Tuberculosis treatment discontinuation and symptom persistence: an observational study of Bihar, India’s public care system covering >100,000,000 inhabitants

The study analyzes data from the Bihar Evaluation of Social Franchising and Telemedicine - Tuberculosis (BEST-TB) survey which received Institutional Review Board review and approvals from Duke University and Stanford University in the U.S. and from Society for the Promotion of Ethical Clinical Trials – Ethics Review Board (SPECT-ERB) in India. Research permissions were also granted by the Governments of India and of the State of Bihar. All subjects provided written informed consent.

The 2011 BEST-TB survey, a cross-sectional survey of TB patients treated in the public healthcare system in 11 districts of Bihar, India, provided data for the present analysis. The survey occurred in 360 rural clusters (Figure 1) whose sampling was determined as part of the baseline for a larger evaluation of interventions for a variety of diseases [34]. To represent patients receiving care in the public TB healthcare system, we constructed the sampling frame using TB patient lists from all Primary Health Centers (PHCs) within the study clusters as India’s Revised National Tuberculosis Control Programme (RNTCP) provides standard TB medications and care free of charge according to the Directly Observed Treatment Short Course (DOTS) protocol to an estimated 60% of TB cases [35‐39]. For potential inclusion in the study, enumerators recorded names and addresses of all TB patients entered on PHC lists within the past year (July 2010- July 2011) living within study clusters (i.e., village catchments representing part of the catchment served by a given PHC). Combining these names for each cluster generated the cluster’s sample list. From each cluster’s list, 6 patients were randomly sampled for study inclusion; all patients were included from cluster lists with fewer than 6 names. While the planned sample size was 2160, a review of all PHC records indicated that 20 clusters had no patients treated in the previous year and living within a study catchment area, and 22% of cluster lists had ≤6 names. Hence, 1691 patients were selected for subsequent household survey. There were no additional exclusion criteria.

Response rates to enumerators visiting and surveying each patient’s household were high. Just 44 households (2.6%) were not surveyed because they could not be located or no competent respondent was found. Of the remaining 1647 households, 5 refused (0.3%) and 68 reported that the TB patient had died (4.0%). Additionally, 28 patients provided incomplete data and were excluded. Household surveys were completed for 1546 patients (91.4% response rate of households contacted for inclusion). Because the present analyses focus on the treatment experience of patients in the RNTCP public system which according to DOTS treatment category should be at least 25 weeks in length, we restricted our analyses to patients initiating treatment at least 6 months prior to the interview (1007 patients).

Based on each TB patient’s responses to the questionnaire, we reconstructed a timeline of his or her illness and treatment relying solely on patient self-report. Because of the high variability of data recorded in the treatment records at the PHCs in our prior piloting of the survey, we did not use these records for our analysis. Instead, we relied solely on patients reporting their detailed experiences of common TB symptoms: persistent cough (≥3 weeks), coughing blood, fever, weight loss, night sweats, chronic fatigue, and loss of appetite [40, 41]. We recorded their reported number of days since the onset of each symptom and the number of days each persisted. We recorded the number of days each patient waited to seek care, before a diagnosis was received, and before beginning treatment. Finally, we recorded the number of days each patient reported remaining on treatment in the public system. We refer to the patient’s most recent episode of TB as the current episode, regardless of whether the patient remained ill at the time of survey. Patients also provided information about treatment for TB prior to the current episode, demographic characteristics, household composition and household assets.

Main outcomes were treatment discontinuation prior to 25 weeks (hereafter simply “treatment discontinuation”) and symptom persistence. Given policy interest in shorter treatment regimens, we repeated our analysis for treatment discontinuation prior to 16 weeks. Given policy interest in shorter treatment regimens and heterogeneity in current treatment regimens, we repeated our analysis for treatment discontinuation prior to 16 weeks for the entire sample and prior to 32 weeks for those with prior TB treatment episodes. Symptom persistence was defined as having ≥1 TB symptom 25 weeks after treatment initiation. We also explored an alternate model using the number of TB symptoms present at 25 weeks.

Main predictors of treatment discontinuation and symptom persistence were reporting treatment for a prior episode of TB and completion of prior TB treatment. It is unclear whether those with prior TB treatment are more or less likely to complete treatment for their current episode. If patients are already familiar with treatment, those choosing to repeat treatment may remain in care longer than first-time patients. However, patients who have undergone prior TB treatment may experience treatment fatigue and discontinue earlier. Failure to complete the prior TB treatment is hypothesized to predict early discontinuation from the current treatment episode. Because repeated treatment can select for MDR particularly if treatment is not completed and MDR is more difficult to cure [32, 42], we also hypothesize that both prior TB treatment and failing to complete prior treatment predict symptom persistence at the end of the current episode of treatment, while noting that there are other reasons for symptoms to persist (residual lesions even after cure; other respiratory ailments) which we cannot distinguish.

