Personal exposure to ultrafine particles: Two-level statistical modeling of background exposure and time-activity patterns during three seasons - Supplementary material Veronika Deffnera, Helmut Küchenhoffa, Verena Maiera, Mike Pitzb, Josef Cyrysc,d, Susanne Breitnerc, Alexandra Schneiderc, Jianwei Guc,d, Uta Geruschkatc, Annette Petersc a Statistical Consulting Unit, Department of Statistics, Ludwig-Maximilians-Universität, Akademiestr. 1, 80799 Munich, Germany b Bavarian Environment Agency (LfU), Augsburg, Germany c Helmholtz Zentrum München, Institute of Epidemiology II, Ingolstädter Landstr. 1, 85764 Neuherberg, Germany d Environment Science Center, Universität Augsburg, Universitätsstr. 1a, 86159 Augsburg, Germany 1. Analysis of comparison measurements The usage of measurement devices over a longer period carries the danger of a gradually decreasing performance of the mobile CPC devices resulting e.g. from degeneration or from soiling. This yields temporally biased measurements. Since measurements from different devices involve a device-specific bias the joint analysis of the data requires the correction of these biases. Comparison measurements were conducted before and after each measurement campaign of our study. Data, analyses and correction procedures are described in this chapter. The following devices used in our study were included in the comparison measurements; they measure with differing temporal resolutions: • Stationary: Condensation Particle Counter (CPC 3025, TSI Inc., USA); 1-min. average (PNC CPC UAS) Twin Differential mobility Particle Sizer (TDMPS)/Aerodynamic Particle Sizer (APS, model 3321, TSI Inc., USA); 10-min. average, updated every 20 minutes (PNC 10-800 UAS) • Mobile: three Condensation Particle Counters (CPC 3007, TSI Inc., USA); 1-min. average The mobile devices operate with batteries limiting their operating time. The length of the comparison periods varied between two and eight days. For the purpose of comparison between the portable CPC 3007 devices and the devices at the measurement station we used the PNC in the size range from 3 nm to 3 µm (measured by CPC 3025) and in the size range from 10 nm until 800 nm (measured by TDMPS). The latter size range corresponds approximately to the size range of particles measured by the portable CPC 3007, however the stationary PNC levels are recorded only every 20 minutes. The different devices recorded the data in varying resolution. Therefore, the data were aggregated appropriately to compare the measurements. Mobile as well as stationary CPC measurements were aggregated to a 30-min. running mean for their comparison. For the comparison of the mobile measurements with the measurements of the appropriate fractions of the particle size spectrometer, i.e. of particles between 10 and 800 nm, only the mobile measurements at the updating time points of the TDMPS/APS were used. Linear models between the stationary and the mobile measurements yielded the correction formulas for the mobile measurements. All comparison periods were analyzed simultaneously because temporally differences were not consistent. The following correction formulas were obtained for the comparison with the stationary TDMPS/APS, which were used for the exposure analysis: Mobile device (X) Stationary measurements (Y) Correction formula CPC 07061001 PNC 10-800 nm UAS Y= EXP(0.397+0.955LOG(X)) CPC 07061002 PNC 10-800 nm UAS Y= EXP(0.278+0.976LOG(X)) CPC 10070001 PNC 10-800 nm UAS Y= EXP(0.553+0.946LOG(X)) Table S1: Correction formulas for 1-minute mobile measurements to reach comparability of the measurements based on 30-minute averaged data. The comparison measurements were conducted under controlled conditions with outdoor background data. We have not adjusted for possible differences between the devices emerging in diverse other situations, e.g. during indoor activities or in traffic. 2. Detailed model formulas 2.1. Stage 1: Background model (before variable selection) Regression equations modeling PNC of individual i at time point t with the maximum set of predictors are presented in the following. The random person intercept is denoted by . represent smooth functions which were modeled with P-splines. Indoors: Outdoors: 2.2. Stage 2: Activity model The regression equations for modeling the residual (resulting from the background model) of realization j of activity k at time point t are described in this paragraph. Indoors: Outdoors: The outcomes in the activity models were the residuals from the background model and can be interpreted as personal PNC levels which were adjusted for the background exposure to PNC. The association between the outcome and the activities was modeled through their main effects. The background adjusted personal exposure to PNC during each realization j of a particular activity k differs from the main effect of activity k; a random intercept was included in the model to represent realization-specific variations of the activities. Moreover, the variance of the outcome differed between the activities resulting in heteroscedastic variances. Therefore, activity-specific residual variances were allowed in the model. Autocorrelation of the residuals of the activity model within a realization of an activity were considered through a rational quadratic correlation structure (see Supplementary Materials 6). 