Ozone exposure assessment in a southern California community.

An ozone exposure assessment study was conducted in a Southern California community. The Harvard ozone passive sampler was used to monitor cohorts of 22 and 18 subjects for 8 weeks during the spring and fall of 1994, respectively. Ozone exposure variables included 12-hr personal O3 measurements, stationary outdoor O3 measurements from a continuous UV photometer and from 12-hr Harvard active monitors, and time-activity information. Results showed that personal O3 exposure levels averaged one-fourth of outdoor stationary O3 levels, attributable to high percentages of time spent indoors. Personal O3 levels were not predicted well by outdoor measurements. A random-effect general linear model analysis indicated that variance in personal exposure measurements was largely accounted for by random error (59-82%), followed by inter-subject (9-18%) and between-day (9-23%) random effects. The microenvironmental model performs differently by season, with the regression model for spring cohorts exhibiting two times the R2 of the fall cohorts (R2 = 0.21 vs. 0.09). When distance from the stationary monitoring site, elevation, and traffic are taken into account in the microenvironmental models, the adjusted R2 increased almost twofold for the fall personal exposure data. The low predictive power is due primarily to the apparent spatial variation of outdoor O3 and errors in O3 measurements and in time-activity records (particularly in recording the use of air conditioning). This study highlights the magnitude of O3 exposure misclassification in epidemiological settings and proposes an approach to reduce exposure uncertainties in assessing air pollution health effects.

The current National Ambient Air Quality Standard (NAAQS) for ozone (03), 0.12 ppm, is based on the health effects due to acute exposures of 1 hr as measured by continuous monitors. However, exposures to 03 for up to 8 hr at less than 0.12 ppm have been shown to result in progressive and significant changes in respiratory function in exercising individuals (1)(2)(3), suggesting that the current 03 standard may not sufficiently protect public health. In response, one of the alternative 03 standards being considered by the EPA is the 8-hr average concentration (4). The use of the integrated passive monitor to obtain exposures greater than 1 hr can greatly enhance our ability to determine the dose-response relationship for acute exposures to 03. The 12-hr 03 measurements are biologically significant when combined with the retention factor of 03 in the deep lung and the ventilation rate to produce the 12-hr delivery dose of 03 (5). Our recent epidemiologic study showed that these 03 dose estimates, not 1-hr maximum 03 measures at the stationary site, were associated with respiratory symptoms and inhaler use among asthmatics (5). Using the 12-hr personal measurements significantly reduces the magnitude of expected exposure misclassification in studies that have relied solely upon 03 measurements from outdoor stationary site monitors to represent personal exposure to 03.
Previous studies (6-9) that examined short-term (12-hr) personal 03 exposures have involved relatively short monitoring periods, generally less than 5 days, or monitored less than five subjects simultaneously. Epidemiological research on the acute and adverse respiratory effects of 03, on the other hand, generally involves repeated daily measurements over several weeks or months in larger cohorts (panel studies). To examine personal 03 exposure and its determinants in a setting directly relevant to epidemiological research, the present exposure assessment study was integrated into two consecutive asthma panel studies. More precisely, this study involved daytime (12-hr) personal 03 monitoring in cohorts of 23 and 18 subjects for two 8-week periods during the spring and fall of 1994, respectively. Extensive outdoor active monitoring throughout the study region was conducted during the fall period. Because of the diverse geographical characteristics of the study area, it was possible to examine variations of 03 concentrations in a three-dimensional domain.
The purpose of this study was to investigate personal 03 exposures among subjects during both spring and fall seasons in the Alpine area and to investigate the feasibility of using ambient 03 measurements from one outdoor fixed site as well as the activity patterns from the subjects to pre-dict personal 03 exposures. This study further examined the influence of outdoor temperature on activity patterns and the effects of activity patterns on personal 03 exposure levels. In addition, the extent of the effects of outdoor 03 spatial variation on the predictive power of personal exposure models was investigated.

Methods
This study was conducted in the Alpine area of San Diego county, California. Alpine (population -12,000) is located approximately 20 miles east of San Diego. Residents of the Alpine community live around or above the base of the average air inversion layer (1,200 ft or 366 m above sea level) (10). High levels of 03 above 120 ppb have been measured on many days per year and a permanent government monitoring site (using a continuous UV photometric 03 analyzer) has been in operation since 1981.
