time stratified case crossover studies

What Is a Case Study?

When you’re performing research as part of your job or for a school assignment, you’ll probably come across case studies that help you to learn more about the topic at hand. But what is a case study and why are they helpful? Read on to learn all about case studies.

At face value, a case study is a deep dive into a topic. Case studies can be found in many fields, particularly across the social sciences and medicine. When you conduct a case study, you create a body of research based on an inquiry and related data from analysis of a group, individual or controlled research environment.

As a researcher, you can benefit from the analysis of case studies similar to inquiries you’re currently studying. Researchers often rely on case studies to answer questions that basic information and standard diagnostics cannot address.

Study a Pattern

One of the main objectives of a case study is to find a pattern that answers whatever the initial inquiry seeks to find. This might be a question about why college students are prone to certain eating habits or what mental health problems afflict house fire survivors. The researcher then collects data, either through observation or data research, and starts connecting the dots to find underlying behaviors or impacts of the sample group’s behavior.

Gather Evidence

During the study period, the researcher gathers evidence to back the observed patterns and future claims that’ll be derived from the data. Since case studies are usually presented in the professional environment, it’s not enough to simply have a theory and observational notes to back up a claim. Instead, the researcher must provide evidence to support the body of study and the resulting conclusions.

Present Findings

As the study progresses, the researcher develops a solid case to present to peers or a governing body. Case study presentation is important because it legitimizes the body of research and opens the findings to a broader analysis that may end up drawing a conclusion that’s more true to the data than what one or two researchers might establish. The presentation might be formal or casual, depending on the case study itself.

Draw Conclusions

Once the body of research is established, it’s time to draw conclusions from the case study. As with all social sciences studies, conclusions from one researcher shouldn’t necessarily be taken as gospel, but they’re helpful for advancing the body of knowledge in a given field. For that purpose, they’re an invaluable way of gathering new material and presenting ideas that others in the field can learn from and expand upon.

• Privacy Policy
• Terms of Service
• © 2023 Ask Media Group, LLC

Browse Econ Literature

• Working papers
• Software components
• Book chapters
• JEL classification

More features

• Subscribe to new research

RePEc Biblio

Author registration.

• Economics Virtual Seminar Calendar NEW!

Analysis of time-stratified case-crossover studies in environmental epidemiology using Stata

• Author & abstract
• Related works & more

Corrections

(Spanish Council for Scientific Research, Barcelona)

(London School of Hygiene and Tropical Medicine)

Suggested Citation

Download full text from publisher.

Follow serials, authors, keywords & more

Public profiles for Economics researchers

Various research rankings in Economics

RePEc Genealogy

Who was a student of whom, using RePEc

Curated articles & papers on economics topics

Upload your paper to be listed on RePEc and IDEAS

New papers by email

Subscribe to new additions to RePEc

EconAcademics

Blog aggregator for economics research

Cases of plagiarism in Economics

About RePEc

Initiative for open bibliographies in Economics

News about RePEc

Questions about IDEAS and RePEc

RePEc volunteers

Participating archives

Publishers indexing in RePEc

Privacy statement

Found an error or omission?

Opportunities to help RePEc

Get papers listed

Have your research listed on RePEc

Open a RePEc archive

Have your institution's/publisher's output listed on RePEc

Get RePEc data

Use data assembled by RePEc

Time-Stratified Case Crossover Study of the Association of Outdoor Ambient Air Pollution With the Risk of Acute Myocardial Infarction in the Context of Seasonal Exposure to the Southeast Asian Haze Problem

Affiliations.

• 1 1 SingHealth Duke-NUS Emergency Medicine Academic Clinical Programme Singapore.
• 2 2 SingHealth Emergency Medicine Residency Programme Singapore.
• 3 3 Cardiovascular & Metabolic Disorders Program Duke-National University of Singapore Medical School Singapore.
• 4 4 Department of Emergency Medicine Singapore General Hospital Singapore.
• 5 6 National Registry of Diseases Office Health Promotion Board Singapore.
• 6 7 Department of Epidemiology and Preventive Medicine School of Public Health and Preventive Medicine Monash University Melbourne Victoria Australia.
• 7 8 Engineering Cluster Singapore Institute of Technology Singapore.
• 8 9 Science and Math Cluster Singapore University of Technology and Design Singapore.
• 9 10 Saw Swee Hock School of Public Health National University of Singapore Singapore.
• 10 11 Health Services Research Centre Singapore Health Services Singapore.
• 11 12 Centre for Quantitative Medicine Duke-NUS Medical School Singapore.
• 12 13 Program in Health Services and Systems Research Duke-NUS Medical School Singapore.
• 13 14 Department of Cardiology National Heart Centre Singapore Singapore.
• 14 5 Department of General Surgery Singapore General Hospital Singapore.
• 15 15 National Heart Research Institute Singapore National Heart Centre Singapore.
• 16 16 Yong Loo Lin School of Medicine National University Singapore Singapore.
• 17 17 The Hatter Cardiovascular Institute University College London London United Kingdom.
• 18 18 The National Institute of Health Research University College London Hospitals Biomedical Research Centre, Research & Development London United Kingdom.
• 19 19 Department of Cardiology Barts Heart Centre St Bartholomew's Hospital London United Kingdom.
• 20 20 Division of Neurology Department of Medicine National University Health System Singapore.
• PMID: 31112443
• PMCID: PMC6475051
• DOI: 10.1161/JAHA.118.011272

Background Prior studies have demonstrated the association of air pollution with cardiovascular deaths. Singapore experiences seasonal transboundary haze. We investigated the association between air pollution and acute myocardial infarction ( AMI ) incidence in Singapore. Methods and Results We performed a time-stratified case-crossover study on all AMI cases in the Singapore Myocardial Infarction Registry (2010-2015). Exposure on days where AMI occurred (case days) were compared with the exposure on days where AMI did not occur (control days). Control days were chosen on the same day of the week earlier and later in the same month and year. We fitted conditional Poisson regression models to daily AMI incidence to include confounders such as ambient temperature, rainfall, wind-speed, and Pollutant Standards Index. We assessed relationships between AMI incidence and Pollutant Standards Index in the entire cohort and subgroups of individual-level characteristics. There were 53 948 cases. Each 30-unit increase in Pollutant Standards Index was association with AMI incidence (incidence risk ratio [ IRR ] 1.04, 95% CI 1.03-1.06). In the subgroup of ST -segment-elevation myocardial infarction the IRR was 1.00, 95% CI 0.98 to 1.03, while for non-ST-segment-elevation myocardial infarction, the IRR was 1.08, 95% CI 1.05 to 1.10. Subgroup analyses showed generally significant. Moderate/unhealthy Pollutant Standards Index showed association with AMI occurrence with IRR 1.08, 95% CI 1.05 to 1.11 and IRR 1.09, 95% CI 1.01 to 1.18, respectively. Excess risk remained elevated through the day of exposure and for >2 years after. Conclusions We found an effect of short-term air pollution on AMI incidence, especially non-ST-segment-elevation myocardial infarction and inpatient AMI . These findings have public health implications for primary prevention and emergency health services during haze.

Keywords: Singapore; air pollution; haze; myocardial infarction; population.

