Comparison of different methods in analyzing short-term air pollution effects in a cohort study of susceptible individuals
- Annette Peters^{1_26}Email author,
- Stephanie von Klot^{1_26},
- Niklas Berglind^{2_26, 3_26},
- Allmut Hörmann^{4_26},
- Hannelore Löwel^{1_26},
- Fredrik Nyberg^{2_26, 5_26},
- Juha Pekkanen^{6_26},
- Carlo A Perucci^{7_26},
- Massimo Stafoggia^{7_26},
- Jordi Sunyer^{8_26},
- Pekka Tiittanen^{6_26} and
- Francesco Forastiere^{7_26}
DOI: 10.1186/1742-5573-3-10
© Peters et al. 2006
Received: 24 May 2005
Accepted: 09 August 2006
Published: 09 August 2006
Abstract
Background
Short-term fluctuations of ambient air pollution have been associated with exacerbation of cardiovascular disease. A multi-city study was designed to assess the probability of recurrent hospitalization in a cohort of incident myocardial infarction survivors in five European cities. The objective of this paper is to discuss the methods for analyzing short-term health effects in a cohort study based on a case-series.
Methods
Three methods were considered for the analyses of the cohort data: Poisson regression approach, case-crossover analyses and extended Cox regression analyses. The major challenge of these analyses is to appropriately consider changes within the cohort over time due to changes in the underlying risk following a myocardial infarction, slow time trends in risk factors within the population, dynamic cohort size and seasonal variation.
Results
Poisson regression analyses, case-crossover analyses and Extended Cox regression analyses gave similar results. Application of smoothing methods showed the capability to adequately model the complex time trends.
Conclusion
From a practical point of view, Poisson regression analyses are less time-consuming, and therefore might be used for confounder selection and most of the analyses. However, replication of the results with Cox models is desirable to assure that the results are independent of the analytical approach used. In addition, extended Cox regression analyses would allow a joint estimation of long-term and short-term health effects of time-varying exposures.
Background
Ambient air pollution has been associated with increases in acute morbidity and mortality [1]. Patients with underlying chronic diseases such as for example diabetes, ischemic heart disease or heart failure may be at particular risk for the effects of ambient air pollution. Series of incident cases may be followed over time to assess the impact of short-term fluctuation on recurrent disease exacerbation. These studies deviate from the classical cohort as well as time-series designs. The major statistical challenge in these studies is the control for changes within the cohort over time due to changes in the underlying risk following disease inception, slow time trends in risk factors within population, dynamic cohort size and seasonal variation. The impact of time is particular important when effects of air pollution are considered, because concentrations of air pollutants vary by season and yearly averages of pollutants change due to weather, air pollution control measures and changes in sources.
In order to evaluate the impact of air pollution on patients with pre-existing diseases, previous hospitalization might be used to define a potentially susceptible subgroup [2–4]. Poisson regression analyses and case-crossover analyses have been used to estimate the impact of variations in daily average air pollution concentrations on the risk of death in previously hospitalized subgroups of the population [5–8].
A multi-city study was designed to follow cohorts of myocardial infarction (MI) survivors in five European cities: Augsburg, Barcelona, Helsinki, Rome and Stockholm (HEAPSS-Study). For entry into the cohort, incident first myocardial infarction cases were considered. The outcomes considered were re-infarctions and other related re-hospitalization or death. Ambient air pollution was characterized based on existing air monitoring networks. In addition, condensation particle counters were set up in each location to measure the ambient particle number concentrations (PNC) and to retrospectively estimate PNC for the entire study period in each location.
This paper describes and compares different statistical approaches for analyzing short-term air pollution health effects in a cohort with ongoing recruitment during the follow-up. Three different methods to analyze the cohort data were considered: (a) Poisson regression analyses on the calendar-time axis, (b) case-crossover analyses and (c) extended Cox regression analyses. As an example the analyses of the association between NO_{2} and any cardiac readmission of MI survivors are presented.
