Search strategy
Automated searches of published literature were carried out on electronic databases (MEDLINE, PsychInfo, ERIC, Dissertation Abstracts, Sociological Abstracts, and Current Contents) using the following search terms: (hepatitis C OR HCV) AND (intravenous drug abuse OR intravenous drug use OR drug misuse OR drug addict OR injecting drug use OR drug abuse OR IDU) AND (prevention OR risk factor OR epidemiology OR prevalence OR incidence OR seroprevalence OR seroincidence OR genotype OR co-infect* OR coinfect*). We performed searches in these electronic databases at six month intervals.
Manual search methods included the retrieval of sources cited by seminal articles about HCV in drug-user populations, and hand searching of journals. We compiled a list of scientific journals that have either published articles on HCV epidemiology and prevention, or might conceivably publish such articles based on a history of publishing articles in a similar field (HIV epidemiology, for example). Internet searches of government websites (including local, state and national public health websites in the US; provincial, national and ministry of health websites in other countries; and websites for international health or drug control organizations) were carried out to locate government reports and unpublished surveillance estimates.
Books of abstracts and proceedings from scientific conferences related to hepatitis, HIV, infectious diseases and harm reduction were also searched for eligible reports, using both the index of key words and by reading through the abstracts of presentations. The NIH CRISP database was also used to identify ongoing or recently completed studies relevant to our meta-analysis; names of investigators identified in CRISP were periodically entered into the electronic databases to identify publications that might conceivably report on measures of interest. Consultants to the study were enlisted to submit reports from their own studies or from other studies they had learned about through professional contacts at conferences or other meetings; these consultants included investigators who were carrying out research related to HCV in drug users in the US, Europe, Asia, Africa, the Middle East and Australia/New Zealand.
Study selection
Throughout this description of study methods, we use the term data report to refer to published and unpublished articles, manuscripts, personal communication, dissertations, abstracts, conference presentations and book chapters that are reviewed for inclusion in the study sample. Data reports available through the end of 2006 were included in the search; study retrieval and coding began in August, 2004.
To be included in our study, data reports must have included in their sample individuals who could conceivably have acquired HCV infection via administration of an illegal drug, i.e., injection drug users, or non-injection drug users who snort, sniff or smoke heroin, cocaine, amphetamines or other drugs using straws or pipes. Marijuana smokers were not included in the study as there are no epidemiologic data to indicate excess HCV in marijuana smokers. Similarly, those who administered amphetamines or other illicit drugs orally (i.e., took pills) were also excluded as there is no biologically plausible route of HCV transmission associated with this practice.
Only those reports from which we could abstract estimates of HCV prevalence, incidence, measures of association, HIV/HCV co-infection or HCV genotype distributions for drug users were included. Further, reports must have given separate estimates for IDUs and NIDUs, as these rates were expected to vary greatly between these two risk groups. Thus, studies that aggregated IDUs and NIDUs in estimates of prevalence, incidence or measures of association were not included in this review. HCV status must have been determined by serologic testing of either sera or saliva; studies that used self-reported HCV antibody status, or those that tested for HCV RNA without reporting the results of anti-HCV testing were excluded.
Each data report could have one or more of the following types of studies associated with it: HCV prevalence studies, HCV incidence studies, HCV genotype studies or studies of HCV co-infection with HIV or other hepatitis viruses. Within a single report, each of these types of studies could appear and be counted as an individual study by our definition. A report was also defined as having multiple studies when epidemiologic estimates were given for subgroups of individuals distinguished by different methodology (e.g., enrollment criteria, sampling location or sampling method) or were reported as separate samples with different demographic characteristics or inclusion criteria. For example, a study in IDUs that spanned five years, and presented all their data (sample demographics and HCV prevalence) for each year of data collection was counted as five prevalence studies. Separating the data into individual studies permitted the collection of sample characteristics or methodological features associated with the subset of subjects being examined.
Screening
A pilot study was carried out to test and develop procedures for screening titles and abstracts to identify data reports that would be eligible to be retrieved, and to estimate an expected sample size (the number of data reports that would provide data to address the aims). In October 2003, we conducted a Medline search using the keywords mentioned above; this search retrieved 1324 articles. Abstracts we obtained from a 5% random sample of these articles (n = 63) and three study staff (the PI, a co-investigator and a research assistant) independently pre-screened each title and abstract. Of necessity, the criteria for screening were broad so as to reduce the likelihood of missing relevant studies, but it was also desirable that screening have relatively high specificity to avoid retrieving a large number of irrelevant studies.
