The impact of the synergistic effect of SO(2) and PM(2.5)/PM(10) on obstructive lung disease in subtropical Taiwan.

SUBJECT:chen-twb-participants

Taiwan Biobank volunteers linked to the Taiwan Air Quality Monitoring Database

2,635 Taiwan Biobank volunteers aged 30-70 with no history of cancer: 1,225 men (46.5%) and 1,410 women (53.5%), mean age 49.80 +/- 10.53 years. By monitoring region, 494 (18.7%) northern, 529 (20.1%) central and 1,612 (61.2%) southern Taiwan. 1,917 (72.8%) never smokers and 718 (27.2%) current or former smokers; 2,371 (90.0%) no or occasional alcohol consumption and 264 (10.0%) current or quit. Mean height 162.94 +/- 8.25 cm, weight 64.38 +/- 12.05 kg, BMI 24.14 +/- 3.4 kg/m2, body adiposity index 28.5 +/- 3.88, body roundness index 3.71 +/- 1.11. Comorbidities: hypertension 275 (10.4%), type 2 diabetes 120 (4.6%), renal failure 4 (0.2%), metabolic syndrome 475 (18.0%), coronary artery disease 27 (1.0%). Each participant provided informed consent and underwent a face-to-face interview, physical examination, blood sampling and a lifestyle questionnaire. The Taiwan Biobank holds no information on occupational exposure to toxic substances, which the authors list as a limitation.

Cross-sectional, population-based study linking the Taiwan Biobank to the Taiwan Air Quality Monitoring Database; exposure assigned from the nearest of 74 monitoring stations as the average concentration over the year leading up to each participant's survey date. Because the design is cross-sectional, the authors state that further prospective studies are needed to elucidate causal effects.

LF:chen-twb-spirometry

Taiwan Biobank spirometry (FEV1 percent-predicted) by obstructive impairment status, with one-year average residential SO2, PM10 and PM2.5

Cross-sectional spirometry in 2,635 Taiwan Biobank participants (mean age 49.80 +/- 10.53 years; 1,225 men (46.5%) and 1,410 women (53.5%)), each linked to the one-year average pollutant concentrations at their nearest air quality monitoring station. Participants were stratified by their own lung function into a normal spirometry control group (72.2%, 1902/2635) and an obstructive impairment group (27.8%, 733/2635, FEV1/FVC < 70%). FEV1 percent-predicted was 58.42 +/- 18.31% in the obstructive impairment group versus 95.09 +/- 13.73% in the normal spirometry group (p < 0.0001), and 95.13 +/- 13.54% in the propensity-matched normal subset (p < 0.0001). No absolute FEV1 or FVC volume and no mean FEV1/FVC ratio is reported anywhere in the paper, so the fev1 slot on each output carries a PERCENT-PREDICTED value. AFTER 1:2 propensity score matching the two groups did not differ in age, gender, smoking, alcohol consumption, anthropometric factors or comorbidities - that sentence of the paper follows its matching paragraph and describes the MATCHED comparison. In Table 1's UNMATCHED columns several of those characteristics do differ: age p = 0.0216, body adiposity index p = 0.0369, type 2 diabetes p = 0.0449 and monitoring region p < 0.0001. So the balance claim should not be read as a property of the raw 733-vs-1,902 contrast that the rest of this file uses. WHOLE-COHORT POLLUTANT LEVELS (Table 1, median (IQR) over all 2,635 observations): PM10 68.12 (17.2) ug/m3, PM2.5 37.72 (10.8) ug/m3, SO2 3.63 (1.19) ppb, CO 0.44 (0.18) ppm, NO 4.09 (3.83) ppb, NO2 14.86 (5.6) ppb, NOX 18.93 (8.71) ppb, O3 30.97 (3.85) ppb; only SO2, PM10 and PM2.5 are carried as structured exposure conditions here because the remaining pollutants play no part in the paper's models. WHOLE-COHORT METEOROLOGY (Table 1, mean (SD)): temperature 24.33 (0.75) degrees C, relative humidity 74.28 (2.45)%, rainfall 0.22 (0.05) mm/day; the obstructive group lived at higher temperature (24.41 +/- 0.84 vs 24.31 +/- 0.72 degrees C, p = 0.0016), higher relative humidity (74.51 +/- 2.37 vs 74.20 +/- 2.47%, p = 0.0028) and lower rainfall (0.21 +/- 0.05 vs 0.22 +/- 0.05 mm/day, p = 0.0039). Meteorological values are recorded here as description text only: ExposureCondition carries exposure_agent, concentration, duration and timing, with no slot for ambient temperature, humidity or rainfall, and their odds ratios are recorded in response_comparisons. 61.2% of participants (1,612 individuals) were from southern Taiwan, which the authors offer as the reason their 27.8% prevalence of obstructive impairment far exceeds Taiwan's reported 6.1% COPD and 5.1% asthma prevalence.

