some listings for which we were unable to find a meaningful counterfactual. As a result, these observations could not be analyzed in our difference-in-differences framework and thus were dropped from the regressions. This explains slight variations in the number of observations in different models (also compared to the complete dataset of listed items). *** p<0.01, ** p<0.05, * p<0.10.

Effect of Escrow on Sales

The above evidence confirms that the choice of using escrow is not random. It depends on the vendors' history of trade on the platform. The resultant endogeneity makes it difficult to estimate the effect of escrow on sales by simply comparing listings requesting the escrow payment or not. Such comparisons could be biased. Most straightforwardly, low reputation or less recurrent clients simultaneously affect the vendors' probability of using escrow and the vendors' number of sales per listing.

We address this problem by comparing the escrow listings to the non-escrow ones while keeping all relevant vendor characteristics constant-both time-variant and timeinvariant ones. Specifically, we regress the log of sales per listing on the type of payment (escrow versus advance payment) and a series of fixed effects:

𝑌 #$% = 𝛼 #$% + 𝛽 * 𝐸𝑠𝑐𝑟𝑜𝑤 # + 𝑉𝑒𝑛𝑑𝑜𝑟 𝐹𝐸 $ + 𝐷𝑎𝑡𝑒 𝐹𝐸 % + 𝜑 # + 𝛾 $% + 𝜀 #$%

Whereby Y is the outcome of interest for auction-listing i, by vendor j, on date g (monthyear). Escrowi is an indicator variable equal to 1 if the item offered for sale could be purchased through the escrow payment (vis-à-vis advance payment). Vendor fixed effects (𝑉𝑒𝑛𝑑𝑜𝑟 𝐹𝐸 $ ) and date fixed effects (𝐷𝑎𝑡𝑒 𝐹𝐸 % ) allow us to exploit variation in the payment method within auctions by the same vendors posted roughly at the same time.

Listing-level characteristics captured by 𝜑 # include: 1) the type of drugs (eight categories), 2) drugs' country of origin (56 countries; see Červený and Ours, 2019) , 3) price per gram, 4) the quantity of grams per sale (see Caulkins, 1994) , and 5) destination market (country dummies). Lastly, vendor's time-variant characteristics captured by 𝛾 $% include: i) the number of recurrent customers and ii) the number of positive and negative reviews. We estimate models with and without price per gram control.

Columns 3 and 4 of Table 2 show the results. In line with Hypothesis 4, we find that escrow payment is associated with a decrease in the number of sales per listing. The effect is sizeable: 5.4 fewer items sold per listing using the escrow payment, compared to the advance payment (column 3). 7 Interestingly, the negative effect of escrow is virtually unchanged in regressions in which we control for the price of drugs, compared to specifications in which we do not use this control (compare columns 3 and 4). We return to the role of prices in Section 5.1.

Heterogeneous Effects

The fact that transactions in escrow result in the lower number of sales is in contrast with theoretical models that predict positive effects of assurance devices. To better understand this finding, we explore hypothesized heterogeneities in the reported effects-related to vendor's reputation (Hypothesis 5), and transaction value (Hypothesis 6). We re-estimate our model including interactions between the escrow variable and the aforementioned moderators:

𝑌 #$% = 𝛼 #$% + 𝛽 * 𝐸𝑠𝑐𝑟𝑜𝑤 # + 𝛽 > 𝑆𝑡𝑎𝑘𝑒𝑠 # + 𝛽 A 𝑅𝑒𝑝𝑢𝑡𝑎𝑡𝑖𝑜𝑛 # + 𝛽 F 𝐸𝑠𝑐𝑟𝑜𝑤 × 𝑆𝑡𝑎𝑘𝑒𝑠 # + 𝛽 H 𝐸𝑠𝑐𝑟𝑜𝑤 × 𝑅𝑒𝑝𝑢𝑡𝑎𝑡𝑖𝑜𝑛 # + 𝑉𝑒𝑛𝑑𝑜𝑟 𝐹𝐸 $ + 𝐷𝑎𝑡𝑒 𝐹𝐸 % + 𝜑 # + 𝛾 $% + 𝜀 #$%

All the interaction terms are statistically significant (column 5 of Table 2 ). First, we find evidence of the moderating effect of reputation. The volume of sales for an auctionlisting with the escrow payment is lower, the more positive reviews the seller has received before posting the offer. The reverse is true for the number of negative comments: it reinforces the negative effect of escrow. These results are consistent with Hypothesis 5 and suggest that once trust in a specific seller has been established due to their reputation, additional assurance devices become redundant.

Second, we find evidence of the moderating effect of transaction value. We find that transaction value attenuates the negative effect of escrow. The effect of escrow has a significant effect on sales for transactions whose overall value is above 1,098 USD, as illustrated in Figure 1 (please note that the figure presents logarithmic transformation of the transaction value variable). The positive interaction between escrow and transaction value suggests that escrow increases sales, but only for very expensive purchases. This finding is consistent with Hypothesis 6.

We probe the robustness of these findings in two ways. First, we replicate our results including the share of item-specific positive reviews as another control. This variable may have an independent effect on sales, capturing the unobserved variation in product quality. The inclusion of item-specific ratings in the model does not alter our previous results (see Table A4 in the Online Appendix).