The screening criteria were: 1) sample included IDU or NIDU, and 2) study reported any parameters of epidemiology (incidence or prevalence of HCV-antibody or HCV-genotype), HIV/HCV co-infection, or measures of association. For this pilot, studies must also have been written in English. Initially, there was 80% agreement as to whether the article should be retrieved. Screening criteria were discussed and revised to reach consensus in their application.
Beginning in August 2004, we followed the process developed in the pilot study to screen abstracts to determine whether a report was eligible in terms of its sample and the data it provided. For the reports whose abstracts deemed them potentially eligible, the full text was reviewed to determine whether it truly met our inclusion criteria. Each abstract or report was screened for eligibility by both a senior research assistant and the project director. Screening sought to eliminate studies that were clearly unrelated to the scope of the project. In cases where it could not be definitively determined whether a report was eligible based on reading the abstract, the full text article was retrieved and reviewed for eligibility. When the full text of an article was not available in English, Spanish, or Italian, the English-language Medline abstract of the article was used for coding.
It was common to find multiple HCV-related reports originating from a single large research project such as ALIVE, VIDUS, CIDUS, RAVEN and many others [23–30]. Thus, several methods were used to identify duplicate or overlapping reports (for example, recording study names and searching our reference manager database for other reports by the same author or set of authors). These potentially-overlapping reports were identified but retained in the database, as some reports included analyses of different sets of factors or subsets of the study sample. Duplicate estimates of the same parameters for the same datasets will be excluded prior to each data analysis project. (Figure 1 shows a decision tree of our study selection and screening process.)
Coding
The coding was carried out by senior research assistants who had graduate training in research methodology, and received additional training in HCV epidemiology, drug use and meta-analytic methods. The content and structure of the coding form was developed by reviewing those used in other meta-analyses, for example, the CDC Prevention Research Synthesis Project [31] and the Self-Report/Biological Measures database of Drug Use [32] The coding form included items such as the type of study or studies included in the body of the report (prevalence, incidence, co-infection or genotype study), study methods (design, inclusion criteria, recruitment method, recruitment locations, method of determining IDU/NIDU status, specimen type, and HCV test method), and demographics and other characteristics of the sample such as duration of drug use, type of drugs used, and frequency of use.
We collected all data related to HCV prevalence and incidence, i.e., all numerators and denominators, including both numbers and proportions of subjects and numbers of person-years, to allow us to combine data from separate studies in subsequent analyses. Anti-HCV prevalence and incidence estimates, relative risks and odds ratios (both crude and adjusted, with their 95% confidence intervals) were also collected in relation to sample characteristics. Simple calculations were made when necessary, and any approximations which were recorded (such as numbers estimated from reading a figure) were marked as such.
In some cases, the statistics recorded were simple measures of overall HCV prevalence or incidence for the entire eligible sample, while in other cases prevalence or incidence measures were reported within highly restricted subgroups defined by a number of variables such as age, gender, race and type of drug used. The coding form was structured so that we could abstract both simple and complicated overall estimates and sub-group comparisons, and therefore accommodate varying degrees of complexity in the data.
Thus, the coding form included fields for a number of pre-defined categories of gender, race/ethnicity, drug use, and risk behavior categories such as receptive and distributive syringe sharing, the shared use of drug preparation equipment (e.g., drug cooker, filtration cotton and rinse water), the use of a syringe to divide drugs, giving or receiving injections, injection in prison or jail, and participation in prevention activities such as drug treatment or syringe exchange programs. Sexual risk behavior data was also collected, for example, number of sexual partners, unprotected sex, commercial sex work and having an IDU sex partner. For each of these risk categories describing behavior, we recorded the referent time period used. There were also fields included in the coding form intended to capture relevant data that did not conform to our preset categories.
We did not record measures such as odds ratios comparing HCV prevalence or incidence in IDUs to that among NIDUs or non-drug users, as there is clear evidence that injection drug use is a highly potent risk behavior for HCV acquisition, and comparisons among such populations would not contribute new knowledge to our understanding of HCV epidemiology.