3. Variable selection via model based boosting Variable selection in the background models was conducted using model-based boosting (Bühlmann and Hothorn, 2007), which is implemented in the R (R Core Team, 2013) package “mboost” (Hothorn et al., 2013). Boosting algorithms estimate coefficients by iteratively approximating the solution with small steps. Adding up the coefficients of a specific covariate of all iteration steps results in the common regression coefficients. As usual, these can be interpreted as conditional effects adjusted for the other covariates included in the model. This approach offers three advantages: (I) the association between exposure and only the relevant covariates is modeled, (II) the multicollinearity problem of the measurement values from the stationary monitor in a usual regression model is avoided and (III) the combination of base–learners for linear, smooth and random effects flexibly models the underlying structure in the data. Categorical covariates were modeled with componentwise ordinary least squares base–learners. The continuous covariates in our study were smoothly estimated using componentwise smoothing splines with three degrees of freedom. Componentwise estimation means that only the predictor with the best improvement is updated in each iteration. The cross–validation criterion to stop the iterative boosting procedure ensured the inclusion of only significantly contributing covariates. 4. Reduced data for the background models Stationary measurements above the 95% quantile and personal measurements above the 90% quantile were excluded for the background model in the first modeling stage. This choice is visualized in the following two figures: Figure S1: Scatterplots between personal indoor measurements of 1-minute PNC (x-axes) and stationary 1minute PM measurements (y-axes); the 90 % quantile of the personal measurements and the 95 % quantile of the stationary measurements are marked with lines. Figure S2: Scatterplots between personal outdoor measurements of 1-minute PNC (x-axes) and stationary 1minute PM measurements (y-axes); the 90 % quantile of the personal measurements and the 95 % quantile of the stationary measurements are marked with lines. Through the exclusion of observations with extreme stationary measurements the smooth effects in the background model at the upper border of the domain could be regarded as reliable because enough information was available. This was of major relevance because the uncertainty of the predictions in the first stage was not included in the second stage models. Using the boosting algorithm as well as other regression methods, the model is calculated through minimizing the size of the residuals. Thus, observations with high residuals potentially have a higher impact on the model calculation than other observations. Extreme personal exposure might have possibly emerged from particle generating activities which were not documented or did not fit in any of the categories of the activity diary (e.g. very high individual exposure of a volunteer while staying in the cafeteria of the university). These observations did not describe individual background exposure and were therefore excluded for the calculation of the background exposure model. Extreme stationary measurements might have emerged from particle generating events which have no bearing on the actual background exposure. Figures S1 and S2 showed that, extreme values were in the majority of observations only measured with the stationary or with the mobile device and not with both devices. The Figures also indicated the appropriateness of the choice of the 90 % quantile for the individual measurements and of the 95 % quantile for the stationary measurements for the exclusion of extreme measurements. 5. Relevance of predictor groups Let denote the predictors in a linear regression model for the outcome variable . Further, it is assumed that the set of predictors into uncorrelated subgroups with explained by a specific can be divided , ,…, as the group index. The fraction of variance predictor group with is determined through . This relation was used to evaluate the relevance of the predictor groups (individual; time including, week, day of the week and time of day; meteorology including temperature, relative humidity and dew point temperature; categorical surrounding variables including at home/not at home and status of windows (indoor models); stationary measurements including PNC and particle mass levels of various size ranges and black carbon concentrations) in the background models. 