Using Respondent-Driven Sampling to Recruit Illegal Drug Purchasers to Evaluate a Drug Market Intervention

Abstract

Background

Violent drug markets are not as prominent as they once were in the United States, but they still exist and are associated with significant crime and lower quality of life. The drug market intervention (DMI) is an innovative strategy that uses focused deterrence, community engagement, and incapacitation to reduce crime and disorder associated with these markets. Although studies show that DMI can reduce crime and overt drug activity, one perspective is prominently missing from these evaluations: those who purchase drugs.

Objectives

This study explores the use of respondent-driven sampling (RDS)—a statistical sampling method—to approximate a representative sample of drug users who purchased drugs in a targeted DMI market to gain insight into the effect of a DMI on market dynamics.

Methods

Using RDS, we recruited individuals who reported hard drug use (crack or powder cocaine, heroin, methamphetamine, or illicit use of prescriptions opioids) in the last month to participate in a survey. The main survey asked about drug use, drug purchasing, and drug market activity before and after DMI; a secondary survey asked about network characteristics and recruitment.

Conclusions

Our sample of 212 respondents met key RDS assumptions, suggesting that the characteristics of our weighted sample approximate the characteristics of the drug user network. The weighted estimates for market purchasers are generally valid for inferences about the aggregate population of customers, but a larger sample size is needed to make stronger inferences about the effects of a DMI on drug market activity.

Keywords

content area crime and justice design and evaluation of programs and policies methodology outcome evaluation survey methodology

Introduction

Violent crime continues to decline in the United States. From 1993 to 2012, the violent crime rate and homicide rate both dropped by roughly 50% (Federal Bureau of Investigation, 2013). There are a number of explanations for the decrease, including the maturation and shrinkage of crack cocaine markets (Blumstein & Rosenfeld, 1998; Levitt, 2004). Although violent crack markets are not as prominent as they once were (Fryer, Heaton, Levitt, & Murphy, 2013), overt markets for crack and other drugs still exist and can create significant harms for participants and those who live near them (Baumer, Lauritsen, Rosenfeld, & Wright, 1998; Blumstein & Rosenfeld, 1998; Weisburd & Mazerolle, 2000).

The drug market intervention (DMI), originally implemented in High Point, North Carolina, is an innovative strategy that uses focused deterrence, community engagement, and incapacitation to shut down overt drug markets and reduce the crime associated with them (Kennedy & Wong, 2009). Although a number of studies evaluating the DMI and related interventions have identified positive effects (Corsaro, Brunson, Gau, & Oldham, 2011; Corsaro, Brunson, & McGarrell, 2013; Corsaro & McGarrell, 2009; Frabutt, Shelton, Di Luca, Harvey, & Kefner, 2009; Hipple Kroovand, Corsaro, & McGarrell, 2010; Kennedy & Wong, 2009; McGarrell, Corsaro, & Brunson, 2010), one perspective is prominently missing from these evaluations: those who purchase drugs in these markets. Although some DMI evaluations have conducted surveys, interviews, or focus groups with community members, we are unaware of any efforts that have specifically targeted drug purchasers in communities that have had a DMI, let alone attempted to approximate a representative sample.

There is a scholarly literature focused on the drug market insights provided by drug purchasers, with many of these studies based on the now-defunct arrestee drug abuse monitoring (ADAM) program (Johnson, Golub, & Dunlap, 2000; Office of National Drug Control Policy, 2012, 2014; Rhodes, Kling, & Johnston, 2007). At its peak, ADAM interviewed booked arrestees in roughly 40 counties and the instrument included detailed questions about drug use and market transactions.¹ Efforts were made to create weights that would allow for county-level inferences about adult male booked arrestees (Hunt & Rhodes, 2008; Taylor, Fitzgerald, Hunt, Reardon, & Brownstein, 2001). The National Survey on Drug Use and Health (NSDUH, annual) also includes a module exclusively focused on marijuana transactions that has yielded important insights about these markets (Caulkins & Pacula, 2006; Office of National Drug Control Policy, 2012, 2014). Unfortunately, representative information about market transactions among NSDUH respondents is unavailable at the subnational level.^2,3

This article contributes to the evaluation, crime control, and DMI literatures by exploring the use of respondent-driven sampling (RDS) to approximate a representative sample of drug users to evaluate the impact of DMIs on changes in drug market dynamics. RDS is a peer-driven chain referral method that employs poststratification weights which can, under certain conditions, generate asymptotically unbiased population estimates for hidden populations (Salganik & Heckathorn, 2004). This study is part of a multisite evaluation across multiple outcome domains, including crime, cost, and public perceptions.

The rest of the article is structured as follows: The second section presents background information about DMI and how RDS is used to learn about hidden populations. The third section describes our methods for implementing RDS, collecting the data, and how the results are weighted. The fourth section presents the results, with the first part describing results of our tests of RDS assumptions and sample stability and the second part presenting our estimates of drug users’ perceptions of changes in the drug market after a DMI. Section 5 discusses our findings with regard to the sample, changes in the drug market as well as some of the limitations of the study design and offers ideas for future RDS work in this field.

Background

The DMI

DMI seeks to reduce the crime associated with overt drug markets (Kennedy, 2009). Enforcement powers are used strategically and sparingly, employing arrest and prosecution only against violent drug dealers and when nonviolent dealers have resisted all efforts to get them to desist and to provide them with help. Through the use of “banked” cases, the strategy makes the promise of law enforcement sanctions extremely direct and credible, focusing the deterrence message on a small group of dealers. The strategy also brings informal social control to bear on dealers from immediate family and community figures and focuses services and support on dealers so that those who are willing are provided with tools with which to change their lives. In the wake of an enforcement operation, the law enforcement–community coalition holds a “call-in” where partnering agencies communicate enforcement threats and make offers of social services to targeted drug offenders. Each operation also includes a maintenance phase, which includes community involvement and increased police responsiveness and enforcement, to prevent the overt drug market from returning.

The Roanoke DMI followed the Bureau of Justice Assistance DMI model,⁴ which consists of four phases: (1) targeting the drug market, (2) working with the community, (3) preparing for the call-in, and (4) maintenance. Roanoke Police analyzed official data to identify drug market areas characterized by high levels of violence and selected one, within a neighborhood called Hurt Park, to be the DMI target area. Within the target area, narcotics officers reviewed drug incident reports, conducted intensive surveillance, and collaborated with probation and parole officers to identify key members of the drug selling network. Drugs were purchased from targeted drug sellers and covertly videotaped. The police identified 15 dealers operating in the market. Ten were arrested immediately and prosecuted; five lower level dealers were offered an alternative to incarceration. These five individuals were invited, along with their families and residents of the neighborhood, to a call-in meeting in November 2012, where they were required to denounce drug dealing and invited participate in social services. In parallel, the DMI team worked to mobilize and provide resources to the Hurt Park community. After the call-in, the Roanoke Police increased their presence in Hurt Park, and the community held several events to boost solidarity in an effort to prevent overt drug dealing from returning.

RDS

RDS is a peer-driven, chain referral statistical sampling method that can efficiently yield large samples of illicit drug users (Abdul-Quader et al., 2006; Draus, Siegal, Carlson, Falck, & Wang, 2005; Frost et al., 2006; Iguchi, Ober, Berry, Fain, & Heckathorn, 2009; McKnight et al., 2006; Wang, Falck, Li, Rahman, & Carlson, 2007). RDS was initially developed by Douglas Heckathorn (Heckathorn, 1997, 2002, 2007) to address biases in nonprobability sampling and barriers to reaching “hidden” or hard-to-reach populations. Hidden populations are those for whom no sampling frame exists (thus the population size and boundaries are unknown) and those for whom there are strong privacy concerns that may prevent certain individuals, such as illicit drug users or the homeless, from participating in research.

