Contextualizing tuberculosis risk in time and space: Comparing time-restricted genotypic case clusters and geospatial clusters to evaluate the relative contribution of recent transmission to incidence of TB using nine years of case data from Michigan, USA.

Abstract

Background.

Novel approaches must address the underlying factors sustaining the TB endemic in the U.S., specifically what maintains new Mycobacterium tuberculosis (Mtb) transmission.

Methods.

Culture-confirmed TB cases reported to the Michigan Department of Health and Human Services (2004-2012) were analyzed for time-restricted genotypic and/or geospatial clustering. Cases with both types of clustering were used as a proxy for recent, local transmission. Modified, multivariate Poisson regression models were fit to estimate this prevalence in relation to various individual- and neighborhood-level demographic and socio-economic variables.

Results.

Those individuals that were spatially clustered were 1.7 times as likely to also be time-restricted genotypically clustered. The prevalence of recent, local transmission was higher among U.S.-born cases, males, and Non-Hispanic Blacks. Moreover, people living in neighborhoods in the highest poverty quartiles had 13.8 times the prevalence of recent, local transmission compared to those in the lowest poverty neighborhoods.

Conclusions.

Our results suggest geographic areas with a high concentration of TB cases are likely driven by ongoing transmission, rather than enclaves of individuals who have reactivated a case of latent TB. Furthermore, efforts to continue reducing Mtb transmission in the U.S., and other low-incidence settings, must better identify community-level sources of risk, manifested through the complex social interactions among people and their environments.

Background

Beginning in 2013, the U.S. experienced a stagnation in the decades-long decline in tuberculosis (TB) incidence, currently at ~3 per 100,000 persons per year (1, 2). This is disconcerting in light of the new World Health Organization (WHO) goal of TB elimination in low incidence countries such as the U.S. (3). It suggests that new approaches are needed to understand and modify the factors that sustain transmission in the U.S., and specifically the role of recent transmission.

TB is caused by infection with Mycobacterium tuberculosis (Mtb), transmitted primarily via airborne droplets (4). An estimated 90-95% of people who are infected with Mtb produce immune responses that force the pathogen into latency, resulting in latent TB infection (LTBI) (4). Individuals with LTBI may later progress to active TB disease, a process that is highly dependent on individual risk factors, the most noteworthy being older age, malnutrition, immunosuppression (especially coinfection with HIV or diabetes) (5), stress, and other factors affecting host immunity (4, 6). Alternatively, active TB disease may arise shortly after initial infection if the initial immune response is inadequate. TB that appears in this manner is considered to reflect a recent transmission event, requiring a different public health strategy for control. Accordingly, a key feature of TB control strategies involves estimating the extent and context of recent transmission in a given population.

As molecular and epidemiologic tools have advanced, so has our ability to distinguish recent transmission from activation of LTBI. Many current studies have relied on molecular markers that show Mtb genotypic similarity as an indicator of recent transmission (7-9). In addition, studies of people in the U.S. and in other developed countries have also used common social links with index cases to classify recent transmission (10). Yet, there are inherent limitations to these methods that involve the discriminatory power of current genotyping technology and of establishing epidemiological links among cases. To address these challenges, TB researchers are increasingly developing more integrative approaches to better classify recent transmission (11-14). One such approach, for example, is the use of whole genome sequencing (WGS), in addition to data on epidemiologic links, to more accurately estimate of chains of recent TB transmission (15).

Traditional paradigms of infectious disease epidemiology emphasize the importance of spatial patterns among individuals and populations in affecting transmission risk, particularly for those diseases in which close contact is necessary for pathogen spread. Accordingly, measures of physical proximity between people have been used in many analyses as an indicator of local transmission (11, 12, 16, 17). However, physical proximity may not always indicate local transmission given the unique natural history of TB. Individuals can be exposed and infected with Mtb long before they develop active TB disease, as the infection can exist in a latent state. Thus, it is plausible that there are concentrations of individuals with LTBI that do not represent ongoing recent transmission. Accordingly, it is imperative to infer whether clusters of individuals with TB disease in physical space correspond to recent, local transmission.