Exposure conditions

Outputs

Evidence

COMP:chen-fev1pred-obstructive-vs-normal

FEV1 percent-predicted, obstructive impairment group vs full normal spirometry group

FEV1 percent-predicted was 58.42 +/- 18.31% in the obstructive impairment group versus 95.09 +/- 13.73% in the normal spirometry group, reported at p < 0.0001 (Table 1, one-way ANOVA). The paper prints only the two group values and the threshold p, never a difference, percent change or ratio, so change_value is omitted rather than computed. The p-value is a threshold, not an exact value, so p_value is omitted too. Note the direction of this contrast is the definition of the groups, not a treatment effect: participants were assigned to groups by their own spirometry.

Statistical test: one-way ANOVA between spirometry groups (Table 1), reported as p < 0.0001

COMP:chen-fev1pred-obstructive-vs-normal-matched

FEV1 percent-predicted, obstructive impairment group vs propensity-score-matched normal spirometry subset

FEV1 percent-predicted was 58.42 +/- 18.31% in the obstructive impairment group versus 95.13 +/- 13.54% in the 1:2 propensity-score-matched normal spirometry subset, again at p < 0.0001 (Table 1). Matching on age category, sex, region, smoking, drinking, BMI, BAI and BRI barely moved the comparison group's value (95.09% unmatched to 95.13% matched), which is what the authors use to argue the spirometry contrast is not driven by those baseline characteristics. change_value and p_value are omitted for the same reasons as the unmatched comparison.

Statistical test: one-way ANOVA between spirometry groups after 1:2 propensity score matching (Table 1), reported as p < 0.0001

COMP:chen-so2-or-model1

Adjusted odds ratio for obstructive impairment per ppb SO2, model 1 (with PM10)

SO2 was an independent predictive factor for obstructive impairment in model 1: adjusted OR 1.25 (95% CI 1.14-1.36), reported as p < 0.001 in the Results text and as p < 0.0001 in Table 3 - the paper is internally inconsistent here, and either way the value is a threshold rather than an exact p, so p_value is omitted. This is the paper's headline SO2 result — higher SO2 at the residence, higher odds of obstructive impairment. The outcome is obstructive impairment dichotomized from FEV1/FVC at 70%, not a change in the FEV1/FVC ratio itself.

Statistical test: stepwise multinomial logistic regression, model 1 (adjusted OR per ppb SO2, fitted with PM10)

COMP:chen-pm10-or-model1

Adjusted odds ratio for obstructive impairment per ug/m3 PM10, model 1

PM10 was an independent predictive factor in model 1 with an adjusted OR BELOW one: 0.99 (95% CI 0.98-0.99), reported as p < 0.001 in the Results text and as p < 0.0001 in Table 3 - the paper is internally inconsistent here, and either way the value is a threshold rather than an exact p, so p_value is omitted. Higher PM10 at the residence went with LOWER odds of obstructive impairment, the opposite of most previous studies. The authors do not treat this as PM10 being protective: they attribute it to the SO2 x PM synergy they report, under which injury occurs at low concentrations of both pollutants together, and state that their finding 'is contrast to most of previous studies'. Recorded exactly as reported.