Second, we examine heterogeneity in the moderating effect of transaction value breaking the analyses by the drug's country of origin and the type of substance. We focus on six countries with more than 20,000 transactions in the data. **** Figure 1 about here **** Notes: The figure plots the marginal effect of the escrow dummy and its 90% confidence interval from the linear regression of the indicated outcomes. The effect of escrow is broken by the log value of transactions in USD.

Mechanisms

We have documented a relationship between the use of escrow and reduction in online sales of illegal drugs. What explains this finding? Below, we evaluate two hypothesized mechanisms: 1) increased prices, and 2) crowding out of trust between traders. Online Appendix A.5 addresses alternative explanations: mistrust in the platform and its failure to adjudicate disputes to the buyers' satisfaction.

Table 3: Escrow and mechanism **** Table 3 about here **** Notes: The table shows point estimates and standard errors of linear regressions of the log of sales outcome on the indicated variables. Standard errors are clustered at the date (month-year) and vendor levels. *** p<0.01, ** p<0.05, * p<0.10.

Increased prices

A possible explanation of the buyers' preference for advance payment over escrow could be related to the fact that escrow is associated with increased prices. We do not find clear evidence for this pattern. Drugs sold in escrow seem to be on average 1.08 USD more expensive per gram (holding time and vendor-level characteristics constant; see column 1 of Table 3 ). However, the estimate is considerably reduced (0.34 USD) while we control for listing-level characteristics (column 2) and the result is no longer statistically significant.

Could thus higher prices explain the negative effect of escrow on sales? Intuitively, the buyers may not want to pay for the escrow service while engaging in transactions that could be regulated by reputation systems. This logic is consistent with the fact that escrow correlates with greater reduction in sales when offered by sellers who have built a good reputation on the platform (the negative interaction between escrow and reputation; column 5 of Table 2 ). Put differently, reputation may provide sufficient assurance against untrustworthy behaviors.

The explanation is also consistent with the positive interaction between escrow and transaction value (column 5 of Table 2 ). From the sellers' perspective, in high-value transactions, a one-shot defection could possibly outweigh the costs of reputation damage related to untrustworthy behavior. As a result, if the stakes are high, reputation feedback alone may not suffice to incentivize the sellers' honest behavior, making the escrow service a desirable safeguard.

Crowding out of trust between traders

In the theory section, we proposed that escrow may backfire by crowding out trust between traders. We find two pieces of evidence that are consistent with this mechanism.

First, we do a sentiment analysis of reviews, comparing comments posted after transactions in escrow and advance payment, using AFINN lexicon (more details on the procedures and supporting qualitative evidence in Online Appendix A.6). Column 4 of Table 3 shows that feedback posted after advance payment transactions is more positive compared to feedback posted after escrow transactions. The differences are statistically significant. The model uses an analogous specification to model 5 in Table 2 , yet the feedback is measured at the transaction (not listing) level, which changes our unit of analysis.

The second piece of suggestive evidence in support of the crowding out mechanism comes from a keyword-based analysis of feedback. We classify all the comments based on whether they include any trust-related keywords. We conservatively rely on the word "trust" and its synonyms taken from the Oxford Dictionary of English (see Online Appendix A.7). Column 5 of Table 3 shows that trust-related keywords appear more frequently in feedback posted after transactions that rely on advance payment (vis-à-vis escrow). Again, the differences are statistically significant, although substantively small (plausibly due to the very conservative selection of keywords).

Discussion and Conclusion

We analyzed the effects of escrow on online sales in a very large cryptomarket for illegal drugs. We find that escrow is associated with reduced sales of drugs, as measured by the number of purchases per auction-listing. We provide suggestive evidence that escrow may be incompatible with informal social control of typical drugs users' communities, crowding out trusting and reciprocal behaviors between traders (see Bohnet and Baytelam, 2007; Holmås et al., 2010) . This finding resonates with criminological research that describes illicit economies as "pre-modern" and strongly reliant on informal social control (see, e.g., Beckert and Wehinger, 2013; Reuter, 1983) .

TABLE 4 ABOUT HERE Before concluding, we briefly outline some limitations. The data on which we conducted our analyses concern illegal transactions in the "dark web." We thus do not know whether and to what extent our results generalize to other settings. Many legal markets rely on analogous assurance devices, including PayPal and Authorize.net services (see González, 2004). If the crowding out mechanism is a dominant channel behind the escrow backlash, the reported effects may not apply to online marketplaces which are not characterized by strong community ties (such as Amazon or eBay). Still, our conclusions may generalize to some online trading communities, such as traders in CD and vinyl records at Discogs platform or similar collectors' communities whose users are linked through strong reciprocity norms. That said, future research should explore the applicability of our mechanisms in different contexts.

Further limitations come from the operationalization of some of our key variables and a few data-related concerns. First, we identified recurrent customers using qualitative information provided in the feedback messages (see Online Appendix A3). Relying on selected keywords to identify recurrent customers certainly yields a rough measurement.

However, as repeated transactions are known to be prevalent in cryptomarkets (Dècary-Héetu and Quessy-Doré 2017), false negatives are more probable than false positives.

Thus, although we may have underestimated recurrent customers, this is likely to result in a conservative bias.

Second, our sentiment analysis of transaction feedback could be imprecise due to the fact that some comments are written in languages other than English and the AFINN lexicon might not recognize the cryptomarket jargon. We address both concerns in Online Appendix A.6. For example, in one supplementary exercise, we only select comments that explicitly mention escrow and advance payment ("finalized early"). We find that comments mentioning the escrow payment have a significantly less positive sentiment scores than comments mentioning advance payment (column 5 in Table 3 ). This finding builds confidence in our sentiment analysis.

