This is a retrospective cohort of TB patients diagnosed from 2011 to 2015, based on surveillance data from the Tuberculosis Cases Notification and Monitoring System (TBWEB) updated in May 2017. This system only covers São Paulo state, representing approximately 25.5% of Brazilian new TB cases [
2]. TBWEB records data collected by health professionals in a standardized TB notification form and allows follow-up on TB outcomes even if patients are transferred among health facilities. As a result, only a few cases in the study period did not have outcome data; when it did happen, it was mainly related to transfers out of São Paulo state.
Study population
We included new TB cases (never treated before or those that previously had received TB treatment for a period ≤1 month), aged ≥15 years, and who started TB treatment during the study period. We intended to focus the study on drug-sensitive TB cases; therefore, we excluded those patients recorded as having TB drug-resistant. From the possible study population of 79,075 patients, 1863 (2.4%) patients did not have a TB outcome recorded. Therefore, in total, 77,212 patients were available for the outcome analysis.
As we included all patient information available during the study period, a sample size was not calculated in advance. Given the number of patients included, the frequency of exposures to substances (Table
1) and an incidence of the outcome in the unexposed group of 14.3%, this study had a power higher than 99% to evaluate associations with a relative risk (RR) of 1.2 or greater for any of the exposure categories (alcohol, drugs or both).
Table 1
Factors associated with unsuccessful tuberculosis treatment outcome, São Paulo-state, Brazil, 2011–2015
Overall | 77,212 | 13,300 | 17 | | |
Alcohol and illicit drug |
Neither | 60,107 | 8607 | 14.3 | 1 | 1 |
Only alcohol disorder | 7487 | 1765 | 23.6 |
1.65 (1.57–1.72)
|
1.48 (1.4–1.56)
|
Only drug use | 5199 | 1472 | 28.3 |
1.98 (1.89–2.07)
|
2.1 (1.98–2.21)
|
Alcohol disorder and drug use | 4418 | 1456 | 33 |
2.3 (2.2–2.41)
|
2.09 (1.97–2.21)
|
Sex |
Female | 22,498 | 3102 | 13.8 | 1 | 1 |
Male | 54,714 | 10,198 | 18.6 |
1.35 (1.3–1.4)
|
1.27 (1.22–1.32)
|
Age (years) |
15–34 | 36,080 | 5370 | 14.9 | 1 | 1 |
34–49 | 21,681 | 4068 | 18.8 |
1.26 (1.21–1.31)
| 0.99 (0.94–1.03) |
50 and more | 19,451 | 3862 | 19.9 |
1.33 (1.28–1.38)
|
1.23 (1.18–1.29)
|
Race |
Non-black | 35,707 | 5727 | 16 | 1 | 1 |
Black | 33,135 | 5951 | 18 |
1.12 (1.08–1.16)
|
1.13 (1.09–1.17)
|
HIV |
No | 69,973 | 10,582 | 15.1 | 1 | 1 |
Yes | 7239 | 2718 | 37.5 |
2.48 (2.4–2.57)
|
2.16 (2.07–2.26)
|
Clinical form |
Pulmonary | 65,271 | 11,364 | 17.4 | 1 | 1 |
Extrapulmonary | 11,903 | 1899 | 16 |
0.92 (0.88–0.96)
|
0.86 (0.82–0.9)
|
Prison |
No | 68,228 | 12,608 | 18.5 | 1 | 1 |
Yes | 8984 | 692 | 7.7 |
0.42 (0.39–0.45)
|
0.51 (0.47–0.56)
|
Homeless |
No | 75,218 | 12,278 | 16.3 | 1 | 1 |
Yes | 1993 | 1022 | 51.3 |
3.14 (3–3.29)
|
2 (1.89–2.13)
|
Directly observed treatment |
No | 16,791 | 4237 | 25.2 | 1 | 1 |
Yes | 54,940 | 7198 | 13.1 |
0.52 (0.5–0.54)
|
0.55 (0.53–0.57)
|
Variables
Using WHO definitions as adapted to TBWEB, the main TB treatment outcomes were classified as treatment success (cured or completed treatment) or unsuccessful TB treatment outcome (death and lost to follow-up) [
10].