With a TB incidence rate of 1.3 per 100,000 persons per year in 2015 (18), the state of Michigan is a relatively low-burden state within the U.S. Yet, burden is not evenly distributed, and there are persistent disparities, both at the individual and neighborhood levels (19, 20). Our study aims were two-fold. First, we developed a method of inferring recent, local Mtb transmission by integrating information on Mtb genotype, time, and location. Second, using nine years of TB case data from Michigan we applied this method of inferring recent, local Mtb transmission to enhance understanding of how individual- and neighborhood-level social factors impact TB disparities, particularly among cases of recent, local transmission.

Methods

Data Sources

There were 1,800 TB cases reported by the Michigan Department of Health and Human Services (MDHHS) from January 1, 2004 to December 31, 2012. Of the 1,800 reported cases, 1,390 (77%) were culture-confirmed, allowing forMtb isolate genotyping. Our study sample was initially comprised of all TB cases reported in Michigan from 2004 through 2012 for which genotyping and residential address data were available. Cases that lacked both spoligotype and 12-locus-MIRU-VNTR results, and/or address information were excluded. The final study sample of 1,235 cases represented 69% of all 1,800 cases recorded by MDHHS, and 89% of the 1,390 culture-confirmed cases.

Demographic characteristics of each case were drawn from the Report of a Verified Case of TB (RVCT) form (21), developed by the U.S. Centers for Disease Control and Prevention (CDC) and used for routine TB surveillance in Michigan. TB case addresses were geocoded using ESRI ArcGIS (v10.4) and then linked to U.S. census block groups. Block group-level characteristics were derived from the American Community Survey (ACS), 2012 5-year estimates (22). We used the linkage to the ACS data to assess neighborhood density and neighborhood poverty. Additional information on the creation of both of these variables is included in the supplementary material.

Cluster Definitions

In this study, we classified individuals as belonging to one of four types of clusters (including non-clustered) based on a combination of genotype (spoligotype and 12-locus MIRU/VNTR), diagnostic year, and spatial location data (Figure 1). Individuals in the study sample were classified as being time-restricted genotypically clustered TB cases only, geospatially clustered TB case only, both time-restricted genotypically and geospatially clustered TB cases, or neither time-restricted genotypically nor geospatially clustered TB cases. Time-restricted genotypic clusters of TB cases were those sharing an identical genotype by spoligotype and 12- locus MIRU/VNTR, as well as diagnosis dates within one year of each other, which is similar to previous studies (14, 23). Genotypically non-clustered TB cases were those not sharing an identical genotype or having diagnosis dates more than one year apart.

Figure 1.

Flowchart showing the assignment of clustering definitions in the study sample.

Geospatial clusters of TB cases were identified through an algorithm based on Kuldorff’ s scan statistic (24) using the standard SaTScan (ver. 9.4.3) software, in which various parameters are specified by the investigator. This analysis "scans" the total study region in user-defined circular sub-areas to detect those areas with a greater-than-expected number of cases per population. Distances between cases were based on the geographic centroid of each census block group. We used the discrete Poisson model for purely spatial clusters, with a spatial cluster maximum radius of 1.0 miles. The “expected” case counts were determined based on the Poisson assumption that the expected number of cases is proportional to the population size of a given area. Both case and population counts per census block group were considered to account for population differences among block groups. Sensitivity analyses were also performed to evaluate whether inferences regarding geospatial clustering would change with different sizes of the circle radius (see Supplementary Table 2).

There were 47 unique spatial clusters of TB cases with an average radius distance of 0.64 miles. The cluster sizes ranged from three to 37 cases. Eight of the clusters likely reflect cases within the same household. Additional information on the spatial clusters of cases is given in Supplementary Table 3.

Using these genotypic clustering and geospatial clustering results, we then defined a case of recent, local transmission as any case that was both genotypically and geospatially clustered with at least one other case. Previous studies have shown that the use of additional data points can improve the accuracy of identifying recent transmission. Notably, using a "gold-standard," field-based epidemiology investigation, France et al. (13) demonstrated that by defining recent transmission through integration of several data types (i.e. genotype, geographic distance, time of diagnosis, source case infectiousness), the accuracy of recent transmission identification was improved compared to that based solely on genotype or space. By the same logic, our approach of analyzing multiple data types to classifying recent-local transmission should also improve the accuracy of inferences regarding recent transmission.

