The Effect of Concurvity in Generalized Additive Models Linking Mortality to Ambient Particulate Matter
Ramsay, Timothy O.1; Burnett, Richard T.2; Krewski, Daniel3
From 1Health Canada,
2McLaughlin Centre for Population Health Risk Assessment, and
3University of Ottawa, Ottawa, Canada.
Address correspondence to: Tim Ramsay, McLaughlin Centre for Population Health Risk Assessment, Room 321, One Stewart St., Ottawa, K1N 6N5, Canada; firstname.lastname@example.org
This research was supported in part by the Toxic Substances Research Initiative of Health Canada, the National Science and Engineering Research Council of Canada, and the Health Effects Institute in Cambridge, Massachusetts.
D. Krewski is the NSERC/SSHRC/McLaughlin Chair in Population Health Risk Assessment at the University of Ottawa.
Submitted 26 March 2002; final version accepted 27 September 2002.
Editors’ note: Invited commentaries on this article appear on pages 11 and 13.
In recent years, a number of studies have applied generalized additive models to time series data to estimate associations between exposure to air pollution and cardiorespiratory morbidity and mortality. If concurvity, the nonparametric analogue of multicollinearity, is present in the data, statistical software such as S-plus can seriously underestimate the variance of fitted model parameters, leading to significance tests with inflated type 1 error. This paper uses computer simulation and analyses of actual epidemiologic data to explore this underestimation of standard errors. We provide a method for assessing concurvity in data and an alternate class of models that is unaffected by concurvity. We argue that some degree of concurvity is likely to be present in all epidemiologic time series datasets and we explore through the use of meta-analysis the possible impact of concurvity on the existing body of work relating ambient levels of sulfate particles to mortality.
Click on the links below to access all the ArticlePlus for this article.
Please note that ArticlePlus files may launch a viewer application outside of your web browser.
Epidemiologic studies have consistently demonstrated associations between exposure to urban air pollution and cardiorespiratory morbidity and mortality. Associations between hospital admissions and short-term increases in the levels of particulate and gaseous pollutants have been explored using time series (Burnett et al.) 1 and case-crossover studies (Lin et al.). 2 Associations between mortality and long-term exposure to fine particulate air pollution have been demonstrated in cohort mortality studies conducted by Dockery et al. 3 and Pope et al. 4
Time series analysts have long recognized the importance of including temporal trends and serial autocorrelation in time series models (Burnett and Krewski). 5 Time series analysis that compensates for both secular trends and autocorrelation in the residuals provides a means of establishing associations between short-term fluctuations in ambient air pollution levels and population health outcomes (Cakmak et al.). 6 Although serial autocorrelation can be explicitly included in time series models, nonparametric local smoothing methods based on generalized additive models (GAMs) have been used as an alternative analytic approach (Schwartz 7; Schwartz et al. 8). Such methods eliminate the need to specify a parametric form for secular trends and allow a greater degree of robustness against model misspecification.
The purpose of this paper is to examine the use of generalized additive models for the analysis of temporal data used in epidemiologic studies of the effects of ambient air pollution on population health. It has recently been noted elsewhere (Dominici et al.) 9 that using the default convergence criteria in S-Plus can also result in biased fitted linear parameters. We show that the variance estimation method currently used in commercial software can underestimate the variance of fitted linear parameters.
In the next section, we describe smoothers and generalized additive models and review the notion of concurvity, a nonparametric analogue of multicollinearity. The consequence of concurvity in the analysis of temporal air pollution data is then examined, using Monte Carlo analysis–based actual epidemiologic data. We show that the inability of the generalized additive model to detect concurvity can lead to misleading statistical inferences, notably an overstatement of the significance of associations between air pollution and health status. All of the simulations and calculations for this paper were executed with S-Plus Version 6. The implications of this observation for population health risk assessment are discussed in the concluding section.
Generalized Additive Models
The Generalized Additive Model
The GAM is a powerful and widely used tool that allows researchers to fit models without specifying parametric relations between the dependent and independent variables (Hastie and Tibshirani). 10 The general form of a GAM is MATHwhere g is the link function and the fi are smooth functions defining the relation between the transformed mean response μ and the independent variables Xi. The functions fi can be chosen to be either parametric or nonparametric and thus allow for a great deal of flexibility in constructing models with nonlinear effects. The gam function in the statistical software package S-Plus makes it easy to fit GAM models with a variety of families of predictor functions fi and different link functions g.
The generalized linear model, or GLM (McCullagh and Nelder), 11 is a special case of the GAM in which none of the functions fi is nonparametric. Linear regression is a special case of the GLM in which the link function is taken to be the identity function. The GAM is therefore a generalization of linear regression and most of the familiar diagnostic tests for fitted linear regression models have analogues applicable to fitted GAMs. One important exception to this rule is concurvity, the nonparametric analogue of multicollinearity. At present, S-Plus provides no diagnostic tools for assessing the impact of concurvity on a fitted GAM.
