Care Cascade for targeted tuberculosis testing and linkage to Care in Homeless Populations in the United States: a meta-analysis

We included studies reporting the results of TTT programs in adolescents and adults experiencing homelessness in the US. Programs needed to target homeless populations specifically, but could include a minority of participants who were stably housed. We defined study populations as homeless if participants were recruited from homeless shelters, health care clinics for homeless populations, or other service agencies primarily serving homeless clientele. We also included studies that described the study populations as homeless or used related terms such as unhoused, unstably housed, itinerant, street youth or other such terms (see Additional file 1).

Eligible studies must have used a biological test such as tuberculin skin test (TST) or interferon gamma release assay (IGRA) to ascertain LTBI. Eligible studies needed to report, at minimum, the numbers of participants with valid LTBI test results and numbers of those testing positive. We excluded studies that used only self-report to assess LTBI, as well as contact investigations, studies where stored specimens were tested for evidence of infection, and studies reporting testing results unlinked to patient identity. We excluded studies that tested participants only for active TB disease (i.e., not LTBI), such as those reporting only data from chest radiographs or sputum testing. We also excluded studies that primarily focused on populations aged 14 or under.

We did not have any eligibility restrictions based on publication status, study design, or language of publication. Single arm and multi-arm studies were eligible. Studies from peer reviewed literature or from conference abstracts and other grey literature were eligible for inclusion.

We developed a comprehensive search strategy that included multiple variations and synonyms of relevant TB and “homelessness” terms as well as National Library of Medicine Medical Subject Heading (MeSH) terms and Embase “Emtree” indexing terms. Because our original conception of this review also included studies addressing HIV, hepatitis B and hepatitis C in homeless populations, our strategy includes terms relevant to those conditions. We subsequently decided to report our review of studies concerned with these blood-borne infections in a separate manuscript [data analysis report provided to the US Centers for Disease Control and Prevention (CDC) on September 21, 2016]. In this review we focus only on TB studies.

We searched PubMed, Embase, Web of Science and the Cochrane Central Register of Controlled Trials, from earliest records to June 13, 2016. Additional file 1 provides our search strategies. We additionally hand-searched CDC’s TB Notes Newsletters and available abstracts from the National TB Controllers’ Association National TB Conferences and searched abstracts from the American Public Health Association’s annual conferences. We reviewed the bibliographies of our included studies as well as references cited in a previous systematic review which examined the prevalence of TB disease in homeless populations [10].

Two persons, working independently, applied eligibility criteria to titles and abstracts of all studies captured in the searches and each identified a selection of potentially-eligible studies. They then compared their respective selections and reached consensus about potential eligibility. A third author stood ready to serve as neutral arbiter in case consensus was not reached, but this was never necessary. Two persons then reviewed full-text articles of records deemed potentially eligible and in an identical process, made final decisions about study eligibility.

Outputs of interest in the TTT cascade were the number of participants who:

We also extracted the number of active cases that were diagnosed upon follow-up of positive TSTs or IGRAs, but did not consider it part of the “cascade” because our cascade ends with attendance for follow-up care, which must occur prior to TB disease or LTBI diagnosis. As this review focuses on testing for LTBI, cases of TB disease diagnosed by chest X-ray or sputum sample done in the absence of a positive TST or IGRA, were not extracted.

One person extracted key data into a pre-piloted and standardized data extraction spreadsheet (Additional file 2). A second person independently extracted data from the TTT care cascade, blinded to the first author’s extraction, and checked the remainder of the first extraction for accuracy. These independent extractions were compared and reconciled via consensus or by decision of the arbiter.

We did not formally assess risk of bias in the individual studies since existing standard instruments for assessing bias risk for intervention efficacy, prevalence, and other epidemiologic studies were not applicable to studies reporting yield of screening programs.

All analyses were conducted using Stata version 14. Proportions of persons proceeding from one step in the testing and linkage to care cascade to subsequent steps and associated 95% confidence intervals (CIs) were calculated using the Wald method. As we wanted to assess the performance of these programs in detecting new cases of LTBI, when studies reported testing persons with a known history of TB infection or TB disease, we subtracted their number from all cascade steps except the number reached. If two or more studies reported data for the same proportion, pooled proportions were calculated using inverse-variance random effects meta-analysis. The Freeman-Tukey double arcsine transformation was used to normalize individual study proportions prior to pooling. Where data existed, we also analyzed proportions stratified by recruitment method (health-care facility based recruitment vs. recruitment from other service agencies).

