Association between ambient air pollutants and upper respiratory tract infection and pneumonia disease burden in Thailand from 2000 to 2022: a high frequency ecological analysis

Disease surveillance data was obtained from Thailand’s Ministry of Public Health (MOPH) disease surveillance system [17] where we obtained reported disease case counts from 2000 to 2022 at the province level. Suspected cases of pneumonia and influenza in provincial public health offices, hospitals and all health stations were reported through the disease surveillance system to the Bureau of Epidemiology, MOPH. Suspected cases were defined as patients who had symptoms matching the clinical criteria. Private hospitals were not covered by the surveillance system. Diagnoses were recorded using 10th revision of the International Classification of Disease (ICD) codes. As Bueng Khan was split from Nong Khai in 2011 to form a separate province, we have merged disease case counts from Bueng Khan back to Nong Khai from 2011 onwards, to allow consistent analysis over the timeframe of the dataset.

Data on annual population size for each province from 2000 to 2022 was obtained from the Official Statistics Registration Systems of Thailand [18]. Similarly, as Bueng Khan was split from Nong Khai in 2011 to form a separate province, we have merged the population numbers from Bueng Khan back to Nong Khai from 2011 onwards, to allow consistent analysis over the timeframe of the dataset.

Climate data was obtained from ERA5, published by the European Centre for Medium-Range Weather Forecasts [19]. Each data point covers a 30 km grid, which we spatially averaged across each province. Mean, median and maximum of total precipitation, vegetation index, air temperature at 2 m and dew point temperature at 2 m was collected. Relative humidity and average humidity were calculated using standard formula.

Ambient air pollutant data was obtained from NASA's Goddard Earth Sciences Data and Information Services Center, GES DISC. The 1-Hourly CO Column Burden, CO Surface Concentration, and 1-Hourly Aerosol diagnostics were obtained from Global Modeling And Assimilation Office [20] to derive PM_2.5 surface concentration. The surface model layer of the 3d 3-Hourly Aerosol Mixing Ratio was obtained from Global Modelling and Assimilation Office [21] to derive PM₁ and PM₁₀ surface concentration. These datasets are a part of Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA 2), which is a reanalysis of the modern satellite data, and is produced by NASA’s Global Modelling and Assimilation Office (GMAO) [22].

Consider reported disease case counts \({y}_{\tau +1}\) which is observed at the discrete time point between \(\tau\) and \(\tau +1\) and supposed that we have additional information arising from a set of V predictors \({{\varvec{x}}}_{\tau }^{(m)}=( {{\varvec{x}}}_{1}^{\left(m\right)}, {{\varvec{x}}}_{\tau }^{\left(m\right)},\dots , {{\varvec{x}}}_{V}^{(m)})\) which are observed m times between \(\tau\) and \(\tau +1\). The variables \({y}_{\tau +1}\) and \({{\varvec{x}}}_{\tau }^{(m)}\) can thus be said to be observed at different frequencies. The traditional and simplest manner of dealing with mixed frequency data would be averaging the high frequency predictors to the same frequency of the dependent variable \({y}_{\tau +1}\). However, this approach may result in omitted variable bias and model mis-specification if the true weighting scheme is not a simple average [11]. The mixed frequency data sampling (MIDAS) framework provides an alternative estimator that can account for different weighting schemes in the high frequency predictors. Specifically, the approach plugs in the high frequency lagged terms of predictors \({{\varvec{x}}}_{\tau }^{(m)}\) in a regression for the low frequency dependent variable \({y}_{\tau +1}\) as follows:where \({y}_{\tau -j}\) denotes past disease case counts for a maximum of \({p}_{y}-1\) lags. We can additionally place \(S\) exogenous variables \({z}_{\tau -j,s}\) recorded at time \(\tau -j\) with the same frequency as \(y\) for a maximum of \({p}_{z}-1\) lags. The high frequency predictors \({x}_{\tau ,v }^{(m)}\) are shrunk to the same frequency as \(y\) and \(z\) by a polynomial term \(B\) to be defined later. \(\alpha\) is an intercept term, \({p}_{y}\) autoregressive terms, \({\gamma }_{j+1 ,s}^{^{\prime}}\) coefficients denoting the effect of past \({z}_{s}\) on \(y\) and \({\beta }_{v}\) are coefficients which capture the overall effect of \(B\left({L}^\frac{1}{m};{{\varvec{\theta}}}_{v}\right)\) on \({y}_{\tau +1}\). We further assume i.i.d and normally distributed errors, with mean zero and finite variance \({\sigma }_{\epsilon }^{2}\):

The polynomial term is given by:where \({L}^{k/m}\) is a lag operator such that \({L}^{1/m}={x}_{\tau -1/m}^{(m)}.{\varvec{\theta}}_{v}\) are parameters which provide the shape of the polynomial term. We use the normalized beta probability density function on the polynomial term \(B\left(k,{{\varvec{\theta}}}_{v}\right):\)where \({x}_{k}=(k-1)/(K-2)\). The beta polynomial was used as it only requires two parameters \({{\varvec{\theta}}}_{v}\in \{{\gamma }_{1},{\gamma }_{2}\}\) to specify and generate a large variety of weighting shapes [23]. In particular, further restrictions such as {\({\gamma }_{1}\) = 1,\({\gamma }_{2}\)} or {1 + \({\gamma }_{1}\), 1 +\({\gamma }_{1}\)  + \({\gamma }_{2}\)} allow the weighting structure to form only downward sloping and hump-shaped weights respectively, which were also explored.