Our article has some important implications. We find that the negative effect of escrow on sales turns positive for high-value transaction (i.e., above 1,100 USD). One possible implication is that criminal organizations who generally trade large quantities are more likely to operate on markets with a higher degree of institutionalization. As noted by Aldrige and Decary-Hetu ( 2014 ), most of the drug deals on cryptomarkets involve small quantities. However, a minor percentage of higher-value transactions generates a huge share of the market revenues (on Alphabay 108 million USD, accounting for 44% of the total revenues; see Table A5 in the Online Appendix), indicating that some business-to-business trading operates in cryptomarkets, at least at the lower level of the distribution network. Therefore, inasmuch as cryptomarkets are capable to develop reliable institutions, these may principally attract drug traffickers.

Endnotes

1 One could describe online exchange as "pseudo-anonymous", given that traders often use pseudonyms, which makes them recognizable to each other to a certain extent (see Appendix A.2).

2 A commitment device is an arrangement through which a person makes it impossible (or non-profitable) for herself to deviate from a promised course of action. An assurance device is an arrangement through which a person protects herself from untrustworthy behaviors of others by making it impossible (or non-profitable) for them to deviate from a promised course of action. Escrow can be seen as a commitment device from the perspective of a seller, and an assurance device from the perspective of a buyer.

3 Note that sometimes sellers may be unable to verify the quality of the products. Purity is not perfectly correlated with the positive effects of substances, and drug quality control is a complex process that requires technical knowledge (Broséus, Gentile, and Esseiva, 2016) . 4 One could wonder whether sellers are able to opt out of escrow only once they have established a history of honest behavior on the platform. In our data (details in section 3.1), we find that no seller traded outside escrow on the first day of their activity on the platform. The earliest sale without escrow involved a seller who, by that time, had been trading for nine days. Yet, during these nine days the seller had completed no other transaction. Thus, Alphabay does not seem to limit the availability of escrow payment to sellers with a history of honest behavior.

5 Our data also include "trust level" and "vendor level" scores, which are reputation indicators assigned by the platform on a rating scale. We did not use these variables, since they correlate with both number of days selling and percentage of positive reviews.

Moreover, these two variables are constant within seller, given that the data was collected through a single "crawl." It is thus not possible to construct longitudinal measures for these variables, as we did for the number of positive and negative reviews. 6 One could argue that established sellers learn that providing extra service to their clientele pays off, and escrow is such a service. Therefore, established sellers may be in a better position to opt out from escrow, but may decide to not do it for the sake of giving clients an additional choice.

A.1. Crypomarkets and the Alphabay

The present section provides information on the context of our study. Cryptomarkets are online marketplaces trading in illegal drugs, weapons, and forged documents (Soska and Christin, 2015) . The most common types are cryptomarkets for illegal drugs (Aldridge and Décary-Hétu, 2016; Demant, Munksgaard, and Houborg, 2018) . Cryptomarkets are located on the "dark web" and are only accessible through an encryption software that allows buyers and sellers to conceal their identities and protect themselves from detection by law enforcement (Martin, 2014; Soska and Christin, 2015) . This is achieved through the TOR network, which anonymizes the location of the marketplace and the Internet traffic of market users (Syverson, Dingledine, and Mathewson, 2004; Moore and Rid, 2016) . Moreover, the traders use virtual currencies that allow them to circumvent traditional means of online payments (Martin, 2014) .

In this study, we focus on one of the largest known cryptomarkets to date:

Alphabay. This cryptomarket operated from March 2015 until January 2017, when it was shut down by an international police operation. According to the Eastern District of California prosecutor ("United States of America v. Alexandre Cazes ALPHA02", 2017).

At the time when the police closed the platform and arrested its creator, the market hosted hundreds of thousands of listings of illegal goods including illicit drugs, fraudulent services, malware, counterfeit documents, and firearms.

In many respects, Alphabay was similar to a conventional e-commerce website.

After sending a security deposit of 300 USD to Alphabay's "wallet," the vendors could create a profile visible to the buyers and offer goods for sale. These profiles provided information about delivery destinations, quantity of the product for sale, description of the item quality and characteristics, and price. In addition, the platform provided information about vendors' feedback from previous customers, as well as "trust level"

and "vendor level" scores, which were platform-assigned metrics based on the vendors' lifespan on the market, their total number of sales, and the content of their reviews.

While selling their products on Alphabay, the vendors could choose between two main methods of payment: advance payment or escrow payment 1 . In the case of advance payment, the buyers sent money directly into the sellers' wallet before receiving the goods. Thus, the buyers had no assurance against opportunistic behavior of the seller. By contrast, in the escrow system, the buyers sent the money to an escrow service, which released the money and transferred it to the seller only after the buyer confirmed the receipt of the item purchased. Transactions were thus overlooked by the platform, which guaranteed to the buyers that any opportunistic or noncompliant behavior of the seller could be redressed (Tzanetakis et al., 2016; Morselli et al., 2017; Bakken, Moeller, and Sandberg, 2018) .

Importantly, in the early days of Alphabay, the platform creators set escrow as the default payment method. As documented in the court files of the prosecutor of the Eastern District of California, users who wanted to complete transactions outside of escrow were provided with Alphabay's services to cover the digital traces of their illicit transactions.