Statistical test: stepwise multinomial logistic regression, model 1 (adjusted OR per ug/m3 PM10)

COMP:chen-temperature-or-model1

Adjusted odds ratio for obstructive impairment per degree C, model 1

Temperature was an independent predictive factor in model 1: adjusted OR 1.24 (95% CI 1.09-1.41), p = 0.0009. Higher ambient temperature at the residence went with higher odds of obstructive impairment. The authors read this through a previous Taiwanese finding of a V/U-shaped temperature-pollutant relationship, in which 24.3-24.9 degrees C corresponds to the lowest pollutant exposure, so temperature may act partly by shifting pollutant exposure.

Statistical test: stepwise multinomial logistic regression, model 1 (adjusted OR per degree Celsius)

COMP:chen-humidity-or-model1

Adjusted odds ratio for obstructive impairment per percent relative humidity, model 1

Relative humidity was an independent predictive factor in model 1: adjusted OR 1.05 (95% CI 1.01-1.10), p = 0.0160. The authors' proposed mechanism is that high relative humidity limits evaporative cooling by perspiration, creating physiological stress that leads to respiratory dysfunction, and that high temperature with high humidity can trigger bronchoconstriction.

Statistical test: stepwise multinomial logistic regression, model 1 (adjusted OR per percent relative humidity)

COMP:chen-rainfall-or-model1

Adjusted odds ratio for obstructive impairment per mm/day rainfall, model 1

Rainfall was an independent predictive factor in model 1 with an adjusted OR far below one: 0.08 (95% CI 0.01-0.68), p = 0.0202. More rainfall went with much lower odds of obstructive impairment. Note the scale: Table 1 reports rainfall as mean (SD) 0.22 (0.05) mm/day, so an OR expressed per whole mm/day is about 4.5 times the cohort mean and roughly 20 cohort standard deviations - far outside the observed range, and not a per-SD effect. The authors attribute the direction to precipitation scavenging, citing a washout effect that reduces particle numbers during rainfall.

Statistical test: stepwise multinomial logistic regression, model 1 (adjusted OR per mm/day rainfall)

COMP:chen-so2-or-model2

Adjusted odds ratio for obstructive impairment per ppb SO2, model 2 (with PM2.5)

SO2 was an independent predictive factor in model 2 as well: adjusted OR 1.28 (95% CI 1.17-1.39), reported as p < 0.001 in the Results text and as p < 0.0001 in Table 3 - the paper is internally inconsistent here, and either way the value is a threshold rather than an exact p, so p_value is omitted. The SO2 estimate is stable across the two models (1.25 with PM10, 1.28 with PM2.5), which is the paper's basis for calling SO2 the pollutant strongly associated with obstructive lung disease.

Statistical test: stepwise multinomial logistic regression, model 2 (adjusted OR per ppb SO2, fitted with PM2.5)

COMP:chen-pm25-or-model2

Adjusted odds ratio for obstructive impairment per ug/m3 PM2.5, model 2

PM2.5 was an independent predictive factor in model 2 with an adjusted OR below one: 0.97 (95% CI 0.96-0.98), reported as p < 0.001 in the Results text and as p < 0.0001 in Table 3 - the paper is internally inconsistent here, and either way the value is a threshold rather than an exact p, so p_value is omitted. As with PM10 in model 1, higher PM2.5 went with LOWER odds of obstructive impairment, which the authors attribute to the SO2 x PM synergy rather than to any protective effect of PM2.5. Recorded exactly as reported.

Statistical test: stepwise multinomial logistic regression, model 2 (adjusted OR per ug/m3 PM2.5)

COMP:chen-temperature-or-model2

Adjusted odds ratio for obstructive impairment per degree C, model 2

Temperature in model 2: adjusted OR 1.31 (95% CI 1.15-1.49), reported as p < 0.001 in the Results text and as p < 0.0001 in Table 3 - the paper is internally inconsistent here, and either way the value is a threshold rather than an exact p, so p_value is omitted - slightly stronger than the 1.24 of model 1.