Several articles by Heckathorn and colleagues (Heckathorn, 1997, 2002, 2011; Salganik & Heckathorn, 2004; Volz & Heckathorn, 2008) explain the theory underlying RDS and expansions of RDS since its inception. In these articles, Heckathorn et al. explain how RDS can reduce biases, such as oversampling from personal social networks (i.e., one’s personal friends and acquaintances who know one another by name) and the nonindependence of network members, typically associated with nonrandom recruitment methods like snowball and venue-based sampling. Drawing upon Markov chains and the theory of biased networks, RDS can theoretically produce a sample that is independent of its initial recruits. RDS also reduces biases that can result from voluntary participation (i.e., self-selection), masking (i.e., nonparticipation bias from individuals reluctant to be recruited by researchers), and differential personal network sizes (i.e., the number of people in one’s social network). RDS’s dual-incentive payment system (participants receive incentives for participating in the study and for recruiting others) results in a robust and efficient recruitment system in which a few individuals yield a large sample of individuals over the course of successive waves. Typically, after about six waves, the RDS sample is independent of the characteristics of initial recruits. If sample sizes are adequate and the tendency for participants to recruit others like themselves is modest (i.e., low “homophily,” a measure of network clustering, in-group affiliation bias, and network segmentation), RDS is thought to adequately address biases associated with nonindependence of network members (Salganik & Heckathorn, 2004). “Adequate” sample size for RDS is roughly the size of a simple random sample (SRS) multiplied by up to a factor of four (Johnston, Chen, Silva-Santisteban, & Raymond, 2012; Wejnert, Pham, Krishna, Le, & DiNenno, 2012).

However, while RDS’ structured recruitment procedures can reduce some biases, other biases, such as nonequivalent network sizes (network sizes are referred to in RDS as “degree”) and participants’ recruiting differentially across groups of people within their networks (such as age and racial/ethnic groups, referred to as “transition probabilities”), may need poststratification weights (Abdul-Quader et al., 2006; Heckathorn, 1997). RDS weights are designed to adjust for sampling probability by correcting for differences in respondent network size and transition probabilities (Heckathorn, 2007).

Finally, in order for RDS to generate asymptotically unbiased population estimates, five key assumptions must be satisfied (Goel & Salganik, 2010; Heckathorn, 2002, 2007). (Note that we list original assumptions of RDS as well as some allowable exceptions as they have been noted in the RDS literature.) The first three assumptions are the conditions under which RDS is thought to be an appropriate sampling method and the fourth and fifth are additionally needed for the RDS weights to produce unbiased estimates. These RDS assumptions are as follows: (1) Respondents know one another as members of the target population, so ties (linkages between two recruits) are reciprocal (i.e., they know the person by name and the person knows them by name and they have seen each other in the past 6 months; however, if only a small number of “strangers” are recruited and show no significant differences from other recruits, they may be included in the sample with no consequence; Volz, Wejnert, Degani, & Heckathorn, 2007). (2) Respondents are linked by a network composed of a single component (e.g., drug use; by recruiting a “linked” network of drug users, for example, asymptotically unbiased estimates may be generated for the larger group of drug users but also for smaller subsets, such as all users who have purchased from a drug market). (3) Sampling occurs with replacement (however, counterfactual, recent analyses have shown that bias from this assumption is insignificant if the sampling fraction is less than 20%, and bias is minimal if the sampling fraction is no more than 40%. Therefore, given that in U.S. urban applications of RDS a sampling fraction in excess of 10% is rare, this assumption is generally not a meaningful contributor to bias (Heckathorn, Cameron, & Yongren, 2013). (4) Respondents can accurately report their personal network size, defined as the number of relatives, friends, and acquaintances who fall within the target population who are known to the participant by name. Support for this assumption is provided by Wejnert’s (2009) study, in which he employed multiple alternative ways for respondents to estimate personal network size, and found strong convergence among the point estimates based on each alternative estimation technique. (5) Recruiters recruit as though they are choosing randomly from within their networks (networks are defined as one’s social connections; Heckathorn, 2007). Under the random recruitment assumption, recruitment is expected to be unbiased because recruiters do not have the incentive or ability to selectively, systematically recruit from a particular group. Notably, the research design determines the viability of this assumption. Participants must feel safe recruiting across groups other than their own (such as recruiting those from outside of their own neighborhoods or those of different ages, races, or ethnicities), must have adequate access (and transportation) to the research site, and should be asked to recruit those they have seen recently (as they would be more likely to recruit those seen recently than those seen long ago; Heckathorn, Salaam, Broadhead, & Hughes, 2002).

RDS recruitment begins with a small number of nonrandomly selected “seeds” who meet the study’s eligibility criteria. Seeds are the first members of a network who recruit others and thereby begin the chain-referral process. As noted previously, seeds are provided dual incentives—incentives for participating in the study and incentives for successful recruitment of people they know who qualify for the study. Recruitment then expands through multiple recruitment waves, with new recruits becoming recruiters and receiving compensation both for participating in the study and for successfully recruiting eligible peers. After multiple waves of recruitment, RDS theoretically reach “equilibrium,” a point at which sample characteristics cease to fluctuate from wave to wave. Under certain conditions, an RDS may be self-weighting. That is, individuals have equivalent network sizes and have recruited equally across groups, so no statistical adjustment to sampling probability is needed. However, given that these conditions frequently do not hold, in general the sample must be weighted using RDS weighting equations (Abdul-Quader et al., 2006; Heckathorn, 1997; Heckathorn et al., 2002).

RDS was originally developed as part of an AIDS prevention intervention in the early 1990s that targeted injection drug users (Broadhead & Heckathorn, 1994). It has since been used in dozens of studies worldwide to recruit and study networks of injecting and noninjecting drug users and men who have sex with men to explore HIV-related risk behaviors (Abdul-Quader et al., 2006; Des Jarlais et al., 2007; Draus et al., 2005; Frost et al., 2006; Iguchi et al., 2009; Kimani et al., 2014; Rhodes et al., 2012; Soliman et al., 2010; Volkmann et al., 2011) and drug use (Daniulaityte, Falck, & Carlson, 2011; Hathaway et al., 2010; Jonas, Young, Oser, Leukefeld, & Havens, 2012; Oteo Perez, Benschop, & Korf, 2012; Wang et al., 2005; Wang et al., 2007). Some recent studies suggest that RDS may have limitations, including unexpectedly large design effects (DEs; Goel & Salganik, 2010; Johnston et al., 2012; Wejnert et al., 2012), misleadingly narrow confidence intervals (Goel & Salganik, 2010), sample bias introduced by convenience sampling of seeds and preferential recruitment (Gile & Handcock, 2010), and bias introduced by the RDS estimators when the sampling fraction is too large (Gile & Handcock, 2010), with some biases leading to differences between RDS and those obtained through other methods (Oteo Perez et al., 2012). Thus, at this point in RDS’ development as a valid statistical sampling method, it must be acknowledged that, in some cases, where the RDS study is not designed or implemented correctly, or where sample sizes have not taken into account a potential DE of 3 or 4 times the size of an SRS, RDS alone may not be sufficient for estimating population proportions. Nevertheless, RDS is a practical and efficient method for recruiting hidden populations and for generating unbiased estimates when key assumptions are met and the sample is properly weighted. Further, although RDS may have limitations, it may be more practical and efficient than other methods for recruiting populations that are truly hidden (Kendall et al., 2008; Platt et al., 2006) and provides multiple strategies for overcoming biases common to other methods. Additionally, some of the limitations listed above draw on simulated samples with structures so extreme that no network sampling method could be successfully employed.

In this paper, we describe our use of RDS to recruit a sample of drug users who purchased from a drug market following a DMI to survey them about their perceptions of changes. We (1) describe our methodology for using RDS to recruit a sample of illicit drug users from within and around a DMI target area, (2) evaluate the degree to which the key RDS assumptions are met in the recruited sample, and (3) provide exploratory estimates of changes in drug use, perceptions of changes in drug purity and price, and perceived changes in drug market characteristics and activity following the implementation of DMI.

Method

Study Design

We used RDS to recruit a regional sample of drug users in Roanoke, VA, between May and July 2012 in order to survey drug users about drug market changes after a DMI. Participants in this study were drug users 18 years of age or older from throughout Roanoke, who reported the use of powder cocaine, crack cocaine, methamphetamine, heroin, or nonprescription use of opioid medications in the past month. Of these, we queried a subsample who had purchased drugs from the target drug market.

The regional sample included drug users from the nine main zip codes across the 43 square miles of the city. From the regional sample, we drew a subset of participants who had purchased drugs within the target area before and after the intervention to examine perceived changes in drug market activity. We chose to recruit a regional sample first, rather than trying specifically to recruit people who purchased in the target neighborhood, for two reasons. First, recruiting a sample of drug users based on whether they purchased drugs from a small drug market that had just undergone an intervention involving arrests (10 of the more chronic and violent drug dealers were arrested) likely would have raised concerns that the study was a sting operation and also could have made drug dealers suspicious of drug users who participated in the study. Second, it was not clear at the outset whether drug users who purchased in the target zone mostly were connected through a single linked network, although this is a key assumption of RDS. Recruiting a drug user network (rather than a drug market network) allowed us to cast a wider net to capture drug users from around the city who had purchased from the target area. Our approach of drawing a broader sample of users and then studying the subset made up of those who purchased from a target market is a standard approach in RDS research. Because RDS provides weights for each respondent, specific groups can be selected for analysis, such as those who purchased from the targeted drug market.

