Weather variables as important clinical predictors of bacterial diarrhoea among international travellers

Abstract

Background

Clinicians and travellers often have limited tools to differentiate bacterial from non-bacterial causes of travellers’ diarrhoea (TD). Development of a clinical prediction rule assessing the aetiology of TD may help identify episodes of bacterial diarrhoea and limit inappropriate antibiotic use. We aimed to identify predictors of bacterial diarrhoea among clinical, demographic and weather variables, as well as to develop and cross-validate a parsimonious predictive model.

Methods

We collected de-identified clinical data from 457 international travellers with acute diarrhoea presenting to two healthcare centres in Nepal and Thailand. We used conventional microbiologic and multiplex molecular methods to identify diarrheal aetiology from stool samples. We used random forest and logistic regression to determine predictors of bacterial diarrhoea.

Results

We identified 195 cases of bacterial aetiology, 63 viral, 125 mixed pathogens, 6 protozoal/parasite and 68 cases without a detected pathogen. Random forest regression indicated that the strongest predictors of bacterial over viral or non-detected aetiologies were average location-specific environmental temperature and red blood cell on stool microscopy. In 5-fold cross-validation, the parsimonious model with the highest discriminative performance had an area under the receiver operator curve of 0.73 using 3 variables with calibration intercept −0.01 (standard deviation, SD 0.31) and slope 0.95 (SD 0.36).

Conclusions

We identified environmental temperature, a location-specific parameter, as an important predictor of bacterial TD, among traditional patient-specific parameters predictive of aetiology. Future work includes further validation and the development of a clinical decision-support tool to inform appropriate use of antibiotics in TD.

Introduction

Travellers’ diarrhoea (TD) is a common illness affecting ~20% of international travellers and contributing to one-third of post-travel medical visits.¹ Aetiologies include bacterial, viral, parasitic or mixed pathogens and can be influenced by external factors such as season,² location or activity. In the context of international travel, laboratory diagnosis of TD may be costly, inconvenient or unavailable. Current joint guidelines from the International Society of Travel Medicine and the US Centers for Disease Control recommend providing travellers with standby antibiotics for self-treatment in the event of moderate to severe diarrheal symptoms.³ Such treatment is currently guided by severity of illness based on functional impact, with limited consideration of aetiology, resulting in potential for inappropriate use of antibiotics for severe viral or protozoal diarrheal infections.⁴ Even among bacterial disease, treatment varies depending on bacterial aetiology. Inappropriate use of antibiotics can lead to side effects, development of resistance⁵ and the potential for other infections including Clostridium difficile.⁶ Clinical decision support tools incorporating clinical prediction rules provide an attractive alternative to microbial testing for identification of cases of bacterial TD when laboratory availability or access is limited.

A clinical prediction rule (CPR) is a tool predicting the probability of a specific outcome. Recent advances in machine learning methods such as random forests (RF) offer new tools for developing clinical prediction models.^7–9 Prior studies using CPRs to predict bacterial diarrhoea among children,^8,10–12 and travellers,¹³ have demonstrated promising results but have been limited by small sample sizes, sub-optimal performance characteristics of predictors and limited pathogen identification. Development of new diagnostic tools such as the TaqMan® array card (TAC) has recently improved identification of diarrheal pathogens.¹⁴ In this study, we incorporated machine learning methods with traditional statistical methods to develop and cross-validate a predictive model with the ultimate goal of developing a parsimonious CPR to assess the aetiology of TD from primarily clinical and epidemiological parameters.

Methods

Ethical approval

The institutional review boards of the University of Utah, the Walter Reed Army Institute of Research, CIWEC Hospital, the Nepal Health Research Council (NHRC) and Bumrungrad International Hospital, approved this study.

Study subjects

We collected de-identified data from studies of international travellers seeking care for acute diarrhoea at two single-hospital sites located in Nepal and Thailand. The Bumrungrad International Hospital in Bangkok, Thailand enrolled 173 travellers from February 2012 to December 2014 as part of a study comparing TD diagnostic tests.¹⁴ The CIWEC Travel Medicine Center in Kathmandu, Nepal enrolled 284 travellers from March 2016 to May 2017. We used clinic presentation as a surrogate for moderate-to-severe diarrheal symptoms.