Methods
Data collection
Incident myocardial infarction survivors were recruited in five European cities (Augsburg, Barcelona, Helsinki, Rome, and Stockholm) during 1992 to 2000 as has been described elsewhere [9]. Briefly, data sources were AMI Registries in Augsburg and Barcelona, administrative databases of hospital admissions in Helsinki, Rome and Stockholm. Enrollment was restricted to residents of the above cities, aged 35 years or more (35–74 in Augsburg, 35–79 in Barcelona) who had their first AMI (index AMI) during the recruitment period. Subsequent first cardiac re-hospitalizations of cohort members within the study area were recorded from day 29 after the index AMI until the individual end of follow-up period defined by death, migration out of the study area, or center specific end of the study. Readmissions of interest were those with primary diagnoses of re-infarctions (ICD-9: 410; ICD-10: i21, i22), acute angina pectoris (ICD-9: 411, 413; ICD-10: i20–i22, i24, i25), dysrhythmia (ICD-9: 427; i46.0, I46.9, i47–i49, r00.1, r00.8) and heart failure (ICD-9:428; ICD-10: i50). Vital status and place of residence at the end of the follow-up period were ascertained for all cohort members.
Statistical analyses
In the study, MI survivors entered the cohort during an extended period. Once the cohort was complete at least one year of follow-up without recruitment was added. As a consequence the following issues had to be considered in the statistical analyses: (a) the number of subjects at risk increases over the recruitment period, (b) the risk of the cohort for a recurrent event on the calendar-time axis varies based on the proportion of recent MI survivors in the entire cohort. In the following we refer to calendar time when we number the dates during follow-up and to cohort-time when we number the person-time followed.
Poisson regression analyses
Epidemiological studies in air pollution research have developed techniques using Poisson regression analyses on the calendar-time axis [10]. In a time series analysis the daily counts of events are regressed on the daily predictor variables such as trend, season, weather, and air pollution. This design assumes that the events are independent and that the event rate is changes only slowly over time. The event rate λ_{t} at the time point t is modeled as follows
ln(λ _{t}) = α + ∑β _{ i } x _{ it }
With λ_{t} = y_{t}/N_{t} being the number of cases y observed at time t divided by the number of subjects N at risk at time t. The model can be rewritten as
ln(y_{t}) = ln(N_{t}) + α + ∑β _{ i } x _{ it }
In time-series analyses it is sometimes assumed that the underlying population at risk does not change, and therefore the population at risk is not modeled explicitly in the regression analyses. However, when applied to a cohort the number of subjects at risk can be explicitly modeled as given in the equation above. In the cohort setting, when analyzing the data on the calendar time axis, the trend captures three different underlying reasons for changes in the rate of MI over time: (a) long-term underlying trends in the study base due to life-style changes, changes in health care or aging of the population, (b) seasonal variation and (c) changing composition of the cohort. Therefore, in the cohort setting proper trend control is very important because it is likely that the effect estimates will be biased if the trend is not correctly specified. Recent discussions on modeling in time-series analyses [11, 12] have lead to a broader use of different smoothing techniques in these analyses and different approaches have been used to assess the trend over time [13]. We selected three approaches to model the trend: (a) natural splines, (b) penalized splines and (c) locally weighted least square (loess) smoothers.
In a hierarchical approach potential confounders were selected, including long term trend, season, days of the week, holidays and meteorology, before adding air pollution concentration as independent variable. Generalized additive models were used to allow for non-parametric functions of the confounders in R (The R Foundation for Statistical Computing Version 1.8.1) using the package "mgcv" (version 0.9–6) [14–16]. All models included the natural logarithm of the number of persons at risk each day as offset and the daily number of events as outcome variable. To allow for possible over- or under-dispersion the quasi-likelihood family was used to estimate the parameters without specifying the underlying distribution function. Penalized regression splines were tested for the continuous confounder variables. The choice of degrees of freedom was left to the algorithm ("magic") in the "mgcv" – package that minimized the Generalized Cross Validation (GCV) criterion. The default of 10 knots as starting value was adjusted to higher values if necessary due to the structure of the data. If the smooth function was not significant or the estimated degrees of freedom were less than two, a linear term was tested instead. Decisions for keeping a covariate in a model were based on judgment using the p value (<.1), GCV score (the smaller the better), and the autocorrelation function (ACF) (the nearer to zero the better).