Publication types

• Multicenter Study
• Air Pollution / adverse effects*
• Cross-Over Studies
• Follow-Up Studies
• Middle Aged
• Myocardial Infarction / epidemiology*
• Myocardial Infarction / etiology
• Particulate Matter / adverse effects*
• Registries*
• Retrospective Studies
• Risk Assessment / methods*
• Singapore / epidemiology
• Time and Motion Studies
• Particulate Matter

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings
• Advanced Search
• Journal List
• Int J Environ Res Public Health

Case-Crossover Method with a Short Time-Window

Numerous epidemiological studies have shown associations between short-term ambient air pollution exposure and various health problems. The time-stratified case-crossover design is a popular technique for estimating these associations. In the standard approach, the case-crossover model is realized by using a conditional logistic regression on data that are interpreted as a set of cases (i.e., individual health events) and controls. In statistical calculations, for each case record, three or four corresponding control records are considered. Here, the case-crossover model is realized as a conditional Poisson regression on counts with stratum indicators. Such an approach enables the reduction of the number of data records that are used in the numerical calculations. In this presentation, the method used analyzes daily counts on the shortest possible time-window, which is composed of two consecutive days. The proposed technique is positively tested on four challenging simulated datasets, for which classical time-series methods fail. The methodology presented here also suggests that the length of exposure (i.e., size of the time-window) may be associated with the severity of health conditions.

1. Introduction

The case-crossover (CC) study design is an approach where each individual case serves as its own control [ 1 ]. The CC method is often and widely used in environmental epidemiology to estimate the risk of a health event related to short-term (acute) exposure to ambient air pollution. Specifically, this method is usually applied for investigating the transient effects of an intermittent exposure on the onset of acute events.

The standard CC technique usually uses one month as its time-window. The strategy to determine for a case corresponding control(s) is usually based on a time-stratified approach [ 2 ]. In the CC method, the day of week is not modeled but is adjusted by the realized design. For the specific case day, the potential controls are generated as multiples of seven from the case day in both directions, pre- and post-event day ( ± w   x   7 , where w = 1 ,   2 ,   3 ,   4 ). The control days are then chosen among the proposed days that belong to the same one common month [ 2 ]. This process results in three or four control days, depending on the length of a month (28, 29, 30, or 31 days) and the day of the week. For example, according to this scheme, for the case of the day of 9 November 2019, there are four control days: 2, 16, 23, and 30 November. In such scenarios, the cases (events) and controls are still related to time. The pattern of sequences of case-control(s) varies in time. We may have various configurations of pre- and post-event control days; for example, one pre and three post (see the example with the case day of 9 November) or two pre and two post. In the original presentation of the CC method [ 1 ], one control was defined, and the control always occurred seven days before the time event. Such a fixed-time configuration of case-control relations usually results in bias from time trends in exposure prior to the occurrence of an adverse event.

To reduce the bias in the CC method, a few authors [ 3 , 4 , 5 ] have presented modifications to the CC technique and have used daily counts rather than individual health events (i.e., case-control relations). Their proposed methods are realized by using clusters (strata) that control for time by the applied hierarchical structures based on the calendar relations <year: month: day of week>. Daily counts are grouped and considered on the defined strata. According to the assumed convention, daily counts on 2, 9, 16, 23, and 30 November all belong to the same stratum that is determined by the structure <year = 2019: month = November: day of week = Saturday>. It is also good to note that, in this approach, the corresponding regression is not affected by time (pre- and post-event scenarios) in the time-window used; rather, it is affected by the concentration levels of the air pollutant considered. Time is controlled by the constructed strata, and it is eliminated from the regression. Thus, the “control is always before case” effect or similar effects, which are usually seen in the CC method and generate bias, are lessened. It is possible to construct strata of various sizes. Among the proposed ones is the CC method that uses a two week time-window, i.e., with the strata of the form <year: two week: day of week>. These strata have only two days [ 3 , 5 ]. This is different from the originally proposed CC method that only uses one control [ 1 ] for each health case. In this situation, the time duration is also two weeks but is interpreted differently. Say that we consider the daily counts on two consecutive Mondays, and the method is realized with the use of the stratum <year: two week: day of week>; in the originally presented CC method, for the case of a Monday, its control is the Monday one week before. In both approaches, two data points are enough to determine a slope (beta) between the air pollution concentration and health outcomes. The authors of [ 6 ] proposed another method to reduce bias in CC methods.

In the time-series study considered here, a few modifications and adaptations of the CC method are proposed and used to control for the bias related to the size of the applied time-window. In one of the time-series study papers by Burr et al., it was stated [ 7 ]: “We additionally showed that the use of 6 df/year for a smooth function of time is not, in general, sufficient to protect estimates from seasonal variation. The use of natural cubic regression splines at higher df/year (e.g., 12) will protect slightly better against seasonal variation than 6 df/year but still suffers from the poor concentration properties of its family.”

This issue was among the reasons for doing the present study. As the authors showed in their publication [ 7 ], a classical time-series methodology did not work for their challenging simulated data. As a solution, they proposed alternative smooth functions. This work presents a simple and effective solution to the bias correction problem by applying a CC method technique that is realized over a short time window.

2. Materials and Methods

This study was conducted by using the simulated data (four sets: Sim1–Sim4) and mortality data from Chicago, IL, USA [ 7 ]. The original Chicago mortality data were provided by the NMMAPSdata database [ 8 ]. Here, the attention was restricted to two air pollutants (trimmed mean daily coarse particulate matter and trimmed mean daily ozone). The mortality data contained daily all-cause mortality counts (death) and daily cardio-vascular mortality counts (CVD), for the period of 1987–2000. These databases are exactly the same as in the publication for the time-series study [ 7 ], and all their details are presented in the original time-series work. The simulated data and code (in R , [ 9 ]) that were used are available at https://github.com/szyszkowiczm/Data2D . The Supplementary Material associated with the publication of Burr et al. [ 7 ] contains the details of the simulated data. In many calculations of short-term risk related to air pollution, smooth functions of time are realized as fixed-df cubic regression splines. In this presentation the constructed strata were used to control time.

In this study, we employed a CC method that uses the hierarchical calendar structure of <year: two days>, noted here symbolically as <year: 2D>; this version of the method is called the CC2D method. For each individual year, consecutive days were grouped in pairs (2D), specifically (1, 2), (3, 4), (5, 6), … etc. For example, the pair of (1, 2) represents the first and second day of January and the pair of (3, 4) represents the next two days of this month. In the constructed models, we adjusted for the day of week. This was the shortest possible time-window in the CC method to analyze daily counts. In this situation, the stratum had two days. Figure 1 illustrates two approaches: the case-crossover realized with events and with counts.

Left panel ( i ) shows variants of the case-crossover (CC) methods, with three control schemes {−7,0}, {−7,0,7}, and {−7,0,7,14,21}, where 0 is an event day, −7 is one week before event, 7 is one week after event, etc. Thus, {−7,0} results in two days; 9 November as an event day and 2 November as a control day. Right panel ( ii ) shows the points used in the realization of the time-stratified CC methods with counts (the CCM method, which has hierarchical clusters of the form <year: month: day of week>). The exposure levels are A–E and can be lagged. The CC2D method (the method that uses the hierarchical calendar structure of <year: two days>) for just one pair—(A, CA), (F, CF)—is shown in the ellipsis; exposures A and F result in counts CA and CF, respectively.