We predicted the likelihood of ending treatment before 25 weeks using a multivariate logistic regression model. We estimated the model using all 1007 patients controlling for prior TB treatment and treatment completion, repeating the analysis separately for 196 patients with prior TB treatment and for 811 with no prior treatment to examine differential effects of other predictors within these subgroups. The Supporting Information also reports a linear probability model of treatment discontinuation prior to 25 weeks and both logistic regression and linear probability models of treatment discontinuation prior to 16 weeks of treatment for the entire sample and prior to 32 weeks for those with prior TB treatment episodes. We likewise assessed determinants of symptom persistence 25 weeks after treatment initiation using a multivariate logistic model. The Supporting Information also reports a negative binomial model for the number of symptoms persisting 25 weeks post treatment initiation. For symptom persistence, we included the same predictors as for default along with duration in treatment covariates. We explored multiple definitions of duration of treatment including an indicator for completing ≥180 days of treatment; number of weeks of completed treatment; and indicators for completing 0-8 weeks, 9-16 weeks, and >16 weeks.

For all analyses, the multivariate models did not employ variable selection techniques but instead included all predictors for which there was a theory-predicted relationship to the outcome. Likewise, for all analyses, we included district-level fixed effects to control for unobserved characteristics of healthcare systems, geographic areas, and patient populations within each district. We also adjusted for survey design using inverse probability sampling weights and calculated robust standard errors clustered by sampling unit [59].

Individuals who did not complete treatment for prior episodes of TB have differentially higher probabilities of discontinuing treatment prior to 25 weeks in their current TB treatment episode compared to those with no prior episodes and those who completed treatment for their prior episodes (Figure 2). More than 69% [95% CI 50-89%] of the 39 patients who reported that they had not completed their previous TB treatment also failed to complete at least 25 weeks of their current treatment (median time to default 17 weeks) compared to 28% [95% CI: 16-40%] of the 157 patients who reported completing their previous treatment and 21% [95% CI: 17-26%] of the 811 patients treated for the first time (median time to default 25 weeks and 24 weeks respectively).

Individuals who had prior TB episodes but did not complete treatment have higher probability of discontinuing their current treatment episodes after adjusting for multiple factors (adjusted Odds Ratio (aOR): 4.77 [95% CI: 1.98 – 11.53]) (Table 2). Discontinuation probabilities of those who completed prior TB treatment are similar to those without prior treatment (approximately 1.05 after combining the aOR for prior TB with the interaction aOR for completing prior treatment: 0.22 [95% CI: 0.08 – 0.59]). Seeing more treatment providers for the current episode of TB predicts higher probabilities of treatment discontinuation (aOR: 3.67 per provider seen [95% CI: 1.94 – 6.95]). When one considers the effects of paying for treatment costs, travel costs, and the amount paid, paying anything predicts higher rates of treatment discontinuation, while those who are able to pay more and choose to do so are less likely to discontinue treatment.

Predictors of treatment discontinuation are distinct when the analysis is repeated separately for those with and without prior TB treatment. Significant predictors of higher probability of treatment discontinuation among those treated for TB previously are having not completed the prior episode of TB treatment along with personal characteristics including having fewer TB symptoms at the time of treatment initiation, younger age, less education, and being non-Hindu. In contrast, for those who had not been treated previously, significant predictors of greater probability of treatment discontinuation were not individual characteristics including number of symptoms as a marker for illness severity but rather care seeking pathways like the number of providers seen for the current episode of TB treatment and payment for treatment and related costs. The direction of these relationships generally remains consistent when analyses are repeated using a linear probability model though at times significance attenuates (see Additional file 1).

The predictors of treatment discontinuation within 4 months of starting treatment are consistent with the results of the main analysis – having had prior TB treatment that was not completed (aOR: 4.63 [95% CI: 1.88 – 11.42]), and seeing more providers (aOR per provider: 5.60 [95% CI: 2.18 – 14.37]) (see Additional files 2 and 3).

Symptoms persisted 25 weeks after treatment initiation for 28% of patients despite remaining in treatment for at least 25 weeks (Figure 3). Among patients discontinuing treatment early, symptoms persisted in 42% of patients completing ≤5 weeks, 33% of patients completing ≤10 weeks, and 30% of patients completing ≤18 weeks. Symptoms most likely to persist despite 25 weeks of treatment include fatigue (18% of patients), appetite loss (16%), night sweats (15%), weight loss (14%) and coughing (11%) (Figure 3). Persistence of symptoms followed different trajectories for those with and without prior TB treatment with symptoms more likely to persist for those with prior TB treatment (Figure 3).

TB symptom persistence 25 weeks post treatment initiation is more likely for those with prior TB treatment (aOR: 5.05 [95% CI: 1.90 – 13.38]), poorer patients (aOR for lowest wealth tertile among TB households: 2.94 [95% CI: 1.51 – 5.72]), and women (aOR of men compared to women: 0.56 [95% CI: 0.33 – 0.93]) (Table 3). For those with prior TB, the main predictors of symptom persistence are the number of treatment providers seen (aOR per provider seen: 3.76 [95% CI: 1.00 – 14.11]) (Table 3). For those with no prior TB, the main predictors of symptom persistence are not actually paying for treatments, medications and travel which is likely a marker of poverty in terms of the inability to pay, and poverty as explicitly measured by being in the lowest wealth tertile among TB households (aOR: 2.91 [95% CI: 1.50 – 5.68]) (Table 3). Notably, number of symptoms at treatment initiation did not predict symptom persistence.

Results for predictors of symptom persistence remain largely consistent when the number of symptoms present 25 weeks after treatment initiation is used as the outcome, when different specifications for time in treatment are used as predictors in both linear probability models and logistic regression models, and when we estimate a model of symptom persistence 16 weeks after treatment initiation (see Additional files 4,5 and 6).