The Harvard 03 passive sampler was used for personal monitoring. The principle of the sampler is oxidation of nitrite (NO2-) by 03 to form nitrate (NO3-), which is quantified by ion chromatography (11). Field blanks (n = 184) and duplicate samples (n = 52) were used for quality assurance and quality control (QA/QC). The limit of detection (LOD), calculated as three times the standard deviation of the field blank values, was 17 ppb of 03 for the 12-hr average samples. The uncertainty, defined as the variance of difference between duplicates divided by N (12), was 3 ppb.
Subjects recruited for the spring study included 9 males (mean age = 18 years; range = 10-38) and 13 females (mean age = 24 years; range = 10-47). Of these subjects, 13 were pediatric subjects and 9 were adults. Informed consent was obtained from all subjects who were monitored simultaneously from 9 May to 3 July 1994. During the fall, 18 subjects were monitored simultaneously from 6 September to 31 October 1994. These subjects included 11 males (mean age = 16 years; range = 9-38) and 7 Articles -Ozone exposure assessment in Southern California females (mean age = 21 years; range = 9-38). Of these, 13 were pediatric subjects and 5 were adults. Fourteen subjects had been previously monitored in the spring. The monitoring duration was approximately 12 hr, starting when subjects awoke, generally between 6 and 8 A.M. Subjects were given clock-shaped time-activity diary forms to record activities (time indoors and outdoors, in Alpine area or outside Alpine area), the level of physical activities, and the use of air conditioning. The time resolution of the diary is 15 min.
The study area is located in a complex terrain, with an altitude ranging from less than 600 ft in the west to over 2,000 ft in the northeast (Fig. 1). During the fall study, in addition to the San Diego County Air Pollution Control District (APCD) monitoring sites at Alpine and El Cajon (west of Alpine), the Harvard 03 active monitor (13) was used at 12 other outdoor locations chosen on the basis of providing a representative range of traffic volume and elevation in the community (Table 1). Passive samplers were collocated with four active monitors. The monitoring duration for the active samplers was 12 hr; samples were taken every other day (n = 22). The active monitor, based on the same chemical principle as the passive sampler, was designed to improve the passive method (13,14). This active sampler used a hollow tube denuder attached to a small personal pump (Model PAS-500, Spectrex Corp., Redwood City, CA). The denuder system consisted of a 1.4 cm (inside diameter) x 10 cm (length) etched borosilicate hollow glass tube attached to a personal pump, which maintained a constant sampling rate of 65 ml/min. Samples were validated by examining the field and laboratory records and were removed when records justified it (e.g., broken or wet samples, unused samples, etc.). The LOD for the active method was 0.5 ppb for 12-hr monitoring during the fall study. The mean difference between the collocated active monitors was 3 ± 9 ppb (uncertainty = 7 ppb; r = 0.83). The mean difference between the collocated active and passive measurements were 0.2 ± 16 ppb (not different from zero; r = 0.48). Although the active monitors were designed to be an improved method, it was speculated that they exhibited leakage at the inlet and outlet connections in the system. Therefore, for data analysis, active measurements were vigorously screened for anomalies by removing outliers, which are defined as measurements over the 90th or under the 10th percentiles of the ratio of active to continuous measurement.
The data analysis was conducted in three major steps: 1. Descriptive statistics were performed for both spring and fall samples. Geometric means and standard deviations were calculated when skewed distributions were observed.
2. Personal exposure modeling was performed using time-activity patterns to account for differences in exposure across various microenvironments. General linear models (GLM) were used to examine random effects of day and subject on personal exposures. Personal exposure models were then developed using the microenvironmental exposure concept and multiple regression techniques. Model 1 assumes negligible 03 exposures while subjects were indoors with air conditioning (A/C) on because of closed windows/doors and air filters known to scavenge 03 (15). The coefficients in the model are based on our earlier study results (9) in which the home outdoor 03 levels were, on average, 80% of those measured at the closest outdoor monitoring sites and the mean indoor-to-outdoor ratio was 0.3.