Scatterplot smoothing functions, commonly referred to as smoothers, are central to GAMs. A smoother is a mathematic technique for approximating an observed variable Y by a smooth function of one (or several) independent variable(s) X. Some examples of commonly used smoothers are smoothing splines, regression splines, LOESS functions (also known as locally estimated polynomial regression) and kernel smoothers. These methods are called nonparametric because they make no parametric assumption about the shape of the function being estimated. Instead, they assume only that the function Y (X) is smooth, or that the value of the variable Y at similar values of X will be similar.
The most common smoothers, including all those mentioned here, are linear smoothers. A linear smoother (see Hastie and Tibshirani) 10p.44 is a smoother in which the smoothed values Ŷ can be expressed in the form Ŷ =AX for some matrix A. In other words, the value of the estimated smooth function at a given point X is always a linear combination of all the observed values of Y. Because different smoothers use different methods for estimating the value of Y at each observed value of X, one can expect different methods to give somewhat different estimates of the smooth function.
Each smoother has a parameter that determines how smooth the resulting function will be. For LOESS functions, that parameter is called span. For smoothing splines and natural splines, the degree of smoothing can be specified through the degrees of freedom parameter. Figure 1 shows some simulated data smoothed by LOESS functions with three different spans. The solid line is the most smooth; the dashed line is the least smooth. In general, the amount of smoothing selected will have more impact on the final function than the type of smoother chosen.
Concurvity, like multicollinearity, can be defined in terms of the eigenvalues of a certain projection matrix associated with the model (Hastie and Tibshirani). 10p.123 This definition does not easily lead to an intuitive understanding of what concurvity means in a practical modeling situation, however, and concurvity is easier to understand by analogy to multicollinearity.
Consider the situation in which one is using linear regression to model the expected value of a variable Y as a function of p variables X1,…, Xp. We say that multicollinearity is present in the data if some subset of the regressors Xi is highly correlated (Montgomery and Peck). 12p.165 More formally, multicollinearity occurs when one of those variables, say Xp, can be approximated by a linear combination of the remaining variables:MATH
If the equality in Eq 2 is exact, then for any constant λ, MATHand the model is nonidentifiable. If, as is more likely to be the case, the equality is only approximate, then the model is identifiable but the parameter estimates associated with the multicollinear variables become both highly unstable and highly correlated.
Concurvity is a nonparametric extension of the concept of multicollinearity. Suppose that one is modeling the expected value of Y as in Eq 1, with each function fi specified to be in some set of functions Hi. Some common choices for the set Hi are linear functions, higher-order polynomials, LOESS (local polynomial regression) functions or regression splines. Concurvity occurs when a function of one of the variables, say Xp, can be approximated by a linear combinations of functions of the other variables with the function hi lying in the set Hi, with MATH
Again, if the equality in Eq 3 is exact, the model is nonidentifiable because the equality MATHholds for any constant λ. We formalize this description of concurvity in the definition below. (Note that this definition is somewhat different from that given by Hastie and Tibshirani, 10 and is equivalent to what they call approximate concurvity.)
Suppose one wishes to fit the GAM model in Eq 1, with fi ∈Hi for some specified function spaces,MATH from observations Y and MATH. We say that there is concurvity in the data if there exists a set of functions MATH such that gi(Xi) is highly correlated with MATH.
As is the case with linear regression (Montgomery and Peck), 12p.190 the parameter estimates of a fitted GAM are highly unstable if there is concurvity in the data. Unfortunately, because of the way variances are estimated in the S-Plus gam function, the variance estimates do not reflect this instability. In other words, the variance estimates produced by the gam function will be biased downwards whenever there is concurvity in the independent variables. This will result in confidence intervals on linear parameters being too narrow and in misleading P-values arising from hypothesis tests based on variance estimates.
Assuming that the smoothers used in a GAM are linear, as are both the spline and LOESS smoothers, there is an exact expression for the asymptotic standard error of the predicted parameters. Unfortunately, the matrices necessary for estimating this expression are not produced by the back-fitting algorithm used to fit GAMs. Furthermore, the algorithm proposed by Hastie and Tibshirani 10p.127 for computing these matrices is computationally expensive for large datasets. Instead, the gam function approximates the standard errors with an admittedly ad hoc method. The technical details of this approximation, also proposed by Hastie and Tibshirani, 10 are discussed in Chambers and Hastie 13 (sections 7.4.3 and 7.4.4). As the simulations of the next section will show, this approximation works well only when there is little or no concurvity present in the data.