Alcohol disorder, the use of illicit drugs and an interaction between them were analysed as the main independent variables. Alcohol disorder is collected in the TB notification form as the presence or not of an associated disorder and is categorized as yes or no. For the definition of alcohol disorder, Brazil adopts the American Psychiatric Association definition [
11]. The use of illicit drugs is also collected in the notification form as a dichotomous variable (yes or no).
Substance exposures were identified by health professionals during clinical interviews. Thus, substance exposures were routinely recorded at the time of diagnosis, which is before the start of treatment.
Statistical analysis
We estimated the RRs and their 95% confidence intervals (95% CI) in a multiple model using Poisson regression with robust variance [
12]. We obtained a model with the following structure:
$$ Ln(y)={\beta}_0+{\beta}_a{X}_a+{\beta}_d{X}_d+{\beta}_{a\wedge d}{X}_{a\wedge d}+\sum \limits_{i=1}^k{\beta}_i{C}_i $$
where
y represents the predicted value of the dependent variable;
β0 is the intercept;
βa,
βd and
βa ∧ d represent the regression coefficients of the dichotomous independent variables (adopting values of 0 [no] or 1 [yes]) that correspond to exposure to alcohol (
Xa), to illicit drugs (
Xd) and to the interaction term defining the concomitant exposures to both alcohol and drugs (
Xa ∧ d) respectively.
This model was adjusted by a number (
k) of covariates (
Ci) with their corresponding coefficients (
βi). We considered the following covariates, which have been recognized as factors associated with TB outcome in previous studies [
5‐
7,
9]: sex; age group (15 to 34; 35 to 49, 50 and older); race (blacks and non-blacks); HIV status; clinical form of TB (pulmonary [includes mixed form] and extrapulmonary); prisoner; homeless population; and directly observed treatment (DOT).
From this model, compared with the category of non-exposed to either alcohol or drug, the adjusted RRs for only alcohol (RRa0), only drug (RR0d) and for the exposure to both substances (RRad) were calculated respectively as: RRa0 = exp(βa); RR0d = exp(βd); and RRad = exp(βa + βd + βa ∧ d).
We calculated the ratio of RRs as a measure of interaction on the multiplicative scale, consisting of the ratio of
RRad to the RR expected from the product of the effects of the two exposures considered separately [
13]. Then,
$$ ratio\ of\ RRs=\frac{RR_{ad}}{RR_{a0}{RR}_{0d}}=\exp \left({\beta}_{a\wedge d}\right) $$
We also calculated the relative excess risk due to interaction (RERI) as a measure of additive interaction [
13] as follows:
RERI =
RRad − (
RRa0 +
RR0d − 1) =
RRad −
RRa0 −
RR0d + 1.
With a ratio of RRs > 1 and RERI> 0, we would consider an interaction as positive in the multiplicative and the additive scales, respectively. However, in the case of a ratio of RRs < 1 and a RERI< 0, we interpreted the interaction as negative and calculated how much lower the RRad was than expected in both scales. Thus, we defined that proportion lower than expected in the multiplicative scale as: \( 1-\frac{RR_{ad}}{\left({RR}_{a0}{RR}_{0d}\right)} \); and in the additive scale as: \( 1-\frac{RR_{ad}}{\left({RR}_{a0}+{RR}_{0d}-1\right)} \).
Subsequently, we simulated the incidence expected from the adjusted RR using the incidence observed in the population not exposed to either alcohol disorder or use of illicit drugs as a reference. These incidences were then compared with those expected in additive and multiplicative scales, based on the sum of attributable risk and the product of RRs respectively. We followed the recommendations made by Knol and VanderWeele [
14] to present interaction analyses.
Finally, we estimated the population attributable fraction (PAF) [
15] of the use of substances for the risk of unsuccessful outcome of TB treatment using the formula:
$$ \mathrm{PAF}={p}^{\hbox{'}}\frac{\theta -1}{\theta } $$
where p’ is the prevalence of exposure to substances in the non-successful treatment population, and θ is the adjusted RR estimated by the regression model. PAF and interaction measures derived from the regression model were calculated using a nonlinear combination of parameter estimates based on the delta method [
16].
To evaluate whether missing outcomes could affect the results, we first compared profile patients with and without treatment outcome registration using the Pearson Chi-square test (Additional file
1: Table S1). Because those populations were slightly different, we ran a sensitivity analysis with all unrecorded treatment outcomes set to either successful or unsuccessful (Additional file
2: Table S2).
All analyses were performed on Stata 12 (Stata Corporation, Texas USA).