In addition, multiple regression models were used to predict personal exposures (E') and to compare with results from Model 1: '= a x Co x F0 +a 2X C0 x Fi + a3x C0x Fa models were examined and included: outdoor spatial variation, indoor 03 concentration variation, and the measurement error resulting from analytical error and from subjects' compliance. Note that there is currently no true gold-standard measure of personal 03 exposures. The passive monitor, in spite of being the most direct measure of personal 03 exposure, has not been validated in natural settings in which the study subject is freely mobile. The personal passive measurements were used in this paper as a reference value for personal exposures. where Fa= fraction of time spent indoors with A/C on. The split sample approach (16), in which samples were randomly split into two groups for model construction, was used to examine model reliability. Colinearity also was examined by calculating the condition index. 3. Factors considered to affect the prediction power of the microenvironmental Easting (one grid = 1 km) Filled circles represent the location of the 18 subjects' homes, the Alpine APCD site (numbered 13), and the 12 additional outdoor monitoring sites (numbered from 1 to 12). (B) The projected twodimensional topography for the Alpine area (northing upward for conventional viewing). In the fall study, the average hourly outdoor 03 concentration was lower than that in the spring. It ranged between 0 and 118 ppb, averaging 45 ± 20 ppb and never exceeding the NAAQS. Table 2 shows the descriptive statistics for both spring and fall 12-hr integrated samples. Outdoor measurements are approximately four to five times higher than personal exposures. Figure 2 demonstrates the variability of the daily average personal exposures across subjects for the spring and fall monitoring periods. Variance among subjects differs by day due to the variation in daily outdoor 03 concentrations and personal activity patterns. Although personal 03 exposures are much lower on average than continuous measurements in both seasons, the contrast is notably greater in the fall (Fig. 2B vs. 2A). The difference between personal exposures and outdoor continuous measurements can be partly explained by the time-activity pattern ( Table 3). The overall activity pattern is comparable in both seasons. The majority of time (-70%) during the daytime monitoring hours was spent indoors at home or at other indoor environments. When indoors, most time was spent at home without air conditioning. This high percentage of time spent indoors explains the substantially lower personal 03 exposures than the outdoor concentrations.
The hours spent outdoors fluctuated by days of the week, resulting in a similar variation in personal exposures. Personal 03 exposures during the spring are highest on Saturdays and lowest on Mondays and Tuesdays. The mean personal exposure on Saturdays is 22.6 ppb, while the mean on other weekdays and Sundays is 17.3 (twosample t-test p<0.001). A similar trend was observed for outdoor 03 measurements at the Alpine APCD site. This weekend effect has been observed in several California cities and has been attributed to weekday to weekend emission reductions in NOX and nonmethane hydrocarbon (due to less commuter driving), which result in reduction in 03 formation (17). Using the GLM for personal exposures and controlling for the subject effect, the effect of days of the week was found to be significant for both personal 03 exposures (p<0.001) and activity pattern (percent time spent outdoors) (p<0.001) in both seasons. Evidently, personal 03 exposures over weekends were elevated due to the greater number of hours spent outdoors and reinforced by higher outdoor 03 levels on Saturdays and Sundays.
Modeling personal exposures. Factors that affect personal 03 exposures were fur-ther examined with random effects GLM in which personal 03 exposures were regressed on day of study and subject. Day-of-study and subject effects were found to be significant in both seasons (p<0.01). The day-of-  Volume 105, Number 1, January 1997 * Environmental Health Perspectives study effect is a manifestation of the daily variation in outdoor 03 concentrations and the average activities of subjects, as well as meteorological factors such as temperature, which impacts photochemical oxidation rates, subject activities, and A/C use. The subject effect reflects the variation in time-activity patterns from one subject to another. The random error accounts for unmeasured determinants of personal 03 exposure not predicted by the subject and day variables, such as variation within activity patterns of individual subjects, compliance, and measurement errors. For spring personal exposures, the random error accounts for 59% of the variance (R2) of individual observations, while intersubject and between-day variations comprise 18% and 23% of the variance, respectively. For fall personal exposures, the random error accounts for most (82%) of the variation in personal exposure measurements (intersubject R2= 9%; inter-day P2= 9%). These large random errors are due partly to the variation in day-to-day individual activities as well as to the low 03 measurements. Springpersonal exposures. The results of the above GLM analysis indicate the importance of the individual activities in personal 03 exposure modeling. This concept is illustrated with microenvironmental models that include the time-activity information as well as outdoor 03 measurements. The predictive equation (Model 1) explains 20% (P2 = 0.20) of the variance in the measured personal 03 exposures. The slope of the regression line for Model 1 is almost unity (0.99 ± 0.02). The mean of the predicted values is 27.6 ppb, while the mean of the measured values is 24.7 ppb. For comparison, the regression model with the outdoor measurements (CO) at the Alpine APCD site as the sole predictor for personal exposures resulted in an R2 of 0.04 only (p<0.001).