Recruitment began 4 months after the call-in component of the DMI was conducted. The call-in is the culmination of DMI, when lower level offenders, their families, and community members are invited to an open meeting during which documentation of offenders’ drug dealing is presented and the offenders are offered the chance to avoid arrest by agreeing to cease their drug dealing and to participate in social service programs. By the time of the call-in, the violent dealers typically already have been arrested. Due to timing issues—because we were evaluating and not conducting the intervention, we did not know exactly when it would take place—and to institutional review board concerns about participant safety if we recruited before or too soon after the intervention and related arrests, we were limited to a postintervention study that could begin no sooner than 3 months post-call-in.

Recruitment

We began recruitment by actively recruiting two seeds from the target neighborhood. (A third seed from a different neighborhood was added later in the study but did not produce any recruits.) The seeds participated in the interview and then were given three RDS “coupons” to give to three people they knew who also knew them (they knew each other’s names), who they had seen in the past 6 months, and who also were drug users over 18. The coupons listed a toll-free number participants could call to inquire about the study.

After participants distributed their coupons, they returned to the office to collect compensation for recruiting other eligible participants and completed the short follow-up interview. Participants were compensated US$25 for completing the main interview and US$10 for each up to three eligible participants they recruited.

Measures

Eligible participants completed a 1 hr audio-assisted computerized self-administered interview that asked about their drug user networks, drug use and drug purchasing behaviors, and drug market activity; during a second office visit, a subset of participants (those who handed out coupons to eligible participants and came back for compensation) also completed a 10-min follow-up interview that captured information about their recruiting behavior. The main interview asked questions about participant demographics, health and access to health care and drug treatment, perceptions of police and community interactions in Roanoke, worry about victimization (Williams, McShane, & Akers, 2000), drug and alcohol use ever and in the past 30 days; about the most recent and typical purchase locations, costs, amounts, purity, and purchase methods of each drug used in the past 30 days, and, for participants who indicated they purchased drugs in the drug market target area, a range of questions about the drug market before and after the DMI call-in.

Because we were unable to collect pre- and postintervention data, we asked participants about the market environment before and after the intervention. To improve recall and to delineate two clear time periods, we utilized a variation of a “bounded recall” method (Neter & Waksberg, 1964; Sudman, Finn, & Lannom, 1984) in which we grounded the participants in recollection of events in their lives during the period before the intervention and asked the drug market questions about that time compared with the market environment “now.” Bounded recall methods can help increase recall of specific time-related events during research interviews (Sudman et al., 1984). The drug market section asked questions that compared attributes of the drug market area now compared with November 2011 (before the call-in), and included questions about the purity and price of each drug purchased, likelihood of being arrested, level of effort to find a dealer (harder, easier, the same) as well as the reasons respondents thought some dealers were no longer dealing.

The final set of questions in the main interview asked about the size and characteristics (i.e., race/ethnicity, age, gender, employment status) of individual drug-user networks, which later were used to calculate RDS weights, and about the nature of the relationship between the participant and the recruiter (i.e., friend, acquaintance, stranger), which was used to check the “reciprocity” assumption. The secondary RDS interview contained questions about the number of coupons accepted and refused and the characteristics (race/ethnicity, age, gender, employment status) of those who accepted and refused. In this article, we describe the demographics of both the full drug user sample and of the subset who had purchased in the drug market, self-reported changes in drug use before and after the intervention, and perceptions of changes in drug purity and price and of the availability of drug dealers in the target area and reasons some dealers were no longer dealing.

Weighting

RDS weights are designed to adjust for sampling probability. They correct for differences in respondent network size (also referred to as degree) and transition probabilities across groups (i.e., the probability that a person will differentially recruit from groups with characteristics different from one’s own, such as race, age, crack users versus noncrack users, etc.) during recruitment (Heckathorn, 2007). The goal of the RDS method is the same as that of other probability sampling methods to make adjustments to raw estimates derived from the recruited sample that account for sampling bias (i.e., each member of the group having a different probability of being included in the sample) in order to make inferences about the population from which the sample was drawn. As with traditional sampling methods, RDS applies weights to sample data that are inversely proportional to the sampling probability of each unit. RDS derives these weights from social network data that are collected within the context of the sample, with a specific focus on the average degree (network size) of those in the sample, and the recruiting patterns across groups (transition probabilities) of survey participants.

A unique feature of RDS is that the estimated sampling probabilities and associated weights are themselves derived from variable-specific recruiting behavior (e.g., the probability that users of crack cocaine will recruit other users of crack cocaine versus the probability that they will recruit noncrack cocaine users). As a result, RDS estimation generates not one vector of sampling weights but several—one for each unique variable of analysis (or alternatively for each combination or simultaneously analyzed variables).

The theory under which these sampling probability weights are generated can be illustrated with a simple example. Consider a proportion of interest, such as the proportion of the drug-using population that uses crack cocaine. If we denote the degree (a measure of average network size) of this group to be D ₁ and the degree of noncrack users to be D ₀, then the total number of ties (linkages) between crack users and noncrack users can be expressed as N ₁ D ₁ S ₁₀, where N ₁ is the number of crack users in the sample and S ₁₀ is the transition—in this case, the probability that noncrack users will be recruited into the sample by crack users.

We can express this quantity as being equal to its mirror:

N_{1} D_{1} S_{10} = N_{0} D_{0} S_{01} .

This equality holds under the assumption of reciprocity—if every link in the social network is bilateral, the total number of ties from crack users to noncrack users must be equal to the total number of ties from noncrack users to crack users. Dividing both sides through by N (the number of crack users and noncrack users), we can reexpress (1) in terms of sample proportions:

P_{1} D_{1} S_{10} = P_{0} D_{0} S_{01} .

With substitution of the equality that P ₁ and P ₀ must sum to 1, (2) can be solved to yield:

P_{1} = \frac{D_{0} S_{01}}{(D_{1} S_{10} + D_{0} S_{01})} .

An equivalent expression for P ₀ can be similarly derived; the sampling weights are generated by combining these estimates with the raw sample proportions ( C _i ), as in:

W_{1} = P_{1} / C_{1} .

This method can be extended to multiclass categorical and continuous variables.

For each estimate presented in this article, we used RDSAT™ (Version 7.1) software (Volz et al., 2012) to generate weights and then compared the weighted estimates with the unweighted estimates. To weight the estimates from the DMI target-area subsample, we created an indicator variable equal to 1 if the participant reported having purchased drugs in the target area in the past 30 days and created weights based on simultaneous partitions of this variable and our different variables of interest. Given slight differences between the unweighted and weighted estimates, we report weighted estimates.

Post hoc Eligibility Verification

Because of the potential for participants to lie about having used certain drugs in order to gain entry into the study, we added a post hoc eligibility verification assessment as a final determination of eligibility for inclusion in the analysis. This consisted of two steps. First, we compared responses given by participants to the eligibility questions given during the in-person eligibility screening. Next we screened the responses to all of the recent drug-use questions for extreme and nonsensical outliers, a potential indication that the participants had not actually used the target drugs. We dropped three participants from the analysis due to inconsistencies between the drugs reported at the initial eligibility screening and those reported in the survey.

Results

RDS Assumptions and Sample Stability

The RDS method provides estimates of population proportions that are asymptotically unbiased provided that (a) the sample size is large enough, (b) homophily bias is modest, and (c) certain assumptions are met (Salganik, 2006; Salganik & Heckathorn, 2004; or allowable variations have been tested and explained in prior published studies). RDS assumptions are as follows: First, ties must be reciprocal—survey participants must know one another; second, respondents must be members of a single linked network; third, sampling occurs with replacement; fourth, respondents must be able to accurately report their network size; and fifth, peer recruitment occurs at random within each recruiter’s network (Heckathorn, 2002).

