Robustness of different regression modelling strategies in epidemiology: a time-series analysis of hospital admissions and air pollutants in Lisbon (1999-2004)

Abstract

a Studies of the acute health effects of air pollution have used exposure windows of different spans and related them to single-day responses. Little is known about whether an increased response window span might be a viable alternative to single-day responses. Our aim is to compare a new model specification where both the exposure and response variables are represented as 7 day moving averages (CMA&CMA model) with the most widely used model specifications in the literature, where the response variable is usually a single-day, in terms of coefficients and their precision and robustness. To this end, daily series of 12 emergency-related hospital admissions and 6 air pollutants spanning 5.5 years in Lisbon were analysed through single-pollutant linear regression and, when necessary M-estimation. With our data, the CMA&CMA model yields coefficients that are very close to models where only the exposure variable is specified as a moving average whether the latter are estimated by OLS or robust M-estimation. In addition, the CMA&CMA model leads to more precise and robust estimates than other model specifications. The new model specification is a straightforward tool for adjusting weekend effects and errors. It is also analogous to robust estimation, with the added advantages of being sensitive to extreme values that are clustered in time, and leading to more precise and robust estimates without loss of high-frequency information. One drawback is the induction of autocorrelation in the residuals.

Association between atmospheric pollutants and hospital admissions in Lisbon

“…1). In HA database, it was observed that hospital admission counts tend to be lower on weekends as compared to working days, as already concluded by Sarmento et al (2009), suggesting that either weekend admissions are registered on Mondays only or patients wait for Monday to go to hospital if the symptoms are not so severe or the number of doctors in hospital is reduced during weekend and people may wait to Monday to be assisted. This is visible in Fig.…”

“…Studies on human health effects usually use daily data; although, recent epidemiological studies have been using exposures in the form of moving averages, distributed lags or data aggregated over several days or even weeks (Schwartz 2000b;Sarmento et al 2009). …”

“…These age groups were adopted for consistency with former studies in Portugal (Sarmento et al 2009;Almeida et al 2014). The studied causes of HA classified according to the World Health Organisation's International Statistical Classification of Diseases by the, 9th Revision (ICD-9) and 10th Revision (ICD-10) endorsed by the Forty-third World Health Assembly in May 1990, came into use in WHO Member States since 1994 (WHO 2014): all circulatory (ICD-9, 390-459; ICD-10, I00-I99), cardiac (ICD-9, 390-429; ICD-10, I30-I52), ischaemic heart (ICD-9, 410-414; ICD-10, I20-I25), stroke (ICD-9, 430-438; ICD-10, Fig.…”

“…The daily data may consist of the actual daily values of health and pollution or some form of average over a number of days. Some studies have modelled single day health events with exposure averages over multiple days (Lipfert 1993;Lumley and Sheppard 2000;Roberts 2005;Sarmento et al 2009;Smith et al 2000) and a minority has modelled the health events also as averages over multiple days, while preserving the days as units of comparison (Schwartz 2000a, b).…”

Exposure and dose assessment to particle components among an elderly population

“…However, epidemiological associations between PM and health outcomes are based predominantly on ambient air measurements where it is assumed that each person in a given region has the same exposure level, which is often obtained from a few air quality monitors and reflects the entire community Freitas et al, 2009a,b;Sarmento et al, 2011). Nevertheless, poor correlations have been found between ambient PM concentrations and personal exposure to PM (Meng et al, 2005) because actual exposure is strongly related to the individual time activity patterns, followed by its distance from each particle source.…”