Trend was included in the model as an obligatory variable with a penalized spline starting with 6 knots per year to control for long term trends, seasonality and changes in the baseline risk. Then current day temperature and the deviation of the current day temperature from the mean temperature of lag day 1–3 were tested as penalized splines. At least one temperature term had to remain in the model. Thereafter penalized splines of air pressure and relative humidity, and then dummies for days of the week and the city specific indicators (holidays, population decrease) were tested one after the other. Finally the model was "fine tuned" changing the parameters where the decision had not been clear and comparing ACF plots and GCV-score to choose the final covariates to include for the further analyses.
Sensitivity analyses were performed to compare the results of the final models with the estimates obtained when modifying the smoothing-functions used or the confounders entered in the Poisson model. Thus the analyses were repeated using natural cubic splines or loess instead of the penalized splines, with comparable effective degrees of freedom to those in the final model with penalized splines. Temperature was replaced by apparent temperature [17] and dew point temperature. Instead of temperature difference the mean of lag day 1 to 3 was used. The sensitivity of the results to the choice made when selecting the confounders was tested by excluding or including borderline significant confounders that had or had not entered the model. At the same time, those that had been included with a linear term were considered with a smooth function in the sensitivity analyses.
Case-crossover analyses
Case-crossover analyses have been developed to study transient effects of acute exposures using a case-only design. This design samples information on exposure status from case and referent periods selected from the person-time of the cases. The period of exposure of the cases is selected as a plausible hazard period immediately preceding the event. Referent periods are chosen to represent the exposure distribution in the non-case time periods at risk.
The referent period selection poses the main challenge to the case-crossover analyses. Different sampling designs for referent periods have been proposed to estimate the effects of air pollution. Exposures that do not change in association with the case-status, such as air pollution can be sampled also from person-time after the case occurred [18]. The stratified approach by Lumley and Levy [19] controls time trends by design.
Case-crossover analyses were performed as an alternative to the Poisson regression using conditional logistic regression models in the S-Plus statistical package version 6.0. We used the "coxph" function with a strata statement. The date of each case contributed a hazard period that was matched with referent periods selected with the stratified approach (stratifying criteria were year, month and weekday). It controls for weekday by design. The same confounder variables were included as in the Poisson regression in order to obtain maximum comparability of the models. P-Splines were used as smoothing method for meteorology variables. The effective degrees of freedom of the smoothers in the Poisson models were translated by adjusting the smoothing parameter of the P-splines.
Proportional hazard models
An alternative possibility is to model the data as cohort data using Cox proportional hazard models. Time t now denotes the time since MI. Extended Cox regression allows for time-varying covariates in survival analyses [20]. The hazard h at time t is given as
where X(t) = (X _{1}, X _{2},....., X _{ p1}, X _{1}(t), X _{2}(t),....., X _{ p2}(t)) and X _{1}, X _{2},....., X _{ p1} are time-invariant variables such as for example age or gender, and X _{1}(t), X _{2}(t),....., X _{ p2}(t) are time-varying variables such as weather or air pollution.
The model makes no assumption about the baseline hazard, and therefore is suitable for the analyses proposed here because the risk for cardiac readmission changes during the follow-up of the MI survivors. While X _{ j }(t) is varying over time, the hazard model provides only one regression coefficient δ _{ j } for each time-varying variable in the model. Thus at time t, there is only one value of the variables X _{ j }(t) that has an effect on the hazard: the value being measured at time t. The model therefore assumes a uniform relative risk for all time points and consequently does not per se address the possibility of effect modification of the air pollution effects by length of follow-up.
The association with daily pollution levels was analyzed in SAS statistical software (SAS Institute Inc., Cary, NC, USA, Release 8.02) PROC PHREG using the counting process style of input. For each subject one record for each day at risk was created. All models included the same covariates as the Poisson models, but instead of smooth functions quadratic functions were used. Trend entered as a linear term. As constant covariates age at entry (in years as quadratic function), diabetes, hypertension and sex (as dummy variables) were considered, since they were identified as predictors of survival in a classical Cox regression.
Results
Description of the HEAPSS Study population, cardiac readmission as a selected outcome, and NO_{2} concentrations in 5 European cities.
