This population-based sentinel study included all incident culture-positive TB patients from sentinel sites from January 1996 to December 2000. In brief, the seven sentinel surveillance sites included the states of Arkansas, Maryland, Massachusetts, Michigan, and New Jersey; Dallas, Tarrant, Cameron, and Hidalgo Counties in Texas; and Alameda, Contra Costa, Marin, San Mateo, Santa Clara, and Solano Counties in California. A detailed description of the study’s design, participants, population, and laboratory and epidemiologic methods is provided elsewhere (6).

All patients included in the study were reported to the CDC national TB case registry on the form Report of a Verified Case of Tuberculosis, a standardized electronic form submitted for TB surveillance to CDC by all state public health reporting areas. Data reported include patient demographics, laboratory test results, drug susceptibilities, information on chest radiographs, and treatment outcomes (7).

Investigators from the sentinel surveillance sites submitted patient isolates to the corresponding regional laboratory for genotyping and conducted routine contact investigations. In addition, participants from the surveillance sites performed detailed epidemiologic investigations on groups of persons with M. tuberculosis isolates that had matching genetic patterns or clusters (see below). The regional genotyping laboratories conducted IS6110 RFLP on isolates from sentinel patients. Since low-copy numbers of IS6110 (i.e., six or fewer copies) reduce test specificity, spacer oligonucleotide typing (spoligotyping) was conducted on such isolates. A cluster, which was identified by analysis of the entire TB genotyping network database, was defined as two or more isolates with either identical RFLP patterns (at least seven copies of IS6110) or identical RFLP and spoligotype patterns for isolates with RFLP patterns that had six or fewer copies of IS6110.

Differences in the proportion of TB patients from the TB genotyping network population living in cities with populations of <100,000, 100,001 to 250,000, 250,001 to 500,000, and >500,000 were evaluated compared with those of the national TB patients for the year 2000 only. Statistics were obtained from the U.S. Census Bureau (available at: URL: http://www.census.gov/population/cen2000/phc-t6/tab04.pdf ).

Correlation of average TB incidence among cases at the seven sentinel sites and percentage of cases with isolates that clustered genetically were examined by year by using the Spearman rank correlation statistic. Clustering was determined by examining each year’s cases independently. A Mantel-Haenszel chi-square or Fisher exact test was used, as appropriate, to ascertain whether the sentinel population was representative of TB patients in the United States in terms of demographic, clinical, behavioral, or outcome characteristics.

We used multiple logistic regression to assess the importance of demographic, clinical, behavioral, or outcome variables in predicting the occurrence of a given genotype for those genetic clusters that occurred most frequently (≥20 isolates). The dependent variable was the presence or absence of a given genotype. The best-fit logistic regression model was determined by the strategy of Hosmer and Lemeshow (8). In brief, a univariate analysis of the categorical independent variables was done by using the Mantel-Haenszel chi-square or Fisher exact test, as appropriate; any variable with a significance value of ≤0.20 was included in a best subset, multivariate logistic regression model. Collinearity of independent variables was assessed by using the variance/covariance matrix from PROC LOGISTIC (SAS Institute, Inc., Cary, NC) generate condition indices and a matrix of variance decomposition proportions to detect dependencies among the variables (9). Backward elimination of independent variables was performed if the probability of the independent variable was ≤0.20. Both the Wald statistic and 95% confidence interval were used on each coefficient to assess the significance of variables in each model; the log-likelihood ratio was used to assess the overall significance of the final models, and the Hosmer-Lemeshow statistic was used to evaluate the fit of each of the final models. Data were analyzed by SAS version 8.0 software (SAS Institute, Inc.) (10).

This population-based study is the largest that has been conducted in the United States to assess risk factors related to specific M. tuberculosis genotypes. Generally, clustered isolates have been considered recently acquired infections (12). However, this assumption may not always be correct. Clustering does not prove that transmission occurred, and its demonstration depends on adequate sampling of the population, incidence of TB, and characteristics of the study population (e.g., age structure, population mobility, duration of residence, and immune status) (1,13) . Only 25%–42% of patients in genetic clusters were shown to have epidemiologic connections with another member of the cluster (14–16). Conventional epidemiologic investigation of these TB patients (including interviews) was conducted, but inclusion in this analysis was outside the scope of this article. Thus, results that indicate clustered genotypes are representative of recent transmission should be interpreted with caution.

Given this caveat, our results nevertheless demonstrate several consistent patterns. Differences in demographic and other risk factors for persons with isolates that clustered corroborated those from smaller studies conducted in the United States and larger surveys in Europe. Extensive surveys from the Netherlands (17) also demonstrated that persons with isolates that clustered genetically were younger than those with unique genotypes. Other risk factors for clustering included being male, born in the United States, non-Hispanic black, or homeless; using drugs and alcohol excessively; and having pulmonary disease and cavitations on chest radiograph, a sputum smear with acid-fast bacilli, and HIV infection. These risk factors have been observed for TB patients in different communities (12,18,19). The heterogeneity and diversity of the study population may account for our failure to produce a multivariate logistic model to predict clustering.