However, the platform trading rules explicitly stated: "Do not send funds outside of escrow. If you do so, it is entirely at your own risk" (retrieved from Alphabay guidelines for traders; 26 April 2017).

1 Other payment solutions were also possible. For example, on Alphabay there was a possibility to negotiate the amount of down-payment with the vendor. However, the use of this alternative payment scheme was extremely limited (4.6 percent). Therefore, we do not discuss this option and exclude the related transactions from our analyses.

A.2. Procedure to identify recurrent clients

We distinguish between recurrent and first-time clients by performing text analysis of transaction feedback posted by the buyers. The content of these comments allows us to approximate recurrent sales thanks to specific keywords. For example, some comments include phrases such as "perfect as usual", "another fine transaction", "[vendor name] always the best", etc. We infer from these phrases that the traders had been involved in a previous transaction together. Below, we provide a full list of keywords that we use to approximate recurrent clients. In sum, we coded 15 percent of clients as recurrent ones (see Table 1 in the main text).

To further improve our identification of recurrent clients, we search for traders who might have known and recognized each other from other cryptomarkets (e.g. Silk Road 2.0). There is evidence that sellers' reputation can be transferred across marketplaces (Décary-Hétu and Quessy-Doré, 2017; Norbutas, Ruiter and Corten, 2020; Ladegaard, 2020) . Buyers could possibly identify their trusted vendors thanks to user names or product descriptions posted on other platforms. Again, we rely on feedback comments to identify these connections. We compile a list of keywords in which buyers refer to other cryptomarkets (e.g. "as reliable as in Silk Road"; see below). We assume that these references are indicative of a prior history of exchange with a vendor. Following this procedure, we identified an additional 2 percent of recurrent clients. Finally, we validate our measurement of recurrent customers through qualitative analysis of 500 comments (randomly selected).

We are aware that relying on selected keywords to identify recurrent customers certainly yields a rough measurement. However, as repeated transactions are known to be rather prevalent in cryptomarkets (Dècary-Héetu and Quessy-Doré, 2017, Norbutas, Ruiter, and Corten, 2020) , false negatives are more probable than false positives. Thus, although we may have underestimated recurrent customers, this is likely to result in a conservative bias. We address this potential measurement bias, comparing lower and upper bound estimates of recurrent clientele (see Table A2 and related discussion in the main text).

Keywords used to identify recurrent clients

"regular", "another", "every time", "as usual", "my vendor", "always", "again", "in the past", "once more", "long time", "all the time", "previously", "never", "come back", "like first time", "second order", "already ordered", "2nd time ordering", "invariably", "without fail", "habitual", "unfailingly", "infallibly", "continue".

Keywords used to identify clients from other platforms

"silk road", "agora", "dream market", "silkroad", "dreammarket", "wallstreet", "wall street", "silkkietie" , "valhalla", "hansa", "atlantis", "c'thulhu", "evolution", "the farmer's market", "sheep marketplace", "utopia", "empire market", "bitcoinpharma", "deepbay", "project black flag", "blackmarket reloaded", "budster", "tormarket", "flomarket", "tortuga", "blackbox market", "greyroad", "cantina", "breaking bad", "black goblin market", "cannabis road", "utopia", "black services market", "red sun marketplace", "buyitnow", "doge road", "torescrow", "white rabbit", "drugslist", "armory vendor market", "freebay", "torescrow", "sanitarium market", "hansa", "exxtacy", "drugslist", "darknet nation", "mr nice guy", "torbay", "darkbay", "pigeon market", "free market", "tortuga 2", "deepzon", "silk street", "underground market", "cannabis road 2", "pandora", "pirate market", "freedom market", "hydra", "cloud-nine", "blue sky", "torbazaar", "topix 2", "alpaca marketplace", "cannabis road 3", "andromeda", "the marketplace", "onionshop", "tom", "area51", "panacea", "ironclad", "kiss", "blackbank market", "tornado", "havana/absolem", "agape", "zanzibar spice", "outlaw market", "middle earth marketplace", "diabolus/sr3", "nucleus marketplace", "abraxas", "mr nice guy 2", "oxygen", "tochka", "darknet heroes league", "east india company", "therealdeal", "haven", "anarchia", "poseidon", "amazon dark", "simply bear", "horizon market". Notes: The figure plots the marginal effects of the escrow dummy from the linear regression of the indicated outcomes using a 2 nd degree polynomial (to allow for nonlinearity). The effect of escrow is broken by the log value of transaction in USD and it is plotted separately for different countries of origin of drugs. Notes: The figure plots the marginal effects of the escrow dummy from the linear regression of the indicated outcomes using a 2 nd degree polynomial (to allow for nonlinearity). The effect of escrow is broken by the log value of transaction in USD and it is plotted separately for different substances.

