The details of participant selection for the present study are shown in Figure 1. The cohort database accumulated 590,976 partic- ipants between 1996 and 2014 (the questionnaire data have been computerized since 1996). We selected 432,433 participants who joined the program between 2001 and 2014, when the 2-y average PM2:5 exposure assessment was available. We computed their esti- mated glomerular filtration rate (eGFR) based on their serum creati- nine level using the equation from theModification of Diet in Renal Disease (MDRD) Study (National Kidney Foundation 2002). We excluded 3,375 participants with an eGFR ≥200 mL=min=1:73m2 or <2 mL=min=1:73m2 because the values suggested that the measurements were probably incorrect due to occasional techni- cal errors (Levey et al. 2009). We further excluded participants
who had made only one medical visit. After these exclusions, 159,944 participants were selected. We then limited the cohort to participants without CKD at baseline by excluding those with an eGFR ≤60 mL=min=1:73 m2 [Kidney Disease: Improving Global Outcomes (KDIGO) Work Group 2013] or who reported physician-diagnosed kidney disease at their first visit. Because uri- nary protein is also an important syndrome for CKD, we also excluded those with urinary protein level ≥2:0 g=L at baseline, leaving 147,658 participants without prevalent CKD who were ≥20 years of age at baseline and had complete data for key varia- bles, including demographics, socioeconomic status, lifestyle, blood tests, and PM2:5 exposure. Finally, we excluded participants with <3 years of follow-up, resulting in a final cohort of 100,629 partici- pants enrolled between 2001 and 2011 for the present analysis. The follow-up duration of the 100,629 participants ranged from 3.0 to 13.0 y (mean, 6.5 y). The number of medical visits ranged from 2 to 18, with a mean of 4.2, totaling 424,455 medical observations. The mean ± SD visit interval was 2.0 ± 1.5 y. The selection pro- cess did not bring in substantial differences among the participants in terms of the distribution of age, sex, and cardiovascular risk fac- tors (Table S1).
Air Pollution Exposure Assessment The details for estimation of PM2:5 air pollution have been described elsewhere (Zhang et al. 2017). In brief, PM2:5 exposure was esti- mated at each participant`s address using a satellite-based spatio- temporal model with a high spatial resolution of 1 km×1 km based on the aerosol optical depth data, which were derived from the two Moderate Resolution Imaging Spectroradiometer (MODIS) instru- ments aboard Terra and Aqua satellites from the U.S. National Aeronautics and Space Administration (Li et al. 2005; Lin et al. 2015). We have validated this model with ground-measured data frommore than 70monitoring stations in Taiwan, and the results are presented elsewhere (Zhang et al. 2017). The correlation coeffi- cients between the average satellite-retrieved and ground-level monitoring PM2:5 concentrations ranged from 0.79 to 0.83 in differ- ent years, and themean percentage errorswere around 20%.
The participants’ residential addresses were usually collected during each medical visit so the medical report could be mailed to them. Some participants provided a company address instead of a residential address. The address was geocoded into latitude and longitude data. and address-specific yearly average PM2:5 concen- trations were then calculated. For participants who provided their company address, PM2:5 was estimated at their company address. We estimated the annual average PM2:5 concentrations for the cal- endar year of each participant`s medical examination and the an- nual average for the previous year. Themean of these two averages (2-y average) was then calculated as an indicator of long-term ex- posure to ambient PM2:5 air pollution in this study. The baseline 2-y average PM2:5 concentrations (hereafter called “baseline PM2:5 exposure”) thus referred to average of the enrollment year and the previous year. The follow-up 2-y average PM2:5 concentrations (hereafter called “follow-up PM2:5 exposure”) referred to average of the follow-upmedical examination year and the previous year.
Health Outcome and Covariates An overnight fasting blood sample was taken in the morning, and the serum creatinine was analyzed using a HITACHI 7150 (before 2005) or a TOSHIBA C8000 (after 2005) analyzer with the uncompensated Jaffe method involving an alkaline picrate ki- netic test (Myers et al. 2006). The eGFR level was calculated based on the following MDRD equation:Figure 1. Flowchart of participant selection.