City | Period of Enrollment | Period of Follow up | Age Range (yrs) | Population Size | Persons followed | Mean Age (yrs) | Any cardiac readmission during follow up | Mean NO_{2} [μg/m^{3}] | Maximum NO_{2} [μg/m^{3}] | |
---|---|---|---|---|---|---|---|---|---|---|
N | λ | |||||||||
Augsburg | 1995–1999 | 1995–2000 | 35–74 | 300,902 | 1560 | 60.4 | 286 | 0.14 | 49.6 | 171.8 |
Barcelona | 1992–1995 | 1992–2000 | 35–79 | 893,601 | 1134 | 61.9 | 296 | 0.09 | 47.7 | 148.0 |
Helsinki | 1993–1999 | 1993–2000 | 35+ | 297,410 | 4026 | 69.2 | 1301 | 0.45 | 30.1 | 173.4 |
Rome | 1998–2000 | 1998–2001 | 35+ | 1,616,356 | 7384 | 67.0 | 1916 | 1.34 | 70.0 | 140.5 |
Stockholm | 1994–1999 | 1994–2000 | 35+ | 548,113 | 7902 | 73.3 | 2856 | 1.23 | 22.8 | 81.2 |
The Effect of NO_{2}
Comparison of the regression coefficients (beta) and standard errors (se) in the analyses of NO_{2} concentrations (average of current and previous day) and any cardiac readmission applying different statistical models and assessing the sensitivity of the results for confounder selection as listed in table 3.
Augsburg | Barcelona | Helsinki | Rome | Stockholm | pooled | Heterogeneity test | |||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Poisson Regression Analyses | beta | se | beta | se | beta | se | beta | se | beta | se | beta | se | p-value |
Final Model | 0.0119 | 0.0044 | 0.0068 | 0.0044 | 0.0015 | 0.0026 | 0.0041 | 0.0022 | 0.0013 | 0.0032 | 0.0039 | 0.0013 | 0.26 |
natural splines^{a} | 0.0119 | 0.0044 | 0.0066 | 0.0044 | 0.0012 | 0.0026 | 0.0037 | 0.0022 | 0.0012 | 0.0032 | 0.0036 | 0.0013 | 0.25 |
loess^{b} | 0.0119 | 0.0044 | 0.0065 | 0.0042 | 0.0015 | 0.0026 | 0.0040 | 0.0021 | 0.0014 | 0.0031 | 0.0038 | 0.0013 | 0.27 |
Removing Confounders | 0.0117 | 0.0041 | 0.0078 | 0.0042 | 0.0019 | 0.0025 | 0.0049 | 0.0022 | 0.0011 | 0.0030 | 0.0044 | 0.0013 | 0.19 |
Adding Confounders | 0.0122 | 0.0045 | 0.0051 | 0.0046 | 0.0024 | 0.0026 | 0.0043 | 0.0022 | 0.0018 | 0.0032 | 0.0041 | 0.0013 | 0.37 |
Apparent Temperature^{c} | 0.0115 | 0.0044 | 0.0067 | 0.0044 | 0.0012 | 0.0026 | 0.0039 | 0.0022 | 0.0013 | 0.0031 | 0.0037 | 0.0013 | 0.28 |
Case-crossover Analyses | |||||||||||||
Stratified control selection | 0.0144 | 0.0060 | -0.0024 | 0.0065 | 0.0031 | 0.0031 | 0.0024 | 0.0026 | 0.0008 | 0.0036 | 0.0029 | 0.0016 | 0.32 |
Cox-Extended Regression | |||||||||||||
Time-varying confounders | 0.0123 | 0.0043 | 0.0046 | 0.0041 | 0.0013 | 0.0025 | 0.0034 | 0.0021 | 0.0017 | 0.0031 | 0.0035 | 0.0013 | 0.24 |
+ individual confounders | 0.0123 | 0.0043 | 0.0047 | 0.0041 | 0.0011 | 0.0025 | 0.0034 | 0.0021 | 0.0016 | 0.0031 | 0.0034 | 0.0013 | 0.23 |
Confounders included in the different models with the following abbreviations: S: penalized spline (* k = 30, otherwise k = 10), Poly: Polynomial with order in brackets, L: linear term, X: dummies, D: by design.