A.3. Additional figures

A.4. Additional tables

Table A1: Determinants of escrow (alternative measure of reputation) (1) (2) Escrow (multi-level) Escrow (clustered SE) Grams (log) -0.007 *** -0.008 * (0.002) (0.004) International sale -0.011 * -0.023 (0.006) (0.018) Transaction value (log) -0.001 -0.000 (0.002) (0.005) Recurrent clients -0.004 -0.000 (0.003) (0.003) Days selling (log) -0.003 ** -0.003 ** (0.001) (0.001) Month-year FE -Yes Vendor FE -Yes Drugs type FE Yes Yes Drugs origin FE No Yes Unit of analysis Listing Listing N 29791 29478 Notes: The table shows point estimates and standard errors of regression of the indicated outcome on the indicated variables. Robust standard errors clustered at the date (monthyear) and vendor levels. *** p<0.01, ** p<0.05, * p<0.10. Table A2: Determinants and effects of escrow (upper-bound estimates for recurrent clients) (1) (2) (3) (4) Escrow (multi-level) Escrow (clustered SE) Sales per item (log) Sales per item (log) Grams (log) -0.006 *** -0.008 * -0.262 *** -0.255 *** (0.002) (0.004) (0.041) (0.033) Transaction value (log) -0.002 -0.000 -0.016 -0.192 *** (0.002) (0.005) (0.036) (0.035) International sale -0.009 * -0.022 0.088 0.078 (0.005) (0.018) (0.056) (0.056) Recurrent clients -0.011 ** -0.015 ** -0.190 *** -0.209 *** (0.005) (0.007) (0.025) (0.025) Positive reviews 0.005 * 0.010 ** 0.012 0.076 ** (0.003) (0.004) (0.017) (0.030) Negative reviews 0.001 0.012 -0.107 *** -0.161 *** (0.009) (0.010) (0.036) (0.049) Escrow -0.256 ** -1.238 *** (0.102) (0.200) Price per gram (log) -0.411 *** -0.402 *** (0.050) (0.046) Escrow x Transaction value (log) 0.212 *** (0.032) Escrow x Positive reviews -0.072 ** (0.027) Escrow x Negative reviews 0.093 (0.060) Month-year FE -Yes Yes Yes Vendor FE -Yes Yes Yes Drugs type FE Yes Yes Yes Yes Drugs origin FE No Yes Yes Yes Unit of analysis Listing Listing Listing Listing N 29791 29478 29441 29441

Notes: The table shows point estimates and standard errors of regression of the indicated outcomes on the indicated variables. Robust standard errors clustered at the date (monthyear) and vendor levels. Note that the above analyses rely on a counterfactual logic. Yet, there are some listings for which we were unable to find a meaningful counterfactual. As a result, these observations could not be analyzed in our difference-in-differences framework and thus were dropped from the regressions. This explains slight variations in the number of observations in different models (also compared to the complete dataset of listed items). *** p<0.01, ** p<0.05, * p<0.10.

Table A3: Effects of escrow (zero sales excluded) (1) (2) Sales per item (log) Sales per item (log) Escrow -0.378 *** -1.015 *** (0.111) (0.180) Grams (log) -0.295 *** -0.314 *** (0.070) (0.063) Price per gram (log) -0.424 *** -0.445 *** (0.074) (0.069) International sale 0.080 0.062 (0.060) (0.057) Recurrent clients -0.172 *** -0.194 *** (0.030) (0.031) Transaction value (log) -0.045 -0.129 ** (0.067) (0.059) Positive reviews 0.024 0.081 ** (0.020) (0.030) Negative reviews -0.074 ** -0.140 *** (0.032) (0.044) Escrow x Transaction value (log) 0.149 *** (0.027) Escrow x Positive reviews -0.062 ** (0.024) Escrow x Negative reviews 0.119 ** (0.054) Month-year FE Yes Yes Vendor FE Yes Yes Drugs type FE Yes Yes Drugs origin FE Yes Yes Unit of analysis Listing (sales>0) Listing (sales>0) N 15217 15217 Notes: The table shows point estimates and standard errors of regression of the indicated outcome on the indicated variables. Robust standard errors clustered at the date (monthyear) and vendor levels. *** p<0.01, ** p<0.05, * p<0.10. Table A4: Effects of escrow (item-specific reviews) (1) (2) Sales per item (log) Sales per item (log) Escrow -0.383 *** -1.021 *** (0.112) (0.182) Grams (log) -0.295 *** -0.314 *** (0.070) (0.063) Price per gram (log) -0.425 *** -0.446 *** (0.074) (0.069) International sale 0.079 0.061 (0.059) (0.056) Recurrent clients -0.174 *** -0.196 *** (0.030) (0.031) Transaction value (log) -0.044 -0.129 ** (0.067) (0.059) Item positive reviews (%) 0.025 0.081 ** (0.020) (0.030) Negative reviews -0.073 ** -0.138 *** (0.032) (0.044) Escrow x Transaction value (log) 0.149 *** (0.027) Escrow x Positive reviews -0.061 ** (0.024) Escrow x Negative reviews 0.117 ** (0.054) Month-year FE Yes Yes Vendor FE Yes Yes Drugs type FE Yes Yes Drugs origin FE Yes Yes Unit of analysis Listing (sales>0) Listing (sales>0) N 15217 15217 Notes: The table shows point estimates and standard errors of regression of the indicated outcome on the indicated variables. Robust standard errors clustered at the date (monthyear) and vendor levels. Note that the analysis excludes items which have not been sold to a single client (see Table A3 above) because these items do not have any reviews. *** p<0.01, ** p<0.05, * p<0.10. Table A5: Market value and the share of large transactions Drug type Market value Market share of transactions >1,100 USD MDA 313,535 51% Hash 3,052,108 30% Ketamine 3,907,987 52% Heroin 4,808,853 23% Meth 10,300,000 51% MDMA 14,700,000 51% Cocaine 26,500,000 40% Buds & Flowers 44,600,000 48% Overall 108,000,000 44% Notes: The table shows the overall market value in USD and the market share of transactions above 1,100 USD by drug type category.