Among both populations, inclusion criteria consisted of adult travellers ≥ 18 years old who were citizens of countries in North America, Europe, Australia, New Zealand, Japan, Taiwan or South Korea and who presented seeking healthcare for acute diarrhoea or gastroenteritis.¹⁴ We included all subjects who met definitions of acute diarrhoea or acute gastroenteritis. Acute diarrhoea was defined as ≥3 loose/liquid stools in the preceding 24 h OR ≥ 2 loose/liquid stools in the preceding 24 h PLUS at least ≥2 associated GI symptoms, including subjective fever/chills, nausea, vomiting, abdominal cramping, abdominal pain, tenesmus, bloating, faecal urgency or gross blood in stool. Acute gastroenteritis was defined as ≥3 loose/liquid stools in the preceding 24 h with ≥1 additional GI symptoms as above OR ≥ 3 vomiting episodes in the preceding 24 h with ≥1 additional GI symptoms OR vomiting (≥2) OR diarrhoea (≥2) within the preceding 24 h with ≥2 additional GI symptoms.^1,3 Travellers were excluded if they were not citizens of the countries listed above or if they had resided outside of their country of citizenship for greater than 1 year. Additional exclusion criteria included symptoms of diarrhoea present for greater than 7 days, an inability to provide a stool sample, or an inability to provide written informed consent.¹⁴

Data collection

After obtaining written informed consent, travellers provided a stool sample and completed a questionnaire covering demographics, clinical history and associated symptoms (Table 1). Stool samples were sent to the laboratory for stool microscopy as well as additional diagnostic testing detailed below. Stool microscopy variables included the presence or absence of red blood cells (RBCs), white blood cells and/or stool mucus.

Table 1.

Demographics, symptoms and stool microscopy results collected from travellers to Nepal and Thailand

Variable	Nepal (% of n = 284)	Thailand (% of n = 169)
Demographic variables:
Male	123 (43)	106 (62)
Median age (IQR)	26 (10)	32 (5)
North American nationality	78 (27)	32 (19)
European nationality	158 (56)	71 (42)
Australia/New Zealand nationality	38 (13)	26 (15)
Asian nationality	10 (3.5)	40 (24)
Symptom variables:
Diarrhea present	280 (99)	169 (100)
Median diarrhoea duration in hours (IQR)	48 (59)	24 (52)
Median total number of diarrheal stools (IQR)	17 (17)	8 (5)
Vomiting	165 (58)	87 (51)
Abdominal cramping/pain	244 (86)	145 (86)
Malaise/fatigue	267 (94)	149 (88)
Febrile	40 (14)	103 (61)
Bloody stool	12 (4.2)	19 (11)
Prior treatment	87 (31)	91 (54)
Stool grade: formed	2 (0.7)	0 (0)
Stool grade: soft	14 (4.9)	12 (7.1)
Stool grade: loose	62 (22)	81 (48)
Stool grade: watery	206 (73)	76 (45)
Weather variables:
Median site temperature °F (30-day average) (IQR)	70.3 (12)	84.2 (3.2)
Median site precipitation (30-day average) in inches (IQR)	0 (0.09)	0 (0.14)
Stool microscopy variables:
Stool microscopy: Mucus	215 (76)	73 (43)
Stool microscopy: WBC	227 (80)	135 (80)
Stool microscopy: RBC	52 (18)	56 (33)
Stool microscopy: Parasite	45 (16)	1 (0.6)

To identify diarrheal pathogens, we relied on multiple diagnostic techniques including stool culture, real-time polymerase chain reaction (PCR) and enzyme-linked immunosorbent assay (ELISA).¹⁴ Using standard laboratory culture techniques, stool culture was used to isolate Aeromonas, Arcobacter, Campylobacter, Escherichia coli (E. coli), Plesiomonas, Salmonella, Shigella or Vibrio species (Table 3). Escherichia coli strains were identified based on stool culture followed by PCR detecting targets for enterotoxigenic (ETEC), enteropathogenic (EPEC), shiga toxin-producing (STEC) or enteroaggregative (EAEC). Reverse transcription PCR identified norovirus and sapovirus. We identified parasitic infections such as Cryptosporidium, Entamoeba histolytica and Giardia using ELISA kits.