Environmental Health Perspectives 107002-2 126(10) October 2018
186:3× ðserum creatinineÞ−1:154 × age−0:203 × ð0:742 for womenÞ
where serum creatinine is in mg/dL. Health outcome is incident CKD in the present study. After
the baseline assessment at the first visit, all participants were fol- lowed up, and the incident CKD was identified by medical assessment (defined as eGFR less than 60 mL=min=1:73m2) in subsequent visits [Kidney Disease: Improving Global Outcomes (KDIGO) Work Group 2013]. The end point was the first occur- rence of CKD or the last visit if CKD did not occur.
In addition to serum creatinine measurement, the participants underwent a number of other medical examinations during their visits. The procedures of the medical examination program in this population have been described in previous publications (Zhang et al. 2017; Wen et al. 2008; Chang et al. 2016; Zhang et al. 2018). All examinations or tests were performed by trained medical professionals, and detailed information, including infor- mation on quality control, can be accessed in the technical reports released by the MJ Health Research Foundation (Chang et al. 2016).
A wide range of potential confounders or modifiers were con- sidered. Weight (to the nearest 0:1 kg) and barefoot height (to the nearest 0:1 cm) were measured with participants wearing light clothes using an auto-anthropometer (KN-5000A, Nakamura). Seated blood pressure was measured using a computerized auto- mercury sphygmomanometer (CH-5000, Citizen). An overnight fasting blood sample was also taken to measure total cholesterol using an auto-analyzer (7150, Hitachi). Urinary protein was an- alyzed using a ROCHE Miditron/ROCHE Cobas U411 semiau- tomated computer-assisted urinalysis system. Urinary protein results were reported at six levels: negative (<0:1 g=L), trace (0:1∼ 0:2 g=L), 1 plus (0:2∼ 1:0 g=L), 2 plus (1:0∼ 2:0 g=L), 3 plus (2:0∼ 4:0 g=L), and 4 plus (>4:0 g=L).
A self-administered questionnaire was used to collect infor- mation on demographic characteristics, occupational exposure, lifestyle, and medical history.
The following covariates were included in the data analysis: age (years), sex (male and female), education level [lower than high school (<10 y), high school (10–12 y), college or university (13–16 y) and postgraduate (>16 y)], smoking (never, ever, and cur- rent), and drinking (<once=week, 1–3 times/wk, and >3 times=wk), body mass index [BMI, calculated as weight (kg) divided by the square of height (m)], systolic blood pressure (mmHg), fasting glu- cose (mg/dL), total cholesterol (mg/dL), self-reported heart disease or stroke (yes or no), and urinary protein (at four levels: negative, trace, 1 plus, and 2 plus; participants with a level of 3 plus or above were excluded, asmentioned above).