Here, we realized conditional Poisson regression as an alternative approach to the conditional logistic regression technique used in the standard CC model. The package “gnm” was used to specify and fit generalized nonlinear models to the defined stratum (here, <year:2D>).

The following model in the R software [ 9 ] was built for the case of the simulated data (four different categories: Sim1–Sim4, with 250 samples each)

• modM <- gnm(y ~x + dowf, data = data,
• family = gaussian, eliminate = factor(stratum)),

where y is response (health outcome), x is exposure (air pollution concentration), and dowf is the day of week as a factor. The stratum (cluster) is defined as data$stratum <- as.factor(data$year:data$2D). In the model, there are hierarchical clusters with two structure levels: two days embedded in a year and 183 clusters per year. In this study, the simulated data used were artificial and represented 10 years (3650 days) of time-series sequences (exposure–response); thus, each year was a series of 365 days. The simulation data were not counts and were generated by the authors [ 7 ] (Burr et al., 2015) to investigate their technique and its properties. Simulation 1 consisted of two periodicities at periods 183 and 75 days. Simulation 2 extended Simulation 1 by including three additional signals. The signals were added at periods of 68 and 105 days. Simulation 3 used a similar signal structure as Simulation 2 but changed the background noise. Simulation 4 was based on the same signal structure as Simulation 3 but was scaled by a factor of the two background noises [ 7 ].

In addition, for comparison purposes, we also considered the following models; CC3D—similar to the CC2D model but with the stratum based on a three-day structure <year:3 days>; CC2W —with a 3-level stratum of the form <year: two week: day of week>; and CC2CW—with a 3-level stratum as in CC2W but “chained,’ as each day was used twice in the neighbor strata, e.g., (1, 2), (2, 3), and (3, 4). [ 5 ]. As the simulated data were challenging for the analysis, we realized these methods, CC2D, CC3D, CC2W, and CC2CW, to observe their performance.

Additionally, to analyze the mortality Chicago data, models with the hierarchical clusters of the form <year: month: day of week> (CCM) were also used. In the constructed models, family = Poisson was set in the R code, and this model also included ambient temperature. The temperature was represented in the form of natural splines with three degrees of freedom. In this situation, we had real epidemiological data.

The simulated data posed rather difficult and challenging problems for the time-series approach [ 7 ]. The true value for the slope was one (beta = 1.0) for each simulated series (Sim1–Sim4). According to the authors of the original publication [ 7 ], the simulated data that used the time-series method, executed with natural cubic regression splines for time with 6 df/year, produced the following slope estimates: 0.284, 0.064, 0.139, and −0.177 (negative) for the data from Sim1 to Sim4, respectively. These were the average values of the estimated slopes with 250 samples for each simulation (as presented in Table 1 in the paper of Burr et al. [ 7 ]).

Estimated parameters obtained for the CC2D and CC3D methods.

The results obtained are summarized in Figure 2 and Figure 3 . Table 1 and Table 2 present the numerical values generated for the simulated data. Figure 1 shows the estimated slope values for each series that used the simulated data (Sim1–Sim4). These series had 250 samples each (4 series x 250 samples), and each has 3650 days, i.e., 10 years.

Estimated slopes (beta, true beta = 1.0) for four sets of simulation data (Sim1–Sim4) with 250 samples each. The panels illustrate the results for the following methods: ( a ) CC2D, ( b ) CC3D (similar to the CC2D model but with the stratum based on a three-day structure <year:3 days>), ( c ) CC2W (with a 3-level stratum of the form <year: two week: day of week>), and ( d ) CC2CW(with a 3-level stratum as in CC2W but “chained,’ as each day was used twice in the neighbor strata). The simulation data are identified by the following symbols: Sim1—square; Sim2—black circle (seen as a solid black line, as the values are almost identical); Sim3—circle; and Sim4—triangle.

Estimated slopes (beta) for mortality data. Chicago, USA, 1987–2000. Note: CC—case-crossover method; M, 2W, and 2D—time-windows of one month, two weeks, and two days, respectively. CVD—cardio-vascular mortality; beta—slope; and CI—confidence interval.

Estimated parameters obtained for the CC2W and CC2CW methods.

The figure illustrates the results for the four methods of CC2D, CC3D, CC2W, and CC2CW in panels (a)–(d), respectively. The CC method on the cluster <year:2D>, i.e., the CC2D method, have very accurate estimates of the slopes. The estimated average slopes from the simulations were 0.9819, 0.9674, 0.9707, and 0.9628, respectively (where the true beta = 1.0) with standard deviations 2.26E-03, 4.81E-05, 2.80E-03, and 4.31E-03, respectively. Table 1 summarizes the numerical results for the CC2D and CC3D methods. Table 2 shows the same statistics of the results from the CC2W and CC2CW methods. The average values of the estimated slopes (beta1–beta4 for simulations (1–4) are shown in bold. Their standard errors are listed under SE1–SE4 for the corresponding simulation data.

Figure 3 represents the results for the mortality data in Chicago and ambient air pollutants. These are real ambient ozone and PM10 (particulate matter with diameters of no greater than 10 microns) air pollution concentration data. Three forms of the CC model were applied to process these data: CCM (one-month time-window), CC2W (two week time-window), and CC2D (two day time-window). The intention was to compare the standard CC method (CCM) with two kinds of two-day methods (CC2W and CC2D).

In the cases of the CCM and CC2W approaches, the day of week was adjusted by the design, as was explained previously, in the form of the clusters used. In this situation, the day of week was one of the three levels of the constructed hierarchical clusters.

4. Discussion

The presented statistical technique, the CC method with two days as the time-window (CC2D), was easy to implement and use in short-term air health effects studies. The proposed method worked very well with the simulated data used here for illustrative purposes. Since in this case we knew the true value of the slope (beta = 1), it was easy to judge and validate the obtained results.

As the results for the simulated data presented in Table 1 and Table 2 indicate, the CC2D method gave the most accurate estimate of the true slope (beta = 1.0) among the other applied methods: CC3D, CC2W, and CC2CW.

As was already mentioned, the CC2CW was realized by using the clusters in the form <year: chained 2 week: day of week>. Two weeks were chained according to the following construction: (first week, second week), (second week, third week), (third week, fourth week), etc., separately for each year. This approach almost doubled the number of observations and narrowed the corresponding confidence intervals. According to the results presented in Table 2 , the estimated slopes were 0.8549, 0.8118, 0.8270, and 0.7656, and their estimated average standard errors were 0.0084, 0.0094, 0.0090, and 0.0138, for Sim1–Sim4, respectively. These standard errors were smaller than those obtained for the CC2W method, as they were 0.0120, 0.0132, 0.0126, and 0.0193, respectively.