A.5. Alternative explanations for the negative effect of escrow on sales:

Mistrust towards the cryptomarket administrators

An alternative explanation of the negative effects of escrow is mistrust toward administrators of cryptomarket platforms, such as Alphabay. This could be the case, for example, if dispute settlement was not handled properly by the platform or if the platform was not perceived to be as reliable as the contract law in a legal market (Horton-Eddison and Di Cristofaro, 2017; Van Buskirk et al., 2017; Moeller, Munksgaard, and Dement, 2017) .

To investigate this possibility indirectly, we study whether transactions in escrow are more likely to result in negative feedback. We regress the likelihood of positive review on the escrow payment as well as the control variables and the fixed effects used in the previous regressions (see columns 2 and 3 of Table 3 ). The results are presented in column Table A6 . We do not find evidence that transactions using escrow result in less positive feedback. On the contrary, customers leave significantly more positive feedback if they purchase products using the escrow payment (a 0.006 increase in the log of sales). This finding is not consistent with the flawed dispute resolution mechanism.

Relatedly, one could also argue that vendors and buyers may be unwilling to use escrow because there is a risk of an exit scam in which the market owners steal all cryptocurrencies deposited in the escrow. Alternatively, the traders may fear that law enforcement authorities could overtake the platform and freeze all the assets, including cryptocurrencies deposited in the escrow. While we cannot rule out this mechanism, it is not consistent with some of our empirical results. For example, if the general mistrust toward Alphabay administrators drove cryptomarket users away from escrow, we should not find that reputation moderates the escrow's effect on sales. Based on the above evidence, we thus tentatively conclude that the platform mistrust mechanism is certainly not the only mechanism at play.

Table A6: Escrow and positive reviews (1) Positive review Escrow 0.006 ** (0.002) Month-year FE Yes Vendor FE Yes Drugs type FE Yes Drugs origin FE Yes Controls Yes Unit of analysis Transaction N 430457 Notes: The table shows point estimates and standard errors of linear regressions of the log of sales outcome on the indicated variable. Standard errors are clustered at the date (month-year) and vendor levels. *** p<0.01, ** p<0.05, * p<0.10.

One could have legitimate concerns related to the use of the AFINN lexicon in the case at hand. As we mentioned in the main text, community members in cryptomarkets develop their unique jargon (e.g. "FE" for advance payment), which the AFINN lexicon cannot recognize. Our analyses may thus miss some contextual details. Yet, the text that we study uses enough common words to make our text analysis meaningful. Figure A4 shows the most common words in feedback comments in our data. The figure underscores the relative commonality of text posted by the buyers at Alphabay. In sum, we find no compelling reasons to not use the AFINN lexicon. In fact, the lexicon has been used for the analysis of customer service chatbox (Feine, Morana, and Gnewuch, 2019) and tweets (Koto and Adriani, 2015) . All these types of text are comparable to online feedback, at least from the lexical perspective as well as in terms of length.

A.7. Trust-related keywords

"credence", "faith", "stock", "acceptance", "assurance", "assuredness", "certainty", "certitude", "conviction", "positiveness", "sureness", "surety", "credit", "reliance", "dependence", "hope". of 1,000 reviews)