Table 3.

Number of detected pathogens among all patient cases

Pathogen	Number detected
Diarrheagenic E. coli	248
ETEC	85
EAEC	73
EPEC	72
Misc. E coli (EIEC, or STEC)	18
Norovirus	144
Campylobacter	114
No pathogens	68
Sapovirus	47
Aeromonas	42
Plesiomonas	40
Salmonella	36
Shigella	34
Rotavirus	15
Vibrio	13
Bacteroides	10
Miscellaneous parasitic	14
Miscellaneous bacteria	10
Miscellaneous viruses	7

In addition to conventional diagnostic methods, data collected from Bumrungrad International Hospital in Thailand included organisms detected on an enteric pathogen panel TAC. TAC testing detects multiple enteropathogens using real-time PCR.^15,16 We used the QiaAmp Fast Stool DNA kit to extract total nucleic acid and used reagents from the Ag-Path-ID One-Step RT-PCR kit. Our TAC panel detects the following 12 species of bacteria: Aeromonas, Bacteroides fragilis, Campylobacter coli/C. jejuni, C. difficile, enteroaggregative E. coli, enteropathogenic E. coli, enterotoxigenic E. coli, Helicobacter pylori, Salmonella, Shigella/enteroinvasive E. coli, Shiga toxin–producing E. coli and Vibrio cholerae. It also detects the following five viruses (adenovirus, astrovirus, norovirus GI/GII, rotavirus and sapovirus), five species of nematodes (Ancylostoma duodenale, Ascaris lumbricoides, Necator americanus, Strongyloides stercoralis and Trichuris trichiuria), five protozoan parasites (Cryptosporidium, Cyclospora cayetanensis, E. histolytica, Giardia lamblia and Cystoisospora belli) and two species of fungi (Encephalitozoon intestinalis and Enterocytozoon bieneusi).¹⁴ We identified pathogen positivity using a positive threshold cycle (Ct) value cutoff of < 35; a negative result consistent of a Ct value >  35. Further details regarding TAC analysis and comparison to conventional methods were described in prior reports.^14,17

In order to investigate the effect of seasonality and local weather patterns on TD aetiology, we constructed variables of temperature and precipitation. We obtained date-specific site temperature and precipitation data from National Oceanic and Atmospheric Administration (NOAA) sites in Bangkok and Kathmandu. To create a temperature variable, we averaged the temperature for the 14 days prior to patient enrolment. The same process was used for the precipitation variable.

Statistical analysis and modelling

In the construction of our models, we used binary variables with the exception of continuous variables age, diarrhoea duration, number of diarrheal episodes, average temperature and average precipitation. Variables with >40% missingness were removed and multiple imputation¹⁸ was used to account for any additional missing data. Four patient cases from the Thai dataset were excluded from analysis due to >40% data missingness leaving a total of 453 cases for analysis. For each case, we categorized the aetiology of diarrhoea as bacterial-only infection, mixed infection with both bacterial and non-bacterial organisms, viral-only infection, protozoal/parasite-only infection or infection with no identified organism.

All analyses were completed using R,¹⁹ and model development/validation was completed in accordance to the TRIPOD checklist (Table S1, Supplementary data are available at JTM online).²⁰ Using the combined dataset from both sites, we used RF regression consisting of 1000 decision trees and the default number of variables considered at each split (p/3, where p is the number of predictors considered) to determine variable importance. This method uses multiple decision trees to determine and calculate the reduction in mean squared prediction error for each variable. We arranged the variables in order of descending importance according to the reduction in mean squared prediction error. As a secondary analysis, to examine the top predictors of each site, given the smaller sample sizes, we used univariable methods (chi-square or Fisher’s exact tests for categorical variables and Wilcoxon rank-sum tests for continuous variables) to determine the relationship between variables and bacterial-only cases at each site.