Statistical analyses

The primary outcome of this study was being classified as a case of recent, local transmission (versus any of the three individual cluster classifications). Again, individuals were classified as a case of recent, local transmission if they were both time-restricted genotypically clustered and spatially clustered. We first examined differences in the demographic characteristics of TB cases by type of cluster using a chi-squared test. Regression analyses were then used to estimate the prevalence of recent, local transmission by a series of individual-and neighborhood-level social factors. Specifically, we used modified, multivariable Poisson regression models with generalized estimating equations (GEE) and a compound symmetry correlation structure and log link. This approach was necessary because the prevalence of recent, local transmission was greater than 10% in the study sample, thus precluding traditional logistic regression in which the odds ratio would overestimate the prevalence ratio. These Poisson regression models were adjusted for individual- and neighborhood-level sociodemographic factors. We used GEE to account for nesting of cases within census block groups (25). Modified Poisson models with GEE allow for accurate parameter estimation and robust variance estimates by accounting for the correlated observations existing among cases in the same block group. The association of individual-level factors (race/ethnicity, sex, nativity, and age) with the prevalence of recent, local transmission was evaluated (Model 1). Model 2 includes the covariates in model 1 in addition to neighborhood-level factors (neighborhood density and neighborhood poverty)

The prevalence ratio of recent, local transmission and corresponding 95% confidence interval (CI) were calculated based on the above Poisson models. We used a two-tailed alpha level equal to 0.05 to determine significance. Statistical analyses were carried out in SAS (ver. 9.4).

The study was approved by the Institutional Review Boards of the University of Michigan and the Michigan Department of Health and Human Services.

RESULTS

Study population

The demographic composition of the 1,235 TB cases included in the study sample is described in Supplementary Table 1. The average annual incidence rate of TB for the state of Michigan during the study period was 1.49 per 100,000 person years (ranging from 1.02 to 2.01). Using the ACS, the cumulative incidence for the total 9-year study period was 16.2 per 100,000 person, ranging from 0 to 6,250 per 100,000 person by census block group. The block groups with incidence rates greater than the average were most concentrated in the Metro Detroit area.

Prevalence of clustering

We then examined the distribution of each cluster type in the sample (see supplementary figure 1). A total of 253 cases (20%) were classified as solely time-restricted genotypically clustered, meaning they shared an identical genotype with at least one other case in the sample and were diagnosed within one year of that case. Another 199 cases (16%) were classified as solely geospatially clustered, which meant that their census block group at the time of diagnosis was within a 1-mile radius of another case’s block group in the study sample. Yet another 224 cases (18%) were classified as cases of recent, local transmission (both time-restricted genotypically and geospatially clustered). The remaining 559 cases (45%) were non-clustered. As compared to non-genotypically clustered cases, those that were time-restricted genotypically clustered were nearly twice as likely to also be spatially clustered (prevalence ratio=1.79, p<0.001).

The distribution of cluster type showed differing patterns by key covariate categories (Table 1). A higher proportion of male cases was classified as recent, local transmission compared to female cases. Over half of the Non-Hispanic White, Asian, and Hispanic cases were classified as non-clustered, while the largest proportion of Non-Hispanic Black cases were classified as recent, local transmission. Nearly 60% of the foreign-born cases were classified as non-clustered compared to 35% of the U.S.-born cases.

Table 1.

Comparison of types of clustering across study year and key socio-demographic characteristics.

	Non-Clustered (Neither Geospatially nor Genotypically clustered)	Geospatially clustered only	Genotypically clustered only	Recent, local Transmission (Geospatially AND Genotypically clustered)
	N (%)	N (%)	N (%)	N (%)	P-value
Total Sampled	559 (45)	199 (16)	253 (20)	224 (18)
Gender
Male	318 (43)	108 (15)	160 (22)	148(20)	0.02
Female	241 (48)	91 (18)	93 (19)	75 (15)
Race/Ethnicity
Non-Hispanic White	182 (64)	29 (10)	55 (19)	19 (7)	<0.0001
Non-Hispanic Black	124 (25)	95 (19)	104 (21)	173 (35)
Asian	179 (59)	47 (15)	60 (20)	18 (6)
Hispanic	62 (51)	23 (19)	25 (20)	12 (10)
Other	12 (43)	5 (18)	9 (32)	2 (7)
Nativity
U.S.-born	235 (35)	91 (13)	166 (24)	188 (28)	<0.0001
Foreign-born	324 (59)	108 (20)	86 (16)	35 (6)

P-value corresponds to a chi-square test examining whether the distribution of cluster type differs by each socio-demographic variable. The proportions reflect row proportions.