Sample size

Early RDS literature suggested that the size of an RDS should be roughly twice the size of what is needed for an SRS due to nonindependence of peer-recruited participants (Salganik, 2006; Salganik & Heckathorn, 2004); however, recent studies of RDS in both real-world and synthetic samples have found even larger DEs than initially recommended and suggest that multiplying the SRS estimate by a factor of 3 or 4 would be more appropriate (Johnston et al., 2012; Wejnert et al., 2012).⁵ Our overall sample size (N = 212 illicit drug users across Roanoke) is roughly double that of what would be needed to make precise estimates in an SRS. To estimate the sample size needed for an SRS, we used existing estimates of drug users provided in the National Survey on Drug Use and Health (NSDUH) (Substance Abuse and Mental Health Services Administration, 2014). However, these are thought to be underestimates, as many hard drug users misreport their use or do not take part in household surveys. Further, because we multiplied by a DE of two, possibly too low for RDS according to recent literature, confidence intervals around our some of our estimates are wide (see Tables 1 –3). In RDS, DE varies based on the variable being analyzed. DE is highest when the variable reflects strong homophily. In these cases, because respondents tend to recruit others like themselves, the interdependence or observation is high. This is often the case for race/ethnicity, and socioeconomic status. In contrast, when homophily is low, so too are DEs, and when homophily is zero, DE falls to one, the same as for an SRS. A DE of three to four in RDS reflects a worst-case scenario (i.e., very high homophily). Thus, for some variables in the present study, DEs were higher than anticipated (and the sample sizes likely too small) to allow for strong inferences about the population.

Table 1.

Sample Demographics and Drug Use (RDS Weighted).

Demographic	Purchased in Target Zone	Did Not Purchase in Target Zone
	n = 74	n = 138
	% [95% CI^a]	% [95% CI^a]
Percentage male	60 [44%, 77%]	64 [53%, 75%]
Race/ethnicity
Black	55 [42%, 70%]	30 [19%, 43%]
White	27 [15%, 40%]	57 [43%, 71%]
Hispanic/Latino	9 [3%, 17%]	7 [1%, 15%]
Education
No formal school or some high school	41 [27%, 58%]	26 [17%, 38%]
Graduated from high school	37 [21%, 49%]	40 [27%, 49%]
Some college	22 [13%, 35%]	27 [18%, 36%]
Employment status
Disabled, not able to work	39 [25%, 53%]	19 [11%, 30%]
Unemployed, receiving unemployment	13 [4%, 25%]	6 [1%, 13%]
Unemployed, not receiving aid	34 [18%, 48%]	31 [20%, 46%]
Working part-time (less than 35 hr/week)	10 [3%, 17%]	27 [14%, 40%]
Working full-time (35 or more hr/week)	4 [0%, 12%]	17 [9%, 25%]
Legal income, past month	US$728 (US$356, US$1,100)^b	US$730 (US$537, US$922)^b
Illegal income, past month	US$134 (US$78, US$189)^b	US$70 (US$36, US$104)^b
Live in target area	29 [17%, 44%]	16 [7%, 26%]
Drug use, past 30 days
Crack cocaine	86 [75%, 96%]	70 [59%, 81%]
Powder cocaine	81 [63%, 92%]	64 [53%, 75%]
Heroin	26 [14%, 40%]	14 [5%, 23%]
Oxycodone (OxyContin, Percocet, or Loritab)	55 [38%, 70%]	44 [33%, 55%]
Methamphetamine	27 [12%, 42%]	15 [8%, 21%]
Marijuana	91 [82%, 97%]	56 [45%, 67%]

Note. CI = confidence interval; RDS = respondent-driven sampling.

^aFor proportions, CIs were obtained from RDSAT™ Version 7.1 software and reflect increased estimator variance due to homophily in the recruiting process. RDSAT generates CIs using a bootstrap resampling procedure. Specifically, RDSAT derives the upper and lower bounds of the 95% CI from the top and bottom 2.5% tails of the distribution of bootstrap replications of the estimate. We specified 20,000 replications for this analysis. ^bRDSAT does not compute estimates of the means of continuous variables. The estimates of the continuous (income) variables reported here were obtained by exporting the vector of individualized weights from RDSAT to Stata and estimated there. As such, the CIs for the continuous variables do not correctly reflect increased estimator variance due to recruiting homophily; the associated standard errors are biased downward, and the true CIs are wider than those reported here.

Table 2.

Self-Reported Changes in Drug Use and Perceived Changes in Purity and Price Since DMI (RDS Weighted).

Drug	Drug Use Since Intervention			Drug Purity Since Intervention			Drug Price Since Intervention
	% [95% CI^a]			% [95% CI^a]			% [95% CI^a]
	More	Less	Same	Higher	Lower	Same	Higher	Lower	Same
Crack n = 63/42/46^b	25 [15%, 40%]	41 [26%, 55%]	34 [18%, 50%]	16 [4%, 30%]	39 [26%, 56%]	45 [26%, 62%]	21 [7%, 38%]	10 [1%, 21%]	70 [51%, 85%]
Powder cocaine n = 38/24/26^b	36 [21%, 60%]	54 [29%, 71%]	10 [3%, 19%]	26 [6%, 39%]	67 [48%, 81%]	5 [1%, 30%]	34 [7%, 62%]	21 [5%, 47%]	46 [19%, 71%]
Marijuana n = 34/31/30^b	19 [8%, 37%]	40 [21%, 66%]	41 [19%, 54%]	25 [11%, 48%]	35 [15%, 62%]	39 [12%, 60%]	49 [17%, 75%]	20 [1%, 47%]	31 [10%, 61%]

Note. CI = confidence interval; DMI = drug market intervention; RDS = respondent-driven sampling.

^a CIs were obtained from RDSAT™ Version 7.1 software and reflect increased estimator variance due to homophily in the recruiting process. RDSAT generates CIs using a bootstrap resampling procedure. Specifically, RDSAT derives the upper and lower bounds of the 95% CI from the top and bottom 2.5% tails of the distribution of bootstrap replications of the estimate. We specified 20,000 replications for this analysis. ^b ns vary for each variable (drug use, drug purity, and drug price) due to the varying missingness within individual. For example, some individuals were able to recall changes in use but not perceived price or purity, while others were able to recall all the three.

Table 3.

Changes in Drug Market Activity Since DMI (RDS Weighted).

Outcome	Crack Cocaine	Powder Cocaine	Marijuana
Outcome	% [95% CI]^a	% [95% CI]^a	% [95% CI]^a
Level of effort to find a drug dealer in target area since DMI	n = 44^b	n = 26^b	n = 27^b
Easier	15 [4%, 29%]	32 [7%, 54%]	4 [0%, 15%]
Harder	35 [21%, 55%]	62 [40%, 88%]	56 [34%, 85%]
Same	50 [28%, 65%]	6 [1%, 13%]	41 [13%, 61%]
Level of effort to find a drug dealer outdoors since DMI	n = 42^b	n = 27^b	n = 28^b
Easier	10 [2%, 18%]	40 [15%, 61%]	25 [2%, 44%]
Harder	42 [24%, 59%]	47 [26%, 72%]	47 [26%, 76%]
Same	49 [32%, 69%]	14 [3%, 29%]	28 [12%, 48%]
Presence of drug dealing selling indoors since DMI	n = 39^b	n = 25^b	n = 26^b
Fewer	24 [10%, 45%]	49 [21%, 75%]	40 [18%, 69%]
More	33 [18%, 52%]	30 [11%, 61%]	27 [0%, 51%]
Same	43 [20%, 61%]	22 [0%, 45%]	33 [8%, 63%]
	n = 49^b	n = 29^b	n = 31^b
Noticed outdoor drug dealers decreased since DMI	52 [36%, 71%]	63 [37%, 80%]	82 [65%, 93%]
Of those who noticed decrease in dealers since DMI	n = 18^b	n = 13^b	n = 17^b
Reason: Dealers moved.	7 [0%, 18%]	14 [0%, 40%]	21 [0%, 54%]
Reason: Dealers got arrested.	85 [66%, 100%]	77 [46%, 100%]	72 [37%, 96%]
Reason: Dealers stopped dealing.	8 [0%, 26%]	0	7 [0%, 18%]
Noticed: Old dealers were replaced with new dealers.	73 [53%, 91%]	62 [32%, 97%]	70 [48%, 97%]
Noticed: Remaining dealers get busier.	84 [67%, 97%]	90 [68%, 100%]	68 [42%, 96%]

Note. CI = confidence interval; DMI = drug market intervention; RDS = respondent-driven sampling.

^aCIs were obtained from RDSAT™ Version 7.1 software and reflect increased estimator variance due to homophily in the recruiting process. RDSAT generates CIs using a bootstrap resampling procedure. Specifically, RDSAT derives the upper and lower bounds of the 95% CI from the top and bottom 2.5% tails of the distribution of bootstrap replications of the estimate. We specified 20,000 replications for this analysis. ^b ns vary for each variable (drug use, drug purity, and drug price) due to varying missingness within individual. For example, some individuals were able to recall changes in use but not perceived price or purity, while others were able to recall all the three.