Trend | Temperature | Temperature difference | Relative humidity | Air pressure | Weekday | Holiday | Population decrease | |
---|---|---|---|---|---|---|---|---|
Augsburg | ||||||||
Poisson Regression : Final | S | - | L | - | - | X | - | - |
Poisson Regression : Remove | S | - | - | - | - | - | - | - |
Poisson Regression : Add | S | S | L | L | - | X | - | - |
Case-crossover | D | - | L | - | - | D | - | - |
Extended Cox Regression | L | - | L | - | - | X | - | - |
Barcelona | ||||||||
Poisson Regression : Final | S | L | - | - | - | X | - | X |
Poisson Regression : Remove | S | - | - | - | - | - | - | X |
Poisson Regression : Add | S | L | L | - | - | X | - | X |
Case-crossover | D | L | - | - | - | D | - | X |
Extended Cox Regression | L | L | - | - | - | X | - | X |
Helsinki | ||||||||
Poisson Regression : Final | S | L | L | L | - | X | - | - |
Poisson Regression : Remove | S | - | L | L | - | X | - | - |
Poisson Regression : Add | S | S | S | L | - | X | - | - |
Case-crossover | D | L | L | L | - | D | - | - |
Extended Cox Regression | L | L | L | L | - | X | - | - |
Rome | ||||||||
Poisson Regression : Final | S* | S | - | L | - | X | X | X |
Poisson Regression : Remove | S* | L | - | L | - | X | - | - |
Poisson Regression : Add | S* | S | - | S | - | X | X | X |
Case-crossover | D | S | - | L | - | D | X | X |
Extended Cox Regression | L | Poly(2) | - | L | - | X | X | X |
Stockholm | ||||||||
Poisson Regression : Final | S | L | - | S | L | X | - | X |
Poisson Regression : Remove | S | - | - | S | L | X | - | X |
Poisson Regression : Add | S | L | L | S | L | X | - | X |
Case-crossover | D | L | - | S | L | D | - | X |
Extended Cox Regression | L | L | - | Poly(2) | L | X | - | X |
Discussion
The major challenge of analyzing short-term health effects of air pollution in a cohort of diseased subjects is to consider simultaneously other time-varying confounders and the changes in the probability of a recurrent event due to the individual characteristics. In the case of myocardial infarction survivors, a major determinant of individual vulnerability is the time since index event due to the underlying healing of the heart tissue.
Three different approaches were considered in the analyses. The first one, Poisson regression analysis summarizes the events on a calendar time-axis. The data in this study demonstrate an example where the underlying probability of observing an event changes with the composition of the cohort. Therefore, time is not only a measure for slow trends and seasonal variation, but also represents the changing fractions of persons at high risk over time. Smoothing methods were used to attempt to model these three different components in the time trend. The recent discussion on smoothing techniques [11, 12] was considered by the selection of three different approaches, namely natural splines, penalized splines and locally weight least square smoothers. All three methods gave quite comparable results and their function is consistent with a decreased risk of hospital readmission as the cohort ages and a seasonal variation in hospital readmissions. In the sensitivity analyses, results were robust to changes in the confounders selected in the final model.
The case-crossover method was chosen because it was designed to control for temporal changes by design. More specifically, the stratified referent period selection approach considered here controls for trends by design [19]. However, it has been noted that case-crossover analyses have reduced power compared to Poisson regression analyses [19]. In this study we observed slightly smaller effect estimates with larger standard errors. The case of a cohort with ongoing recruitment and dynamically changing composition has not been methodologically considered before. One may assume that the underlying changes in rates are constant within each stratum [21]. However, as observed in figure 1, panel C this assumption might be violated at the beginning of the study. Therefore, the changing number of subjects at risk might be responsible for the small differences observed between Poisson regression analyses, which explicitly consider the varying number of subjects at risk on a given day, and the case-crossover analyses. Nevertheless, case-crossover analyses were considered for these analyses because they quite elegantly allow analyses of subgroups.
Extended Cox regression analyses, on the other hand, were formulated to consider the present study design in the correct way. Here, the underlying risk is modeled by a hazard function h _{0}(t) which is variable over time and considers both the changing hazard due to the recovery from the incident MI as well as changes in the composition of the cohort with respect to season and calendar year. The results are remarkably consistent with those obtained in the Poisson regression analyses. The practical disadvantage of the method is that the analyses are time-consuming, in particular if these models are used for the selection of time-varying confounders. Consideration of individual confounders did not change the association between NO_{2} and hospital readmissions. In contrast to a recently published similar approach, we chose time of follow-up instead of subject's age as time axis [22]. While these two approaches should give similar effect estimates we favor the model on the follow-up time axis as it more appropriately considers the changing risk levels within the cohort.
