**4. Discussion**

Due to the limited time most people spend outside, the amount of ambient concentration of PM2.5 that people are directly exposed to is likely to be di fferent based on variation in people's behavior and the performance characteristics of the buildings they are occupying [55]. Consequently, spatial variability, time-activity and losses due to outdoor-to-indoor transport are all sources of exposure uncertainty in the epidemiological analysis, when fixed-site monitor concentrations are used as surrogates for exposure to air pollution. In this work we established a more comprehensive understanding of population exposure concentration and the impact that di fferent exposure metrics can make on all-cause mortality predictions. We showed that the I/O ratios and individual's patterns of movement play a key role in estimating exposure to PM2.5 and that transportation-MEs, predominately the highly polluted London Underground, are important in accurately establishing exposure. We demonstrated that subway and Indoor MEs make a significant contribution to the exposure misclassification and therefore mortality change predictions. Azimi and Stephens [31] highlighted the importance of including indoor MEs when estimating the total exposure and the need for a better understanding of how the infiltration factors vary by building type in order to improve the exposure estimates and reduce the uncertainty. Based on field measurements, they found that exposure to PM2.5 of outdoor origin inside the residence contributed around 67% to the total U.S. mortality burden. In our analysis, we found that the Indoor environment contributed approximately 83% to the total mortality burden in London. The di fference in our results may be explained by the di fferent MEs considered in each study. As our aim was to quantify the misclassification and give an insight into how the absence of significant MEs from an exposure assessment could increase the uncertainty, we mainly focused on the di fferent infiltration factors of home types and the LU. Martins et al. [56] determined the PM2.5 exposure and estimated the daily PM2.5 dose during Barcelona subway commuting. They estimated that the PM2.5 dose received by an adult in the subway contributed approximately 46% to the total daily dose in the respiratory tract. In our study, LU contributed approximately 15% to the total health burden. Due to the di fferent methods used and di fferent health endpoints, their results cannot directly be compared to ours. However, their outcomes indicate the non-trivial contribution from subway ME on health e ffects estimates. Several studies have compared static (home address-based) with more dynamic air pollution methods and proved that there is a reduction in average total exposure levels in urban areas with related characteristics as GLA [32,57]. Tang et al. [57] used a staged modelling approach to evaluate the use of static ambient concentrations as exposure estimates and examined the impact of dynamic components on estimated air pollution exposure. They found that the mean population exposures in Hong Kong for their full dynamic model were approximately 20% lower than the ambient baseline estimates of the static approach. Smith et al. [32] combined a dispersion modelling approach with building infiltration factors and travel behavior in order to create the London Hybrid Exposure Model (LHEM). They found that their model's estimates were around 37% lower for PM2.5 than the static approach (residential address-based). Similarly, by adopting a staged modelling approach to evaluate the e ffect of including dynamic components to our exposure models we found that the absence of mobility and infiltration factors in the static Tier-model 1 led to an overestimation of annual PM2.5 population exposure. Overall, the exposure estimates of our most complex model (Tier 5) were around 34% lower than those of the static baseline model (Tier 1). These findings were di fferent from Tang et al.'s [57] study but very similar to the LHEM study, mainly because the study population was the same and similar travel behavior data was used. Recently, Singh et al. [58] quantified the population exposure to PM2.5 concentrations in London and assessed the importance of including movement and indoor infiltration to total population exposure. They found that their refined exposure assessment predicted 28% lower total population exposure than the traditional static exposure method. As in this study, the time-activity data were derived from the LTDS [39] and the study area was London. However, the small di fference between their results and ours could be explained by the di fferent datasets used for the infiltration factors and the di fferent concept used for the key MEs (e.g., the London Underground). Results from other similar studies are di fficult to find as we compare di fferent exposure estimates during the same time period (2017) in an e ffort to examine the e ffect on all-cause mortality predictions. We showed that using a static exposure metric instead of a more dynamic approach (based on time-activity data and indoor infiltration) to predict the mortality in the GLA population would lead to an overestimation of 1174–1541 mean predicted estimates of mortality attributed to PM2.5. Ebelt et al. [59] found for several health outcomes associated with cardiopulmonary diseases, analyses with ambient exposures resulted in larger e ffect estimates. These results strongly supported their original hypothesis that the reduced exposure misclassification resulting from the utilization of ambient exposures instead of ambient concentrations provide more precise estimates of e ffects in epidemiology.

This work provides further understanding as to the impact of an exposure assessment on the mortality predictions and helps to mitigate the uncertainty in health risk assessments of air pollution. As a result, it would be possible to increase the e fficiency of regional or local air quality managemen<sup>t</sup> strategies.

#### *Limitations and Future Work*

The current study contains several limitations. Only some of the deep and subsurface underground lines were monitored and only for a small sampling period. In this study, we assumed that these measurements also represented the corresponding lines that were not measured. Moreover, only 23 monitoring stations were available for PM2.5 and their locations were not uniformly spread across the study area. Consequently, this may have a ffected the simulation accuracy and the interpolated ambient concentration estimates in the unmonitored areas that were far from the stations that might have contained higher uncertainty. Furthermore, another limitation was the assumption that the Indoor microenvironment and the average dwelling I/O ratios also represented the o ffice and commercial buildings. The toxicity of PM2.5 was not included, but mainly because it was out of the scope of the study to investigate the toxicity of the particles.

The space–time–activity data is based upon the London Travel Demand Survey for the period 2005–2010 and may not be fully accurate locally, spatially and temporally, for the year 2017. Moreover, the annual average of the time-activity data that we used, assuming that people followed the typical daily mobility patterns for the whole year, may have increased the uncertainty in our models, because those data might not have accurately represented a part of the population. Since the main body of our study was based on averages and the population was not divided into di fferent age groups, our health burden predictions may be less accurate for special groups of people that have di fferent behaviours (e.g., ill or elderly that spend most of the day inside their residence).

Parameters that could a ffect particle infiltration, such as di fferences in indoor-outdoor air pressure due to the impact of the surrounding micro-environment, and the existence and e fficiency of mechanical filtration, were not the focus of the current study and were therefore not investigated.

In the future, this study could be improved by conducting further measurements in the London Underground and for larger periods of time. Simple sensitivity tests could be made in order to check each model's response and how the misclassification a ffects our estimates. As the next stage of this work we could investigate how this framework applies to other cities with higher ambient PM2.5 concentrations and di fferent indoor characteristics (such as interventions-PACs, HVAC). Taking into consideration that each urban area may have di fferent characteristics, it is important to examine how the incorporation of the local urban or building features could make an impact on exposure concentration estimates and health burden predictions.