Homophily bias

Homophily bias for variables for which we might expect strong homophily (i.e., race/ethnicity and age) is modest, ranging from 0.148 (the tendency of Hispanic participants to recruit others who were Hispanic) to 0.413 (the tendency of participants 25 and older to recruit others of that age), suggesting that there was little tendency for recruits to recruit others like themselves. (When homophily is large, especially when it exceeds 0.7, then RDS loses efficiency.)

Assumption 1: Reciprocal ties

Respondents must know each other as members of the target population (Heckathorn, 2007). We follow the example of previous studies and assess this assumption by examining the proportion of respondents reporting that their recruiter was a stranger (Iguchi et al., 2009; Ramirez-Valles, Heckathorn, Vázquez, Diaz, & Campbell, 2005).

Recruiters recruited primarily those with whom they had reciprocal ties. The proportion of respondents identifying as stranger recruited in our study was 4.3%. Prior work suggests that stranger-recruited observations may be retained if they are not found to differ significantly from other recruits (Volz et al., 2007). To examine whether inclusion of these stranger-recruited observations made a difference in our estimates, we generated RDS population estimates both for demographic characteristics (age, race, gender) and for indicators for powder cocaine and crack usage in the past 30 days under two alternate scenarios: first with stranger-recruited participants retained and second with stranger-recruited participants dropped. (Note that when dropping participants from an RDS, the sample must be “re-seeded” to account for recruiters being dropped. Re-seeding refers to the process of treating recruits as seeds if their own recruiter was dropped.) We found that dropping the participants who reported they were recruited by a stranger did not alter estimates of any of these variables in a statistically significant way, thus we retained the stranger-recruited observations.

Assumption 2: Single linked network

Under this assumption, ties between network members must be dense enough to sustain the chain-referral process (Heckathorn, 2007). Drug users are a hidden population; as such, whether the subpopulation in any given city consists of a single network or of multiple disjointed networks (i.e., several separate social networks that are not linked to one another) is not provable. However, one may posit that if disjointed drug user networks did exist, they might partition either geographically (e.g., users on the east side form one network; users on the west side form another) or by drug use (heroin users are not linked to meth users). We tested these hypotheses by examining the affiliation matrices in our sample for evidence of cross-group recruitment by both ZIP code and drug choice. Affiliation matrices are derived from recruitment data and indicate the probability of recruiting within or outside the group, within the variables of interest.

We found that recruitment behavior by geographic residence and by drug choice supports the hypothesis of a single linked network of drug users in Roanoke. With respect to ZIP code, respondents showed some homophily—in general, they were more likely to recruit residents of their own ZIP code. However, recruitment chains eventually extended away from the ZIP codes of the original seeds and into neighboring ZIP codes, consistent with the hypothesis of a single linked network. This is illustrated in Figure 1, which displays the recruitment chains by the ZIP code of the respondent. The three large squares represent the three seeds that began the recruitment chains and the smaller squares represent recruits. The different colors represent the zip codes reported by each recruit, as shown in the legend. Of note, this assumption is about the population of drug users and not the sample itself. For example, if participant B, recruited by participant A, is a drug user and purchases from the target drug market, and participant C, recruited by participant B, is a drug user but purchases from a different drug market than B and does not personally know participant A, participants A and C are still part of a single linked network of drug users, even if they were recruited by different drug users and purchased drugs from different markets.

Figure 1.

Respondent-driven sampling (RDS) by respondent ZIP code.

Our analysis of cross-group recruitment by drug choice revealed first that a majority of drug users in Roanoke (79.7%) were poly-drug users, meaning that they reported the use of more than one of the eligibility drugs in the previous 30 days, and, second, that recruitment across classes of drug users was evident in the sample. Following partition into six groups—one for each of the five eligibility drugs and one poly-drug class, the only group in the Roanoke sample to show homophily was the poly-class. These results are also consistent with the notion that users of different drugs are joined in a single linked network.

Assumption 3: Sampling with replacement

Sampling is assumed to occur with replacement so the set of respondents available for recruitment is not depleted (Heckathorn, 2007). However, for practical reasons, RDS studies typically are administered without replacement so individuals do not participate in the research interview more than one time. Sampling without replacement is appropriate when the sampling fraction is small (i.e., smaller than 20% but even up to 40%; Heckathorn, 2007); when recruited samples are a small proportion of the overall target population, bias resulting from the violation is negligible (Salganik & Heckathorn, 2004). To estimate the sampling fraction without knowing the size of the population of drug users in Roanoke, we examined data estimating the percentage of the population of the state of Virginia to have used illicit drugs other than marijuana as 6.46% for the population aged 18–25 and 1.86% for the population over 25 and applied these percentages to the population of Roanoke (n = 95,793; USDHHS, 2014).⁶

Based on these data, we estimate the total number of adult illicit drug users in Roanoke to be at least 1,841; thus our overall sample (N = 212) represents at most 11.5% of the target population. Although there are no precise guidelines defining what constitutes a small sampling fraction, prior research has concluded that bias from this assumption is insignificant if the sampling fraction exceeds 20%, and bias is minimal if the sampling fraction is no more than 40% (Wejnert & Heckathorn, 2008). (Of note, NSDUH data are thought to underestimate the population of illicit drug users, however, for the purpose of estimating the sampling fraction, using the lower estimate provides us with a larger sampling fraction. Thus, doubling or even tripling the estimate of the number of illicit drug users would further reduce the sampling fraction.)

Assumption 4: Accurate reporting of network size

The fourth RDS assumption is that respondents can accurately report the number of their own relatives, friends, and acquaintances who are members of the target population (Heckathorn, 2007). This assumption may be partially addressed by comparing reported network size to actual recruiting behavior, with the expectation that reported network size should not be exceeded by either the total number of recruits (capped at three) or the total number of those offered coupons (as reported in the follow-up interview).

Sixteen of 60 participants who returned for their follow-up interview reported offering coupons to a number of drug-using peers who exceeded the number they originally reported in the main interview. For example, one respondent indicated a network of six drug-using peers in the main interview but later indicated that nine peers had refused coupons while three had accepted, suggesting an actual network size of 12. These results could suggest inaccurate reporting of network size; however, there are alternative explanations for this inconsistency. First, peer networks are not necessarily stable over time, and it is possible that the networks of some respondents expanded during the interval between their participation in the main and follow-up interviews. Another possible explanation is that respondents accurately reported their network size but also attempted to recruit strangers into the study. Although some degree of concern exists for the ability to accurately report one’s own network size and has been previously noted by others (Frost et al., 2006; Kochen, 1989), there is also evidence that self-reported network size is an accurate network indicator (Marsden, 1990). Our data do not strongly suggest inaccuracies in drug users’ reporting of network size.

Assumption 5: Random recruitment within network

The random recruitment assumption posits that individuals recruit as though they are choosing randomly from their personal networks (Heckathorn, 2007). The underlying assumption is that individuals lack an incentive or ability to intentionally and selectively recruit any particular group (Heckathorn, 2007). To assess whether this assumption was met, we used the estimation method described by Heckathorn (Heckathorn, 2007) and created estimators of population proportions in which self-reported network proportions were substituted for the RDS transition probabilities. Consider Equation 3 (p. 12):

Equation 3 expresses the estimated proportion of a subset of a binary group, such as males (P_0), as a function of the recruitment transition probabilities of that group (S_10 and S_01)—the probability of recruitment of males by females and of females by males. This is how the default RDS estimate is calculated. To test the assumption of random recruitment, we (1) obtained estimates of network composition ratios for observable attributes (such as the percentage of drug-using networks that was male), and (2) generated new estimates with those ratios that essentially substituted for the actual recruitment transition probabilities. These alternative estimators theoretically control for differential recruitment because they assume random recruitment according to the proportions defined by self-described networks. By comparing the confidence intervals for this new class of estimators with those of the original RDS estimators, we examined whether differential recruitment (resulting either from nonrandom solicitation by recruiters or from nonrandom refusal by network members) biased the RDS population estimates. We compared these adjusted population estimates with the ordinary RDS estimates for two attributes: race/ethnicity and gender.

The shift from RDS estimates to estimates derived from self-reported network composition did not produce a statistically significant shift in the proportion of drug users that were male or White (see Figure 2). Note that the reliability of the comparison is directly related to the ease with which the network attribute can be observed by the individual respondent. In this sense, the comparison for gender is probably most reliable, followed by White versus non-White.

Figure 2.

Analysis of differential recruitment bias.