An alternate strategy for estimating the variance is to use a Monte Carlo method such as the parametric bootstrap. Our simulations suggest that the variance estimates produced by the parametric bootstrap, although less biased than those produced by S-Plus, are still biased downwards. A better solution is to abandon the nonparametric model altogether by using a parametric smoother such as natural splines in place of the nonparametric smoother. This approach, which effectively turns the GAM into a GLM and thus produces correct variance estimates, is discussed further in Dominici et al. 9
Consequence of Concurvity
Example 1: Real Data
In this section, we present the results of a simulation to illustrate the variance inflation that can arise from concurvity. These results will show that because of the inability of the gam function to detect concurvity, significance tests based on gam variance estimates can lead to understated P-values and therefore to inflated type I errors. The simulation is based on an analysis done by Burnett et al. 14 relating nonaccidental deaths in Toronto to PM2.5. Before describing the simulation, we will describe the original analysis.
The dependent variable in this study was the daily count of nonaccidental deaths (dt) in Toronto for the years of 1980 through 1994. The independent variables were time (t), daily minimum temperature (mint), daily dewpoint temperature (dewt), and a daily measure of suspended particulate matter with diameter less than or equal to 2.5 microns (PM2.5t). These data were fitted with a Poisson GAM model of the form MATHwhere t = 1,…,5479 represents the day and f, g, and h were LOESS functions. The error term, εi, is assumed to have mean 0 and variance φμt. The overdispersion parameter φ was estimated from the data.
This model was obtained by Burnett et al. 14 through stepwise variable selection from a larger set of potential predictor variables. The linear effect of suspended particulate matter on mortality, β̂, was estimated by the S-Plus gam function to be 1.13 × 10−3 with a standard error of 2.66 × 10−4. Based on these estimates, it was concluded that the effect of PM2.5 was statistically substantial.
Upon reanalysis, however, we noticed some concurvity present in the data. The concurvity can be seen from the fact that the correlation between PM2.5 and the predicted values obtained from fitting the model E (PM2.5t) =f (t) +g (mint) +h (dewt) is 0.59. To test the impact of this concurvity, we executed the following simulation 1000 times. First, we simulated overdispersed Poisson data of the form MATHwhere all the parameters were obtained by fitting to the real data. We then fit to this simulated data and recorded both the slope estimate and the estimated standard error of the slope.
Using the convergence criteria suggested by Dominici et al., 9 the average estimated standard error of the slope parameter given by S-Plus was 2.68 × 10−4. By contrast, the standard deviation of the thousand estimated slopes was 3.13 × 10−4. In other words, the true standard error is about 16% higher than the S-Plus output would lead us to believe. This bias is a result of the concurvity in the data. A standard normal test of the true null hypotheses that β = 1.13 × 10−3, tested at the 0.05 level of significance, was rejected 176 times out of 1,000. This represents a type I error rate of about 18%, rather than the 5% that one should observe with a test at the 0.05 level of significance. Clearly, it is unwise to conduct significance tests based on the S-Plus error estimate for this type of model.
Using the default convergence criteria built into S-Plus rather than the stricter criteria introduced by Dominici et al. 9 resulted in an even larger bias. When the simulation just described was run with the default criteria, the empiric estimate of standard error was 3.30 × 10−4, or 23% higher than the S-Plus estimate. The S-Plus estimate of standard error did not change. The larger empiric error was presumably a result of the failure of the back-fitting algorithm to converge to the maximum likelihood estimate, thus adding an extra source of variation to the parameter estimates. Note that, because the original study used the default convergence criteria, 23% is probably a more realistic estimate of the standard-error bias in the Burnett et al. 14 study than 18%.
Example 2: Simulated Data
One can easily explore the relation between concurvity and biased standard error estimates further by generating data with various degrees of concurvity. In this section, we provide a simple example of how to do this in S-Plus, together with the results of a simulation based on the example. The example is based on the Poisson model:MATH
To create concurvity, we let both X1 and X2 be uniformly distributed on the interval [0,π], and we let X3 = 0.1 sin(X1) + 0.1 sin(X2) + τ. The error term τ is normally distributed with mean 0 and standard deviation 0.05. This data can then be fit with a Poisson GAM using LOESS functions for X1 and X2 and a linear function of X3. The S-Plus code for a simulation of this model is available with the electronic version of this article at www.epidem.com. When we ran this simulation, we found the average S-Plus standard error estimate to be 1.08 whereas the sample standard deviation of the hundred slope estimates was 1.37. This discrepancy can be made larger by making the standard deviation of the error τ smaller (ie, by making the concurvity worse).
Concurvity in Time Series Data
In the GAM models used by researchers to explore the effect of air pollution on population health, a Poisson model is typically used with time series data to explain log counts of an outcome such as mortality or hospital admissions as a function of air pollution plus a sum of functions of other confounding variables. One of the confounding variables is usually time, which is used to control for seasonal and other temporal trends. Because many variables display temporal trends, it is very likely that at least some degree of concurvity will be present whenever time is used as an independent variable in a GAM. For example, it is reasonable to assume that some of the variation in any continuously varying air pollution variable can be explained by a smooth function of time.