Statistical test: stepwise multinomial logistic regression, model 2 (adjusted OR per degree Celsius)

COMP:chen-humidity-or-model2

Adjusted odds ratio for obstructive impairment per percent relative humidity, model 2

Relative humidity in model 2: adjusted OR 1.04 (95% CI 1.00-1.09), p = 0.0372. The lower confidence bound sits at 1.00 as printed, so this is the weakest of the meteorological associations.

Statistical test: stepwise multinomial logistic regression, model 2 (adjusted OR per percent relative humidity)

COMP:chen-rainfall-or-model2

Adjusted odds ratio for obstructive impairment per mm/day rainfall, model 2

Rainfall in model 2: adjusted OR 0.08 (95% CI 0.01-0.67), p = 0.0197 - essentially identical to model 1's 0.08 (0.01-0.68). The same scale caveat applies: cohort rainfall is 0.22 (0.05) mm/day as mean (SD), so an OR per whole mm/day spans roughly 20 cohort standard deviations and sits far outside the observed range.

Statistical test: stepwise multinomial logistic regression, model 2 (adjusted OR per mm/day rainfall)

COMP:chen-so2-persd-or-model1

Adjusted odds ratio for obstructive impairment per SD increase in SO2, model 1

Per standard deviation increase in SO2, the model 1 adjusted OR was 1.30 (95% CI 1.17-1.44), reported at p < 0.0001 (so p_value is omitted). The per-SD scaling makes the SO2 effect directly comparable with the per-SD PM estimates: SO2 1.30 up versus PM10 0.79 down.

Statistical test: stepwise multinomial logistic regression, model 1, adjusted OR per standard deviation increase in SO2

COMP:chen-so2-persd-or-model2

Adjusted odds ratio for obstructive impairment per SD increase in SO2, model 2

Per standard deviation increase in SO2, the model 2 adjusted OR was 1.34 (95% CI 1.21-1.48), reported at p < 0.0001 (so p_value is omitted) - the largest SO2 estimate in the paper.

Statistical test: stepwise multinomial logistic regression, model 2, adjusted OR per standard deviation increase in SO2

COMP:chen-pm10-persd-or-model1

Adjusted odds ratio for obstructive impairment per SD increase in PM10, model 1

Per standard deviation increase in PM10, the model 1 adjusted OR was 0.79 (95% CI 0.71-0.88), reported at p < 0.0001 (so p_value is omitted). On the per-SD scale the inverse PM10 association is much larger than the per-ug/m3 OR of 0.99 makes it look, because one SD of PM10 spans many ug/m3. The paper never reports pollutant standard deviations - Table 1 gives PM10 as median (IQR) 68.12 (17.2) ug/m3 - so the SD behind this rescaling is not stated; back-calculating from the two ORs puts it somewhere in the high teens to low twenties ug/m3 depending on how much the 0.99 is rounded.

Statistical test: stepwise multinomial logistic regression, model 1, adjusted OR per standard deviation increase in PM10

COMP:chen-pm25-persd-or-model2

Adjusted odds ratio for obstructive impairment per SD increase in PM2.5, model 2

Per standard deviation increase in PM2.5, the model 2 adjusted OR was 0.73 (95% CI 0.65-0.81), reported at p < 0.0001 (so p_value is omitted) - the strongest of the inverse PM associations on the per-SD scale.

Statistical test: stepwise multinomial logistic regression, model 2, adjusted OR per standard deviation increase in PM2.5

COMP:chen-so2-pm10-interaction

SO2 x PM10 crude (unadjusted) interaction coefficient for the relative risk of obstructive impairment

SO2 modified the relationship of PM10 with the relative risk of obstructive impairment: interaction beta = 0.01 with SE 0.00, p for interaction = 0.0052. This is the CRUDE, UNADJUSTED interaction term: in Table 3 the 'beta (SE), P for interaction' column sits inside the crude block, immediately after the crude OR and its p-value, and before the model 1 and model 2 adjusted OR columns. The model-adjusted counterpart from the same table row is recorded separately in COMP:chen-so2-pm10-interaction-adjusted and is NOT significant (beta = 0.00, SE 0.00, p = 0.3423). The paper headlines this crude term in its abstract and Results text, which is why it is recorded, but it should not be read as an adjusted effect. The value is a coefficient on the log-odds scale for the SO2 x PM10 product term, not a change in any lung function measurement, and its positive sign means the PM10 association strengthens as SO2 rises. change_type is absolute_difference with a dimensionless unit because that is what a regression coefficient is; change_direction reflects the sign of beta only.