Statistical Analysis We applied the Cox proportional hazard regression to investigate the association between PM2:5 and the incidence of CKD. The time scale used in the models was time-in-study (i.e., follow-up time). Five models were developed with the use of covariates from base- line visits: a) a crude model, with no adjustment; b) Model 1, adjusted for age, sex, education level, smoking, and drinking; c) Model 2, further adjusted for BMI, systolic blood pressure, fasting glucose, total cholesterol, and self-reported heart disease or stroke; d) Model 3, further adjusted for baseline eGFR level (a major de- terminant of kidney outcomes) (Tangri et al. 2011); and d) Model 4, further adjusted for the urinary protein level to observe its effects on the associations, because the urinary protein level is an impor- tant indicator for evaluation of renal function and the progress of CKD (Wen et al. 2008). We used the results from Model 2 as our main model, because both eGFR and urinary protein levels might
be precursors of CKD. The hazard ratio (HR) and 95% confidence interval (CI) were calculated to indicate the PM2:5 effects. The par- ticipants were categorized into five groups based on quintiles of baseline PM2:5 exposure and those with PM2:5 in the lowest quin- tile served as the reference group. A trend test was performed with the PM2:5 quintile treated as a numeric variable (an ordinal variable coded as 1–5) in the models. We also treated PM2:5 as a continuous variable and effect estimates were reported for each 10 lg=m3
increase in the PM2:5 concentration. We also conducted stratified analysis by baseline age (<65 y vs
≥65 y), baseline sex (male versus female), baseline smoking (never smoker versus ever smoker), baseline BMI (<25 kg=m2 versus ≥25 kg=m2 ), baseline hypertension (defined as systolic blood pres- sure ≥140mmHg , diastolic blood pressure ≥90mmHg or self- reported physician-diagnosed hypertension: yes versus no), baseline diabetes (defined as fasting glucose ≥126 mg=dL or self-reported physician-diagnosed diabetes: yes versus no), and baseline self- reported heart disease or stroke, because previous studies had sug- gested that these factors could amplify the adverse effects of PM air pollution (Brook et al. 2010). We introduced an interaction term [PM2:5 ðcontinuous variableÞ× each factor ðdichotomous variableÞ] in the Cox regression model to investigate the potential modifying effects. Each factor was examined separately. P values were calcu- lated for the product terms.
A series of sensitivity analyses were performed by a) including only participantswho used a residential address (16,574 participants with 226 cases were excluded because they provided a company address); b) including only participants whose eGFR was measured with a HITACHI 7150 before 2005 (36,634 participants with 805 cases were excluded, because their eGFR were measured with a TOSHIBA C8000 since 2005); c) including all 147,658 nonCKD participants 20 years of age or older at baseline (i.e., the 47,029 par- ticipants with a follow-up duration less than 3 y plus the 100,629 participants with a follow-up duration greater than 3 y); and d) using a time-varying Cox model with PM2:5 and covariates treated as time-varying variables; e) conducting a sensitivity analysis (for comparison) using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula (Stevens et al. 2011) for eGRF calculation.
All statistical analyses were performed using R (version 3.2.3; R Core Team). A two-tailed p-value<0:05 was considered to indicate statistical significance.
Results Table 1 summarizes the participants’ general characteristics at baseline. The distribution of these characteristics was generally similar across the participants grouped by PM2:5 quintiles. Among the 100,629 participants, 4,046 incident CKD cases developed dur- ing the follow-up (the number of cases in the categories of eGFR of 45∼60, 30∼45, 15∼30, and ∼ 15mL=min=1:73m2 were 3,944, 92, 8, and 2, respectively). The incidence rate was 6.24 per 1,000 per- son-years.
The locations of the participants are shown in Figure S1. The southwestern areas were generally the most heavily polluted and the middle-eastern areas were the least heavily polluted. The spa- tial pattern of exposure contrast throughout the island generally remained stable during the study period. Figure 2 shows the dis- tribution of baseline PM2:5 exposure. The baseline PM2:5 expo- sure increased slightly from 2001 to 2004 (the mean baseline PM2:5 exposure was 25.0, 26.7, 29.1, and 29:9 lg=m3 , respec- tively, for participants enrolled in year 2001, 2002, 2003, and 2004), and then declined and remained relatively stable from 2005 to 2011 (the mean baseline PM2:5 exposure was 27.5, 27.2, 27.4, 27.5, 28.1, 26.4, and 25:2 lg=m3, respectively for year
Environmental Health Perspectives 107002-3 126(10) October 2018
2005, 2006, 2007, 2008, 2009, 2010, and 2011). The overall mean ± SD was 27:1 ± 8:0 lg=m3 with an IQR of 10:4 lg=m3.