Sample stability: Equilibrium

Sample equilibrium in RDS is the point in the recruitment process at which the sample stabilizes, that is, the sample becomes independent of the choice of initial seeds and theoretically reflects the composition of the population being studied. To assess equilibrium for the recruited sample, we examined equilibrium to assess whether there were an adequate number of waves in the study to allow the sample to reach equilibrium for race/ethnicity. (Waves are the number of levels in a recruitment chain that follow from an initial seed.)

Our assessment of equilibrium for race/ethnicity suggests that the full sample had enough waves to reach equilibrium on this characteristic and that the sample stabilized (see Figure 3). Sample stability suggests that the number of waves was sufficient to erase the effects of the convenience sample of seeds and yield a recruitment matrix that provided a valid measure of the cross-ties among groups; these cross-ties define an important feature of the population’s network structure (Heckathorn, 2007).

Figure 3.

Full sample (N = 212) equilibrium for race/ethnicity.

Exploratory Outcomes

Purchaser characteristics

The sample that purchased drugs (crack, powder cocaine, methamphetamine, heroin, or nonprescription oxycodone medications) in the Hurt Park target area was 60% male and 55% African American, 27% White, 9% Hispanic (see Table 1). Seventy-three percent of the target-area sample reported they were disabled and unable to work or unemployed and not receiving government assistance. Monthly legal income was similar between the two groups (US$728 for target-area purchasers and US$730 for nontarget-area purchasers). Illegal income was slightly higher for target-area purchasers. A higher proportion of drug users who purchased in the target area reported living in the target area (29%) compared with those who did not purchase in the target area (16%). Higher proportions of target-area purchasers reported having used crack cocaine (86%), powder cocaine (81%), heroin (26%), prescription opioids (55%), methamphetamine (27%), and marijuana (91%) in the past 30 days compared with those who did not purchase in the target area (70%, 64%, 14%, 44%, 15%, and 56%, respectively).

Purchased in the drug market in last 30 days

Ten percent of the full sample (N = 212) purchased crack cocaine in the target area in the past 30 days, 5% purchased powder cocaine, 8% purchased marijuana, and under 5% purchased methamphetamine (0%), oxycodone (3%), and heroin (2%). Due to the low number of participants who purchased methamphetamine, oxycodone, and heroin in the target area, we do not report on the reported effects of DMI on use, purity, price, or availability of dealers of those drugs.

Changes in drug use since DMI

Higher proportions of target-area purchasers reported using less crack cocaine and powder cocaine after DMI (41% and 54%, respectively) than those reported using more (25% and 36%, respectively) or the same amount (34% and 10%, respectively; see Table 2). Among marijuana users, 40% said they used less of the drug after DMI, 19% said they used more after, and 41% said the amount they used was the same before and after the intervention.

Perceived changes in drug purity and price since DMI

More than two thirds (67%) of the participants who purchased powder cocaine in the target area reported they thought the purity of powder cocaine was lower after DMI, while 26% said the purity was higher after, and 5% said the purity was the same before and after DMI (see Table 2). Among crack purchasers, 39% thought the purity was lower after, 16% thought it was higher, and 45% thought the purity was the same before and after.

In terms of changes in drug price after DMI, 21%, 34%, and 49% of target-area drug purchasers thought the prices of crack, powder cocaine, and marijuana, respectively, were higher after DMI, while 10%, 21%, and 20%, respectively, thought prices were lower (see Table 2).⁷ (We asked if they thought prices had gone up, down, or stayed the same.) The majority (70%) of crack purchasers thought that crack prices stayed the same, while 46% of powder cocaine purchasers thought the price of powder cocaine stayed the same and 31% thought marijuana prices stayed the same.

Perceived changes in drug market activity and drug availability since DMI

Higher proportions of target-area powder cocaine purchasers thought it was harder to find a cocaine dealer in general (62%) and harder to find a cocaine dealer outdoors (47%) in the target area after DMI than those who said it was easier to find a dealer at all and easier to find a dealer outdoors after DMI (32% and 40%, respectively); fewer said the level of effort was the same (6% and 14%, respectively; see Table 3). About one half of those who purchased crack in the target area thought that the level of effort to find a crack dealer at all and to find a crack dealer outdoors was the same before and after DMI (50% and 49%, respectively), 35% and 42% thought it was harder to find a crack dealer at all and in general after DMI, and only 15% and 10% thought it was easier to find a crack dealer after DMI than before. Fifty-six percent and 47% of marijuana purchasers thought it was harder to find a marijuana dealer at all and outdoors after DMI, while 41% and 28% thought the level of effort was the same. Fewer marijuana purchasers thought it was easier to find a marijuana dealer at all (4%) and outdoors (25%) after the intervention.

About one third of target-area purchasers thought there were more crack (33%), powder cocaine (30%), and marijuana (27%) dealers selling each drug indoors after DMI than before, while 24% thought there were fewer crack dealers selling indoors before DMI, 49% thought there were fewer powder cocaine dealers selling indoors before, and 40% thought there were fewer marijuana dealers selling indoors before the intervention.

Of those who purchased crack in the target area, about one half (52%) noticed the regular dealers had stopped dealing outdoors since DMI, while the majority of powder cocaine (63%) and marijuana (81%) purchasers said that regular dealers had stopped dealing outdoors. Of those who noticed the dealers had stopped dealing outdoors, the majority of crack purchasers (85%), powder cocaine purchasers (77%), and marijuana purchasers (72%) said the reason the dealers stopped purchasing was because they had been arrested. Very few said that dealers had moved to another neighborhood or that the dealers stopped dealing altogether. The majority of those who said that the usual dealers had stopped also said that the old dealers were replaced with new dealers (73%, 62%, and 70% of crack, powder cocaine, and marijuana purchasers, respectively) and that the dealers who did not stop dealing got busier (84%, 90%, and 68%, respectively).

Conclusions

The study is innovative in its use of RDS to recruit a sample of target-area purchasers to evaluate a criminal justice intervention and for its inclusion of drug purchasers in a DMI evaluation. We successfully implemented RDS to recruit a sample of drug users from in and around a drug market following DMI and we addressed three questions: (1) Is RDS a practical and effective approach for recruiting drug users from in and around a DMI target zone? (2) Can RDS be used to approximate a representative sample of illicit drug users from a DMI target zone? (3) What can we learn from drug users about the effect of DMI on drug use and drug purchasing behavior and on perceptions of availability and drug dealer behavior?

In response to the first question, consistent with prior RDS studies, we had few problems recruiting drug users to participate despite beginning recruitment after the large-scale police crackdown that preceded the DMI call-in. Our concerns about drug users being unwilling to participate due to worries about another undercover operation or being unwilling to share information about the drug market were unfounded. (Of note, some participants did initially express concern about the study being a police sting operation, but the worries apparently disappeared quickly and did not seem to affect recruitment.) Further, respondents did not report any negative consequences, such as harassment or violence from drug dealers, from their participation in the study. Finally, our recruitment was efficient—we recruited over 200 participants in 3 months.

With regard to the second question, we found that our sample showed little homophily bias (the indicator of recruitment across groups) and met or fell within an acceptable margin of error of RDS assumptions. By partitioning and weighting the sample of those who purchased drugs from the target area, we generated theoretically unbiased estimates of the perceptions of target-area purchasers about the changes in the drug market after the DMI. However, our wide confidence intervals suggest that a larger sample size is needed to make stronger inferences about the population of purchasers, thus our results must be considered exploratory. Further, because we are dealing with a hidden population, it is difficult to validate our findings about purchaser characteristics, especially since Roanoke was never an ADAM site. Substance use treatment admissions data for the Roanoke Metropolitan Service Area in 2012 show that “cocaine/crack” was the most popular primary substance of abuse after excluding alcohol and marijuana; “Other Opiates and Synthetics” were the second-most popular (USDHHS, 2014). While it is reassuring that the prevalence rates for cocaine, crack, and prescription opioids given in Table 1 match these rankings (after excluding marijuana), caution must be used when comparing the prevalence rates and treatment admissions. In addition to not being perfectly correlated since the probability of dependence differs by substance (Anthony, Warner, & Kessler, 1994), the Treatment Episode Data Set is nonunique (i.e., the same individual can account for multiple admissions), it only covers facilities that are funded with public dollars, and it includes a geographic area larger than the nine ZIP codes targeted by our RDS study.

The third research question is more difficult to answer because of the research design we ultimately adopted to provide the greatest amount of participant safety and because of our small sample size; thus, again, our insights should be considered exploratory. Because we were unable to conduct a preintervention survey, we can only report on the postintervention perceptions of target-area purchasers about changes in the market. The insights from drug users in the present study about the reduced availability of drug dealers are generally consistent with anecdotes and district-level crime data reported by the Roanoke Police Department.⁸ These insights are also consistent with our forthcoming findings of community focus group discussions. Three months after the DMI in Roanoke, community members thought that loitering, harassment, crime, and violence associated with dealing went down in the target area; perceptions of these changes were sustained 6 and 15 months after DMI (Saunders, Ober, Barnes-Proby, & Brunson, In Press).