In order to implement MIDAS, we consider the Bayesian paradigm and estimate parameters based on Markov chain Monte Carlo (MCMC) methods, which alleviates certain issues in frequentist parameter estimation. First, uncertainty and standard errors can be characterized easily using credible intervals. Second, the MCMC approach facilitates easy modification to allow shrinkage and sparsity, thereby allowing unimportant variables to be shrunk closer to zero.

Briefly, we first placed diffuse, conjugate priors on both the parameters of the observation equation \({\{p}_{j+1},{\gamma }_{j+1 ,s}^{^{\prime}},{\beta }_{v}\}\) and error variance \({\sigma }_{\epsilon }^{2}\). Gibbs sampling was used as conditional posterior distributions for these parameters were well defined. We also modified the baseline procedure of sampling parameters in the observation equation to induce shrinkage, through incorporating the structure of Bayesian Lasso [24]. This comprised changing the prior for the variance of the error term, to enable more probability weight to be placed on 0 for the parameters of the observation equation.

Following Ghysels [23], a Gamma distribution is placed as a prior on the parameters of the polynomial terms {\({\gamma }_{1},{\gamma }_{2}\}\) with both shape and scale parameters being 1, which amounts to a flat weighting scheme that puts equal weight on the high frequency data. The polynomial terms were sampled using a random walk Metropolis-in-Gibbs step. The Metropolis step is an accept-reject step which takes a candidate draw from some proposal distribution, with acceptance for that draw is given by a probability that depends on the likelihood, parameter’s prior distribution and the proposal density. In this case, we use the Gamma proposal distribution as a proposal density, as it corresponds to the functional form of the MIDAS weighting polynomial. Full details for the estimation strategy are outlined in the supplementary information.

Convergence of MCMC chains was first assessed by visual inspection of trace plots and Gewecke convergence diagnostic checks. Residual autocorrelation is computed for up to 20 week lags to ensure that the transmission dynamics are properly accounted for in each specification, across models. Quantile–quantile plots are used to see whether the specifications adequately account for the data structure. We additionally computed 4 other statistics on in-sample model fit, namely, the \({R}^{2}\) and in-sample mean-squared error to look at whether overall variations in data are captured, the adjusted \({R}^{2}\) to look at whether overall variations in data are captured while penalizing larger models. Lastly, the deviance information criterion [25] look at whether overall variations in data are captured while penalizing larger models with larger uncertainty in posterior distributions.

After adjustment for confounders, SO_2, PM_2.5, and CO over a 40-day period were estimated to be positively associated to monthly URTI case counts in 41, 15 and 31 of 76 provinces, with their 95% credible intervals away from zero. In other provinces, ambient air pollutants were negatively associated to monthly URTI case counts. On average, greatest influence of ambient air pollutants were estimated to be around the 20^th day mark across provinces (Fig. 1A1-3), but the range of values differed greatly, where the 7 – 36, 6 – 41 and 7 – 40 day measurements for SO₂, CO and PM_2.5 (Fig. 1) were ascribed the greatest influence on case counts across provinces respectively. Exposure–response shapes were also highly heterogenous across provinces but were mainly hump-shaped across the 40-day period (See supplementary material).

Whereas PM_2.5, SO₂ and CO over a 40-day period were positively associated to monthly pneumonia case counts in 15, 31 and 49 of 76 provinces (Fig. 2A1-3, B1-3, C1-3). Similar to URTI, we found on average that the greatest influence of ambient air pollutants were estimated to be in the 20^th day mark across provinces, but the range was noticeably narrower compared to case of URTIs, at the 0 – 32, 3 – 30 and 7 – 33 day mark for SO₂, CO and PM_2.5 respectively. Similarly, exposure–response shapes were highly heterogenous across provinces for ambient air pollutants and pneumonia case counts. These shapes were however mainly hump-shaped across the 40-day period (See supplementary material).

Controlling for all other variables, on average, a 1 μg/m³ rise in SO₂ over the past 40 days would lead to an expected 1624.1 (Fig. 3, A2, B2, C2) increase in URTI case counts. However, an expected change of -15.1 (Fig. 3, A1, B1, C1) and -406.3 (Fig. 3, A3, B3, C3) in URTI case counts respectively are expected to occur with a 1 ppb rise in CO and 1 μg/m³ rise in PM_2.5 surface concentrations over the past 40 days respectively across provinces.

