Background/aim Air pollution is a major public health concern. Multi-pollutant models are usually used in air pollution studies to identify the independent health effects of more than one pollutant. However, these model estimates can be unstable and biassed due to correlations between the exposures and exposure measurement error. Our goal is to assess and quantify this bias for PM_2.5 and NO₂ using simulations.

Methods A systematic review and meta-analysis on the differences between ambient concentrations and personal exposures only from outdoor origins was conducted. It has provided plausible values for the error structures to use as simulation input variables. Studies on building infiltration rates and activity patterns were reviewed and applied to total personal exposures for the calculation of personal exposure from outdoor sources. Then, hypothetical true and error-prone exposures of classical, Berkson and mixture types were created and the appropriate multi-pollutant models were fitted. Also, as a sensitivity analysis, we applied measurement error correction formulas (Regression Calibration and SIMEX) to check the effects of measurement error on real-life concentration-response functions.

Results Review results indicate that ambient concentrations of PM_2.5 are greater than personal exposure from ambient sources, by an average of 5μg/m³. However, results present heterogeneity based on the area, climate and participants’ age. NO₂ work is under investigation. Regarding the simulation results, we confirm the findings from the literature. For classical error, bias is observed in our preliminary results, especially when the variance of the errors is relatively high. For Berkson type, the effect estimates were, as expected, not statistically significantly different from the true ones. We will update our results with better informed input variables from our review, to identify the true independent effects of the pollutants. Also, we will compare the results of SIMEX and Reg Calibration.

Conclusion Simulations can lead to the quantification of the consequences of measurement error and adjusting for it can result in better model estimates. It may be inferred that certain potential interpretations are more unlikely than others. The ultimate aim of this work is to apply new understanding to the selection of concentration-response functions for health impact assessment.