In the case of the mortality data in Chicago, we do not know the true effects of the exposures on the daily mortality counts. In the models applied, we adjusted for the ambient temperature by using natural splines with three degrees of freedom. The CC2D method indicated a positive association between ozone and death, as well as a statistically significant association for CVD mortality. The same types of the associations were obtained by the authors in [ 7 ] (see Figure 5 in [ 7 ]). The results appear very reliable because the CVD counts were a subset of the death counts. In the case of particulate matter exposure, the CCM and CC2W methods indicated positive statistically significant associations for death, but they only indicated a positive non-significant association for CVD deaths. The same type of relations was reported by the authors in [ 7 ]. It is strange that the results were not the same nature for CVD deaths as they were for death. The CC2D method did not show the associations for CVD deaths, all deaths and coarse particulate matter concentrations. It is difficult to make conclusions here, as the methods we used were different in nature. The CCW and CC2W methods used the same approach as case-control relationships, while the CC2D method used counts and was closer to the time-series technique than to the case-crossover methodology. The CC2D method estimated the slopes by using two neighboring days, so it may have been more related to acute events (see Figure 1 ).

5. Conclusions

The presented CC2D method performed very well on the four simulated datasets. This simple technique enabled an accurate estimate of the slope (beta = 1). In addition, we have a few conclusions: (a) The Conditional Poisson model is a flexible and reliable alternative to the conditional logistic case cross-over model; (b) using counts reduces the number of records in the realization of the CC technique [ 10 ]; and (c) the strata applied to control time may have various sizes. We also conclude that it is reasonable to run the CC2D model, at least to verify the associations suggested by other approaches.

Acknowledgments

The author thanks Wesley Burr for the simulated data used in this study.

Supplementary Materials

The following are available online at https://github.com/szyszkowiczm/Data2D —the simulated data and the corresponding codes in R.

This research received no external funding.

Conflicts of Interest

The author declares no conflict of interest.

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here .

Loading metrics

Open Access

Peer-reviewed

Research Article

Using Bayesian time-stratified case-crossover models to examine associations between air pollution and “asthma seasons” in a low air pollution environment

Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Visualization, Writing – original draft

* E-mail: [email protected]

Current address: Environmental Health Department, Boston University School of Public Health, Boston, Massachusetts, United States of America

Affiliation Department of Public Health Sciences, Medical University of South Carolina, Charleston, South Carolina, United States of America

Roles Conceptualization, Methodology, Supervision, Writing – review & editing

Roles Conceptualization, Methodology, Resources, Writing – review & editing

Roles Methodology, Resources, Supervision, Writing – review & editing

• Matthew Bozigar, 
• Andrew B. Lawson, 
• John L. Pearce, 
• Erik R. Svendsen, 
• John E. Vena

• Published: December 8, 2021
• https://doi.org/10.1371/journal.pone.0260264
• Reader Comments

Many areas of the United States have air pollution levels typically below Environmental Protection Agency (EPA) regulatory limits. Most health effects studies of air pollution use meteorological (e.g., warm/cool) or astronomical (e.g., solstice/equinox) definitions of seasons despite evidence suggesting temporally-misaligned intra-annual periods of relative asthma burden (i.e., “asthma seasons”). We introduce asthma seasons to elucidate whether air pollutants are associated with seasonal differences in asthma emergency department (ED) visits in a low air pollution environment. Within a Bayesian time-stratified case-crossover framework, we quantify seasonal associations between highly resolved estimates of six criteria air pollutants, two weather variables, and asthma ED visits among 66,092 children ages 5–19 living in South Carolina (SC) census tracts from 2005 to 2014. Results show that coarse particulates (particulate matter <10 μm and >2.5 μm: PM 10-2.5 ) and nitrogen oxides (NO x ) may contribute to asthma ED visits across years, but are particularly implicated in the highest-burden fall asthma season. Fine particulate matter (<2.5 μm: PM 2.5 ) is only associated in the lowest-burden summer asthma season. Relatively cool and dry conditions in the summer asthma season and increased temperatures in the spring and fall asthma seasons are associated with increased ED visit odds. Few significant associations in the medium-burden winter and medium-high-burden spring asthma seasons suggest other ED visit drivers (e.g., viral infections) for each, respectively. Across rural and urban areas characterized by generally low air pollution levels, there are acute health effects associated with particulate matter, but only in the summer and fall asthma seasons and differing by PM size.

Citation: Bozigar M, Lawson AB, Pearce JL, Svendsen ER, Vena JE (2021) Using Bayesian time-stratified case-crossover models to examine associations between air pollution and “asthma seasons” in a low air pollution environment. PLoS ONE 16(12): e0260264. https://doi.org/10.1371/journal.pone.0260264

Editor: Daniel Dunea, Valahia University of Targoviste, ROMANIA

Received: February 19, 2021; Accepted: November 5, 2021; Published: December 8, 2021

Copyright: © 2021 Bozigar et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: Emergency department visit records cannot be shared publicly because they are protected health information. However, they can be requested from the South Carolina Revenue and Fiscal Affairs - Data Oversight Council, but research protocols are subject to their approval. Inquiries can be sent to Chris Finney at [email protected] . Weather estimates are publicly available from https://prism.oregonstate.edu/ . Inquiries about access to 12 km resolution air pollution estimates from the fusion model can be directed to Georgia Institute of Technology (Georgia Tech), specifically Armistead (Ted) Russell: [email protected] . Cleaned, non-identifiable analytic data and related code to reproduce the main statistical analyses are available on GitHub: https://github.com/mbozigar/asthma-seasons .

Funding: Manuscript funding was graciously provided by the Environmental Health Department in the Boston University School of Public Health, chaired by Dr. Jon Levy.

Competing interests: The authors have declared that no competing interests exist.

1.0 Introduction

Emergency department (ED) visits for asthma have complex drivers influenced by disease severity, access to and utilization of various preventive care services [ 1 ], and numerous environmental factors [ 2 ]. Asthma exacerbations and ED visits have been associated with multiple ambient air pollutants, complex mixtures, and temporal lags [ 3 – 6 ]. While there is evidence of seasonality, previous studies have used several common season definitions, such as astronomical seasons defined by equinoxes and solstices (i.e., winter, spring, summer, fall), meteorological seasons (e.g., cool, warm), and 3-month simplifications (e.g., winter: December January, and February) to name a few. When exacerbations were linked with non-environmental factors, such as body mass index (BMI), researchers employed traditional astronomical seasons [ 7 ]. In studies focused on ambient air pollution exposures, scholars tend to define time periods by warm and cold seasons, reflective of temporal patterns in the exposures [ 8 – 10 ]. In addition, some acute health effects of air pollution studies used simplified 3-month seasonal blocks [ 11 , 12 ]. Air pollution, pollen, and viruses are several examples of seasonally-varying environmental exposures exacerbating asthma symptoms [ 13 – 17 ]. Some of these exposures may co-occur over time, while others may only partially overlap, if at all. Viral infections tend to increase in cooler periods such as the fall and winter [ 18 ]. Seasonal associations between rain events or thunderstorms and asthma exacerbations have been found in warmer months, such as the spring and summer in the US, with a likely mechanism being the release of grass pollen [ 19 – 21 ].

Air pollution and other place-based ambient environmental factors can interact in complex ways among people with asthma [ 2 ]. Most short-term analyses of air pollution and asthma ED visits focus on highly urbanized areas [ 8 , 10 , 22 , 23 ] that tend to have higher air pollution levels than suburban or rural areas and are monitored more thoroughly [ 24 ]. Acute health effects studies of air pollution have been predominantly designed to estimate associations and dose-response relationships in such urban environments. They assume that temporal variation in asthma ED visits are driven by seasonal patterns among ambient pollutants, themselves, despite evidence of temporally-misaligned patterns in the asthma ED visits. At the national level, aggregated asthma ED visits peak in the fall, but there are differing regional and local seasonal burdens [ 25 – 27 ]. Thus, study type, objectives, location, exposures, exposure levels, and exposure timing are elements that should be considered when defining seasons for an asthma outcome, such as emergency department visits, that can exhibit large temporal differences within a given year.

