We documented the stability of NO2 measurements and predicted levels of exposure by two LUR models twelve years apart. The NO2 concentrations measured in the same locations had a good agreement, showing the stability of spatial contrast. This study shows that 74 % and 70 % of the NO2 variability in Rome in 1995/96 and in 2007, respectively, is explained by land use and traffic variables. The 2007 LUR model was able to predict NO2 levels measured 12 years before in a better way than what the 1995/96 model was able to do prospectively (r = 0.83 vs. 0.73). When we applied the estimated NO2 from the two LUR models to residential addresses of the large cohort members, the correlation was extremely high (0.96). The association between NO2 exposure at residence and mortality from 2001 to 2006 was very similar using the 1995/96 and the 2007 models.
The choice of sites for measurements was driven by the 1995/96 campaign which aimed to investigate the level of exposure of children at school. The schools, randomly selected, were placed mostly in urban background sites (44/67). We did notice that the prediction from the LUR model underestimated to some extent the highest levels of NO2. On the other hand, 17 % of the residents in the city live at less than 50 metres from a high traffic road, which means that a significant proportion of residents is not represented by urban background locations . Therefore, to better characterize the variability of NO2 in the city we added 11 sites (7 high traffic locations) in the second measurement campaign. Potential consequences of this choice were a lower R2 of the 2007 LUR model and a poorer ability of the 1995/96 model to predict the 2007 measurements compared to the performance of the 2007 model to predict the observations retrospectively. On the other hand, the R2 of the models were comparable to those developed in other settings that went from 0.51 to 0.90 , and were better than the one we developed a few years ago (R2 was 0.69) when data on vehicular traffic were not available .
A slight decrease in air pollution was noticed when comparing the results of the two surveys, and it occurred both in traffic and in urban background sites. Data from fixed monitors of the Regional Environmental Protection Agency show in the period 1999–2008 a small decrease in NO2 concentrations in traffic sites, and a stability in urban background sites .
Few studies have been conducted on the stability of small-area spatial contrasts over time using LUR models and they corroborate our findings. In Oslo, Madsen and colleagues found a good agreement in spatial contrasts over a three year period.  In the Netherlands, Eeftens and colleagues showed good agreement in measured and modelled levels of NO2 eight years apart .
The prediction of NO2 levels to the studied population indicates that traffic-related air pollution in Rome is not evenly distributed throughout the population: we found higher levels of exposure in the oldest age group, and in the higher socioeconomic compared to the lower socioeconomic census blocks. This is not surprising as other local studies have shown the same pattern, which is related to the urbanization history of Rome, with the elderly and well-off population more likely to live in the prestigious and well-travelled city centre than in the periphery [25, 29]. Although the common belief is that low socioeconomic groups of the population are disadvantaged from all points of view, air quality included, this idea has been contradicted also by studies conducted in Canada and the Netherlands [6, 30].
The mortality - NO2 association was identical using the two LUR models in all circumstances: when using quintiles of the distribution, categories of NO2 concentration, and the inter-quartile range. However, the effect was slightly lower with the 2007 model compared to 1995/96 when we calculated the association with natural mortality for a 10 μg/m3 increase in NO2 and this is attributable to the lower estimated variability in 1995/96. Nevertheless, it is worth noting that our estimates of the strength of the association between NO2 and mortality were similar to findings in other settings [1, 3, 5].
This study has some limitations. We used two different methods to measure NO2 concentrations: Palmes tubes and Ogawa badges. Although we took it into account, calibrating the 1995/96 levels, we introduced some level of error: 16 % of variance in measurements taken using Ogawa samplers was unexplained by Palmes tube measurements. Moreover, we assigned a correction factor based only on one week of measurements, while it could change in different weather conditions. In each campaign we used as a single measure of NO2 concentration the mean of the three measurements in the same site, without taking account of variation during the year. The mean concentration of NO2 during the first measurement period in urban background fixed monitors was 47.0 μg/m3, comparable to our mean of 45.5 μg/m3. Data were not available to adjust 1995/96 measurements for temporal trends and for this reason we have not done it neither for 2007. However, during the three weeks of measurements in 2007 at the city background monitoring station the average concentration of NO2 was 44.9 μg/m3 (SD 14.6) and was comparable to the annual average concentration we estimated (43.9 μg/m3, SD 14.6). An important limit of our models is that traffic data were available only for major roads (i.e. >4000 vehicles/day), and they were estimated using an integrated data monitoring and evaluation system developed under the Heaven European project (heaven.rec.org). Moreover, period-specific data on traffic were not available as we used for both models the same 2005 traffic data. The only time-specific variables were those related to population. For the cohort analysis, we did study subjects who did not move since five years before the enrolment to the end of the follow-up. This is a strength of the study and in Rome the majority of adult population is rather stable and the percentage of those aged 45–80 years who change the address during ten years is 15 % and decrease with increasing age. For these reasons we believe that our mortality analysis is valid. The cohort we used is based on administrative data, and information on important risk factors and potential confounders such as obesity, smoking, diet are not available. However, we were interested on comparing the effect of the two estimates from LUR models on the same population, than measuring the effect estimates of air pollution exposure on mortality. In fact we wanted to investigate to what extent we can use exposure models based on measurements taken after the settlement of the epidemiologic study.