Table 2 shows the association between the PM2:5 and the inci- dence of CKD. A higher level of PM2:5 was associated with a higher risk of developing CKD. In the main model (Model 2), in contrast with the participants with the first quintile exposure, par- ticipants with the fourth or fifth quintiles were significantly asso- ciated, with an HR (95% CI) of 1.11 (1.01, 1.22) or 1.15 (1.05, 1.26) in CKD development, respectively. A significant concentra- tion–response trend was observed (p<0:001). Every 10lg=m3 increment in the PM2:5 level was associated with a 6% increased
risk of developing CKD (HR: 1.06, 95% CI: 1.02, 1.10). An addi- tional adjustment for baseline eGFR or urinary protein produced only marginal changes in the results.
Table 3 shows the results of the stratified analyses for potential modifiers. No significant effect modifications were observed when data analyses were stratified by age, sex, BMI, hypertension, diabetes, and cardiovascular disease history (all p values >0:05).
The results of sensitivity analyses 1 to 4 are presented in Table 4. Overall, the associations were consistent by excluding participants using company address, using TOSHIBA C8000
Table 1. Baseline characteristics of study participants stratified by PM2:5 quintiles between 2001 and 2011.
(N=100629) 1st quintile (N=20119)
2nd quintile (N=20130)
3rd quintile (N= 20122)
4th quintile (N=20136)
5th quintile (N=20122)
Mean 2-year PM2:5 lg=m3
(range) 27:1± 8:0 (5.8–49.6)
19:1± 1:8 (5.8–21.1)
22:2± 0:6 (>21:1–23:3)
24:2± 0:6 (>23:3–25:5)
29:0± 3:4 (>25:5–36:1)
41:1± 2:9 (>36:1–49:6)
Age mean ± SD, (years) 38:9± 11:3 39:9± 11:6 38:8± 11:2 38:1± 11:0 38:6± 11:3 39:0± 11:3 Male 52,837 (52.5) 10,280 (51.1) 10,694 (53.1) 10,669 (53.0) 10,774 (53.5) 10,420 (51.8) Education Lower than high school 13,354 (13.3) 3,074 (15.3) 2,472 (12.3) 2,271 (11.3) 2,722 (13.5) 2,815 (14.0) High school 20,351 (20.2) 4,320 (21.5) 3,901 (19.4) 3,631 (18.0) 4,081 (20.3) 4,418 (22.0) College 25,534 (25.4) 5,289 (26.3) 5,003 (24.9) 5,074 (25.2) 4,856 (24.1) 5,312 (26.4) University 29,777 (29.6) 5,389 (26.8) 6,046 (30.0) 6,470 (32.2) 6,029 (29.9) 5,843 (29.0) Postgraduate 11,613 (11.5) 2,047 (10.2) 2,708 (13.5) 2,676 (13.3) 2,448 (12.2) 1,734 (8.6) Cigarette smoking Never 74,172 (74.2) 14,855 (73.8) 14,791 (73.5) 14,923 (74.2) 14,903 (74.0) 15,240 (75.7) Former 5,441 (5.4) 1,136 (5.6) 1,098 (5.5) 1,049 (5.2) 1,123 (5.6) 1,035 (5.1) Current 20,476 (20.3) 4,128 (20.5) 4,241 (21.1) 4,150 (20.6) 4,110 (20.4) 3,847 (19.1) Alcohol drinking Never 84,024 (83.5) 16,758 (83.3) 16,904 (84.0) 16,849 (83.7) 16,787 (83.4) 16,726 (83.1) Former 2,292 (2.3) 470 (2.3) 430 (2.1) 459 (2.3) 433 (2.2) 500 (2.5) Current (≥1 time=week) 14,313 (14.2) 2,796 (14.4) 2,796 (13.9) 2,814 (14.0) 2,916 (14.5) 2,896 (14.4) Body mass index mean ± SD, (kg=m2) 22:9± 3:5 22:9± 3:4 22:9± 3:4 22:8± 3:5 23:0± 3:5 22:9± 3:5 Systolic blood pressure mean ± SD, (mmHg) 117:0± 16:4 116:9± 16:4 117:0± 16:3 116:9± 16:1 116:8± 16:6 117:2± 16:7 Fasting glucose mean ± SD, (mg/dL) 98:1± 17:7 98:2± 18:4 98:1± 17:3 98:2± 17:1 98:1± 18:0 97:8± 17:7 Total cholesterol mean ± SD, (mg/dL) 190:1± 34:6 189:9± 34:1 190:4± 34:7 190:5± 35:0 191:1± 34:9 188:7± 34:4 eGFR mean ± SD, (mL=min=1:73m2) 87:0± 14:9 86:5± 14:5 87:5± 15:0 87:9± 15:1 86:5± 14:9 86:7± 14:8 Self-reported heart disease or stroke 2,122 (2.1) 481 (2.4) 390 (1.9) 440 (2.2) 406 (2.0) 405 (2.0) Hypertension 12,603 (12.5) 2,684 (13.3) 2,534 (12.6) 2,393 (11.9) 2,432 (11.1) 2,560 (12.7) Diabetes 3,242 (3.2) 682 (3.4) 626 (3.1) 584 (2.9) 670 (3.3) 680 (3.4) Address Residential address 84,055 (83.5) 18,839 (93.6) 16,755 (83.2) 15,560 (77.3) 15,515 (77.1) 17,386 (86.4) Company address 16,574 (16.5) 1,280 (6.4) 3,375 (15.8) 4,562 (22.7) 4,621 (22.9) 2,736 (13.6)