Statistical test: multinomial logistic regression CRUDE (unadjusted) interaction term (SO2 x PM10), beta with standard error and p for interaction, from Table 3's crude block

COMP:chen-so2-pm25-interaction

SO2 x PM2.5 crude (unadjusted) interaction coefficient for the relative risk of obstructive impairment

SO2 also modified the relationship of PM2.5 with the relative risk of obstructive impairment: interaction beta = 0.01 with SE 0.00, p for interaction = 0.0155. As with the SO2 x PM10 term, this is the CRUDE, UNADJUSTED interaction from Table 3's crude block, not a model estimate - and note that PM2.5 is the model 2 pollutant, so labelling it 'model 1' would be wrong twice over. The model-adjusted counterpart is recorded in COMP:chen-so2-pm25-interaction-adjusted and is NOT significant (beta = 0.05, SE 0.05, p = 0.3423). Same log-odds scale and caveats as the SO2 x PM10 interaction above.

Statistical test: multinomial logistic regression CRUDE (unadjusted) interaction term (SO2 x PM2.5), beta with standard error and p for interaction, from Table 3's crude block

COMP:chen-so2-pm10-interaction-persd

SO2 x PM10 crude (unadjusted) interaction coefficient, per SD increase

On the per standard deviation scale the same CRUDE SO2 x PM10 interaction has beta = 0.11 with SE 0.04 and the same p for interaction = 0.0052 (Table 3, 'Per SD increasing' block, crude column). The p-value is identical to the per-unit version because rescaling an exposure does not change the test; only the coefficient's size does. The model-adjusted per-SD counterpart is recorded in COMP:chen-so2-pm10-interaction-persd-adjusted and is NOT significant (beta = -0.00, SE 0.00, p = 0.5086).

Statistical test: multinomial logistic regression CRUDE (unadjusted) interaction term (SO2 x PM10) per standard deviation increase, beta with standard error and p for interaction, from Table 3's crude column

COMP:chen-so2-pm25-interaction-persd

SO2 x PM2.5 crude (unadjusted) interaction coefficient, per SD increase

On the per standard deviation scale the CRUDE SO2 x PM2.5 interaction has beta = 0.09 with SE 0.04 and the same p for interaction = 0.0155 (Table 3, 'Per SD increasing' block, crude column). The model-adjusted per-SD counterpart is recorded in COMP:chen-so2-pm25-interaction-persd-adjusted and is NOT significant (beta = -0.03, SE 0.04, p = 0.5086).

Statistical test: multinomial logistic regression CRUDE (unadjusted) interaction term (SO2 x PM2.5) per standard deviation increase, beta with standard error and p for interaction, from Table 3's crude column

COMP:chen-so2-pm10-interaction-adjusted

SO2 x PM10 model-adjusted interaction coefficient for the relative risk of obstructive impairment

The model-adjusted SO2 x PM10 interaction from Table 3's rightmost column ('Adjusted beta (SE), P for interaction'): beta = 0.00 with SE 0.00, p for interaction = 0.3423 - NOT significant. Compare the crude term in COMP:chen-so2-pm10-interaction (beta = 0.01, p = 0.0052), which the paper headlines in its abstract. The synergy claim in the abstract and Conclusion therefore rests on the crude interaction term and on the bivariate thin-plate smoothing splines of Table 4 (both p < 0.0001), not on this adjusted interaction coefficient. The paper does not comment on the discrepancy. change_direction is recorded as no_change because the coefficient is 0.00 as printed and its p-value is far from significance.