Though controlled exposure studies of air pollutants are important for understanding disease pathways [ 28 , 29 ], they are ethically untenable for a relatively severe outcome such as ED visits. As such, epidemiologists must contend with various observational designs that attempt to infer disease patterns in study populations. Case-crossover studies inherently control for time-invariant characteristics and are therefore useful for studying short-term, time-varying exposures affecting health [ 30 ]. They are additionally useful when only case data are known, such as in administrative health datasets (e.g., ED visits), in which there are no contrasting non-case events to provide outcome variability to develop a data model. With care, case-crossover models can be developed for individuals in case-only datasets because non-case events (i.e., referents) can be strategically selected and temporally matched to the case events for each person. Case-crossover designs do not need to be aggregated over space and therefore permit spatially explicit exposure estimates, unlike time-series designs [ 31 ]. While time series designs serve as a primary option for case-only data in environmental epidemiology, they apply only one exposure estimate to all study participants in each time period. Bayesian case-crossover models have been equally or more accurate than frequentist versions [ 32 ], and are attractive for their robustness to model misspecification, efficiency, flexibility, and the ability to include informative prior information. Despite their potential, at the time of writing we found only two studies led by Li et al. (2013) and Guo et al. (2014) that had developed Bayesian case-crossover models for studies of acute health effects of air pollution [ 32 , 33 ].

According to the US Environmental Protection Agency (EPA), annual average particulate matter <2.5 μm (PM 2.5 ) levels are deemed “safe” when they are below 12 μg/m 3 [ 34 ]. For particulate matter <10 μm (PM 10 ), the average annual safety standard is 50 μg/m 3 [ 34 ]. Many rural areas with low air pollution levels (below these EPA regulatory limits) can have a relatively high burden of asthma ED visits [ 35 – 37 ], but the environmental drivers of asthma ED visits across urban-rural areas are understudied. In addition, few researchers have reconsidered seasonal definitions to focus on the intra-annual periods of relative burden, “asthma seasons”, and the unique ambient environmental drivers of those burdens across urban-rural subpopulations.

Our specific objectives are to 1) detail trends in pediatric asthma ED visits in a large and diverse study population and geography, and subsequently to 2) define asthma seasons to identify ambient air pollutants associated with seasonal burdens. We hypothesize that associations between ambient air pollution and asthma ED visits vary by specific asthma seasons in a low air pollution environment. To address our hypothesis, we estimate associations in South Carolina’s (SC) asthma seasons between EPA criteria air pollutants, weather, and asthma ED visits for children living in South Carolina from 2005 to 2014 using a Bayesian time-stratified case-crossover design.

2.0 Methods

2.1 health outcome.

This research was approved (Pro00068172) by the Medical University of South Carolina institutional review board as a part of the SocioEnvironmental Associations with Asthma Increased Risk (SEA-AIR) study. Consent for study participation was not required as the health outcomes used in this study were obtained as secondary, anonymized administrative health data. The health outcome data consisted of 66,092 ED visits with a primary diagnosis of asthma (International Classification of Disease 9, ICD9, codes 493.XX) among children ages 5–19 years residing in South Carolina from 2005 to 2014. The South Carolina Revenue and Fiscal Affairs (SCRFA) office linked records from multiple payor sources. To the best of our knowledge, the data have population wide coverage, capturing all pediatric ED visits for asthma in SC during the 10-year study period. Basic demographic information, diagnostic codes, dates of admittance and discharge, and geographic identifiers were included. Records included geographic identifiers of both ZIP codes and census tracts. Records with missing census tract identifiers (>20%) were assigned to a census tract using a novel geographic identifier assignment algorithm [ 38 ]. ED records were assigned exposure and weather estimates of their billing code census tract, respectively.

There are 1,103 census tracts in SC (2010 US Census geography), and 1,085 of them are not water-only (i.e., off the coast) or institutional-only (e.g., correctional facility) [ 39 ]. The children in this study lived in 1,079 of these 1,085 regular census tracts. Census tracts in SC average 70.6 sq km, or approximately half of a 12 by 12 km (144 sq km) grid cell utilized by the EPA Community Multiscale Air Quality (CMAQ) model [ 40 ]. However, census tracts vary widely in size because they are proportional to population density, as evidenced by their range in SC from 0.42 to 819 sq km with a standard deviation of 106.8 among the 1,085 regular census tracts.

2.2 Air pollutant estimates

Estimates of air pollutant exposure were “fused” from a chemical transport model, the CMAQ model, and monitored values at a 12 km resolution for the US by other researchers [ 40 , 41 ]. Daily estimates were available for six EPA criteria pollutants: carbon monoxide (CO), nitrogen oxides (NO x ), ozone (O 3 ), and sulfur dioxide (SO 2 ), PM 2.5 , and PM 10 . We included NO x over nitrogen dioxide (NO 2 ) because it incorporated both nitrogen oxides (NO and NO 2 ) that have been previously linked with asthma [ 42 ]. Furthermore, we calculated the fraction of coarse particulate matter (PM 10-2.5 ) by subtracting the PM 2.5 estimates from the PM 10 estimates, which we included in statistical modeling instead of PM 10 that incorporates fine particulates as well.

Using ArcGIS (Environmental Systems Research Institute, Redlands, CA), the national daily gridded air pollutant estimates were first clipped to a grid encompassing SC and a surrounding 12 km buffer extending into neighboring North Carolina (NC) and Georgia (GA) to leverage nearby grid points across state lines. The daily gridded estimates were then spatially interpolated to population weighted census tracts in SC using inverse distance weighting (IDW). While estimates interpolated from a 12 km grid were likely spatially smoothed by this procedure, IDW estimates have previously been employed for the purpose of capturing temporal (i.e., daily) variation [ 43 ], which is consistent with the objectives of our acute health effects study design.

2.3 Weather estimates

Air temperature and dewpoint temperature data were obtained from the PRISM Climate Group in the form of daily national smooth surfaces [ 44 ]. The daily national surfaces were first clipped to the spatial extent or SC, and we then calculated the daily spatial average for each census tract by block kriging. Temporal trends in census tract air pollution and weather estimates were assessed using box and whisker plots by season over time. Seasonal correlation patterns among air pollutants were assessed by collapsing air pollution and weather estimates over the entire study period by day of the year, sub-setting by season, calculating Spearman correlations for estimate rankings, and visualizing in the form of heat maps.