A third of the foreign-born cases were recent immigrants to the United States, and overall, the percentage of clustered isolates from foreign-born persons was lower than the percentage from nonimmigrants (Figure 3), indicating that at least a portion of these cases resulted from reactivation of latent disease or recent infection in the country of origin. In addition, for foreign-born persons, clustering of M. tuberculosis increased with the duration of residence in the United States. These results suggest that recently imported strains of M. tuberculosis from foreign-born persons may not commonly spread to U.S. residents or that transmission may be occurring after a lag time before the imported strains manifest as disease in contacts. Similar observations have been published in studies from San Francisco, New York, Switzerland, and Norway (20–24). These data may also reflect gaps in our knowledge of M. tuberculosis genotypes in circulation; a comparison of the U.S. TB genotyping network results with other databases worldwide may be warranted.

Logistic regression analysis of the most commonly occurring strains demonstrated that different risk factors were associated with specific genotypes. Several genotypes were associated with ethnic origin (e.g., Asian or Pacific Islander and Hispanic patients with six and three genotypes, respectively; Table 4). A recent study in Norway showed that several clusters consisted of patients of the same ethnic origin (23). An association has also been observed between the patient’s ethnic origin and IS6110 copy number (25). These results, in conjunction with additional epidemiologic data, may be useful in tracking the geographic origin and spread of M. tuberculosis strains of public health importance (26).

A small proportion of clustered isolates were from persons from more than four sites spanning 5 years of study (Table 4). Although an in-depth analysis of epidemiologic links was not possible in this study, we found no evidence of recent transmission between patients with identical genotypes from the different states (data not shown); this lack of transmission was also noted in a smaller study in the United States (27). Since TB transmission is generally considered a local event, these ubiquitous genotypes may be widespread because of social factors (e.g., homelessness or alcohol or drug abuse; Table 4). In addition, these genotypes may represent older, endemic domestic strains that have been in the United States for centuries and have dispersed more widely throughout the United States than the more recently imported strains. Further molecular characterization of these genotypes may show additional differences not detected by RFLP. Nonetheless, the effect of M. tuberculosis virulence or host factors on the distribution of these genotypes cannot be ascertained.

The proportion of strains that were classified into clusters of identical genotypes (48%) was comparable with proportions in the Netherlands and Denmark (50%) (2,28), but the proportion was considerably higher than in two other countries (17% in Switzerland [29]; 20% in Norway [23]). The cumulative percentage of clustered strains reached a plateau by the end of the study’s second year (Figure 1), a finding consistent with other molecular epidemiologic TB studies (2). Increases in maximum cluster size were anticipated because, as sample sizes increase with time, the number of isolates in each cluster would be expected to increase. In addition, higher proportions of clustered cases were observed for low-band number patterns (Figure 3), which had the maximum cluster size and may indicate that the low-copy IS6110 patterns are not specific, even with the addition of spoligotyping.

The sensitivity and specificity of IS6110 RFLP in molecular epidemiologic studies have not been quantified and represent a potential limitation of this study. Although the stability of IS6110 is relatively high, the half-life of IS6110 RFLP is estimated to be 3–10 years (29–31) based on typing of serial isolates from individual patients. A study of isolates from patients in confirmed chains of transmission showed little change in IS6110 patterns (32). Calculation of these rates may be influenced by the duration between time of disease onset and time of sampling and may be proportional to the effectiveness of the TB control program (30). Because genotyping results were not available for 10% of TB cases in this study, estimates of the degree of clustering and the size of clusters are conservative. Some unique isolates might have clustered if some of the missing isolates had been aaavailable or if other cases with the same strain were present outside the study area (33).

Sentinel surveillance sites defined by artificial boundaries (i.e., state lines) not entirely representative of TB patients from the United States were included in this study. More than 90% of the isolates from patients from the surveillance areas were genotyped, and these isolates were representative of those culture-positive patients from the sentinel surveillance areas. However, 16% of all TB case-patients reported in the United States were included in these sentinel surveillance sites during the 5-year study period. In addition, the sentinel surveillance population had higher proportions of foreign-born persons than the national average. Because of the propensity of foreign-born persons to have isolates with unique genotypes, the actual rate of clustering may have been underestimated. Nonetheless, sentinel surveillance of TB cases has provided a useful method for documenting genotypes in circulation in the United States and for identifying risk factor correlates of common genotypes.

Annual declines in TB incidence were paralleled by similar declines in the proportion of cases with genotypes in clusters (Figure 2), a finding consistent with the hypothesis that decreased clustering is expected with declining incidence (20). Since effort was similar each year, this association is not likely to be an artifact related to sample size (i.e., as sample size or number of cases becomes smaller, the probability of detecting clusters decreases). These findings underscore the importance of long-term longitudinal molecular studies and the potential usefulness of these methods in evaluating program effectiveness and improving program management.