The multiple linear regression equation (Model 2), which includes indoor exposures with A/C on, results in an P2 of 0.21 ( Table  4). The coefficient for outdoor exposure was 0.52, smaller than the 0.8 used in Model 1. This difference may be due to the variation in elevation and traffic counts in the community and thus the differences in outdoor 03 concentrations between home and the Alpine APCD site. The coefficient for indoor exposures without A/C is comparable to the mean Toronto indoor/outdoor ratio (0.3). The coefficient for indoor exposures with A/C on (0.22; p<0.001) implies that indoor exposure exists even when windows are dosed and the A/C is on. The X2 value for Model 2 is low; however, when Model 2 is constructed for each subject, the resultant X2 values range between 0.07 and 0.85. This indicates that the variation in the ability of the model to predict individual exposures is large and may be attributable to the performance of the sampler as well as the subjects' compliance to the monitoring protocol. The reliability of Model 2 is further validated using the split sample approach (16). The difference in P2 values between the models that use two randomly split data sets is 0.03, and the difference in P2 between Model 2 and the models that use either split data set is less than 0.02.
FaUpersonal exposures. The same modeling effort was carried out for fall personal exposures. First, outdoor measurements, CO at the Alpine APCD site, are used to predict personal exposures (s), resulting in an R2 of 0.07 (E = 0.23 CO; p<O.OO1). A similar P2 of 0.06 (p<0.001) is found when Model 1 is used to predict personal exposures, with model predictions generally overestimating personal exposures (estimated = 20.5 ± 7.9 ppb; measured = 12.7 ± 10.2 ppb; n = 663). The slope of the regression line for the predicted versus measured values is 1.06 ± 0.03. This higher random error in the fall model as opposed to the spring model may be attributable to the lower personal O3 exposures, which may result in a higher measurement error.
The multiple linear regression model (Model 2) results in an P2 of 0.09 ( Table  4). The split sample approach also was used to examine the model reliability. The difference in R2 values between the split data sets is 0.01, demonstrating the reliability of the model. For Model 2, the regression coeffi- cients from the fall data are smaller than those used for Model 1. In particular, the coefficient for indoor exposure with A/C off (El) is smaller than that with A/C on (EJ).
The adequacy of the regression model regarding potential colinearity problems was further examined by calculating the condition indices. Results show small condition index values for all independent variables (<8), indicating that colinearity is not a problem in the models and that coefficients are stable; however, we suspect that the regression coefficients for indoor exposures with A/C on or off are inaccurate. It was observed during the study that despite instructions, young subjects might not have accurately recorded the use of A/C at school. The following sections investigate further possible modeling errors for the fall study results.

Exposures
The results of our study show that peak O3 concentrations occur at times when subjects are likely to be outdoors, as shown in Figure 3. Therefore, peak exposures should occur between 11 A.M. and 1 P.M., when subjects were likely to be outdoors. However, even at the peak of the outdoor activity profile, less than 35% of the time was spent outdoors. Therefore, subjects were generally not exposed to the high outdoor O3 concentrations as would have been measured by the outdoor monitors. In addition to the hour of the day, whether subjects spent their time outdoors also depended on the temperature. As shown in Figure 4, outdoor activities increase with temperature between 50 and 70°F. When the temperature is higher than I4--Ambienti3| 70°F, outdoor activities reach a plateau and decrease slightly at very high temperatures except for one high outdoor-activity outlier, which occurred primarily (95% of the observations) on Saturdays. Subjects tend to remain indoors to escape high temperatures; thus they unintentionally avoid high 03 exposures. We expected that the percentage of A/C usage would be high while subjects were indoors and outdoor temperatures were high, especially during school hours. Although the reported A/C usage (percent of time A/C on while subjects are indoors) and the corresponding hourly outdoor temperature are well correlated (R2 = 0.88), no more than 35% of the subjects reported using A/C even when the temperature exceeded 900F. While this may be true at home when electricity costs are a concern, it should not be the case at the four schools in the area where the A/C is operated either manually by teachers or automatically. Reporting errors could have contributed to the low percentage of A/C use, which in turn results in an overestimation of personal 03 exposure in the above models. To adjust for this potential bias, we developed a model by assuming that there is a linear relationship between A/C use and outdoor temperature and that most A/Cs are on at temperatures greater than 80°F. However, with this A/C model, the predictive power (RJ2) of Model 1 was not improved, although the mean of the predicted values is closer to that measured. We also further examined the first term in Model 1 (0.8 x CO x Fo), i.e., the spatial variation in 03 exposure in relation to the stationary site levels.