Quality assurance methods
A number of strategies were used in the course of this study to ensure reliable, valid and consistent coding of data. For example, each data report coded by a research assistant was subsequently reviewed for accuracy and completeness by the Project Director and the Principal Investigator. The coders made any necessary changes to the coding before the report was considered complete. Any differences of opinion among the study group members were settled by discussion at a weekly study meeting convened for this purpose. A study manual was developed to guide coding and to record special cases and their resolution.
Backcoding
Once a significant portion of the published data had been coded (October 2005), the research assistants conducted a data-coding reliability sub-study. A 10% sample of eligible coded reports was selected and re-coded by a different coder and all discrepancies in coding were noted and summarized. This project was undertaken to establish whether (1) any changes in coding rules had occurred over the course of the study, (2) there were any systematic differences between coders or (3) there were any specific items that may have been inconsistently coded. This reliability project revealed that two items (recruitment method and recruitment location) were inconsistently coded. Our review of the text of the reports revealed that variation between coders was principally due to a lack of clarity in the presentation of this information across a significant number of reports. Coding rules for these two variables were discussed and re-defined, and these items were back-coded for all articles.
Contacting Authors
When the text or abstract of an article or conference presentation indicated that measures of interest to our study had been collected but were not reported, we contacted authors to request the information. These requests were limited to simple descriptive data such as separate estimates of HCV prevalence or incidence for IDUs vs. NIDUs, or to ask for the numerator or denominator used to estimate incidence or prevalence. We did not ask for additional analyses of data. Authors were contacted only if the report had been published or presented less than five years before the coding date. Overall, we contacted just over 100 authors, and 42% of authors we contacted sent a reply to the request for data. Of those who replied, 70% provided data that could be included as part of the meta-analysis.
Unpublished data
We also searched subject indexes of conference proceedings (abstracts) on paper, CD-ROM or various websites. The conferences included were annual meetings of the American Association for the Study of Liver Diseases, the Society for Epidemiologic Research, the American Public Health Association, the International AIDS Society, the National Harm Reduction Conference, the International Conference on the Reduction of Drug Related Harm, the College on Problems of Drug Dependence, International Society for Infectious Diseases, and the American Sociological Association. Investigators from the study also attended a number of these conferences and collected copies of presentations and other unpublished reports. These abstracts were screened and the data were coded by the same process we used for published abstracts and articles.
Substantial effort was devoted to ascertaining whether the data in the unpublished literature duplicated any reports that had been previously coded. This step involved cross-checking of authors of unpublished studies against our database and in PubMed. When matches were found, the unpublished literature was compared to the published report to identify whether they overlapped in terms of location, number of subjects, and year of data collection. Approximately half of the conference abstracts that appeared to be codeable were subsequently eliminated because there was a published article with overlapping data.
Study quality measures
We followed the recommendation of the MOOSE group [21] that meta-analyses of observational studies evaluate certain elements of each included study indicative of their quality. Thus, we devised a number of 'quality' items to assess whether each study's design and implementation was clearly explained and appropriate for the claims made in the reports. Our scale rated several study factors likely to affect findings of HCV prevalence, incidence and co-infection in drug user populations, including the rigor with which the investigators handled the problem of misclassification with respect to route of drug administration. Additionally, the quality scale assessed some aspects of the overall design and methodology of the studies, such as sample size, reporting of demographics, presence of inclusion or exclusion criteria, participation rate, adjustment for confounding, and consistency in reporting data. Sources of bias in study methods have been demonstrated to account for heterogeneity in results in meta-analyses [21].
An initial list of quality criteria was developed by the research team to assess NIDU studies included in a paper published by our group [13]. The quality criteria for the IDU sample were modeled on the NIDU quality scale. A detailed description of the NIDU quality scale can be found in Scheinmann et al. [13]. For the complete list of quality items used for evaluating IDU studies, see Table 1.
To determine whether the items in our quality scale were internally consistent we performed Cronbach's alpha reliability analyses using two-thirds of the data from the HCV Synthesis Study. The 19-item scale had an alpha value of 0.73, reflecting adequate scale coherence. Removal of items with low item-total correlations did not improve the alpha. The distribution of the quality scale scores was bimodal. Quality was then recoded dichotomously by splitting between the modes.