The problem created by concurvity can be understood intuitively by thinking about temporal trends in the data. If the smooth function of time is left out of the model, the residuals tend to be autocorrelated (they display a temporal trend). This is thought to be because some unknown confounding variables, each with some temporal trend of its own, have been left out of the model. The smooth function of time is included to filter out the effect of these unknown variables. If another covariate, say PM10, displays a temporal trend and has a linear effect on log mortality, then the effect of that covariate will also have a temporal trend. The gam function is then given the rather ill-defined job of extracting the temporal trend from log mortality and apportioning some of that trend to time and some to PM10. It is precisely this ambiguity that leads to a larger variance in the estimates. The problem with the variance estimates will also occur with normally distributed data (as well as for any other distribution) and, as far as we know, with all other current software that fits nonparametric GAMs.
The general problem of assessing concurvity, which has been studied by Donnell et al. 15 and, to a certain degree, by Gu, 16 has not been adequately solved. The “semiparametric” GAMs (containing at least one linear parameter) used in epidemiologic investigations, however, have a particular structure that makes concurvity detection relatively straightforward. Suppose that one is fitting a model of the form MATHwith each function fi specified to belong to a set of functions Hi. Suppose further that one is only really interested in the value of the linear parameter β. The variance of β̂ will only be affected by concurvity if there exists a relation of the form MATHwhere the function gi is in the function space Hi and the variance of ε is small relative to the variance of X1. To assess the degree of concurvity in the data, one should fit Eq 5 and then compute the correlation between X1 and X̂1, the fitted values from Eq 5. Although it is difficult to say exactly what magnitude of correlation coefficient should be a cause for concern, our experience suggests that there is certainly reason to expect problems with concurvity if the correlation coefficient is greater than 0.5 (ie, R2 = 0.25).
If one does suspect a problem with concurvity, we recommend the use of a parametric smoother such as the natural spline (ns) in S-Plus. Substituting natural splines with the same degrees of freedom as a nonparametric smoother turns a nonparametric GAM into a fully parametric model that can fit similar trends to the GAM model. This is discussed further in Dominici et al., 9 where it is proposed that using natural splines rather than nonparametric smoothers also removes the problem of biased parameter estimates attributable to lack of convergence of the GAM-estimation algorithm. Given the fact that time series data based on air pollution are unlikely ever to be completely free of concurvity, the best course of action for analysts is certainly always to use fully parametric models. Nonparametric smoother-based models may, however, still be valuable for exploratory analyses. They may well be able to suggest adequate parametric forms. For example, it may well be the case that the effect of a certain variable might be described well by a simple quadratic function or, in the case of a periodic seasonal effect, by a sum of trigonometric terms. If this were the case, it would make sense to use the quadratic or trigonometric function rather than a natural spline in the parametric model.
Air Pollution and Mortality
We observed in the introduction that the epidemiologic literature contains a wealth of reports of a statistical association between air pollution and cardiorespiratory morbidity and mortality. Given that many of these results are based on analyses using generalized additive models and that most of these models are likely to be affected by concurvity, we conducted a meta-analysis to investigate the potential impact of inflated standard errors in these studies. We did this by reanalyzing a meta-analysis by Stieb et al., 17 which examined a collection of comparable epidemiologic studies on the relation between mortality and air pollution. In the reanalysis, we assumed that the pollution effect in every study had been underestimated.
Stieb et al. 17 identified 31 epidemiologic studies, conducted throughout the world, that compared daily variations in PM10 concentrations and daily numbers of nonaccidental deaths. The estimate of the log-relative risk, β̂i, in the daily mortality rate associated with a μg/m3 change in PM10 was recorded for each study in addition to the standard error of this estimate, v̂i. Stieb et al. 17 used a linear random effects model to pool the β̂i among studies, assuming that the true association varied at random. Specifically, it was assumed that the β̂i were normally distributed with mean β and variance θ + v̂i, with θ representing the true variation in the log-relative risks between studies. They estimated β to be 2.0 × 10−2 with a standard error of 3.0 × 10−3. Their estimate of θ was 1.0 × 10−4.
Many of these studies used nonparametric smoothers to model both time and weather variables. Therefore, the reported standard errors of the air pollution association with mortality may have been underestimated. Our analysis of the association between PM2.5 and mortality in Toronto, using the default convergence criteria provided by S-Plus, indicated that the true standard error was 23% greater than the estimated standard error. We repeated the random effects analysis of the 31 studies examining PM10 effects on mortality by inflating the standard error reported in these studies by 23%. The estimate of the pooled log-relative risk was then calculated to be 1.8 × 10−2 with a standard error of 3.3 × 10−3. The estimated variation in log-relative risks between studies, v̂, was 5.3 × 10−5.