Spatial Variation
Continuous measurements. Because of the high altitude, 03 concentrations at the Alpine APCD site (622 m, mean 03 = 54 ppb) are always higher than those at the nearby El Cajon APCD site (200 m, mean 03 = 41 ppb). In addition, 03 levels at these two locations did not follow the same trend. The correlation between these two continuous measurements depends on the averaging time period. The R2 for hourly 03 concentrations at these two sites is 0.38. The J2 decreases when longer averaging durations for 03 concentrations are used, especially when averaging duration extends into early morning (before 7 A.M.) or late evening (after 7 P.M.). It is possible that the radiation inversion layer may at times be lower than the Alpine APCD site, especially after sunset and before sunrise. In this case, 03 aloft did not mix with the air underneath, resulting in higher 03 concentrations at the higher altitude Alpine APCD site than at the El Cajon APCD site. This observation may explain in part Volume 105, Number 1, January 1997 * Environmental Health Perspectives Articles -Ozone exposure assessment in Southern Califoria the low R2 for exposure models based on stationary measurements. Active measurements. The spatial pattern of outdoor 03 concentrations in the Alpine area is further examined using the active 03 measurements at the additional monitoring sites (Fig. 1). The contour plot for the average outdoor 03 concentrations in fall (Fig.  5) shows that sites near the town of Alpine have the lowest 03 concentration while sites at the highest elevations have the highest 03 concentration. Table 5 summarizes 03 concentrations at different elevations and traffic conditions. The mean outdoor 03 concentrations at elevations less than 600 m are comparable. The mean concentration for sites located at elevations greater than 600 m is on average 10 ppb higher than others. Ozone concentrations at locations in low to medium traffic areas are on average 5 to 6 ppb higher than those located in heavy traffic areas. The GLM results indicate that after controlling for the effect of day of study, both traffic (p<0.01) and elevation (p<0.001) trends are significant. To predict 03 concentrations at the active sites, continuous measurements at the Alpine APCD site are first used as the only predictor in a regression model (R2 = 0.49). However, the multiple regression model, which indudes traffic conditions (1, light; 2, medium; 3, heavy), distance between the active and Alpine APCD sites, and elevation, only improves the predictions slighdy (R2 = 0.53).
Personal measurements. The cross-sectional correlation was calculated to further compare personal exposures with the Alpine APCD site measurements. The R2 varies substantially by subject, ranging between 0 and 0.25. The low R2 may be attributed in part to the spatial variability in outdoor 03 concentrations. Figure 6 shows the relationship between individual 18 values and the distance between the subject's home and the Alpine APCD site. A decreasing correlation with distance for possibly two groups of subjects is observed. The first group, above the fitted curve, exhibits a higher correlation with the Alpine measurements even at distance as far as 12 km. The second group shows a rapid decrease in 18 with distance. Most subjects in the first group (five out of six) live in areas with low to medium traffic, which are comparable to the traffic conditions at the Alpine APCD site. This contrasts with the second group, with subjects living in areas with medium to heavy traffic.
To examine the effect of this spatial variability, the microenvironmental model is modified by adding predictive factors, including traffic conditions, three-dimensional distance from the home (or school for students during weekdays) to the Alpine APCD site, and the difference in  (Table 6). 040 Because subjects tended to be outdoors 0.40 during 03 peak hours (Fig. 3 measurements as well as the spatial interpola- Figure 6. R2 for personal exposures and the tion model predictions, including kriging Alpine air pollution control district (APCD) site and inverse distance predictions, were used to measurements for individual subjects as a funcpredict outdoor 03 concentrations. The pre-tion of the three-dimensional distance between dicted values were then used in Models , and the subjects' home and the Alpine APCD site. d(cted values were then used i Models 5 and (Note that subjects did not always stay home and 6 ( Table 6) to replace C0. However, these that the distances are only an approximation.) models did not improve the R2 values.
Colinearity was examined for all models. The highest correlation between the independent variables is 0.64 between distance and elevation difference. The tolerance values for all independent variables are greater than 0.1 and the condition index values are smaller than 20. Therefore, colinearity is not evident for any models. For the spring model, when additional variables for distance, elevation, and traffic conditions were added to predict 12-hr personal exposures, the R2 (adjusted 2 0.22; significant predictors include outdoor and indoor exposures, traffic, and distance) does not significantly differ from that ofModel 2 (R2 = 0.21).