References

1. References
2. A Afilipoaie P Shortis Crypto-Market Enforcement-New Strategy and Tactics 2018 54
3. Not an 'Ebay for Drugs': The Cryptomarket 'Silk Road' as a Paradigm Shifting Criminal Innovation Judith Aldridge David Ddcary-Hhtu 10.2139/ssrn.2436643 SSRN 2436643 SSRN Electronic Journal SSRN Journal 1556-5068 2014 Elsevier BV
4. Determinants of escrow service adoption in consumer-to-consumer online auction market: An experimental study Solomon Antony Zhangxi Lin Bo Xu 10.1016/j.dss.2006.04.012 Decision Support Systems Decision Support Systems 0167-9236 42 3 2006 Elsevier BV
5. Coordination Problems in Cryptomarkets: Changes in Cooperation, Competition, and Valuation S A Bakken K Moeller S Sandberg European Journal of Criminology 15 4 2018
6. Centralized Sanctioning and Legitimate Authority Promote Cooperation in Humans D Baldassarri G Grossman Proceedings of the National Academy of Sciences 108 27 2011
7. The Impact of Negotiated Exchange on Trust and Trustworthiness D Barrera Social Networks 29 4 2007
8. The Social Mechanisms of Trust. Sociologica D Barrera Italian Journal of Sociology 2 2008. 2008
9. Introduction: The Social Organization of Illicit Markets J Beckert M Dewey The Architecture of Illegal Markets: Towards an Economic Sociology of Illegality in the Economy Jens Beckert Matías Dewey Oxford Oxford University Press 2017
10. J Beckert F Wehinger the shadow: Illegal markets and economic sociology 2013 11
11. I Bohnet Y Baytelman Institutions and Trust: Implications for Preferences 2007 19
12. Efficiency in the Trust Game: An Experimental Study of Precommitment J Bracht N Feltovich International Journal of Game Theory 37 1 2008
13. The cutting of cocaine and heroin: a critical review J Broséus N Gentile P Esseiva Forensic Science International 262 2016
14. Rational Choice Research on Social Dilemmas: Embeddedness Effects on Trust V Buskens W Raub The Handbook of Rational Choice Social Research Rafael Wittek Tom Snijders Victor Nee Stanford Stanford University Press 2013
15. Experimental Tests of a Sequential Equilibrium Reputation Model C Camerer K Weigelt Econometrica 56 1 1988
16. What is the average price of an illicit drug? J P Caulkins Addiction 89 7 1994
17. Cannabis Prices on the Dark Web J Červený J C Van Ours European Economic Review 120 103306 2019
18. Evolving and Diversifying Selling Practices on Drug Cryptomarkets: An Exploration of Off-Platform “Direct Dealing” Andrew Childs 0000-0002-1956-3212 Ross Coomber Melissa Bull Monica J Barratt 10.1177/0022042619897425 Journal of Drug Issues Journal of Drug Issues 0022-0426 1945-1369 50 2 2020 SAGE Publications
19. Foundations of social theory J S Coleman 1990 Harvard University Press Cambridge (MA
20. Criminals and Signals: An Assessment of Criminal Performance in the Carding Underworld D Décary-Hétu A Leppänen Security Journal 2939 2013
21. Are Repeat Buyers in Cryptomarkets Loyal Customers? Repeat Business Between Dyads of Cryptomarket Vendors and Users D Décary-Hétu O Quessy-Doré American Behavioral Scientist 61 11 2017
22. Reputation Formation and the Evolution of Cooperation in Anonymous Online Markets A Diekmann B Jann W Przepiorka S Wehrli American Sociological Review 79 1 2014
23. Building Them Up, Breaking Them Down: Topology, Vendor Selection Patterns, and a Digital Drug Market's Robustness to Disruption S W Duxbury D L Haynie Social Networks 52 2018
24. Third-Party Punishment and Social Norms E Fehr U Fischbacher Evolution and Human Behavior 25 2 2004
25. Motivation Crowding Theory Bruno S Frey Reto Jegen 10.1111/1467-6419.00150 Journal of Economic Surveys Journal of Economic Surveys 0950-0804 1467-6419 15 5 2001 Wiley
26. PayPal: The Legal Status of C2C Payment Systems A G González Computer Law & Security Review 20 4 2004
27. Reputation in the Internet Black Market: An Empirical and Theoretical Analysis of the Deep Web R A Hardy J R Norgaard Journal of Institutional Economics 12 3 2016
28. Does Monetary Punishment Crowd Out Pro-Social Motivation? A Natural Experiment on Hospital Length of Stay T H Holmås E Kjerstad H Lurås O R Straume Journal of Economic Behavior & Organization 75 2 2010
29. An Exploration of the Factors Affecting the Advertised Price for Stolen Data T J Holt Y T Chua O Smirnova APWG eCrime Researchers Summit IEEE 2013
30. Examining Signals of Trust in Criminal Markets Online T J Holt O Smirnova A Hutchings Journal of Cybersecurity 2 2 7 2016
31. Hard Interventions and Innovation in Crypto Drug Markets: The Escrow Example M Horton-Eddison M Di Cristofaro Policy Brief 11 2017
32. Hope or Hype: On the Viability of Escrow Services as Trusted Third Parties in Online Auction Environments X Hu Z Lin A B Whinston H Zhang Information Systems Research 15 3 2004
33. The Emergence of Exchange Structures: An Experimental Study of Uncertainty, Commitment, and Trust P Kollock American Journal of Sociology 100 2 1994
34. Open Secrecy: How Police Crackdowns and Creative Problem-Solving Brought Illegal Markets out of the Shadows I Ladegaard Social Forces 99 2 2020
35. Efficiency of Purchasing and Selling Agents in Markets with Quality Uncertainty: The Case of Illicit Drug Transactions C B Lakhdar H Leleu N G Vaillant F C Wolff European Journal of Operational Research 226 3 2013
36. Nice People Doing Shady Things': Drugs and the Morality of Exchange in the Darknet cryptomarkets K Masson A Bancroft International Journal of Drug Policy 58 2018
37. Darknet Market Archives M Mckenna S Goode 2017. 2013-2015
38. Flow My FE the Vendor Said: Exploring Violent and Fraudulent Resource Exchanges on Cryptomarkets for Illicit Drugs K Moeller R Munksgaard J Demant American Behavioral Scientist 61 11 2017
39. Putting a Price on Drugs: An Economic Sociological Study of Price Formation in Illegal Drug Markets K Moeller S Sandberg Criminology 57 2 2019