Because crime and violence tend to be associated more with open-air markets than with indoor markets, the reduction in outdoor dealing reported by the drug purchasers may serve to lower crime and violence and increase safety in DMI target areas. Further, about one half of the crack purchasers and the majority of the powder cocaine and marijuana purchasers noticed that the regular dealers of these drugs had stopped dealing since the intervention. The top two reasons that these users said the regular dealers had stopped dealing was that they had been arrested or that they moved to a different neighborhood. This suggests that, at least in the short term, DMI could have had a noticeable effect on drug dealers in the target area.

Further research should consider ways to employ longitudinal designs that use RDS to recruit purchasers before and after the intervention in a way that minimizes concerns about safety to human subjects and should, if using RDS, aim to recruit a larger sample of drug users. Although safety concerns regarding the recruitment of drug users immediately prior to an intervention involving the arrest of drug dealers are serious and potentially consequential, they are not insurmountable. Other studies may not have the restriction of having to evaluate an intervention without knowing the implementation date. Knowledge of the implementation plan of the intervention under study would allow evaluators to conduct the preintervention study well in advance (perhaps up to a year) of the intervention, which likely would reduce any perceptions of research participants snitching on drug dealers or tipping off drug dealers about attention on the drug market. Finally, future RDS studies targeting drug purchasers should take into account underestimates of illicit drug users and RDS DEs potentially higher than two (possibly up to four) when calculating sample size.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: National Institute of Justice, Office of Justice Programs (2010-DJ-BX-1672).

Notes

References

Abdul-Quader

A. S.

Heckathorn

D. D.

McKnight

Bramson

Nemeth

Sabin

… Des Jarlais

D. C.

(2006). Effectiveness of respondent-driven sampling for recruiting drug users in New York City: Findings from a pilot study. Journal of Urban Health, 83, 459–476.

Anthony

Warner

Kessler

(1994). Comparative epidemiology of dependence on tobacco, alcohol, controlled substances, and inhalants: Basic findings from the National Comorbidity Survey. Experimental and Clinical Psychopharmacology, 2, 244–268.

Baumer

E. P.

Lauritsen

J. L.

Rosenfeld

Wright

(1998). The influence of crack cocaine on robbery, burglary, and homicide rates: A cross-city, longitudinal analysis. Journal of Research in Crime and Delinquency, 35, 316–340.

Blumstein

Rosenfeld

(1998). Explaining recent trends in U.S. homicide rates. The Journal of Criminal Law and Criminology, 88, 1175–1216.

Broadhead

Heckathorn

(1994). AIDS prevention outreach among injection drug users: Agency problems and new approaches. Social Problems, 41, 473–495.

Caulkins

J. P.

Pacula

R. L.

(2006). Marijuana markets: Inferences from reports by the household population. Journal of Drug Issues, 36, 173–200.

Corsaro

Brunson

R. K.

Gau

Oldham

(2011). The Peoria pulling levers drug market intervention: A review of program process, changes in perceptions, and crime impact. Washington, DC: U.S. Department of Justice.

Corsaro

Brunson

R. K.

McGarrell

E. F.

(2013). Problem-oriented policing and open-air drug markets: Examining the Rockford Pulling Levers deterrence strategy. Crime & Delinquency, 59, 1085–1107.

Corsaro

McGarrell

E. F.

(2009). An evaluation of the Nashville Drug Market Initiative (DMI) pulling levels strategy: Drug market intervention (Working Paper). School of Criminal Justice, Michigan State University, East Lansing, Michigan.

10.

Daniulaityte

Falck

Carlson

R. G.

(2011). Illicit use of buprenorphine in a community sample of young adult non-medical users of pharmaceutical opioids. Drug Alcohol Depend, 122, 201–207.

11.

Des Jarlais

D. C.

Arasteh

Perlis

Hagan

Heckathorn

D. D.

McKnight

… Friedman

S. R.

(2007). The transition from injection to non-injection drug use: Long-term outcomes among heroin and cocaine users in New York City. Addiction, 102, 778–785.

12.

Draus

P. J.

Siegal

H. A.

Carlson

R. G.

Falck

R. S.

Wang

(2005). Cracking the cornfields: Recruiting illicit stimulant drug users in rural Ohio. The Sociological Quarterly, 46, 165–189.

13.

Federal Bureau of Investigation. (2013). Crime in the United States 2012. Retrieved January 25, 2014, from http://www.fbi.gov/about-us/cjis/ucr/crime-in-the-u.s/2012/crime-in-the-u.s.-2012/tables/1tabledatadecoverviewpdf/table_1_crime_in_the_united_states_by_volume_and_rate_per_100000_inhabitants_1993-2012.xls

14.

Fendrich

Johnson

T. P.

Wislar

J. S.

Hubbell

Spiehler

(2004). The utility of drug testing in epidemiological research: Results from a general population survey. Addiction, 99, 197–208.

15.

Frabutt

J. M.

Shelton

T. L.

Di Luca

K. L.

Harvey

L. K.

Kefner

M. K.

(2009). A collaborative approach to eliminating street drug markets through focused deterrence (Final Report submitted to the National Institute of Justice (#2006-IJ-CX-0034)). Washington, DC: National Institute of Justice.

16.

Frost

S. D.

Brouwer

K. C.

Firestone Cruz

M. A.

Ramos

M. E.

Lozada

R. M.

… Strathdee

S. A.

(2006). Respondent-driven sampling of injection drug users in two U.S.-Mexico border cities: Recruitment dynamics and impact on estimates of HIV and syphilis prevalence. Journal of Urban Health, 83, i83–i97.

17.

Fryer

Heaton

Levitt

Murphy

(2013). Measuring crack cocaine and its impact. Economic Inquiry, 51, 1651–1681.

18.

Gile

K. J.

Handcock

M. S.

(2010). Respondent-driven sampling: An assessment of current methodology. Sociological Methodology, 40, 285–327.

19.

Goel

Salganik

M. J.

(2010). Assessing respondent-driven sampling. Proceedings of the National Academy of Science of the United States of America, 107, 6743–6747.

20.

Harrison

L. S.

Martin

S. S.

Enev

Harrington

(2007). Comparing drug testing and self-report of drug use among youths and young adults in the general population ((SMA)07-4249). Rockville, MD: Department of Health and Human Services, Substance Abuse and Mental Health Services Administration, Office of Applied Studies.

21.

Hathaway

A. D.

Hyshka

Erickson

P. G.

Asbridge

Brochu

Cousineau

M. M.

… Marsh

(2010). Whither RDS? An investigation of respondent driven sampling as a method of recruiting mainstream marijuana users. Harm Reduction Journal, 7, 15.

22.

Heckathorn

(1997). Respondent driven sampling: A new approach to the study of hidden samples. Social Problems, 44, 174–199.

23.

Heckathorn

(2002). Respondent driven sampling II: Deriving valid population estimates from chain-referral samples of hidden populations. Social Problems, 49, 11–34.

24.

Heckathorn

(2007). Extensions of respondent-driven sampling: Analyzing continuous variables and controlling for differential recruitment. Sociological Methodology, 37, 151–207.

25.

Heckathorn

D. D.

(2011). Snowball versus respondent-driven sampling. Sociological Methodology, 41, 355–366.

26.

Heckathorn

D. D.

Cameron

Yongren

(2013). Respondent-driven sampling: Origins and current developments. Paper presented at the 2013 International Methodology Symposium, Statistics, Ottawa, Ontario, Canada.

27.

Heckathorn

Salaam

Broadhead

Hughes

(2002). Extensions of respondent-driven sampling: A new approach to the study of injection drug users aged 18-25. Aids and Behavior, 6, 55–67.

28.

Hipple Kroovand

Corsaro

McGarrell

E. F.

(2010). The high point drug market initiative: A process and impact assessment. East Lansing: Michigan State University, School of Criminal Justice.

29.

Hunt

Rhodes

(2008). Arrestee drug abuse monitoring program II in the United States, 2008. In Inter-university Consortium for Political and Social Research (Ed.), Technical documentation report (pp. 10–21). Ann Arbor: University of Michigan.

30.

Iguchi

Ober

Berry

Fain

Heckathorn

(2009). Simultaneous recruitment of drug users and men who have sex with men in the United States and Russia using respondent driven sampling: Sampling methods and implications. Journal of Urban Health, 86, 5–31.

