The study setting and survey instrument have been described previously.^51,52 Briefly, a cross-sectional survey of injection drug users in Winnipeg, Manitoba (population: 675 000) was conducted from December 2003 to September 2004. Potential participants were recruited through advertisements placed at local community health centers and meeting places (the latter as identified by key informants) and via word of mouth. To be eligible, individuals had to report use of illicit injection drugs in the 6-month period preceding the interview and had to be 15 years or older.

Participants self-initiated telephone contact with the study nurse, who administered all surveys in person. Interviews took place in a private setting of the participant’s choosing. A total of 435 people were interviewed, and an honorarium of Can $40 (approximately US$52 at the time of the study) was provided to all participants. The questionnaire was divided into 3 sections. The first section consisted of questions based on the respondent’s own characteristics, the second elicited information on the respondent’s egocentric network, and the third included questions on the respondent’s injection drug use risk network. The first section was of primary interest in this study.

Agglomerative hierarchical cluster analysis was used to cluster respondents.^45,46,53 Ward’s linkage⁵⁴ was used, with the matching coefficient^55,56 specified as the similarity measure. Numerous proximity or similarity measures and algorithms are available for use in cluster analysis, although none can be considered as the gold standard.^45,57 We chose Ward’s linkage because it has been shown to be a useful clustering algorithm for binary data^58–60 and is more likely than other techniques to produce approximately equal group sizes, thus facilitating statistical comparisons. We used the matching coefficient as a similarity measure as it has been shown to be effective when used with Ward’s algorithm.⁶¹

Other similarity measures are also suitable for use with Ward’s algorithm as well⁶²; therefore, following Finch,⁶² we also used the Dice⁶³ and Jaccard^64,65 similarity measures to analyze our data. Because cluster solutions were similar (i.e., there were only slight differences in the optimal number of clusters and the number of individuals within each cluster), we describe only the results derived from use of the matching coefficient. Following the recommendation of Finch to use stopping rules that are readily available and have been proven to be effective in past research,⁶² we used Duda’s pseudo T² statistic⁶⁶ and Calinski’s pseudo F statistic⁶⁷ to determine ideal cluster sizes.

The original data set was based on interviews with 435 participants. Complete data on each of the clustering variables were available for 414 participants (235 male and 179 female participants), and we used these data in our analyses. Stopping rules suggested that 7 clusters were ideal for both male and female participants. Table 1 shows the characteristics of each cluster based on the clustering variables used. For brevity and clarity, only the top 2 drugs of choice are listed for each cluster (or in some cases only 1). These drugs were those reported by almost all members of a cluster; reported use of other drugs was relatively sporadic.

With type of drug(s) injected as a nominal cluster descriptor, the following clusters were identified (for both genders with the exception of cluster 5): cluster 1, exclusively cocaine; cluster 2, cocaine and crack; cluster 3, cocaine and crystal methamphetamine; cluster 4, morphine and cocaine; cluster 5, a second morphine and cocaine cluster among male participants and a Dilaudid and morphine cluster among female participants; cluster 6, Talwin and Ritalin in combination with cocaine; and cluster 7, exclusively Talwin and Ritalin.

Clusters were further differentiated by the remaining 2 clustering variables: syringe sharing and ethnicity. Among male clusters, the percentage of those who had shared a syringe in the preceding 6 months ranged from 0% to 35%, with an average of 14%; the percentage of Aboriginals ranged from 0% to 100%, with an average of 53%. Among female clusters, syringe sharing ranged from 0% to 69%, with an average of 21%; the percentage of Aboriginals ranged from 8% to 100%, with an average of 79%.

Tables 2 and 3 summarize the postclustering comparisons for each cluster. Only variables for which there were significant differences between the clusters are shown (for Tables 2 and 3, and for all summations that follow, the significant differences noted between clusters use cluster 1 as the reference). Table 2 includes variables (continuous) analyzed using linear regression, whereas Table 3 includes dichotomous variables assessed in logistic regression analyses (hence, unstandardized linear regression parameter estimates are presented in Table 2 and odds ratios are presented in Table 3). We summarize these results for the male and female clusters.

HBV and HCV prevalence in clusters marked by Talwin and Ritalin use was high. We have previously observed high HCV and HBV prevalence among Talwin and Ritalin users in this population, regardless of syringe-sharing behavior.⁵² Communal drugs or filters may explain this observation, in that Talwin and Ritalin are prepared at room temperature.^31,32,85 The association found between Talwin and Ritalin use and injection in hotel rooms was demonstrated in a previous study as well,⁵¹ further validating our methodology.

Another notable result was the existence of a cluster of female injection drug users with a very low prevalence of blood-borne pathogens. This group was composed predominantly of non-Aboriginal crystal methamphetamine and cocaine users. Despite having the highest percentages of injection drug users who shared syringes (69%), binged (75%), and injected on the street (63%) in the 6 months prior to the study, this cluster did not include any HIV- or HBV-positive members, and significantly fewer of its members were HCV positive (relative to the reference cluster). Further analyses showed that members of this group were significantly younger than were the other female injection drug users and, subsequently, had injected for fewer years on average.