40. Undermining trust and cooperation: The paradox of sanctioning systems in social dilemmas Laetitia B Mulder Eric Van Dijk David De Cremer Henk A M Wilke 10.1016/j.jesp.2005.03.002 Journal of Experimental Social Psychology Journal of Experimental Social Psychology 0022-1031 42 2 2006 Elsevier BV
41. Uncertainty and Risk: A Framework for Understanding Pricing in Online Drug Markets R Munksgaard M Tzanetakis International Journal of Drug Policy 101 103535 2022
42. Better Bang for the Buck? Generalizing Trust in Online Drug Markets R Munksgaard J A Ferris A Winstock L J Maier M J Barratt British Journal of Criminology 2022
43. Building a Case for Trust: Reputation R Munksgaard Institutional Regulation and Social Ties in Online Drug Markets. Global Crime 2022
44. Believe it When You See It: Dyadic Embeddedness and Reputation Effects on Trust in Cryptomarkets for Illegal Drugs L Norbutas S Ruiter R Corten Social Networks 63 2020
45. R Nozick Anarchy, state, and utopia New York Basic Books 1974 5038
46. Markets as Governance Environments for Organizations at the Edge of Illegality: Insights from Social Network Analysis M Odabaş T J Holt R L Breiger American Behavioral Scientist 61 11 2017
47. Temporal Embeddedness and Signals of Trustworthiness: Experimental Tests of a Game Theoretic Model in the United Kingdom, Russia, and Switzerland W Przepiorka A Diekmann European Sociological Review 29 5 2013
48. Order without Law: Reputation Promotes Cooperation in a Cryptomarket for Illegal Drugs Wojtek Przepiorka Lukas Norbutas Rense Corten 10.1093/esr/jcx072 European Sociological Review 0266-7215 1468-2672 33 6 2017 Oxford University Press (OUP)
49. Hostage Posting as a Mechanism of Trust: Binding, Compensation, and Signaling W Raub Rationality and Society 16 3 2004
50. How Credibility Affects eWOM Reading: The Influences of Expertise, Trustworthiness, and Similarity on Utilitarian and Social Functions J Reichelt J Sievert F Jacob Journal of Marketing Communications 20 1-2 2014
51. Disorganized Crime: The Economics of the Visible Hand P Reuter 1983 MIT press Cambridge, MA
52. Illegal 'Lemons': Price Dispersion in Cocaine and Heroin Markets P Reuter J P Caulkins Bulletin on Narcotics 56 1-2 2004
53. Job Market Signaling M Spence Quarterly Journal of Economics 87 3 1973
54. J Van Buskirk S Naicker R Bruno L Burns C Breen A Roxburgh Drugs and the Internet 2016
55. The Recovery of Online Drug Markets Following Law Enforcement and Other Disruptions J Van Buskirk R Bruno T Dobbins C Breen L Burns S Naicker A Roxburgh Drug and Alcohol Dependence 173 2017
56. Trust: The Evolutionary Game of Mind and Society T Yamagishi 2011 New York, Springer
57. Hidden wholesale: The drug diffusing capacity of online drug cryptomarkets Judith Aldridge 0000-0002-9402-8682 David D Décary-Hétu 10.1016/j.drugpo.2016.04.020 International Journal of Drug Policy International Journal of Drug Policy 0955-3959 35 2016 Elsevier BV
58. Coordination Problems in Cryptomarkets: Changes in Cooperation, Competition, and Valuation S A Bakken K Moeller S Sandberg European Journal of Criminology 15 4 2018
59. The Moral Economy: Why Good Incentives Are No Substitute for Good Citizens S Bowles 2016 Yale University Press New Haven
60. Are Repeat Buyers in Cryptomarkets Loyal Customers? Repeat Business Between Dyads of Cryptomarket Vendors and Users D Décary-Hétu O Quessy-Doré American Behavioral Scientist 61 11 2017
61. Personal Use, Social Supply, or Redistribution? Cryptomarket Demand on Silk Road 2 and Agora J Demant R Munksgaard E Houborg Trends in Organized Crime 21 1 2018
62. Measuring Service Encounter sStisfaction with Customer Service Chatbots Using Sentiment Analysis J Feine S Morana U Gnewuch 2019
63. Hard Interventions and Innovation in Crypto Drug Markets: The Escrow Example M Horton-Eddison M Di Cristofaro Policy Brief 11 2017
64. A Comparative study on Twitter Sentiment Analysis: Which Features Are Good? F Koto M Adriani International Conference on Applications of natural language to information systems Cham Springer 2015
65. Open Secrecy: How Police Crackdowns and Creative Problem-Solving Brought Illegal Markets out of the Shadows I Ladegaard Social Forces 99 2 2020
66. Drugs on the Dark Net: How Cryptomarkets are Transforming the Global Trade in Illicit Drugs J Martin 2014 Springer New York
67. Flow My FE the Vendor Said: Exploring Violent and Fraudulent Resource Exchanges on Cryptomarkets for Illicit Drugs K Moeller R Munksgaard J Demant American Behavioral Scientist 61 11 2017
68. Cryptopolitik and the Darknet Daniel Moore Thomas Rid 10.1080/00396338.2016.1142085 Survival Survival 0039-6338 1468-2699 58 1 2016 Informa UK Limited
69. Conflict Management in Illicit Drug Cryptomarkets C Morselli D Décary-Hétu M Paquet-Clouston J Aldridge International Criminal Justice Review 27 4 2017
70. Undermining Trust and Cooperation: The Paradox of Sanctioning Systems in Social Dilemmas L B Mulder E Van Dijk D De Cremer H A M Wilke Journal of Experimental Social Psychology 42 2006
71. Believe it When You See it: Dyadic Embeddedness and Reputation Effects on Trust in Cryptomarkets for Illegal Drugs L Norbutas S Ruiter R Corten Social Networks 63 2020
72. Hostage Posting as a Mechanism of Trust: Binding, Compensation, and Signaling W Raub Rationality and Society 16 3 2004
73. Measuring the Longitudinal Evolution of the Online Anonymous Marketplace Ecosystem K Soska Christin N Usenix Security 15 2015
74. Tor: The Second-Generation Onion Router P Syverson R Dingledine N Mathewson 2004 Usenix Security
75. The Transparency Paradox. Building Trust, Resolving Disputes and Optimising Logistics on Conventional and Online Drugs Markets M Tzanetakis G Kamphausen B Werse R Laufenberg International Journal of Drug Policy 35 2016