Note: Data are presented as mean± SD for continuous variables and number (percentage) for categorical variables. Data are complete for all variables.
Figure 2. Distribution of baseline PM2:5 exposure of the participants by year. Boxes cover the 25–75th percentile (interquartile range: IQR) with a center line for the median concentration. Whiskers extend to the highest observation within 3 IQRs of the box, with more extreme observations shown as circles.
Environmental Health Perspectives 107002-4 126(10) October 2018
analyzer since January, 2005, or including participants with a follow-up duration of less than 3 y. Using time-varyingCox regres- sion model also yielded similar results. The results of sensitivity analysis 5 using the CKD-EPI formula are presented in Table S2. We observed results very similar to those in Table 2.
Discussion Our study shows that long-term exposure to ambient PM2:5 was associated with an increased risk of incident CKD (based on eGFR <60 mL=min=1:73m2 at a follow-up visit) among adult residents of Taiwan. Participants in this study with the fourth or the fifth quintiles of PM2:5 were significantly associated with increased risk of developing CKD, with an HR (95% CI) of 1.11 (1.01, 1.22) or 1.15 (1.05, 1.26), respectively, in comparison with the participants with the first quintile of PM2:5. Every 10 lg=m3
increase in the 2-y average of PM2:5 was associated with a 6% increase in the risk of CKD (95% CI: 2%, 10%).
The association between PM2:5 exposure andCKDdevelopment remained robust after adjustment for a wide range of potential
confounders and modifiers. It is well documented that demographic and lifestyle factors are associated with the risk of CKD. We used time-in-study as time scale in this study, adjusting for age at base- line, which is the typical approach in a Cox model. Although some alternative time scales (such as attained age and calendar time) were proposed to reduce bias, there is no option that can be assumed with certainty to be “the best” (Griffin et al. 2012). We observed slightly stronger associations after adjusting for these factors. In addition to demographic factors and lifestyles, cardiovascular risk factors are also closely associated with CKD development. Emerging literature has also suggested that air pollution is associated with cardiovascu- lar risk factors, such as hypertension and diabetes. However, adjust- ment for cardiovascular risk factors and diseases in this study did not significantly change the association, suggesting that they may not have significant intermediate effects in the association between PM exposure and CKD development. In addition, interaction tests did not show significant intermediate effects in subgroup analysis (Table 3). We also explored the potential effects of baseline eGFR and urinary protein. Our results showed that they had no significant influences on the association.