Example: Males are more likely to be non-clustered (43%) compared to geospatially clustered alone (15%).

Prevalence of local transmission

We developed and compared models of recent, local transmission in relation to various individual- and neighborhood-level covariates (Table 2). Using only individual-level characteristics (Model 1), race/ethnicity, nativity and age were all significant independent predictors of recent, local transmission. After controlling for all other covariates, the prevalence of recent, local transmission in Non-Hispanic Blacks was 4 times greater than that of non-Hispanic Whites. Hispanics and Asians were not statistically significantly different from Non-Hispanic Whites in prevalence of recent, local transmission, although a greater risk was observed for both Hispanics (2.5 times) and Asians (1.9 times), as compared to the Non-Hispanic Whites (Table 2).

Table 2.

Estimates of the adjusted prevalence ratio of recent, local transmission by neighborhood-level and individual-level factors.

Factor	Model 1 PR (95%)	P	Model 2 PR (95%)	P
Race/Ethnicity
Non-Hispanic Black	3.97 (1.80, 8.73)	<0.0001	1.77 (0.97, 3.22)	0.0722
Asian	1.87 (0.64, 5.45)		1.66 (0.72, 3.83)
Hispanic	2.47 (0.87, 7.04)		1.25 (0.59, 2.64)
Non-Hispanic White	Ref.		Ref.
Gender
Male	1.19 (0.86, 1.64)	0.3000	1.12 (0.87, 1.45)	0.3811
Female	Ref.		Ref.
Nativity
U.S.	3.54 (1.87, 6.72)	0.0001	2.14 (1.27, 3.60)	0.0044
Foreign-B	Ref.		Ref.
Age Category
< 18 years	0.31 (0.12, 0.71)	0.0013	0.37 (0.18, 0.76)	<0.0001
65+ years	0.73 (0.37, 1.43)		0.84 (0.46, 1.52)
18-64 years	Ref.		Ref.
Neighborhood Density
Q4			1.28 (0.86, 1.89)	0.5300
Q3			1.09 (0.72, 1.65)
Q2			1.08 (0.70, 1.67)
Q1			Ref.
Neighborhood Poverty*
Q4			13.78 (4.76, 39.87)	<0.0001
Q3			8.07 (2.82, 23.10)
Q2			3.25 (1.07, 9.93)
Q1			Ref.

Results based on modified Poisson models with GEE. P-values reflect the type 3 analysis effects.

^*Higher quartiles reflect greater poverty or higher density.

Model 1 considers only individual-level demographic characteristics. In Model 2, neighborhood density and poverty are added as covariates.

The prevalence of recent, local transmission among U.S.-born individuals was 3.5 times that of foreign-born persons (P=0.0001), after controlling for all other covariates (Table 2). Furthermore, the prevalence of recent, local transmission was lower in younger age groups compared to those in the 18-64 year age group. People <18 years of age had a 69% reduced prevalence of recent, local transmission, which was significantly lower compared to those in the 18-64 year old group.

In Model 2, we added neighborhood density and neighborhood poverty to the analysis. Using both individual- and neighborhood-level covariates, results showed that nativity, age, and neighborhood-level poverty were significant predictors of recent, local transmission (Table 2). The prevalence of recent, local transmission among U.S.-born was 2.14 times that of foreign-born (P=0.004), controlling for all other covariates. Those who were <18 years old had a 63% lower prevalence of local transmission compared to the 18-64 year age group which was statistically significant.

Individuals residing in neighborhoods in the highest quartile of poverty had 13.8 times the prevalence of recent, local transmission compared to those in the lowest poverty neighborhoods (Q4 vs Q1) (P < 0.0001), controlling for all other covariates (Table 2). Moreover, there was a monotonic relationship of increasing neighborhood poverty and increasing prevalence of recent, local transmission.