6. Categorical effects in highly correlated time series Different modeling strategies for categorical effects in longitudinal data are evaluated in this chapter with an example data set for exposure and activity data. Since the data generating process is very complex, the modeling approaches were compared using real data instead of simulated data. A small data set was used in order to explain the characteristics of models in combination with a visual imagination of the data. Three exemplary time series from the validation study, depicted in Figure S3, were chosen, one without any activity and two further time series with cooking and candle lighting. Figure S3: Three exemplary time series of personal exposure to 1-min. PNC (in 1/cm³), adjusted for background exposure. Each temporal coherent period with the same activity (or non-activity) of an individual was called the realization of an activity. Because the model contained only categorical covariates, Figure S3 lead to suspect heteroscedastic residuals, i.e. the variance of the residuals within a realization of an activity varied between the different realizations. Also realization-specific correlation structures were considerable. Furthermore, categorical effects in highly autocorrelated continuous time series could be biased if an AR(1) process for the residuals was assumed because mainly the marginal observations of the activities were used to calculate the effect. Therefore, also other correlation structures were examined. These approaches were evaluated with the following models concerning their coefficient estimates, model fit and compliance of the assumptions using the example data: M1: Linear mixed model (M1a) and additional realization-specific variance of the error term (M1b) M2: Linear mixed model with AR(1) structure for the error term (M2a) and additional realization-specific variance of the error term (M2b) M3: Linear mixed model with AR(1) structure for the error term and realization-specific correlation coefficients (M3a) and additional realization-specific variance of the error term (M3b) M4: Linear mixed model with rational quadratic correlation structure (Pinheiro and Bates, 2000) for the error term (M4a) and additional realization-specific variance of the error term (M4b) M5: Linear mixed model with rational quadratic correlation structure for the error term and realizationspecific correlation coefficients (M5a) and additional realization-specific variance of the error term (M5b) M6: Additive mixed models with few (every tenth observation) (M6a) and many (every third observation) (M6b) knots The rational correlation structure is defined as with denoting the nugget, the distance of the time points and the range (Pinheiro and Bates, 2000). The exposure to PNC during single realizations of activities was assumed to be independent resulting in a blockdiagonal correlation structure; the blocks were determined through the realizations of the activities. A realization-specific random intercept was included in all models. Also realization-specific correlation structures were considered, because the level of autocorrelation may vary depending on the way of conducting a certain activity. If the model was correctly specified, the linear mixed model M1 provided consistent estimates of the effect coefficients. Therefore, models accounting for heteroscedastic and autocorrelated residuals should yield similar estimates for the coefficients. Further reference criteria for the evaluation of the different approaches referred to the homoscedasticity (Breusch-Pagan test (BP test, Breusch and Pagan, 1979)) and independence of the residuals (AC-coef.). In addition, the models were compared concerning their fit via the Akaike information criterion (AIC). The results of the model comparison are depicted in Figure S4. The linear mixed model (M1a) with random person intercept showed a bad model fit and strong violations of the model assumptions regarding highly autocorrelated and heteroscedastic residuals, which was only partially eliminated through the realization-specific residual variances (M1b). Including an overall as well as a realization-specific correlation structure (M2 and M3) fractionally alleviated the problem of autocorrelation and improved the model fit but did not identify the visually apparent effects of the activities. The models with an autoregressive process of order one for the residuals exhibited a strong bias in one or even both activity effect estimates of about 40 000 particles. Instead, using the rational quadratic correlation structure (M4) provided adequate estimates for the coefficients, a comparably good model fit and met the model assumptions. Individual-specific rational quadratic correlation coefficients (M5) degraded the model fit due to the convergence of the algorithm to a local maximum. Smoothing splines (M6) for modeling the temporal dependencies were not able to remove the autocorrelation of the residuals; furthermore, the AIC of these models was high. Model M4b showed the best properties and was used for the analyses. Figure S4: Model comparison regarding coefficients, residual autocorrelation (AC-coef.), heteroscedasticity (BP test, asterisks mark significance) and model fit (AIC). 