The standard error of the mean effect was only 8% larger in the analysis that used inflated study-specific standard errors (23%) compared with the analysis that was based on the data reported in the literature. The estimate of the standard error of the mean effect is a function of the true variation in effects among studies and the estimation of the study-specific effect. If the study-specific estimation error is increased, then the estimate of the true variation among studies is reduced as in our example. Thus, the impact of underestimating the study-specific standard error will be mitigated somewhat through the meta-analysis.
The conclusions based on these two analyses are similar in that there is strong evidence of a statistical association between daily variations in PM10 concentrations and nonaccidental mortality. However, the estimated variation in estimates between studies was cut almost in half. Although the statistical evidence to suggest that there exists true variation in the PM10 association with mortality among studies is reduced from P = 0.000003 to P = 0.009116, one would still likely conclude that such variation is real based on both sets of analyses. Of course, this analysis depends on the assumption that the standard error of each of the excess risks was underestimated by 23%. Even though the true degree of underestimation will certainly not be uniformly equal to 23%, these results suggest that a meta-analysis based on correctly estimated excess risk would still show a statistically substantial PM10 effect.
1. Burnett RT, Brook J, Dann T, et al. Association between particulate and gas phase components of urban air pollution and daily mortality in eight Canadian cities. Inhal Toxicol (in press).
2. Lin M, Chen Y, Burnett RT, Villeneuve P, Krewski D. The influence of ambient coarse particulate matter on hospitalization in children: case-crossover and time series analyses. Environ Health Perspect (in press).
3. Dockery DW, Pope CA, Xu X, et al. An association between air pollution and mortality in six U.S. cities. N Engl J Med 1993; 329: 1753–1759.
4. Pope CA, Thun MJ, Namboodiri MM, et al. Particulate air pollution as a predictor of mortality in a prospective study of U.S. Adults. Am J Respir Crit Care Med 1995; 151: 669–674.
5. Burnett RT, Krewski D. Air pollution effects on hospital admission rates: a random effects modeling approach. Can J Stat 1994; 22: 441–458.
6. Cakmak S, Burnett RT, Krewski D. Adjusting for temporal variation in the analysis of parallel time series of health and environmental variables. J Exp Anal Environ Epid 1998; 8: 129–144.
7. Schwartz J. Nonparametric smoothing in the analysis of air pollution and respiratory illness. Can J Stat 1994; 22: 471–488.
8. Schwartz J. Air pollution and hospital admissions for heart disease in eight U.S. counties. Epidemiology 1999; 10: 17–22.
9. Dominici F, McDermott A, Zeger S, Samet J. On the use of generealized additive models in time-series studies of air pollution and health. Am J Epidemiol 2002;156:3:1–11.
10. Hastie T, Tibshirani R. Generalized Additive Models. London: Chapman and Hall, 1990.
11. McCullagh P, Nelder J. Generalized Linear Models. 2nd ed. New York: Chapman and Hall, 1989.
12. Montgomery DC, Peck EA. Introduction to Regression Analysis. 2nd ed. New York: John Wiley and Sons, 1992.
13. Chambers J, Hastie T, eds. Statistical Models in S. Pacific Grove, Calif: Wadsworth and Brooks, 1992.
14. Burnett RT, Dales R, Krewski D, Vincent R, Dann T, Brook JR. Associations between ambient particulate sulfate and admissions to Ontario hospitals for cardiac and respiratory diseases. Am J Epidemiol 1995; 142: 15–22.
15. Donnell DJ, Buja A, Stuetzle W. Analysis of additive dependencies and concurvities using smallest additive principal components. Ann Stat 1994; 22: 1635–1673.
16. Gu C. Diagnostics for nonparametric regression with additive terms. J Am Stat Soc 1992; 87: 1051–1058.
17. Stieb DM, Judek S, Burnett RT. Meta-analysis of time series studies of air pollution and mortality: effects of gases and particles and the influence of cause of death, age and season. J Air Waste Manage Assoc 2002; 52: 470–484.
This article has been cited 99 time(s).
