Population impact of fine particulate matter on tuberculosis risk in China: a causal inference

Abstract

Background: Previous studies have suggested the potential association between air pollution and tuberculosis incidence, but this association remains inconclusive and evidence to assess causality is particularly lacking. We aimed to draw causal inference between fine particulate matter less than 2.5 μm in diameter (PM_2.5) and tuberculosis in China.

Methods: Granger causality (GC) inference was performed within vector autoregressive models at levels and/or first-differences using annual national aggregated data during 1982-2019, annual provincial aggregated data during 1982-2019 and monthly provincial aggregated data during 2004-2018. Convergent cross-mapping (CCM) approach was used to determine the backbone nonlinear causal association based on the monthly provincial aggregated data during 2004-2018. Moreover, distributed lag nonlinear model (DLNM) was applied to quantify the causal effects.

Results: GC tests identified PM_2.5 driving tuberculosis dynamics at national and provincial levels in Granger sense. Empirical dynamic modeling provided the CCM causal intensity of PM_2.5 effect on tuberculosis at provincial level and demonstrated that PM_2.5 had a positive effect on tuberculosis incidence. Then, DLNM estimation demonstrated that the PM_2.5 exposure driven tuberculosis risk was concentration- and time-dependent in a nonlinear manner. This result still held in the multi-pollutant model.

Conclusions: Causal inference showed that PM_2.5 exposure driving tuberculosis, which showing a concentration gradient change. Air pollutant control may have potential public health benefit of decreasing tuberculosis burden.

Keywords: Causality; Eco-driver; Empirical dynamic modeling; PM2.5; Tuberculosis.

Fig. 1

Methodology flowchart of the causal inference study on PM_2.5 and TB.

Fig. 2

Cross-map causality of PM_2.5 on tuberculosis. (A) Cross-map causality beyond shared seasonality of ambient PM_2.5 on tuberculosis based on univariate SSR. The box-and-whisker plots show the null distributions for cross-map skill (ρCCM) expected from random surrogate time series which share the same seasonality as the true PM_2.5 concentration. Red circles demonstrate the unlagged ρCCM for observed TB predicting purported PM_2.5. Filled circles indicate the significant ρCCM (P ≤ 0.1). Provinces are ordered according to their latitudes. (B) Forecast improvement with multivariate SSR is quantified using ΔρCCM = ρCCM (with PM_2.5) - ρCCM (without PM_2.5). Wilcoxon signed-rank exact test reveals a significant difference. (C) Effect of PM2.5 on TB (ΔTB/ΔPM_2.5) for each province. In the scenario analysis, PM_2.5 shows a positive effect on TB incidence for 22 provinces (P ≤ 0.1). (D) Range of ΔTB/ΔPM_2.5 as a function of PM_2.5 grouped over all provinces. SSR, state-space reconstruction; CCM, convergent cross-mapping; ρCCM, the Pearson correlation between observations and CCM prediction

Fig. 3

Exposure-response relationship between PM_2.5 and tuberculosis incidence in single-pollutant DLNM model. (A) Three-dimensional plot: the height of the hexahedron represents RR for the association between TB incidence and ambient PM_2.5 exposure, while two bottom edges represent the full range of monthly mean PM_2.5 concentration and the number of months delayed. (B) Contour plot: the red color gradient represents RR > 1, and the blue gradient represents RR < 1. (C) Cumulative effects of PM_2.5 exposure for 15 months. (D-E) Pooled and cumulative effects with 10 µg/m³ increase in PM_2.5 throughout 0–15 months. The reference level of PM_2.5 is set as 15 µg/m³. Monthly mean temperature, precipitation and sunshine duration, and annual population density, GDP per capita, certified doctors and beds of medical institutions are added as time-varying local control variables. TB, tuberculosis; RR, relative risk