7. Relative durations of activities Winter Spring Public City Sum m er Public City Public City Scenario Car transport center Car transport center Car transport center indoors, at hom e 57.0 53.3 0.1 56.9 54.1 0.0 60.4 56.1 0.0 indoors, not at hom e 20.4 18.6 30.5 18.1 17.7 25.6 18.2 18.3 30.6 opened 4.3 3.1 0.9 14.1 12.1 7.7 8.4 11.3 3.8 closed 65.8 64.6 25.4 48.8 39.8 13.0 57.5 36.2 25.1 tilted 5.8 2.8 0.7 10.8 20.3 2.2 13.0 27.1 1.3 air conditioning/ventilation 0.4 0.0 0.6 0.0 0.0 0.0 0.0 0.0 0.0 passive sm oking 0.6 1.4 2.4 0.3 0.1 1.5 0.0 0.3 1.3 sm oke of w ood 0.6 0.8 0.0 2.0 0.0 0.0 0.0 0.0 0.0 sm oke of candles 6.1 4.4 1.3 0.5 1.6 0.1 3.0 0.6 0.0 cooking 6.0 4.2 0.7 6.3 6.2 0.0 5.4 4.7 0.0 cleaning 2.9 2.2 0.0 4.5 2.7 0.0 3.8 2.0 0.0 ironing 1.0 5.4 0.0 0.0 3.3 0.0 3.0 1.6 0.0 spray 0.6 0.4 0.0 1.2 0.3 0.0 1.1 0.8 0.0 dust 5.4 3.3 0.0 8.2 7.2 0.0 5.4 7.2 0.0 smell of food 1.7 0.0 0.6 1.2 2.0 0.1 1.9 2.5 0.1 smell of smoke 0.6 0.3 0.2 0.4 0.4 0.5 0.1 0.1 0.3 smell of vehicle exhaust 0.0 0.0 0.4 0.0 0.0 0.1 0.0 0.0 0.0 building site 0.0 0.0 0.0 0.0 0.0 0.1 0.0 0.0 1.1 steam 0.4 0.2 0.0 0.4 0.4 0.0 0.5 0.1 0.0 laser printer 0.7 0.0 0.0 0.0 0.0 0.0 0.1 1.0 0.0 1.6 0.9 5.4 1.9 1.4 6.4 1.7 2.0 5.8 12.8 20.2 52.3 13.6 19.0 58.4 11.8 16.8 55.8 11.9 0.0 0.0 11.7 1.3 0.0 10.8 0.0 0.0 pedestrian 0.8 2.6 52.3 1.9 1.3 58.4 1.0 2.0 55.8 public transport (bus/tram/train) 0.0 17.6 0.0 0.0 16.5 0.0 0.0 14.8 0.0 w indow sm ell abnorm alities outdoors, not in traffic outdoors, in traffic m eans of transport car Table S2: Percentage frequency of activities (recorded every minute) during the three scenarios for each measurement period. 8. Daily variations of 5–min. averaged PNC Figure S5: Mean diurnal variation of 5-minute averaged PNC; solid: personal PNC levels, dashed: PNC levels at the stationary monitor. Mean personal exposure to PNC was higher than the concentration measured at the station. Increased indoor exposure to PNC was mainly measured at about noon during scenarios “car” and “public transport”, when the volunteers prepared their food. Mean personal exposure to PNC was similar to stationary measured PNC levels during scenario “city center”. 9. Correlations between personal and stationary measured PNC levels Indoors Outdoors Winter Spring Summer Winter Spring Summer Air temperature -0.037 * -0.122 * -0.164 * -0.138 * -0.048 * 0.010 Rel. humidity -0.164 * 0.059 * -0.055 * -0.126 * -0.039 * -0.018 Dew point temperature -0.214 * -0.095 * -0.155 * -0.323 * -0.073 * -0.017 PM2.5 0.069 * 0.173 * 0.003 -0.047 * 0.078 * 0.208 * PM10 0.120 * 0.199 * 0.067 * 0.010 0.126 * 0.211 * NC CPC 0.164 * 0.150 * 0.228 * 0.347 * 0.413 * 0.294 * BC 0.093 * 0.122 * 0.103 * 0.169 * 0.273 * 0.154 * NC 10-800 0.210 * 0.190 * 0.229 * 0.365 * 0.414 * 0.334 * Table S3: Spearman rank correlations between 1-minute personal PNC levels and stationary concentration levels and personally measured meteorology; coefficients which significantly differ from zero are marked with *. The correlations (Spearman rank correlation) of personal exposure to PNC with mobile meteorology measurements and particulate matter concentrations from the stationary monitor are presented in Table S3. Indoor personal exposure to PNC was not correlated with any of the examined variables (|r| < 0.23). Correlations of outdoor personal PNC levels with the number concentration of the corresponding particle size range and the stationary PNC levels were slightly stronger with highest values in spring. Stationary PNC measurements weakly correlated with PM2.5 (r = 0.374) and PM10 (r = 0.431) recorded at the stationary monitor. 10. Exposure ratio: logarithmically scaled abscissa Figure S5: Ratio between the median 1-minute exposure to PNC of each volunteer during a particular activity and the median 1-minute exposure to PNC of all individuals stratified on the season, separately for indoor and outdoor activities (dots; blue: winter, green: spring, red: summer). Exposure ratios for more than four persons are summarized by a boxplot. Bibliography Breusch TS, Pagan AR. A simple test for heteroscedasticity and random coefficient variation. Econometrica 1979;47(5):1287–94. Bühlmann P, Hothorn T. Boosting algorithms: Regularization, prediction and model fitting. Stat Sci 2007;22(4):477–505. Hothorn T, Bühlmann P, Kneib T, Schmid M, Hofner B. mboost: Model-based boosting, R package version 2.2-3, 2013, http://CRAN.R-project.org/package=mboost. Pinheiro JC, Bates DM. Mixed-Effects Models in S and S-PLUS. New York: Springer; 2000. R Core Team. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria 2013. URL http://www.R-project.org/.