Journal of Toxicology and Environmental Health-Part A-Current IssuesMethods for bias reduction in time-series studies of particulate matter air pollution and mortalityJournal of Toxicology and Environmental Health-Part A-Current Issues
Environmental Health PerspectivesShort-Term Mortality Rates during a Decade of Improved Air Quality in Erfurt, GermanyEnvironmental Health Perspectives
Environmental Health PerspectivesA temporal, multicity model to estimate the effects of short-term exposure to ambient air pollution on healthEnvironmental Health Perspectives
Canadian Journal of Public Health-Revue Canadienne De Sante Publique
Air pollution and daily hospital admissions for cardiovascular diseases in Windsor, Ontario
Canadian Journal of Public Health-Revue Canadienne De Sante Publique, 96(1):
Regulatory Toxicology and PharmacologyCounterpoint: Time-series studies of acute health events and environmental conditions are not confounded by personal risk factorsRegulatory Toxicology and Pharmacology
Computer Methods and Programs in Biomedicine
An S-Plus function to calculate relative risks and adjusted means for regression models using natural splines
Computer Methods and Programs in Biomedicine, 84(1):
Environmental ResearchAssociations between ambient air pollution and daily mortality among persons with diabetes and cardiovascular diseaseEnvironmental Research
Journal of Toxicology and Environmental Health-Part A-Current IssuesMortality and long-term exposure to ambient air pollution: Ongoing analyses based on the American Cancer Society cohortJournal of Toxicology and Environmental Health-Part A-Current Issues
Journal of Occupational and Environmental MedicineDo subject characteristics modify the effects of particulate air pollution on daily mortality among the elderly?Journal of Occupational and Environmental Medicine
Journal of the Air & Waste Management AssociationEstimating Mortality Effects of Fine Particulate Matter in Japan: A Comparison of Time-Series and Case-Crossover AnalysesJournal of the Air & Waste Management Association
Occupational and Environmental MedicineOzone modifies associations between temperature and cardiovascular mortality: analysis of the NMMAPS dataOccupational and Environmental Medicine
Statistics in MedicineA comparison of methods for the analysis of recurrent health outcome data with environmental covariatesStatistics in Medicine
Environmental Health Perspectives
Cox models: Lepeule et al. Respond
Environmental Health Perspectives, 114():
Journal of Toxicology and Environmental Health-Part A-Current IssuesA new longitudinal design for identifying subgroups of the population who are susceptible to the short-term effects of ambient air pollutionJournal of Toxicology and Environmental Health-Part A-Current Issues
Art of Semiparametrics
A relaxed iterative projection algorithm for rank-deficient regression problems
Art of Semiparametrics, ():
Science of the Total EnvironmentInfluence of set-up conditions of exposure indicators on the estimate of short-term associations between urban pollution and mortalityScience of the Total Environment
Statistics in MedicineEffect measures in non-parametric regression with interactions between continuous exposuresStatistics in Medicine
Journal of Toxicology and Environmental Health-Part A-Current IssuesRevised analyses of the National Morbidity, Mortality, and Air Pollution Study: Mortality among residents of 90 citiesJournal of Toxicology and Environmental Health-Part A-Current Issues
Environmental Health PerspectivesExploring bias in a generalized additive model for spatial air pollution dataEnvironmental Health Perspectives
Journal of Environmental Health
Relationship between particulate matter measured by optical particle counter and mortality in Seoul, Korea, during 2001
Journal of Environmental Health, 71(2):
Environmental and Ecological StatisticsApplication of the chi-plot to assess for dependence in environmental epidemiology studiesEnvironmental and Ecological Statistics
Journal of Toxicology and Environmental Health-Part A-Current IssuesAmbient air pollution and population health: OverviewJournal of Toxicology and Environmental Health-Part A-Current Issues
Social Science & MedicineParticulate air pollution, social confounders, and mortality in small areas of an industrial citySocial Science & Medicine
Journal of Toxicology and Environmental Health-Part A-Current IssuesBayesian model averaging in time-series studies of air pollution and mortalityJournal of Toxicology and Environmental Health-Part A-Current Issues
Advances in Statistical Methods for the Health Sciences: Applications to Cancer and Aids Studies, Genome Sequence Analysis, and Survival Analysis
Longitudinal analysis of short-term bronchiolitis air pollution association using semiparametric models
Advances in Statistical Methods for the Health Sciences: Applications to Cancer and Aids Studies, Genome Sequence Analysis, and Survival Analysis, ():
International Journal of EpidemiologyCommentary: Comparison of time series and case-crossover analyses of air pollution and hospital admission dataInternational Journal of Epidemiology
International Journal of BiometeorologyExcess deaths during the 2004 heatwave in Brisbane, AustraliaInternational Journal of Biometeorology
Journal of Agricultural Biological and Environmental StatisticsA new variance estimator for parameters of semiparametric generalized additive modelsJournal of Agricultural Biological and Environmental Statistics
Statistical Methods in Medical ResearchPenalised regression splines: theory and application to medical researchStatistical Methods in Medical Research