31.

Johnson

Golub

Dunlap

(2000). The rise and decline of hard drugs, drug markets, and violence in inner-city New York. In Blumstein

Wallman

(Eds.), The crime drop in America (pp. 164–206). New York, NY: Cambridge University Press.

32.

Johnston

L. G.

Chen

Y. H.

Silva-Santisteban

Raymond

H. F.

(2012). An empirical examination of respondent driven sampling design effects among HIV risk groups from studies conducted around the world. AIDS and Behavior, 17, 2202–2210.

33.

Jonas

A. B.

Young

A. M.

Oser

C. B.

Leukefeld

C. G.

Havens

J. R.

(2012). OxyContin(R) as currency: OxyContin(R) use and increased social capital among rural Appalachian drug users. Social Science & Medicine, 74, 1602–1609.

34.

Kendall

Kerr

L. R.

Gondim

R. C.

Werneck

G. L.

Macena

R. H.

Pontes

M. K.

… McFarland

(2008). An empirical comparison of respondent-driven sampling, time location sampling, and snowball sampling for behavioral surveillance in men who have sex with men, Fortaleza, Brazil. AIDS and Behavior, 12, S97–S104.

35.

Kennedy

D. M.

(2009). Deterrence and crime prevention: Reconsidering the prospect of sanction. New York, NY: Routledge.

36.

Kennedy

D. M.

Wong

S. L.

(2009). The high point drug intervention strategy. New York, NY: Center for Crime Prevention and Control John Jay College of Criminal Justice, National Network for Safe Communities.

37.

Kilmer

Caulkins

J. P.

Midgette

Dahlkemper

MacCoun

R. J.

Pacula

R. L.

(2013). Before the grand opening: Measuring Washington state’s Marijuana market in the last year before legalized commercial sales. Santa Monica, CA: RAND.

38.

Kimani

S. M.

Watt

M. H.

Merli

M. G.

Skinner

Myers

Pieterse

… Meade

C. S.

(2014). Respondent driven sampling is an effective method for engaging methamphetamine users in HIV prevention research in South Africa. Drug Alcohol Depend, 143, 134–140.

39.

Kochen

(1989). The small world. Norwood, NJ: Ablex.

40.

Levitt

S. D.

(2004). Understanding why crime fell in the 1990s: Four factors that explain the decline and six that do not. The Journal of Economic Perspectives, 18, 163–190.

41.

Marsden

P. V.

(1990). Network data and measurement. Annual Review of Sociology, 16, 435–463.

42.

McGarrell

E. F.

Corsaro

Brunson

R. K.

(2010). The drug market intervention approach to overt drug markets. Journal of Criminal Justice and Security, 12, 397–407.

43.

McKnight

Des Jarlais

Bramson

Tower

Abdul-Quader

A. S.

Nemeth

Heckathorn

(2006). Respondent-driven sampling in a study of drug users in New York City: Notes from the field. Journal of Urban Health, 83, i54–i59.

44.

Neter

Waksberg

(1964). A study of response errors in expenditures data from household interviews. Journal of the American Statistical Association, 59, 18–55.

45.

Rhodes

Dyous

Hunt

Luallen

Callahan

Subramanian

(2012). What America’s Users Spend on Illegal Drugs, 2000-2006. Washington, DC: Office of National Drug Control Policy, Executive Office of the President.

46.

Kilmer

Everingham

Caulkins

Midgette

Pacula

Reuter

… Lundberg

(2014). What America’s users spend on illegal drugs, 2000-2010. Washington, DC: Office of National Drug Control Policy, Executive Office of the President.

47.

Oteo Perez

Benschop

Korf

D. J.

(2012). Differential profiles of crack users in respondent-driven and institutional samples: A three-site comparison. European Addiction Research, 18, 184–192.

48.

Platt

Wall

Rhodes

Judd

Hickman

Johnston

L. G.

… Sarang

(2006). Methods to recruit hard-to-reach groups: Comparing two chain referral sampling methods of recruiting injecting drug users across nine studies in Russia and Estonia. Journal of Urban Health, 83, i39–i53.

49.

Ramirez-Valles

Heckathorn

D. D.

Vázquez

Diaz

R. M.

Campbell

R. T.

(2005). From networks to populations: The development and application of respondent-driven sampling among IDUs and Latino gay men. Aids and Behavior, 9, 387–402.

50.

Rhodes

S. D.

McCoy

T. P.

Hergenrather

K. C.

Vissman

A. T.

Wolfson

Alonzo

… Eng

. (2012). Prevalence Estimates of Health Risk Behaviors of Immigrant Latino Men Who Have Sex With Men. The Journal of Rural Health, 28(1), 73–83. http://doi.org/10.1111/j.1748-0361.2011.00373.x

51.

Rhodes

Kling

Johnston

(2007). Using booking data to model drug user arrest rates: A preliminary to estimating the prevalence of chronic drug use. Journal of Quantitative Criminology, 23, 1–22.

52.

Salganik

M. J.

(2006). Variance estimation, design effects, and sample size calculations for respondent-driven sampling. Journal of Urban Health, 83, i98–i112.

53.

Salganik

M. J.

Heckathorn

D. D.

(2004). Sampling and estimation in hidden populations using respondent-driven sampling. Sociological Methodology, 34, 193–239.

54.

Saunders

Ober

Barnes-Proby

Brunson

R. K

. (In Press). Police Legitimacy and Disrupting Overt Drug Markets. in Policing: an International Journal of Police Strategies & Management.

55.

Soliman

Rahman

I. A.

Shawky

Bahaa

Elkamhawi

Sattar

… Sayed

(2010). HIV prevalence and risk behaviors of male injection drug users in Cairo, Egypt. AIDS, 24, S33–S38.

56.

Substance Abuse and Mental Health Services Administration. (2014). Results from the 2013 National Survey on Drug Use and Health: Summary of National Findings, NSDUH Series H-41, HHS Publication No.(SMA) 11-4658. In: Administration SAaMHS, ed. Rockville, MD: Substance Abuse and Mental Health Services Administration.

57.

Sudman

Finn

Lannom

(1984). The use of bounded recall procedures in single interviews. Public Opinion Quarterly, 48, 520–524.

58.

Taylor

B. G.

Fitzgerald

Hunt

Reardon

J. A.

Brownstein

H. H.

(2001). ADAM preliminary 2000 findings on drug use and drug markets—Adult male arrestees. Washington, DC: National Institute of Justice. Retrieved from https://www.ncjrs.gov/txtfiles1/nij/189101.txt

59.

Volkmann

Lozada

Anderson

C. M.

Patterson

T. L.

Vera

Strathdee

S. A.

(2011). Factors associated with drug-related harms related to policing in Tijuana, Mexico. Harm Reduction Journal, 8, 7.

60.

Volz

Heckathorn

(2008). Probability based estimation theory for respondent driven sampling. Journal of Official Statistics, 24, 79–97.

61.

Volz

Wejnert

Degani

Heckathorn

D. D.

(2007). Respondent-driven sampling analysis tool (RDSAT) version 5.6. Ithaca, NY: Cornell University.

62.

Wang

Carlson

Falck

Siegal

Rahmen

(2005). Respondent-driven sampling to recruit MDMA users: A methodological assessment. Drug Alcohol Depend, 78, 147–157.

63.

Wang

Falck

R. S.

Rahman

Carlson

R. G.

(2007). Respondent-driven sampling in the recruitment of illicit stimulant drug users in a rural setting: Findings and technical issues. Addictive Behaviors, 32, 924–937.

64.

Weisburd

Mazerolle

(2000). Crime and disorder in drug hot spots: Implications for theory and practice in policing. Police Quarterly, 3, 331–339.

65.

Wejnert

(2009). An empirical test of respondent-driven sampling: Point estimates, variance, degree measures, and out-of-equilibrium data. Sociological Methodology, 39, 73–116.

66.

Wejnert

Heckathorn

D. D.

(2008). Web-based network sampling efficiency and efficacy of respondent-driven sampling for online research. Sociological Methods & Research, 37, 105–134.

67.

Wejnert

Pham

Krishna

DiNenno

(2012). Estimating design effect and calculating sample size for respondent-driven sampling studies of injection drug users in the United States. AIDS and Behavior, 16, 797–806.

68.

Williams

F. P.

McShane

M. D.

Akers

R. L.

(2000). Worry about victimization: An alternative and reliable measure for fear of crime. Western Criminology Review 2. Retrieved from http://www.westerncriminology.org/documents/WCR/v02n2/williams/williams.html