Sensitivity Analyses

Several sensitivity analyses were performed to further interrogate our findings. First, we tested whether time-restricted genotypic clusters were more likely to be spatial clusters. Results showed that those individuals who were classified as time-restricted genotypic clusters were more likely to be spatial clusters (Supplemental Table 4) (prevalence ratio: 1.71; 95% CI: 1.48, 1.97).

We also further examined the group of individuals who were time-restricted genotypically clustered, but not spatially clustered. On average, those genotypic clusters that were time-restricted (but not spatial clusters) were less likely to be non-Hispanic Black and U.S.-born. These individuals were also more likely to be older (65+) and live in less dense and lower poverty neighborhoods. Finally, they were less likely to report homelessness in the past 12 months, injection drug use, HIV positive status, and pulmonary TB. Full results of this additional analysis are reported in the supplement (Supplemental Table 5).

We also examined the distribution of race/ethnicity and nativity by neighborhood-level covariates. That analysis showed significant differences in the distribution of both race and nativity by level of neighborhood poverty. Full results are provided in the supplement (Supplemental Table 6).

DISCUSSION

Using Michigan TB surveillance data from 2004 through 2012, we applied an integrative approach for classifying recent, local Mtb transmission that accounted for bacterium genotype, geographic location of cases, and time to examine socio-demographic disparities in disease. There were two key findings from these analyses. First, our more complex classification of recent transmission highlighted ongoing disparities in TB risk, and particularly the need to address Mtb transmission in high poverty neighborhoods. Second, as compared to previous studies, we developed more robust inferences concerning risk factors associated with recent, local transmission by using an integrative approach involving multiple data types. We believe that the more conservative spatial and temporal parameters we employed are well-suited to infer recent transmission in urban environments where the prevalence of Mtb infection (i.e. LTBI) is unknown.

One aim of our study was to develop a method that explores how integrating multiple types of data could improve identification of ongoing transmission in an urban community setting. Previous studies have demonstrated that the use of additional data sources to classify recent transmission may produce inferences that are closer to the "gold standard", i.e. shoe-leather epidemiological contact investigations (11, 13). Using this logic, we sought to make use of data on Mtb genotype, time of diagnosis, and spatial location of cases to provide a more robust proxy of recent transmission. We tested whether time-restricted genotypically clustered cases were more likely to be spatially clustered than non-genotypically clustered cases, and applied sensitivity analysis to assess the robustness of our working definition of spatially clustered cases. The results indicated that time-restricted genotypically clustered cases were nearly twice as likely to be spatially clustered than non-genotypically clustered cases. The value of this approach is that two complementary types of information were combined in hopes of increasing the accuracy of classifying recent transmission. Combining multiple lines of data improves our ability to understand or infer recent transmission, and thereby improves policies for prevention.

Several recent reports have documented the value of integrative approaches to classifying recent Mtb transmission. Moonan et al. used a method that first defined genotypically matched cases, and then identified geographic areas with higher than expected prevalence of genotypic clustering, which they defined as geospatial clustering (12). However, residence was represented by zip code, which would have missed finer-resolution spatial patterns. The report by France et al. that established the plausible-source case approach was a rigorous assessment of several different methods of classifying recent transmission (13). Notably, they compared their approach to the gold standard of epidemiologic field evaluation by optimizing sensitivity and specificity given varying geographic distances and disease prevalences. However, that study did not consider whether the geographic distance that is relevant for predicting transmission of Mtb may differ depending on other contextual features. For example, in places where much of the TB burden is concentrated in urban areas, a 10-mile radius may be too large a distance to recognize meaningful geospatial clusters of individuals with a shared environment conducive to transmitting Mtb. In comparison, our approach is likely more conservative using smaller geographic distances to determine geospatial clusters which may more accurately reflect the shared spatial environment of individuals leading to ongoing transmission of Mtb. Moreover, our analysis extends beyond these previous studies by performing both multivariable, individual-level and neighborhood-level risk factor analyses.

Notably, using WGS Guthrie et al. found the estimated transmission rate among pediatric cases was considerably lower than using MIRU-VNTR alone (15). WGS is a major development in TB molecular epidemiology and may represent the new standard for incorporating genomic data into studies of recent transmission of TB.