2.4 Case and referent window selection

We used a time-stratified case-crossover design. Time-stratification of case events by year, month, and day of the week helps control for short-term, seasonal, and long-term temporal trends [ 32 , 45 ]. Relative to other strategies, time-stratification has been shown to be the least biased referent selection strategy in case-crossover models because of lower time trend and “overlap” biases, respectively [ 46 , 47 ]. We calculated a 3-day moving average (3DMA) over lag days 0 (day of), 1 (1 day prior), and 2 (2 days prior) for each of the pollutant exposures and the weather variables for every ED visit (i.e., case), respectively [ 8 , 48 , 49 ]. These 3DMAs represented the case windows. 3DMA referent windows were created by matching to each 3DMA case window on year, month, and day of the week, per the time-stratified design. If a 3DMA case window had more than three separate 3DMA referent window matches available (depending on the count of a particular day of week within a given month), we randomly sampled three 3DMA referent windows from those available. For example, the 3DMAs for respective pollutants and weather factors at the respective patient’s billing address for a hypothetical pediatric asthma ED visit on the second Tuesday of February in 2010 (case window) would be matched to three of the 3DMAs for the remaining Tuesdays in February of 2010 (referent windows). The 3DMA referent windows provided contrasts to the 3DMA case windows for times when each child was not admitted to the ED for asthma, respectively.

2.5 Asthma seasons

To differentiate local asthma seasons from each other and from other commonly-used definitions (e.g., astronomical seasons), we graphed individual ED visits over time. To further elucidate intra-annual patterns indicative of relative ED visit burdens, we collapsed all ED visits over 2005–2014 ED by day of the year into a single graph. Visual patterns in the graphs were used to identify short-term and long-term trends, and seasonal means were also calculated.

2.6 Statistical analysis

2.7 Model fitting and building

We fit models in the NIMBLE package in R [ 52 – 54 ]. NIMBLE conducts Markov chain Monte Carlo (MCMC) sampling by recompiling models written in Bayesian Using Gibbs Sampling (BUGS) language (e.g., WinBUGS) into C++ language, which greatly increases computational efficiency and stability. We fit models on all the data and season subsets, using the timing of the case day to determine the season. We removed variables that induced variable inflation due to high collinearity during model fitting. For instance, in all models CO was removed because it was highly collinear with NO x , likely because their main respective sources are both fuel combustion emissions [ 55 ]. In addition, CO estimates were calibrated using only one monitor in all of SC and was prone to temporal gaps [ 56 ], which may have introduced more error in its estimates relative to those of other pollutants that were monitored more comprehensively. In the overall, summer, and fall models, dewpoint temperature was removed due to its high collinearity with temperature, respectively. In the summer model, SO 2 was additionally removed because it was highly collinear with NO x . We reported exponentiated β coefficients [ 57 ] in interquartile ranges (IQR) overall and by season, and interpreted them as the odds of an IQR increase/decrease either overall or by season (OR IQR ), respectively. Analytic data and code to reproduce the analyses using non-identifiable data are available at https://github.com/mbozigar/asthma-seasons .

2.8 Sensitivity analyses

We assessed numerous ways to control for potentially unmeasured confounding factors via random effects. The inclusion of spatial (structured and unstructured) and individual (i.e., linking records over time) factors did not improve model fit by deviance information criterion (DIC). Defining seasons and strategies to control for seasonal trends in the overall model was challenging. We assessed multiple temporal cut points, indicators, non-linear effects, prior distributions, and random effects. Many referent day strategies, including unidirectional asymmetric, bidirectional symmetric, and time-stratified designs were assessed. There were minor differences across the strategies, but main findings were consistent. Results were similar for PM 10 and PM 10-2.5 . We opted for the latter in statistical modeling to better differentiate it from PM 2.5 and to capture changes in the PM 10-2.5 fraction over time and space from differential local sources and contexts that have previously been found to occur in a similar geographic area [ 58 ].

3.0 Results

3.1 descriptive results.

Fig 1 shows the study area of SC, its regions, main urban areas, 12 km CMAQ grid point locations, and population weighted census tract centroids. Regions in SC generally delineate unique geologic and geographic features. The Lowcountry encompasses the low, coastal plain, while the Midlands region is characterized by a somewhat sandy, hilly landscape conducive to agriculture, and the Upstate consists of foothills of the Blue Ridge Mountains.

• PPT PowerPoint slide
• PNG larger image
• TIFF original image

https://doi.org/10.1371/journal.pone.0260264.g001

Table 1 shows the asthma ED visit and SC populations for children ages 5–19 years. Contrasted with the SC pediatric population of the same age range, the asthma ED visit population was more male (58.5 to 51.0%), younger (46.9 to 32.1%), African American (68.0 to 28.4%), and lived more in the Midlands region (52.4 to 49.9%) than other regions. The ED visit population was also predominantly on public insurance such as Medicaid (58.3%) and visited an ED in the fall (47.5%). For age-sex subgroups in the ED visit population, there were 20,034 males and 10,989 females ages 5–9 years, 11,494 males and 7,547 females ages 10–14 years, and 7,133 males and 8,895 females ages 15–19 (tabular/graphical results not shown).

https://doi.org/10.1371/journal.pone.0260264.t001

Fig 2 showed that ED visits may not mirror commonly used seasonal definitions, such as astronomical seasons. SC’s asthma seasons tended to start and end earlier, with the exception of winter, and were thus misaligned with astronomical seasons. We defined a medium burden (16.5 visits/day/year) winter asthma season from January 1 st to the end of February that paralleled the much cooler mid-winter months when children were in school. The relative increase in ED visits indicative of medium-high burden (18.9 visits/day/year) from March 1 st through May 31 st , mirroring rising, albeit fluctuating, temperatures and highly variable conditions typical of the spring allergy season when children were still in school, helped us define the SC spring asthma season. The summer asthma season was characterized by low burden (9.4 visits/day/year) and an earlier June 1 st start and August 19 th end when children were generally not in school. Exhibiting the largest seasonal disparity (23.5 visits/day/year), the fall asthma season was defined as starting on August 20 th , the approximate beginning of the school year in SC. This period begins with warm temperatures and then cools during a secondary allergy season ending December 31 st .

Seasonal mean daily ED visits were 16.5 visits/day/year in winter (medium burden), 18.9 visits/day/year in spring (medium-high burden), 9.4 visits/day/year in summer (low burden), and 23.5 visits/day/year in fall (high burden).

https://doi.org/10.1371/journal.pone.0260264.g002

The top center panel of Fig 3 shows the repeating seasonal pattern and a slowly increasing annual average of daily ED visits for asthma. In the remaining panels, large disparities in ED visits were observed for males, the youngest children (ages 5–9 years), African Americans, children on public insurance, children living in the Midlands, and seasonally in the fall asthma season. The fall asthma season had the greatest comparative burden, including a spike effect during the “back-to-school” period at the end of August and beginning of September. For most groups, disparities appeared to be increasing over the study period.

https://doi.org/10.1371/journal.pone.0260264.g003

Except for O 3 and potentially PM 10-2.5 , air pollutant levels seemed to decrease over the study period as shown in Fig 4 . Each pollutant exhibited variation across seasons, with NO x , O 3 and, to a lesser extent the PM measures, having the greatest between asthma season variation. Many correlations between the pollutants differed by asthma season for all years aggregated ( Fig 5 ). For example, O 3 was positively correlated with the NO x in the spring and summer asthma seasons, but they were negatively correlated in fall and winter asthma seasons.