A recent review by the American Heart Association has highlighted the emerging evidence for the biological plausibility between ambient air pollution concentrations in urban areas and cardiovascular disease exacerbation [23]. However, effect estimates obtained from the general population might underestimate the risk of susceptible subpopulations, which have also higher baseline risks [24]. Cohort studies assessing the risk of susceptible populations are highly recommended to provide better estimates for risk assessment. For example, the age at first MI, socio-economic status as well as the time since the first MI might modify the risk of short-term air pollution exposures for an individual. All three methods might be used to assess the susceptibility of subgroups. The extended Cox regression analysis is the only method that would allow the estimation of the main effect of the considered effect modifier, but computation of the models is time-consuming.
Recent research indicated that spatial and temporal variability of long-term exposures to ambient particles may be important factors to consider [25, 26]. Furthermore, future research might consider short-term fluctuations as well as individualized estimates of long-term exposures to ambient particles in assessing the health impact of environmental exposures. For these studies, Extended Cox regression analyses would be the method of choice. In its most simplistic version, one may estimate jointly the effect of homes exposed to high traffic together with air pollution concentrations from a central monitoring side. However, also more sophisticated approaches of exposure assessment building on spatio-temporal models such as for example described by Gulliver and Briggs [27] or Sahu and colleagues [28] can be foreseen.
Conclusion
Of the three methods considered for the analyses of the HEAPSS Study, the Poisson regression approach and the extended Cox regression analyses gave similar results. Case-crossover analyses might underestimate the strength of the association in this specific setting, but the differences were small. Further methodological investigation may be warranted. From a practical point of view, Poisson regression analyses are less time-consuming, and therefore might be used for confounder selection and most of the analyses. However, replication of the results with Cox models is desirable to assure that the results are independent of the analytical approach used. For the identification of susceptible subgroups within a cohort of susceptible populations, case-crossover analyses might be the least time consuming approach, however, Extended Cox regression analyses would allow a joint estimation of the main effects and the effect modification.
Declarations
Authors’ Affiliations
References
- Brunekreef B, Holgate ST: Air pollution and health. [Review] [147 refs]. Lancet 2002, 360:1233–1242.View ArticlePubMedGoogle Scholar
- Zanobetti A, Schwartz J, Dockery DW: Airborne particles are a risk factor for hospital admissions for heart and lung disease. Environ Health Perspect 2000, 108:1071–1077.PubMedGoogle Scholar
- Zanobetti A, Schwartz J, Gold DR: Are there sensitive subgroups for the effects of airborne particles? Environ Health Perspect 2000, 108:841–845.PubMedGoogle Scholar
- Annesi-Maesano I, Agabiti N, Pistelli R, Couilliot MF, Forastiere F: Subpopulations at increased risk of adverse health outcomes from air pollution. [Review] [53 refs]. Eur Respir J Suppl 2003, 40:57s-63s.View ArticlePubMedGoogle Scholar
- Sunyer J, Schwartz J, Tobias A, Macfarlane D, Garcia J, Anto JM: Patients with chronic obstructive pulmonary disease are at increased risk of death associated with urban particle air pollution: a case-crossover analysis. Am 2000, 151:50–56.Google Scholar
- Goldberg MS, Burnett RT, Bailar JCIII, Tamblyn R, Ernst P, Flegel K, Brook J, Bonvalot Y, Singh R, Valois MF, Vincent R: Identification of persons with cardiorespiratory conditions who are at risk of dying from the acute effects of ambient air particles. Environ 2001, 109 Suppl 4:487–494.Google Scholar
- Kwon HJ, Cho SH, Nyberg F, Pershagen G: Effects of ambient air pollution on daily mortality in a cohort of patients with congestive heart failure. Epidemiology 2001, 12:413–419.View ArticlePubMedGoogle Scholar