Environmental ResearchSocial disadvantage, air pollution, and asthma physician visits in Toronto, CanadaEnvironmental Research
Journal of Toxicology and Environmental Health-Part A-Current IssuesAir quality risk assessment and managementJournal of Toxicology and Environmental Health-Part A-Current Issues
Statistics in MedicineUsing generalized additive models to reduce residual confoundingStatistics in Medicine
Journal of the American Statistical AssociationImproved semiparametric time series models of air pollution and mortalityJournal of the American Statistical Association
Environmental ResearchWinter air pollution and infant bronchiolitis in ParisEnvironmental Research
Human and Ecological Risk AssessmentAn integrated framework for risk management and population healthHuman and Ecological Risk Assessment
Computational Statistics & Data AnalysisParametric and semi-parametric approaches in the analysis of short-term effects of air pollution on healthComputational Statistics & Data Analysis
International Statistical Review
Health effects of air pollution: A statistical review
International Statistical Review, 71(2):
Inhalation ToxicologyThe health impact of common inorganic components of fine particulate matter (PM2.5) in ambient air: A critical reviewInhalation Toxicology
Cancer Epidemiology Biomarkers & PreventionModeling human metabolism of benzene following occupational and environmental exposuresCancer Epidemiology Biomarkers & Prevention
Inhalation ToxicologyThe health relevance of ambient particulate matter characteristics: Coherence of toxicological and epidemiological inferencesInhalation Toxicology
Archives of Pediatrics & Adolescent Medicine
The relationship of fall school opening and emergency department asthma visits in a large metropolitan area
Archives of Pediatrics & Adolescent Medicine, 159(9):
American Journal of EpidemiologyGaseous air pollutants and asthma hospitalization of children with low household income in Vancouver, British Columbia, CanadaAmerican Journal of Epidemiology
Archives of Environmental Health
Associations between short-term changes in nitrogen dioxide and mortality in Canadian cities
Archives of Environmental Health, 59(5):
Environmental ResearchEffect of short-term exposure to low levels of gaseous pollutants on chronic obstructive pulmonary disease hospitalizationsEnvironmental Research
Environmental Health PerspectivesAmbient ozone concentrations cause increased hospitalizations for asthma in children: An 18-year study in Southern CaliforniaEnvironmental Health Perspectives
International Journal of BiometeorologyTemperature modifies the health effects of particulate matter in Brisbane, AustraliaInternational Journal of Biometeorology
Journal of the Royal Statistical Society Series A-Statistics in Society
Model choice in time series studies of air pollution and mortality
Journal of the Royal Statistical Society Series A-Statistics in Society, 169():
Annual Review of Public HealthTime-series studies of particulate matterAnnual Review of Public Health
Journal of Statistical Software
Analyzing Temperature Effects on Mortality Within the R Environment: The Constrained Segmented Distributed Lag Parameterization
Journal of Statistical Software, 32():
Statistics in MedicineSeasonal confounding in air pollution and health time-series studies: Effect on air pollution effect estimatesStatistics in Medicine
Journal of the Royal Statistical Society Series A-Statistics in Society
Model choice in time series studies of air pollution and mortality - Comments on the paper by Peng, Dominici and Louis
Journal of the Royal Statistical Society Series A-Statistics in Society, 169():
Journal of the Air & Waste Management Association
Meta-analysis of time-series studies of air pollution and mortality: Update in relation to the use of generalized additive models
Journal of the Air & Waste Management Association, 53(3):
Environmental Health PerspectivesUsing supervised principal components analysis to assess multiple pollutant effectsEnvironmental Health Perspectives
Inhalation ToxicologyAir pollution and postneonatal mortality in a tropical city: Kaohsiung, TaiwanInhalation Toxicology
EnvironmetricsSeasonal confounding and residual correlation in analyses of health effects of air pollutionEnvironmetrics
Environmental ResearchHeat acclimatization and sunshine cause false indications of mortality due to ozoneEnvironmental Research
Environmental Health Perspectives
Estimating the exposure-response relationships between particulate matter and mortality within the APHEA multicity project
Environmental Health Perspectives, 113(1):
Inhalation ToxicologyInfluence of relatively low level of particulate air pollution on hospitalization for COPD in elderly peopleInhalation Toxicology
Inhalation ToxicologyAssociation between ozone and respiratory admissions among children and the elderly in Vancouver, CanadaInhalation Toxicology
EnvironmetricsSemiparametric penalized generalized additive models for environmental research and epidemiologyEnvironmetrics
Environmental Health PerspectivesAcute Effects of Ambient Particulate Matter on Mortality in Europe and North America: Results from the APHENA StudyEnvironmental Health Perspectives
Journal of the Air & Waste Management AssociationA new multipollutant, no-threshold air quality health index based on short-term associations observed in daily TimeSeries analysesJournal of the Air & Waste Management Association
Journal of Exposure Analysis and Environmental EpidemiologyAssociation between particulate matter and emergency room visits, hospital admissions and mortality in Spokane, WashingtonJournal of Exposure Analysis and Environmental Epidemiology