Our finding that crystal methamphetamine use was associated with younger age and risky syringe behavior is consistent with the results of other studies⁸⁶; however, to the best of our knowledge, our study is the first to demonstrate the existence of high-risk female injection drug users with a low prevalence of blood-borne pathogens who inject crystal methamphetamine. Social network analyses of sexually transmitted infections have demonstrated the importance of mixing patterns in promoting transmission.^87–89 Therefore, the potential opportunity for pathogens to affect the members of this vulnerable group will depend on the extent to which they are sociobehaviorally connected to other injection drug users, whose prevalence of blood-borne pathogens may be higher.^{2,7,23,29,90,91}

The observation of this high-risk cluster also has significant policy implications. Although structural interventions are important in decreasing attributable risk,^4–6 structural changes coupled with interventions that target more narrowly defined groups may be more effective in reducing transmission of blood-borne pathogens.^15,20

The dramatic increase in HIV incidence seen in Vancouver, British Columbia, during the 1990s may be particularly illustrative of the importance of identifying and assessing heterogeneity among injection drug users. The unique profile and high prevalence of cocaine injectors^2,69 were important factors in the observed increase in HIV.^18,44,81,92 SEPs were designed so that clean syringes were provided only as used syringes were returned in a one-to-one ratio, and services were limited during the evening. This policy made access to clean syringes difficult for cocaine injectors because they frequently injected during hours when no service was available.^11,13,44 Our findings suggest that similar attention to intervention design may be warranted for young female crystal methamphetamine users, to ensure that appropriate structural interventions are available to lower their infection risk.

Although we used cluster analysis specifically to explore aspects of injection drug use in our study population, our results also help to illustrate the utility of this approach for public health investigations in general. Various authors^35,93–95 in the social sciences have discussed the need to use contextual methods in data analyses, in addition to simply using contextual ideas (theories, models, and frameworks). In particular, Luke³⁵ has argued that a relatively narrow array of analytic techniques have been traditionally used in the social sciences. According to Luke, although these statistical approaches have clearly generated usable and useful results, they may not reveal the full complexity of the effects of the physical, social, economic, political, and cultural environments on human behavior and health.

As one of the social contextual methods discussed by Luke,³⁵ cluster analysis is a technique that has developed useful subtypes and generated understanding of patterns within multivariate data for a diverse range of fields.^{45,57,96–99} For example, an early classification of alcohol abuse by Goldstein and Linden¹⁰⁰ influenced a rich literature on alcoholic subtypes.^101–103 This literature has subsequently informed knowledge around diagnosis, etiology, and treatment options for alcoholics.^104,105

With respect to public health, the development of objective and empirically derived typologies has been one of the more useful applications of cluster analysis. As many studies suggest, once subtypes of formerly homogeneous groups are hypothesized and (most important) validated, targeted prevention, intervention, and treatment efforts can be mobilized.^{37,99,106–108} As noted by Houck et al., “prevention efforts are most successful when they are compatible with their target audiences.”^37(p627) In addition, although in theory other multivariable techniques (e.g., logistic regression) could have been used to analyze our data set, it is unlikely that our results could have been produced without strong a priori hypotheses, such as those generated by this technique. We endorse the viewpoint that cluster analysis can serve a complementary, as opposed to competing, role alongside traditional statistical tools.

However, some caution is warranted in the use of cluster analysis for typology development. As mentioned by Howell et al.,¹⁰⁹ cluster analysis as a technique will always reveal clusters in data; validation in other samples, substantive theoretical support, and sound expert opinion should ultimately determine whether identified clusters have clinical significance. For example, the validity of our clusters is supported by their consistency with the plausibility of the patterns identified (e.g., the low rate of infectivity among participants in the crystal methamphetamine cluster was consistent with the length of time they had injected drugs, despite their high potential for infection) and by the identification of patterns consistent with those identified through other techniques (e.g., the association between Talwin and Ritalin use, hepatitis prevalence, and injection in hotel rooms^51,52).

When plausible and consistent results are obtained, such as those identified here, the need to validate clusters does not necessarily mean that public health interventions should be withheld until that validation occurs. In our study area, with reference to female crystal methamphetamine users, the imminent and precarious infection risk posed by a combination of high-risk behaviors should be sufficient justification for an urgent public health outreach response.

There were some limitations of our study. First, because the sampling methodology used was nonrandom, the generalizability of our data to injection drug users in Winnipeg or other locales may be limited. Second, all information was self-reported and therefore subject to recall or social desirability biases. Third, the data were cross-sectional, and thus causality cannot be determined.

Fourth, small cell sizes may have resulted in unstable statistical outcomes. Fifth, there is a limit to how much can be inferred from quantitative data; for example, low prevalence rates of blood-borne pathogens may be attributable to conscientious partner restriction.³⁰ Qualitative studies may be warranted to more fully explore the underlying social context of some of the patterns identified through techniques of this type.

Sixth, we acknowledge subjectivity in the choice of clustering variables. It can be argued, for example, that blood-borne pathogen infection status is an important predictor of risk behavior and thus could be considered a clustering variable. However, from a public health investigation perspective, assessment of infection status may not necessarily be rapid.

Finally, using different linkage methods may result in different clusters being discovered.^45,53,61,62 We recommend the use of a few different similarity measures in cluster analyses. The similar pattern of results we obtained using multiple measures increases the robustness of our findings. As well, the potential artificiality of our cluster solutions must be emphasized; we do not purport to have discovered “true” clusters in the sense of all members of a given cluster forming a single connected social network. We sought to determine whether cluster analysis could provide new knowledge by discerning patterns in our sample through the use of important clustering variables.

Studies have suggested that blood-borne pathogen infections are increasing in younger populations.^69,110 Using a novel technique (with respect to research on blood-borne pathogens), we discovered a cohort of noninfected female crystal methamphetamine users with a high potential for pathogen transmission. Studies that further expand understanding of the timing, progression, and escalation of drug use and sharing behaviors in young people³⁸ may be essential in informing interventions designed to prevent the bridging of blood-borne pathogens into this and similar cohorts.