- Bateson TF, Schwartz J: Who is Sensitive to the Effects of Particulate Air Pollution on Mortality?: A Case-Crossover Analysis of Effect Modifiers. Epidemiology 2004, 15:143–149.View ArticlePubMedGoogle Scholar
- Klot S, Peters A, Aalto P, Bellander T, Berglind N, D'Ippoliti D, Elosua R, Hörmann A, Kulmala M, Lanki T, Löwel H, Pekkanen J, Picciotto S, Sunyer J, Forastiere F: Ambient air pollution is associated with increased risk of hospital readmissions of myocardial infarction survivors in European cities. Circulation 2005, 112:3073–3079.View ArticleGoogle Scholar
- Kwon HJ, Cho SH, Nyberg F, Pershagen G: Effects of ambient air pollution on daily mortality in a cohort of patients with congestive heart failure. Epidemiology 2001,12(4):413–419.View ArticlePubMedGoogle Scholar
- Dominici F, Zeger S, Samet J: A measurement error model for time-series studies of air pollution and mortality. Biostatistics 2000, 157–175.Google Scholar
- Ramsay TO, Burnett RT, Krewski D: The effect of concurvity in generalized additive models linking mortality to ambient particulate matter. Epidemiology 2003, 14:18–23.View ArticlePubMedGoogle Scholar
- Schwartz J, Zanobetti A, Bateson TF: Mortality and morbidity among eldery residents of cities with daily PM measurements. Revised Analyses of Time-Series Studies of Air Pollution and Health Boston, Health Effects Institute 2003, 25–58.Google Scholar
- Wood SN, Augustin NH: GAMs with integrated model selection using penalized regression splines and applications to environmental modelling. Ecological Modelling 2002, 157:157–177.View ArticleGoogle Scholar
- Wood SN: Modelling and smoothing parameter estimation with multiple quadratic penalties. J R Statist Soc B 2000, 62:413–428.View ArticleGoogle Scholar
- Wood SN: Thin plate regression splines. J R Statist Soc B 2003, 65:95–114.View ArticleGoogle Scholar
- O'Neill MS, Zanobetti A, Schwartz J: Modifiers of the temperature and mortality association in seven US cities. Am J Epidemiol 2003, 157:1074–1082.View ArticlePubMedGoogle Scholar
- Navidi W: Bidirectional case-crossover designs for exposures with time trends. Biometrics 1998, 54:596–605.View ArticlePubMedGoogle Scholar
- Lumley T, Levy D: Bias in the case-crossover design: implications for studies of air pollution. Environmetrics 2000, 11:689–704.View ArticleGoogle Scholar
- Kleinbaum DG: Survival Analyses Springer 2004.Google Scholar
- Janes H, Sheppard L, Lumley T: Case-crossover analyses of air pollution exposure data: referent selection strategies and their implications for bias. Epidemiology 2005, 16:717–726.View ArticlePubMedGoogle Scholar
- Lepeule J, Rondeau V, Filleul L, Dartigues JF: Survival analysis to estimate association between short-term mortality and air pollution. Environ Health Perspect 2006, 114:242–247.View ArticlePubMedGoogle Scholar
- Brook RD, Franklin B, Cascio WE, Hong Y, Howard G, Lipsett M, Luepker RV, Mittleman MA, Samet JM, Smith SCJ, Tager IB: Air Pollution and Cardiovascular Disease: A statement of the health care professionals from the expert panel on population and prevention science of the American Heart Association. Circulation 2004, 109:2655–2671.View ArticlePubMedGoogle Scholar
- Peters A: Susceptible subgroups: the challenge of studying interactions. Epidemiology 2004, 15:131–132.View ArticlePubMedGoogle Scholar
- Laden F, Schwartz J, Speizer FE, Dockery DW: Reduction in Fine Particulate Air Pollution and Mortality: Extended Follow-up of the Harvard Six Cities Study. Am J Respir Crit Care Med 2006, 173:667–672.View ArticlePubMedGoogle Scholar
- Jerrett M, Burnett RT, Ma R, Pope CAIII, Krewski D, Newbold KB, Thurston G, Shi Y, Finkelstein N, Calle EE, Thun MJ: Spatial analysis of air pollution and mortality in Los Angeles. Epidemiology 2005, 16:727–736.View ArticlePubMedGoogle Scholar
- Gulliver J, Briggs DJ: Time-space modeling of journey-time exposure to traffic-related air pollution using GIS. Environ Res 2005, 97:10–25.View ArticlePubMedGoogle Scholar
- Sahu SK, Gelfand AE, Holland DM: Spatio-temporal modeling of fine particulate matter. J Agricult Biol Environ Stat 2006, 11:61–86.View ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.