Environmental ResearchDifference in the relation between daily mortality and air pollution among elderly band all-ages populations in southwestern FranceEnvironmental Research
Journal of Exposure Science and Environmental EpidemiologyDaily mortality and particulate matter in different size classes in Erfurt, GermanyJournal of Exposure Science and Environmental Epidemiology
Risk AnalysisImproving the Predictive Accuracy of Hurricane Power Outage Forecasts Using Generalized Additive ModelsRisk Analysis
Journal of Exposure Science and Environmental EpidemiologyCharacterization of PM2.5, gaseous pollutants, and meteorological interactions in the context of time-series health effects modelsJournal of Exposure Science and Environmental Epidemiology
Journal of Toxicology and Environmental Health-Part A-Current IssuesSources of uncertainty in calculating mortality and morbidity attributable to air pollutionJournal of Toxicology and Environmental Health-Part A-Current Issues
American Journal of Respiratory and Critical Care MedicineAcute effects of ozone on mortality from the "Air pollution and health: A European approach" projectAmerican Journal of Respiratory and Critical Care Medicine
American Journal of Respiratory and Critical Care Medicine
Acute effects of air pollution on admissions: Reanalysis of APHEA 2
American Journal of Respiratory and Critical Care Medicine, 169():
Journal of Epidemiology and Community Health
Predicted health impacts of urban air quality management
Journal of Epidemiology and Community Health, 58(2):
Environmental Health PerspectivesDoes particulate matter modify the association between temperature and cardiorespiratory diseases?Environmental Health Perspectives
EnvironmetricsA comparative study of the use of GAM and GLM in air pollution researchEnvironmetrics
Inhalation ToxicologyThe effect of coarse ambient particulate matter on first, second, and overall hospital admissions for respiratory disease among the elderlyInhalation Toxicology
American Journal of EpidemiologyComparison of a spatial perspective with the multilevel analytical approach in neighborhood studies: The case of mental and behavioral disorders due to psychoactive substance use in Malmo, Sweden, 2001American Journal of Epidemiology
Environmental Health PerspectivesBootstrap-after-Bootstrap Model Averaging for Reducing Model Uncertainty in Model Selection for Air Pollution Mortality StudiesEnvironmental Health Perspectives
International Journal of BiometeorologyTemperature, air pollution and total mortality during summers in Sydney, 1994-2004International Journal of Biometeorology
Journal of Toxicology and Environmental Health-Part A-Current IssuesHealth effects associated with exposure to ambient air pollutionJournal of Toxicology and Environmental Health-Part A-Current Issues
Science of the Total EnvironmentA vulnerability study of the low-income elderly in the context of high temperature and mortality in Seoul, KoreaScience of the Total Environment
Journal of Epidemiology and Community HealthA bootstrap method to avoid the effect of concurvity in generalised additive models in time series studies of air pollutionJournal of Epidemiology and Community Health
Environmental ResearchThe role of air pollution in the relationship between a heat stress index and human mortality in TorontoEnvironmental Research
Bmc Infectious DiseasesEvaluation of alternative respiratory syndromes for specific syndromic surveillance of influenza and respiratory syncytial virus: a time series analysisBmc Infectious Diseases
Journal of Exposure Analysis and Environmental EpidemiologyDo associations between airborne particles and daily mortality in Mexico City differ by measurement method, region, or modeling strategy?Journal of Exposure Analysis and Environmental Epidemiology
International Journal of EpidemiologyA note on temperature effect estimate in mortality time series analysisInternational Journal of Epidemiology
Archives of Environmental & Occupational Health
Temperature Enhanced Effects of Ozone on Cardiovascular Mortality in 95 Large US Communities, 1987-2000: Assessment Using the NMMAPS Data
Archives of Environmental & Occupational Health, 64(3):
Regional Environmental ChangeHuman impacts on fire occurrence: a case study of hundred years of forest fires in a dry alpine valley in SwitzerlandRegional Environmental Change
Environmental Science & TechnologyQuantile-Based Bayesian Maximum Entropy Approach for Spatiotemporal Modeling of Ambient Air Quality LevelsEnvironmental Science & Technology
International Journal of Environmental Health ResearchRelationship between ambient ozone concentrations and daily hospital admissions for childhood asthma/atopic dermatitis in two cities of Korea during 2004-2005International Journal of Environmental Health Research
European Journal of EpidemiologyGeneralized Additive Models applied to analysis of the relation between amount and type of alcohol and all-cause mortalityEuropean Journal of Epidemiology
Journal of Toxicology and Environmental Health-Part ARejoinder: Reanalysis of the Harvard Six Cities Study and American Cancer Society study of particulate air pollution and mortalityJournal of Toxicology and Environmental Health-Part A
Archives of Environmental Health
Association between particulate air pollution and first hospital admission for childhood respiratory illness in Vancouver, Canada
Archives of Environmental Health, 59(1):
Environmental Health PerspectivesAir pollution and mortality in Chile: Susceptibility among the elderlyEnvironmental Health Perspectives
concurvity; generalized additive model; sulfate particles; air pollution; PM2.5; PM10
Supplemental Digital Content
© 2003 Lippincott Williams & Wilkins, Inc.
Highlight selected keywords in the article text.