Statistical test: multinomial logistic regression ADJUSTED interaction term (SO2 x PM10), beta with standard error and p for interaction, from Table 3's adjusted interaction column

COMP:chen-so2-pm25-interaction-adjusted

SO2 x PM2.5 model-adjusted interaction coefficient for the relative risk of obstructive impairment

The model-adjusted SO2 x PM2.5 interaction from Table 3's rightmost column: beta = 0.05 with SE 0.05, p for interaction = 0.3423 - NOT significant, and with a standard error equal to the coefficient itself. Compare the crude term in COMP:chen-so2-pm25-interaction (beta = 0.01, p = 0.0155). Table 3 prints the same p-value (0.3423) for both the PM10 and PM2.5 adjusted interaction terms.

Statistical test: multinomial logistic regression ADJUSTED interaction term (SO2 x PM2.5), beta with standard error and p for interaction, from Table 3's adjusted interaction column

COMP:chen-so2-pm10-interaction-persd-adjusted

SO2 x PM10 model-adjusted interaction coefficient, per SD increase

The model-adjusted per-SD SO2 x PM10 interaction from Table 3's 'Per SD increasing' block: beta = -0.00 with SE 0.00, p for interaction = 0.5086 - NOT significant, and NEGATIVE as printed (the paper writes it with a minus sign, which is why change_value carries -0.00 rather than 0.00). Compare the crude per-SD term in COMP:chen-so2-pm10-interaction-persd (beta = 0.11, p = 0.0052): adjustment reverses the sign and removes the significance.

Statistical test: multinomial logistic regression ADJUSTED interaction term (SO2 x PM10) per standard deviation increase, beta with standard error and p for interaction

COMP:chen-so2-pm25-interaction-persd-adjusted

SO2 x PM2.5 model-adjusted interaction coefficient, per SD increase

The model-adjusted per-SD SO2 x PM2.5 interaction from Table 3's 'Per SD increasing' block: beta = -0.03 with SE 0.04, p for interaction = 0.5086 - NOT significant and negative. Compare the crude per-SD term in COMP:chen-so2-pm25-interaction-persd (beta = 0.09, p = 0.0155): as with PM10, adjustment reverses the sign and removes the significance. Table 3 prints the same p-value (0.5086) for both per-SD adjusted interaction terms.

Statistical test: multinomial logistic regression ADJUSTED interaction term (SO2 x PM2.5) per standard deviation increase, beta with standard error and p for interaction

COMP:chen-so2-or-crude

Crude odds ratio for obstructive impairment per ppb SO2

Before adjustment, SO2 carried a crude OR of 1.08 (95% CI 1.01-1.16), p = 0.0268 - in the same direction as the adjusted estimates (1.25 and 1.28) but much weaker, so adjustment for temperature, relative humidity and rainfall strengthens the SO2 association considerably.

Statistical test: multinomial logistic regression, crude (unadjusted) OR per ppb SO2

COMP:chen-pm10-or-crude

Crude odds ratio for obstructive impairment per ug/m3 PM10

Crude OR for PM10 0.99 (95% CI 0.99-0.999), p = 0.0115 - the inverse PM association is present before adjustment as well as after it (model 1 adjusted OR 0.99, 95% CI 0.98-0.99). Note the asymmetric upper bound printed as 0.999 in Table 3.

Statistical test: multinomial logistic regression, crude (unadjusted) OR per ug/m3 PM10

COMP:chen-pm25-or-crude

Crude odds ratio for obstructive impairment per ug/m3 PM2.5

Crude OR for PM2.5 0.99 (95% CI 0.98-0.99), p = 0.0015 - again inverse before adjustment, and strengthened by adjustment (model 2 adjusted OR 0.97, 95% CI 0.96-0.98).