Whereas a 1 ppb rise in CO surface concentrations over 40 days would lead to an expected 39.3 change (Fig. 4, A1, B1, C1) in monthly pneumonia case counts, a 1 μg/m³ increase in SO₂ and PM_2.5 surface concentration would lead to an expected -943 (Fig. 4A2, B2, C2) and -401.8 (Fig. 4) change in monthly pneumonia case counts across provinces respectively. Comparing the range of posterior mean effect sizes and quantile values across provinces for major ambient air pollutants on monthly pneumonia and URTI case counts showed that associations between ambient air pollutants on pneumonia were far more variable compared to that of URTI (Figs. 3 and 4).

Across provinces, controlling for past disease case counts and the effects of ambient air pollutants, increases in climate measurements the past two months have mixed associations to monthly URTI and pneumonia case counts. In particular, an increase in mean absolute humidity over the past 1 and 2 month was associated to decreases in contemporaneous URTI case counts in 22 and 9 of 76 provinces respectively. Whereas an increase in mean relative humidity 1 and 2 months prior was associated to decreases in contemporaneous URTI case counts in only 2 and 1 provinces respectively. However, in 8 provinces, increases in mean relative humidity 1 month prior was associated to increases in contemporaneous URTI case counts. In all provinces, higher temperature 1 and 2 months prior were associated to decreases in URTI in all provinces. Lastly, increases in total precipitation the past month were associated to increases and decreases in contemporaneous conjunctivitis case counts in 21 and 10 provinces respectively.

Similarly, increases in temperature for the past 1 and 2 months were associated to decreases in pneumonia case counts in 74 and 73 provinces respectively. Increases in absolute humidity the past 1 and 2 months were associated to decreases in contemporaneous pneumonia case counts in 16 and 11 provinces respectively. Whereas in only 3 provinces, changes in relative humidity were associated to any increase or decrease in contemporaneous pneumonia case counts.

Comparing the contributive effect of past disease case counts, ambient air pollutants and climate on disease case counts suggests that the mean impact of past climate variables across provinces on contemporaneous URTI case counts is minor, ranging from -0.7 – 0.17% (Fig. 5A2), while the mean impact of past climate variables across provinces on pneumonia case counts is higher at -27.8% – 15.8% (Fig. 5B2). Whereas the impact of ambient air pollutants were greater in influencing contemporaneous URTI and pneumonia case counts at -14.8% – 4.96% (Fig. 5A1, B1) and -79.78% – 107.5% (Fig. 5A3, B3) respectively. Lastly, the impact of past case counts in influencing contemporaneous URTI and pneumonia case counts at -17.4% – -0.16% (Fig. 5A1, B1) and -118.4% – 17.9% (Fig. 5A3, B3) respectively. For the case of pneumonia, wide estimates for contributive effect were partially attributed to the timepoints where case counts were low in some provinces (See supplementary material) and model predictions were pulled in separate directions by case count and ambient air pollutant measurements.

All analysis was repeated using case counts normalized by population size. That is, disease case counts per 100,000 individuals was used as the dependent variable of interest for the analysis. Similar to the analysis using raw disease case counts as the dependent variable of interest, we found that CO and SO₂ surface concentrations the past 40 days were more likely to be positive associated to increases in disease case counts for the contemporaneous month. Mean duration where the pollutants were taken to be most important by MIDAS weights were also around the 18 – 21th day mark. The contributive burdens of different confounders on disease case counts were also similarly found to be higher on past disease case counts and ambient air pollutant measurements, rather than climate (See supplementary material).

Quantile–quantile plots showed that models considered, with or without MIDAS terms, in general could capture distributional characteristics of disease case counts well. Visual diagnostics of trace-plots also showed that MCMC procedures have converged for all parameters in each model.

Overall, we found that incorporating MIDAS weights provided overall improvements over the baseline model where no MIDAS weights were incorporated, with overall lower model mean absolute errors, deviance information criterion and coefficients of determination. Furthermore, the model incorporating no restrictions on MIDAS weights worked better compared to models incorporating downward sloping or hump-shaped weights. These models had better model fit as demonstrated by the mean-squared error as well as the deviance information criterion in 40.8% and 37.5% of provinces across all models considered. Imposing restrictions for MIDAS weights to be downward sloping weights also performed reasonably well, as demonstrated by their comparatively better mean-squared error as well as the deviance information criterion in 34.9% and 40.8% of provinces across all models considered.

Incorporating variable shrinkage to the overall regression specification also improved model performance in terms of mean-squared error, unadjusted and adjusted coefficients of determination as well as the deviance information criterion in almost all provinces. These tests are reported in the supplementary information section.

Given these diagnostic tests, the model incorporating shrinkage in regression parameters, as well as no restrictions on MIDAS polynomials was taken as representative to infer exposure–response curves between ambient air pollutants and URTI/pneumonia case counts across all provinces.