In sensitivity analyses, we examined the 20% of the sample that were time-restricted genotypically clustered, but not spatially clustered. Based on our findings, we believe these likely represent cases linked by remote transmission. For example, the transmission event may have happened sometime in the past, individuals dispersed and then reactivated at a later time point. However, given the unique natural history of TB infection and the complexity of TB pathogenesis, as well as the limitations in the discriminatory power of the genotyping methods, it is difficult to determine with certainty when transmission occurred.

Nearly half of cases in our study were neither geographically nor genotypically clustered, likely due to reactivation of LTBI. Notably, more foreign-born (59%) than U.S.-born cases were classified as non-clustered, which was confirmed in multivariable analyses. This finding is consistent with the hypothesis that foreign-born persons are more likely to be infected with Mtb in their country of origin, and then progress to active TB disease sometime after their arrival in the U.S. (26-28). Other studies have reported similar findings regarding foreign-born populations (11, 12, 16). One implication may be that TB control efforts should continue to focus on reducing activation of LTBI among foreign-born populations.

One of the key findings of our study was the importance of social context in sustaining Mtb transmission in the community. Markers of both individual- and neighborhood-level social context are critical, and interact with one another. One of our previous studies reported that U.S.-born, non-Hispanic Blacks were the highest risk racial/ethnic group for TB incidence (19), and that the association with race was largely explained by markers of neighborhood poverty (20). However, that report did not incorporate geospatial data. In the present analysis, we tested whether this trend held true when using a more nuanced measure of recent transmission that simultaneously also accounted for spatial context. Indeed, our findings have affirmed the importance of the neighborhood social environment in determining recent, local transmission.

Despite the strengths in our approach to classifying recent, local transmission, there are several limitations. First, the residential address may not be accurate for all cases. Additionally, while the use of GEE can account for correlations between individuals in the same block group, there could also be some residual dependency among blocks group captured within the same Sat Scan spatial cluster. This could result in biased standard errors—smaller than what would be expected if such dependency were corrected.

Although a more specific designation of recent, local transmission was devised, we are still not able to infer the direction of pathogen transmission. Also, while we believe that adding a time restriction to the genotypic cluster definition improved accuracy, some misclassification may still have occurred, particularly for those cases diagnosed at the beginning or end of the study period. There were also 535 individuals excluded from the study sample because of either missing genotypic information or address information. Notably, among the 535 missing, 16% were among individuals less than 18 years of age. Pediatric TB is widely considered a useful proxy for recent transmission as children (particularly those less than 5 years of age) have necessarily only been infected for short time (29-32). Thus, excluding those individuals likely leads to an underestimate of the prevalence of recent transmission in our sample.

Finally, recent studies have shed light on the limits of current genotyping methods in inferring recent transmission. Studies find that transmission clusters involving the Beijing family are particularly poorly resolved using MIRU-VNTR typing (33, 34). The Beijing genotype is the major component of the L2 lineage (35). In the current study, 60% of cases were from the EuroAmerican lineage (L4) and only 14% from the East Asian lineage (L2). Thus, we believe that based on genotyping alone, the MIRU-VNTR may be performing better than in other studies which have higher proportions of the L2 lineage. However, more importantly, these studies further highlight the necessity of approaches such as the current one. Incorporating multiple sources of data to infer recent transmission is critical, especially in contexts where WGS was not available.

CONCLUSION

As the U.S. moves towards the laudable goal of TB elimination, it is imperative to understand which factors contribute, and in which ways, to recent Mtb transmission. Our findings highlight the critical role that neighborhood social context continues to play in creating and/or maintaining disparities in TB incidence in the U.S. Improved TB control will require a deeper understanding of how complex social interactions occurring among individuals and as part of their environments are driving transmission.

List of Abbreviations:

ACS

American Community Survey

CDC

Centers for Disease Control and Prevention

CI

confidence interval

GEE

generalized estimating equations

LTBI

latent TB infection

MTB

Mycobacterium tuberculosis

RVCT

Report of a Verified Case of TB form

TB

Tuberculosis

U.S.

United States

WGS

whole genome sequencing