Statistical test: multinomial logistic regression, crude (unadjusted) OR per ug/m3 PM2.5

KER:chen-oxidative-stress-to-lung-function

Increased oxidative stress is associated with decreased lung function

The authors' proposed mechanism linking their epidemiological findings to lung injury: SO2 exposure affects the respiratory tract and causes oxidative stress and DNA damage that further damages the lungs, and the SO2 x PM synergy they report is attributed, citing Yun et al., to PM10 and SO2 jointly inducing reactive oxygen species production and NF-kB activation at low concentrations of both. This paper measures no oxidative stress marker itself; the oxidative stress limb comes entirely from cited literature in the Discussion, while the decreased lung function limb is this paper's own cross-sectional association (adjusted OR for obstructive impairment 1.25 and 1.28 per ppb SO2). The design is cross-sectional, so the authors state that prospective studies are needed to elucidate causal effects.

PROTOCOL:chen-spirometry-001

Taiwan Biobank spirometry with ATS/ERS FEV1/FVC < 70% definition of obstruction

Pulmonary function parameters including forced expiratory volume in one second (FEV1), forced vital capacity (FVC), the FEV1/FVC% ratio, the FVC-predicted value and the FEV1-predicted value were recorded in the Taiwan Biobank. Technicians used MicroLab spirometers and Spida 5 software (Micro Medical Ltd., Rochester, Kent, UK) to perform the spirometry measurements. Obstructive lung diseases including asthma, COPD and bronchiectasis were defined as patients with FEV1/FVC < 70%, according to the American Thoracic Society and European Respiratory Society guidelines. The paper does not state a reference equation for the predicted values, nor whether spirometry was pre- or post-bronchodilator.

Equipment: MicroLab spirometer (Micro Medical Ltd., Rochester, Kent, UK); Spida 5 software (Micro Medical Ltd.)

PROTOCOL:chen-air-monitoring-001

Taiwan Air Quality Monitoring Database: 74-station daily ambient pollutant monitoring

Daily ambient pollutant concentrations came from 74 air quality monitoring stations located throughout Taiwan, as recorded by the Taiwan Air Quality Monitoring Database (TAQMD) of the Taiwan Environmental Protection Administration. Measurement methods by pollutant: PM2.5 and PM10 by beta-ray attenuation; SO2 by ultraviolet fluorescence; CO by nondispersive infrared; O3 by ultraviolet absorption; NOx by chemiluminescence. All air pollutant data are stored in the cloud every hour and are free to access, and the average concentrations of air pollutants in a selected year were obtained before analysis.

Equipment: 74 Taiwan EPA air quality monitoring stations; Beta-ray attenuation monitors (PM2.5, PM10); Ultraviolet fluorescence analyser (SO2)

PROTOCOL:chen-exposure-assignment-001

Three-step geocoding of residential address to nearest monitoring station, one-year average exposure

A three-step procedure assigned each participant an exposure. First, Google geocoding determined the exact geoposition of each residential address. Second, the interpolation point between each residential address and the nearest air quality monitoring station was determined. Lastly, data from that station recorded during the year leading up to the survey date were selected and the average values of PM2.5, PM10, CO, NO, NO2, NOx, SO2 and O3 for the chosen year were calculated. The exposure metric is therefore a ONE-YEAR AVERAGE at the nearest station, not a personal measurement; the authors list this as a limitation, since personal exposure, travel exposure and indoor air quality are not captured.

PROTOCOL:chen-statistics-001

ANOVA, propensity score matching, multinomial logistic regression and generalized additive modelling

Descriptive analysis between spirometry groups used one-way ANOVA and the chi-square or Fisher's exact tests as appropriate; air pollution factors were compared between the obstructive impairment group and the normal spirometry comparison group by independent t-test. Propensity score matching (1:2) balanced baseline characteristics (age categories, sex, live region, smoking, drinking, BMI, BAI and BRI). Multinomial logistic regression estimated crude odds ratios and 95% CIs, and stepwise multinomial logistic regression calculated adjusted ORs and 95% CIs. Pearson's correlation analysis evaluated relationships between temperature, relative humidity, rainfall, PM10, PM2.5 and SO2. Because SO2 correlated with both PM2.5 and PM10, a generalized additive model (GAM) was used to fit equations and splines and to explore linear and nonlinear effects on obstructive impairment, with a semiparametric model taking temperature, relative humidity and rainfall as the linear part. All analyses used SAS version 9.4.