We calculated pooled cumulative proportions for each step in the cascade for the subset of studies that used TST for testing. Cumulative proportions were products of adjacent-step pooled proportions. The cumulative proportion tested of those reached was equal to the product of the proportion recruited of those reached and the proportion tested of those recruited. The cumulative proportion with valid test results of those reached was equal to the product of the cumulative proportion tested of those reached and the proportion with valid test results of those tested. Each subsequent cumulative proportion was the product of a single step proportion and the cumulative proportion that immediately proceeded it. We also calculated cumulative proportions stratified by test type, using test specific values for the proportion with valid test results of those tested and the proportion testing positive of those with valid test results. Confidence intervals for cumulative proportions were calculated using a simulation method. Each pooled proportion and confidence limit closest to 0.5 were normalized using the Freeman-Tukey double arcsine transformation, with a sample size equal to the harmonic mean of the individual study sample sizes and the number of successes equal to the product of the pooled proportion and the harmonic mean. For each pooled proportion, we took 50,000 draws from a normal distribution with a mean equal to the transformed pooled proportion and a standard deviation equal to the absolute value of the difference between the transformed mean and the transformed confidence limit divided by 1.96. Draws less than the transformed value of zero were replaced with the transformed value of zero, and draws greater than the transformed value of one were replaced with the transformed value of one. Draws were then reverse transformed into proportions to build probability distributions for each pooled proportions. Draws for each proportion were then multiplied with draws from the next proportion in the cascade in a sequential manner, and the 95% confidence intervals were taken from the 2.5th and 97.5th percentile of the resulting product distributions.

The time trend for the proportions testing positive of persons with valid results for studies using TST that included dates of data collection was analyzed using the Spearman’s rank correlation coefficient.

From a total of 3566 peer-reviewed citations identified through article databases and tracking of cited references of included studies in Scopus, 40 met inclusion criteria for either the TB or the blood-borne disease review, 21 of these reported outcomes of TB testing. Figure 1 details this process, and Additional file 3 lists reasons for exclusion after full-text review. In addition to 21 indexed publications, we found one eligible study cited in CDC’s peer-reviewed newsletter “TB Notes” and one conference abstract via our grey literature search. Descriptions of the included studies are in Table 1. Studies were conducted in 12 states and at least 15 cities (including six in New York City). With the possible exception of one study conducted in an unnamed county in Indiana [11], all studies took place in urban areas. Studies were published or presented between 1986 and 2014, with data collection reported to be between 1982 and 2009. Of 15 studies [12‐26] that reported gender distribution, 8 [12, 14, 17, 18, 21, 24‐26] had exclusively male study populations, the remainder had between 4% and 22% female study populations. In most studies that reported race, the majority of the populations were African American. Two studies used IGRA [27, 28], and one study used a combination of TST and IGRA [11] for testing, the remainder relied exclusively on TST.

Twenty studies [12‐26, 29‐34] used TST for testing. Additional file 4: Table S1 reports proportions of persons proceeding from each cascade step to subsequent steps, and the number of studies contributing data on each proportion. Of persons reached, 93.7% (95% CI 72.4 to 100%) were recruited; of persons recruited, 97.9% (89.3 to 100%) had tests placed; of those with tests placed, 85.5% (78.6 to 91.3%) returned to have tests read; of those with tests read, 99.9% (99.6 to 100%) had valid test results; and of those with valid test results, 24.7% (21.0 to 28.5%) tested positive. Of four studies [12, 14, 19, 26] that reported number given a referral for follow-up, all persons who tested positive were referred, and 99.8% of persons who agreed to be referred to care attended at least once. Among the eleven studies [12, 14, 15, 19, 21, 23, 25, 29, 31‐33] that reported number of TB disease cases diagnosed among persons tested for TB infection, 1.2% (0.0 to 3.4%) of persons testing positive were diagnosed with TB disease. The proportion of variability in the effect estimates that is due to heterogeneity (I²) was greater than 75% for all proportions where I² was calculable, except the proportion referred to treatment of those testing positive (I² = 0), for which all study estimates were 100%, and proportion attending treatment of those referred to treatment (I² = 59.1%).

Adjacent step and cumulative proportions in the testing and linkage to care cascade are shown in Fig. 2. For a hypothetical cohort of 1000 homeless persons reached in studies using TST, we estimate that 784 (95% CI 585 to 880) would have valid test results, 194 would test positive (140 to 234), all 194 (139 to 234) would be referred to follow-up and 193 (134 to 231) would attend follow-up.

The proportions testing positive declined over time (Spearman’s ρ = − 0.57, p = 0.046). Figure 3 displays this finding graphically.

Two studies that relied exclusively on IGRA [27, 28], recruited participants from health care facilities in San Francisco. Neither study reported numbers reached, recruited, or referred to follow-up. The pooled proportion of persons with valid test results of those tested was 97.0% (95% CI 96.8 to 97.3%) and the pooled proportion testing positive of those with valid test results was 8.9% (8.4 to 9.4%). One of these studies reported that 274 (67%) of 411 persons testing positive attended follow-up care [27].

One study [11] relied on a combination of TST and IGRA for testing without stratifying results by test type, but appears to have used mostly IGRA. In this study, of 1421 persons with valid test results, 185 (13%) tested positive, and 158 attended follow-up care.