https://doi.org/10.1371/journal.pone.0260264.g004

https://doi.org/10.1371/journal.pone.0260264.g005

3.2 Statistical results

Results from the overall model of all SC ED visits for asthma among children showed increased odds of an ED visit from elevated levels of NO x (OR IQR : 1.018, 95% CI: 1.002, 1.032) and PM 10-2.5 (OR IQR : 1.054, 95% CI: 1.026, 1.063), controlling for other time-varying and invariant factors ( Table 2 ). Results changed by asthma season, as air pollutants were not significantly associated with increased ED visit odds within the medium-burden winter and medium-high-burden spring asthma seasons. However, recognizing a small magnitude (often an order of magnitude lower than either’s independent association, respectively; results not shown) but statistically significant interaction between temperature and dewpoint temperature in each model, increased temperature was associated with asthma ED visits overall and in the spring and fall asthma seasons, respectively. Lower temperatures were significantly associated with increased ED visits in the low-burden summer asthma season. In the summer asthma season, there were statistically significant associations with two air pollutants: a negative association for NO x (OR IQR : 0.954, 95% CI: 0.915, 0.991) and a positive association for PM 2.5 (OR IQR : 1.162, 95% CI: 1.105, 1.222). In the high-burden fall, we found statistically significant associations with asthma ED visits that were positive for NO x (OR IQR : 1.034, 95% CI: 1.009, 1.060), negative for PM 2.5 (OR IQR : 0.970, 95% CI: 0.942, 0.998), and positive for PM 10-2.5 (OR IQR : 1.144, 95% CI: 1.114, 1.177). Thus, the magnitude of the association between asthma ED visits and PM 10-2.5 was thus nearly three times greater in fall relative to an entire year.

https://doi.org/10.1371/journal.pone.0260264.t002

4.0 Discussion

Our first objective was to detail trends in asthma ED visits in SC. The trends we found are indicative of disparities over time and for specific subpopulations. Disparities generally increased for particular groups, including males, young children, African Americans, children on public insurance, and children living in the Midlands in SC from 2005 to 2014 ( Fig 3 ). These results are partially consistent with national trends, as there are large disparities in asthma rates and outcomes across numerous factors including race, urban-rural status, socioeconomic status (SES), and others [ 59 – 61 ]. But the increasing disparities in SC over time are inconsistent with recent evidence that disparities may be plateauing in the US [ 62 ]. Furthermore, asthma ED visit disparities existed for males in the youngest age group (5–9 years) and among females in the oldest age group (15–19 years), which reinforces evidence that puberty and sex hormones likely play a role in asthma differences as children age [ 63 , 64 ]. In addition, the specific subpopulation that visited the ED for asthma was notably different than the SC population of the same age 5-19-year group ( Table 1 ). Sociodemographic and geographic disparities in asthma ED visits mirror broader health disparities in SC [ 65 – 69 ]. Further attention to the drivers of health disparities, including for asthma, are needed in places such as SC that may not follow national patterns.

To address our second objective of detailing the seasonal ambient environmental drivers of asthma ED visits, we introduced the concept of asthma seasons, defined by intra-annual periods of asthma ED visit burden. Our study location spanned both rural and urban areas, and we sought to avoid assuming that ambient air pollutants and weather patterns were necessarily the key seasonal influences. Overall, ED visits seemed to be increasing over time ( Fig 3 ). ED visit patterns in SC seem to have four discernable asthma seasons that are similar to but still distinct from astronomical seasons, with particularly high burdens in the fall and spring asthma seasons (Figs 2 and 3 ). We were especially interested in the environmental drivers of fall and spring ED visits, given the relative disparities.

Asthma seasons should hypothetically differ by location based on geography, atmospheric chemistry, weather, land use, ecology, viral oscillation, and other factors, but not all acute health effects studies of air pollution adequately detail intra-annual temporal patterns of asthma outcomes locally. Those studies that do, for example in Shanghai, found that the intra-annual asthma burden similarly peaks over October, November, and December [ 9 ], but asthma ED visits there exhibit differing patterns in the remaining months of the year when contrasted with SC. Though not a study of environmental drivers, researchers identified a back-to-school peak when detailing asthma-related primary care provider visits by week and by subpopulation groups in Israel [ 70 ]. Back-to-school and other intra-annual patterns in asthma hospitalizations, in addition to exacerbation triggers, were detailed in a study in Texas [ 71 ]. Furthermore, when using SC’s asthma seasons, pollutants similarly varied by season (Figs 4 and 5 ).

We observed statistically significant, positive overall associations between both NO x and PM 10 and asthma ED visits in our overall model ( Table 2 ), even though SC is generally considered a low air pollution state [ 72 ] and area, globally [ 73 ]. Previous studies have linked NO x with overall asthma incidence [ 74 ], ED visits in cool seasons [ 3 ], and repeated visits [ 42 ]. Though we removed CO from statistical models because of its high collinearity with NO x , the main sources of both NO x and CO are assumed to be emissions from fuel combustion [ 55 ]. As such, results suggest overall, season-invariant risks from fuel combustion in SC, even at relatively low overall pollutant levels.

That we found an overall, season-invariant association with PM 10-2.5 and not PM 2.5 in SC ( Table 2 ) is somewhat surprising given the wide literature linking PM 2.5 with asthma [ 4 ]. Neither daily nor long-term exposure to coarse PM were statistically significantly associated with respiratory outcomes among elderly Medicare patients [ 75 , 76 ]. However, long-term exposure to coarse PM was linked to asthma in a nationwide pediatric Medicaid cohort [ 77 ]. Research has found that the commonly used CMAQ model has limitations in predicting ground-level PM 10 from biogenic sources [ 78 , 79 ]; consequently, studies that used CMAQ-only estimates for PM 10 may have mischaracterized relationships with health outcomes. This study addresses the limitations by using estimates from the fusion model that provide more accurate estimates of daily PM 10 [ 41 ], which may explain the improved ability for this analysis to detect health effects for this pollutant, particularly at lower levels.

We hypothesized that associations between asthma ED visits and air pollution varied by asthma season, and we generally found supporting evidence to corroborate the hypothesis. Relative to the overall model, the relationship between PM 10-2.5 and asthma ED visits was stronger in the fall. While SC’s fall asthma season is, on average, neither consistently cool nor warm ( Fig 4 ), others have linked increased PM 10 to asthma ED visits in the cool season in nearby urban Atlanta, but not the warm season [ 8 ]. In urban Shanghai, associations between PM 10 and asthma ED visits were null overall and within both warm and cool seasons, respectively. Given the results from this study, future research identifying the seasonally varying sources and PM 10 components may help elucidate those that are key drivers of seasonal differences in asthma ED visits across urban-rural areas.

A potential seasonal allergenic component of PM 10 could be respirable antigenic particles smaller than 10μm in diameter from larger pollen grains [ 80 , 81 ]. Allergenic plants, such as ragweed, release seasonal pollen during the fall season in SC and other eastern and midwestern states, particularly in more rural and agricultural regions like the SC midlands. Allergenic particle levels tend to increase after rain events [ 21 ]. Agriculture and biomass burning, also common in the Midlands, should also be further studied as potentially important seasonally varying sources of PM 10 , as seasonal variation in PM 10 fractions of elemental and organic carbon have been attributed to seasonal agricultural activity in other contexts [ 82 ]. However, allergens and other airborne irritants differ in mechanistic triggering of asthma exacerbations and may interact in complex ways to affect asthma pathogenesis across individuals [ 2 ], warranting additional research.

