Investigating the Consequences of Measurement Error of Gradually More Sophisticated Long-Term Personal Exposure Models in Assessing Health Effects: The London Study (MELONS).

INTRODUCTION

Cohort studies have been widely used to estimate the effects of long-term exposure to air pollutants on health outcomes. The nature of the exposure (i.e., personal exposure to outdoor-generated pollution) and the large number of participants in cohorts preclude measuring individual exposure longitudinally. Thus, surrogate measures, such as exposure models, are increasingly used in epidemiological studies to estimate individualized long-term exposures. We evaluated whether increasingly detailed estimates of long-term individual exposure in large-scale studies yield better estimates of the health effects of exposure to outdoor air pollution. We utilized several personal exposure measurement campaigns, which were implemented before the start of MELONS, the uniquely dense monitoring network and surrogate measures previously developed for London.

METHODS

Data from 344 participants in four personal measurement campaigns, two measuring particulate matter ≤2.5 μm in aerodynamic diameter (PM), nitrogen dioxide (NO), and ozone (O), and two measuring black carbon, covering 12,901 person days during 2015-2019, were used. The total personal exposure measurements were separated into exposures from outdoor and indoor sources by estimating appropriate infiltration factors and behaviors. The exposures were extrapolated from the measurement period per subject (from a few days to >9 months) to annual exposures, taking ambient concentration, infiltration, and behavior variability into account. These annual exposures were defined as true exposures, although it is acknowledged that several assumptions involved in their estimation introduce uncertainty. Surrogate measures of exposure were assigned based on the nearest fixed-site monitor to the residence or the prediction from combined dispersion, machine learning, and land use regression models at the participants' residence. The models were adjusted for age-group and area-specific time-activity patterns based on a large survey. Measurement errors (MEs) were calculated between "true" and surrogate exposures and used as input in a simulation study to investigate the resulting bias in health effect estimates, using total mortality as a health outcome. We estimated the amount of classical and Berkson error in the ME. In addition, we tested, in several theoretical error scenarios, the effectiveness of two correction methods: simulation extrapolation (SIMEX) and regression calibration (RCAL). Finally, we applied the different surrogate exposure methods using data from the UK Biobank London cohort (~62,000 subjects) to assess associations with several mortality and morbidity outcomes in Cox regression models adjusted for multiple covariates and applied correction methods.

RESULTS

Exposure to outdoor-generated pollution accounted for at least 50% of total personal exposure, even in subjects spending almost all of their time indoors. We found large MEs, possibly due not only to the nature and uncertainty of using surrogate measures but also to several uncertainties incorporated in the "true" exposure assessment. The resulting bias in health effect estimates from ME was large and almost always toward the null (i.e., the health effects are underestimated, sometimes by as much as 100%). Larger total ME and larger proportion of classical ME led to more underestimation of effects. SIMEX and RCAL were effective methods for bias correction. Furthermore, the different scale (magnitude) of measurement of surrogate exposure estimates of ambient concentrations introduced additional systematic ME, which was addressed by expressing the effects per interquartile range and not per fixed increment of the pollutant. The application to the UK Biobank cohort data showed hazard ratios above 1 for a few outcomes and surrogate exposures, which were corrected, leading to larger estimated effects.

CONCLUSIONS

Our results underline the importance of exposure to ambient air pollution ME in estimating health effects and the difficulty in obtaining an accurate estimate of the "true" personal exposure to outdoor-generated pollutants. The common use of surrogate measures of exposure introduces ME, which can be substantial and largely classical, leading to a large underestimation of effects on health. Researchers should consider correcting for ME when reporting results from epidemiological studies on the effects of long-term air pollution exposures and plan ahead by designing appropriate validation studies.