Discussion
The ability to predict personal exposure to 03 is important in that large-scale epidemiological studies on the adverse effects of 03 could be conducted with greater accuracy but without the considerable expense of personal sampling. The present study demonstrates the methodology needed to carry out such an endeavor, as well as the limitations of the resultant predictive models. Although the models reported here are unlikely to ulfill the need for accurate exposure assessment in an epidemiological setting, they do provide valuable insight into the determinants of 03 exposure and the relevance of outdoor stationary site measurements to exposure.
For instance, personal exposures to 03 were found to differ by the days of the week. More time is spent outdoors on weekends, when outdoor 03 concentrations are higher, than on weekdays. During the daytime period, more time is spent outdoors during the period when outdoor 03 peaked. Personal exposures are therefore enhanced by such activity patterns. Nevertheless, subjects do not spend a significant percentage of time outdoors (<35% at the peak of the activity profile). Since most of the time is spent indoors, cumulative indoor 03 exposures are important. Indoor 03 concentrations, however, vary with the ventilation conditions and home characteristics. Although deposition of 03 on clothes may result in a reduced personal measurements on the passive sampler (12), the analysis on time-activity pattern suggests that the low personal exposure is not an artifact but a manifestation of the low fraction of time spent outdoors.
The R2 between the Alpine and El Cajon APCD sites ranged between 0.37 and 0.53, while the R2 for Models 1 through 6 ranged between 0.08 and 0.22. The lower R2 values for these microenvironmental models can be explained by the fact that the outdoor monitors are fixed and thus easier to predict while subjects move around different microenvironments with various 03 concentrations. In addition, personal exposure measurements were not made at fixed times and durations. The analyzed personal measurements include sampling durations between 9.6 and 14.4 hr (within a 20% range of 12 hr), starting between 6 and 10 A.M. Furthermore, the low R2 may also result from the measurement error due to different effective collection rates of the passive sampler in various environments (12). Correlations between personal exposures and the outdoor measurements varied among subjects due to various activity patterns and geographical and demographic characteristics near the home, school, or workplace. Intrasubject variance and monitoring error probably account for most of the variability in personal exposures. Less than 40% of the variance in personal 03 exposure was attributed to intersubject and day-ofweek variabilities.
Based on outdoor concentrations obtained from the Alpine APCD site and time-activity information collected from the subjects, simple personal exposure models predict 20% and 6% of the variability in the measured personal exposures in spring and fall, respectively. Although such predictive powers are low, the models are reliable when assessed by the split sample method. Inaccurate time-activity information on the use of air conditioning might have contributed to the low R2.
Efforts will have to be made in the field to correct this A/C reporting error, such as taking daily air exchange rate measurements or collecting information on A/C use from workplaces and the school. In addition, the spatial variation of 03 also resulted in variation in personal 03 exposures in the study area. By adding factors that are associated with the spatial variation to the simple personal exposure models, the R2 was markedly improved from 0.06 to 0.19 for the fall data. Similar modeling efforts, however, only resulted in a minimal improvement of the predictions for the spring data. It is possible that measurement error from the personal monitor and the time-activity records accounts for most of the variance in personal measurements.
Conclusions Key findings of this study include the following: * Personal 03 exposure differs dramatically (by fourfold on average) from outdoor stationary 03 measurements (Fig. 2) and is not predicted well by these outdoor measurements (R2 = 0.07 or less). * There is considerable intersubject variability in personal 03, which is partially explained by differences in percent of time spent outdoors. * Models perform differently by season; microenvironmental Model 1 and multiple regression Model 2 for spring cohorts have two to three times the R2 of the fall cohorts. * The fall, but not the spring, model was markedly improved by adjusting for distance from the outdoor stationary site. Recent studies (8,9,18,19) that examine associations between personal exposures and outdoor measurements found similar values for R2. Thus, microenvironmental models including outdoor 03 concentrations and simple personal activity information are improvements in estimating personal exposures over models including only outdoor stationary site measurements. Predicting personal exposures was further improved by the inclusion of microenvironmental variables such as spatial variation and workplace (e.g., school) exposures in the present and past 03 modeling studies (9). However, it remains to be shown to what degree such predictive models enhance the ability to detect the adverse respiratory health effects of 03 in an epidemiological study that would otherwise rely on outdoor stationary site 03 measurements.