Table 2 displays pooled proportions stratified by recruitment type for studies using TST for testing. Persons recruited at health care facilities (usually clinics for homeless persons) were retained at higher proportions for the first three steps of the cascade than persons recruited from other types of facilities. This difference is only statistically significant at α = 0.05 for recruited of reached.

Meta-analyses often include a formal assessment of study bias. The Cochrane recommended tool for bias assessment was developed for the purpose of evaluating studies that compare an intervention with a control condition. This tool was considered inappropriate for the evaluation of this study because this analysis does not seek to compare one condition to another. We are unaware of any widely used and well accepted instrument for the evaluation of biases in studies that seek to estimate the yield of disease screening programs, therefore we did not conduct a formal bias assessment of individual studies. However, we can speculate on the types of biases that may have affected our results. A likely major source of bias is publication bias. The results of TTT programs for homeless persons are rarely published because assessments done by busy public health workers are intended only to inform the program. It is possible programs that publish or otherwise disseminate their results differ from programs that do not. For example, programs may publish about exceptionally successful programs, or conversely, publish about programs that faced exceptional challenges. Funnel plots of proportion with tests read of those with tests placed and proportions with positive tests of those with valid results for TST studies (Additional file 5) did not reveal any obvious asymmetry, but funnel plot asymmetry is largely due to the use of null-hypothesis significance testing. Because the included studies were not comparative in nature, there was no significance testing for the individual study proportions and publication bias may shift the entire distribution rather than causing asymmetry. Another consequence of incomplete reporting of targeted testing programs is that it is not possible to know what proportion of the homeless population is currently reached by these programs, which in turn makes it difficult to estimate the possible effects of scaling up testing in this population.

Misclassification is another potential source of bias; numbers of persons who were retained at each step of the cascade may not have been recorded correctly, and mistakes may have been made in interpreting test results. Studies did not generally provide sufficient evidence to evaluate the risk of bias due to misclassification.

In 2015, 66.4% of reported cases of TB disease in the United States were in persons who were born in another country, and TB incidence was nearly 13 times higher in persons born outside the US than those born in the US. Per data from the Online Tuberculosis Information System, 30.5% of recently homeless persons with TB diseases between 2011 and 2015 were also born outside the US [41]. However, we found very little information in our included studies regarding the intersection of these two critical risk populations in our included studies. Only one study stated the proportion of the study population that was born abroad; this study found that 17.3% of 260 homeless persons tested for TB in Los Angeles in the early 1990’s were non-US born [16].

There are also issues that limit the generalizability of our results. We included 17 studies conducted wholly or in part prior to 2000. The preponderance of older studies limits generalizability due to possible changes in LTBI prevalence. While the proportion of reported cases of TB disease in persons experiencing recent homelessness has remained at about 6% since the mid-1990s, the number of incident cases of TB disease in the United States have decreased in both the general public and homeless populations, with only 495 cases reported in 2015 in persons experiencing homelessness in the last year vs. 795 cases reported in 2005 [2]. The decline in cases in persons experiencing homelessness may be partially due to the active case-finding activities that have been advocated for these populations since the 1990’s [34]. We also found that studies were mainly conducted in coastal urban areas, meaning that our results may generalize poorly to testing programs for homeless persons in other settings.

Most studies in this review relied on TST, rather than IGRA for testing [42], and the studies that did use IGRA reported a limited number of cascade steps. In addition, the two studies that used IGRA exclusively were both conducted in San Francisco, and both recruited from healthcare facilities, limiting generalizability of results. Given these limitations, we do not think it is appropriate to use the results of this review to compare relative performance of TST with IGRA. The two studies which relied exclusively on IGRAs for testing did not report the number of persons (either those testing positive or overall) who received their test results, but in a separate review [unpublished] of targeted blood-borne infection testing for homeless persons, we estimated that only 57.9% of persons testing positive for HCV infection and 50.8% of persons testing positive for HIV infection were given test results. Additional studies are needed to assess LTBI test positivity and referral for follow-up care using the newer IGRA diagnostics.

Almost all of the studies included in this review recruited participants from health care facilities, shelters, or other types of service agencies. Only two [11, 16] included persons recruited through outreach or street venue based sampling. Because some homeless persons cannot or do not access services, most of the programs we reviewed likely left a large subpopulation of homeless persons without access to testing. We have conducted some stratified analyses by the type of service agency where recruitment was done, but failure to find a significant difference between proportions should be interpreted with caution because tests for heterogeneity are low-powered and frequently fail to reject false hypotheses.

We found high heterogeneity in most of our pooled proportions, indicating a high probability that study results differed from one another due to underlying program factors, rather than random variability. The random effects model that we used to calculate pooled proportions explicitly allows pooling of results sampled from differing underlying probability distributions. However, since our results indicate that real-world performance is likely to differ substantially between programs, program planners, modelers, and others who wish to incorporate our results into their work should consider our confidence intervals, in addition to our point estimates.