Summer had the lowest asthma burden in all of SC’s asthma seasons, yet two statistically significant air pollutant associations were found: a negative association with NO x and a positive association with PM 2.5 . The estimated positive association with PM 2.5 , is a relationship that others have found in warm seasons in different locations [ 3 , 83 ]. NO x , reaches its lowest and least variable levels in SC’s summer asthma season, and it may further be involved in complex interactions beyond the scope of this study. The primary source of NO x is fuel combustion emissions [ 55 ] and NO x is subsequently monitored mainly in SC’s urban areas characterized by high vehicle traffic density [ 56 ]. As such, there is little non-urban monitored data for calibrating modeled NO x estimates in SC’s suburban and rural areas, which may be reflected by the fusion model’s somewhat mediocre NO x performance relative to other pollutants [ 41 ]. Furthermore, air temperature had a relatively strong negative association, which is highly correlated with dewpoint temperature in summer in SC. The association with temperature, and by extension dewpoint temperature, indicates increased ED visit odds from cooler and drier conditions that are common in SC after weather fronts, as is similarly seen following thunderstorms [ 20 ].

By defining seasons relative to the outcome, we found that neither criteria pollutants nor weather were associated with asthma ED visits in a large portion of the year aligned with the winter and spring asthma seasons (January 1 st –May 31 st ). Additional research is needed to tease apart drivers of ED visits in the medium-burden winter and medium-high-burden spring asthma seasons, respectively. However, it is important to contextualize the null associations with pollutants and weather covariates: outdoor ambient factors may simply play less of a role during these asthma seasons.

Common viruses, including influenza, usually peak in the coldest months, usually when people spend more time indoors [ 18 ]. Periods of high viral transmission encompass the winter and well into parts of the fall and spring in many places in the US [ 18 ]. In SC, temperatures are usually mild through fall, with appreciable cold temperatures often beginning only around early January each year, defined in this study as the winter asthma season. Others have found that daily viral transmission, annually peaking during the back-to-school portion of the fall and later in the winter, was the key predictor of asthma hospitalizations among children, and influenza prevalence was the key predictor of the annual winter surge in adult asthma hospitalizations [ 71 ]. We could not formally test the hypothesis that viruses are the main driver of the back-to-school surge in SC because we lacked viral transmission data. However, we saw a large spike each year around the time children went back to school in SC (Figs 2 and 3 ) that is suggestive of viral transmission as a potential contributor. Of additional interest, recent studies of the COVID-19 pandemic have found that asthma ED visit rates during this event were significantly reduced [ 84 ]. Researchers hypothesize that adoption of behaviors to reduce spread of the COVID-19 virus, such as wearing masks, are mechanisms that reduced transmission of many other types of viruses shown to exacerbate asthma [ 84 , 85 ]. Given the mounting evidence of seasonal viral influences on asthma exacerbations, particularly during cold weather periods, future studies of the environmental drivers of seasonal differences in asthma should prioritize inclusion of such data.

Neither PM 2.5 nor PM 10 were significantly associated with ED visits in the spring asthma season. That no associations were found suggests that antigenic pollen particles from common spring-blooming plants such as grass and trees may have a minor or negligible role in asthma ED visits in a state like SC during its spring asthma season. This result contrasts with findings from many studies that have identified positive associations between asthma exacerbation events, such as ED visits, and both grass and tree pollen counts [ 12 , 14 , 16 , 17 , 19 ]. To more robustly assess hypotheses related to pollen in SC, inclusion of spatio-temporally varying pollen counts, species types, and particle size distributions need to be included.

4.1 Limitations

This research did not measure personal exposures, as it relied on interpolated census tract estimates of daily air pollutants and weather. Some air pollutants tend to vary greatly by location and across and within days (i.e., spatio-temporally). This research could not incorporate individual sensitivities to specific allergens, nor other individual asthma case-related information. We were unable to incorporate spatio-temporally resolved pollen counts or estimates, nor similar data for viruses. Time-varying indoor exposures such as environmental tobacco smoke (ETS) usage was not included. Similarly, children are not static in one location daily, but spend time indoors and outdoors at home, school, and many other locations. Consequently, there was potential for exposure misclassification. In addition, we found a few implausible protective associations that were potentially indicative of bias introduced by other complex environmental factors and synergisms, inaccurate estimates, uncontrolled confounders, or other factors. It indicated that even with relatively precise spatio-temporal resolution exposure estimates, teasing out independent effects of single environmental factors among many in a low-ambient air pollution using administrative health data setting remains highly challenging. It makes a case for incorporating environmental mixtures as opposed to single pollutant study designs. Finally, though we relied on 3-day moving averages guided by previous research findings, other lag structures may indicate significant associations between the same pollutants we identified by season, or different combinations.

4.2 Conclusion

We identified increasing asthma disparities across several socio-demographic factors in SC from 2005–2014, a departure from the plateauing national trend. We uniquely outlined the concept of asthma seasons that were defined by local intra-annual periods of relative burden. From such a perspective, the fall asthma season was the most burdensome for ED visits among children, followed by the spring and winter asthma seasons. The summer asthma season was the least burdensome. Results from Bayesian case-crossover analyses supported our hypothesis that associations between air pollution, weather, and asthma ED visits varied by asthma season. Across urban-rural areas characterized by generally low air pollution levels, there were acute health effects associated with NO x , particulate matter, and weather. But, associations differed by asthma season and PM size. With discretion, these results may be somewhat generalizable to geographically, demographically, and climatically-similar southern states. Our Bayesian methodology is reproducible for any location and can be tailored to any spatio-temporally-varying exposures for identifying and elucidating the acute health effects of local environmental exposures.

Acknowledgments

This research used air pollution estimates developed in a collaboration between Emory University and the Georgia Institute of Technology. Dr. Armistead (Ted) Russell and Nirupama Senthilkumar, were integral in facilitating our data access. We also acknowledge the South Carolina Revenue and Fiscal Affairs (SCRFA) office that provided high spatio-temporal resolution asthma ED visit outcome data.

• View Article
• PubMed/NCBI
• Google Scholar
• 31. Kedem B, Konstantinos F. Regression Models for Time Series Analysis. Hoboken, NJ: John Wiley & Sons, Inc.; 2002.
• 39. US Census Bureau. South Carolina. 2010.
• 44. PRISM Climate Group. PRISM Climate Group. 2004 [cited 1 Jul 2017]. Available: http://www.prism.oregonstate.edu/ .
• 50. Lawson AB. Bayesian disease mapping: Hierarchical modeling in spatial epidemiology. 3rd ed. Bayesian Disease Mapping: Hierarchical Modeling in Spatial Epidemiology. 2008.
• 52. R Core Team. R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing; 2019.
• 57. Hosmer DW Jr., Lemeshow S, Sturdivant RX. Applied logistic regression. Third Edit. Hoboken, NJ: John Wiley & Sons, Inc.; 2013.
• 72. American Lung Association. Report card: South Carolina. [cited 11 Nov 2019]. Available: https://www.lung.org/our-initiatives/healthy-air/sota/city-rankings/states/south-carolina/ .
• 73. World Health Organization. Ambient air pollution: a global assessment of exposure and burden of disease. Geneva, Switzerland; 2016.