To assess predictive performance, we used 5-fold repeated cross-validation with 100 iterations using both logistic regression (LR) and RF models. In each iteration, we randomly divided data combined from all sites into an 80% training set and a 20% testing set. We calculated the area under the receiver operator curve (AUROC or AUC) for both the LR and the RF models using a range of variables from 1 to 35 in varying increments. We fit models with varying input parameter sets and outcomes to determine the best discriminative performance based on the AUC. For each cross-validation test set, we estimated the calibration intercept and slope, respectively, by first fitting a LR with the predicted value of each test observation as an offset in an intercept-only model, and second by fitting a LR model with the predicted value of each test observation as the regressor.

Results

We collected demographic, symptom and stool microscopy data from the single-hospital site in Thailand from Feb 2012 to Dec 2014 and from the single-hospital site in Nepal from March 2016 to May 2017 (Table 1). Among the 457 cases, we detected the following diarrheal pathogens; 195 bacterial cases, 125 cases with mixed pathogens, 68 cases without detected pathogen, 63 viral cases and 6 cases of protozoal disease (Table 2). These aetiologies are further broken down into number per month of data collection in Thailand and Nepal (Figures S1 and S2, respectively, Supplementary data are available at JTM online), and compared with  30-day   moving   averages   of   environmental   temperature

Table 2.

Stool analysis results showing categories of infection types and number of infections per site in each category

Infection category	Nepal (% of n = 284)	Thailand (% of n = 169)
Bacterial only	99 (35)	93 (55)
Viral only	50 (18)	11 (6.5)
Protozoa/parasite only	5 (1.8)	0 (0)
Mixed infection	75 (26)	52 (31)
No organism identified	55 (19)	13 (7.7)

and precipitation (Figures S3 and S4, Supplementary data are was diarrhoeagenic E. coli (248 organisms detected among 187 patients) followed by norovirus (144 patients) and Campylobacter (114 patients) (Table 3).

We analysed multiple model variations to identify predictors of bacterial TD. Using the same variable selection process, variations comparing bacterial-only aetiologies to viral plus non-detected infections outperformed models using total bacterial cases (bacterial-only infections plus mixed infections with bacterial and non-bacterial pathogens) in cross-validation. Given our goal of identifying the strongest predictors of bacterial diarrhoea as well as the challenge of identifying the predominant pathogen among mixed infections, we subsequently excluded cases of mixed infections from analysis. Figure S6 (Supplementary data are available at JTM online) includes the results of analysis including mixed infections.

Using RF regression, we identified predictors of bacterial diarrhoea and ranked them from strongest to weakest predictors based off of their reduction in mean squared prediction error (Table 4). A lower average environmental temperature [odds ratio (OR) 0.99, 95% confidence interval (CI) 0.98–1.0], and RBC on stool microscopy (OR 1.27, CI 1.13–1.43) were the top predictors of bacterial aetiology. When separated by site, using univariable analysis, we found that the variables with the most evidence of an association with bacterial aetiology at the Nepal site included RBCs on stool microscopy, absence of fever and a lower average site temperature (Table S2, Supplementary data are available at JTM online), and at the Thai site included the presence of mucus on stool microscopy, an absence of vomiting and the presence of RBC on stool microscopy (Table S3, Supplementary data are available at JTM online).

Table 4.

The top 10 predictive variables based on RF ranked in descending order based on mean squared prediction error (IncMSE)

	Variable	Mean square-error (IncMSE)	Odds ratio	P value
1	Average environmental temperature	22.39	0.99 (0.98–1.00)	0.001
2	Stool study: RBC	16.19	1.27 (1.13–1.43)	<0.001
3	Total number of diarrheal stools at time of evaluation	6.66	1 (0.99–1.00)	0.11
4	Nationality: North American	3.87	0.95 (0.83–1.10)	0.505
5	Age	3.24	1.01 (1.00–1.01)	0.009
6	Average environmental precipitation	2.64	1.18 (0.80–1.74)	0.406
7	Nationality: European	1.91	1.12 (0.99–1.27)	0.078
8	Stool study: WBC	1.84	1.06 (0.91–1.23)	0.444
9	Stool grade: loose	1.45	1.04 (0.94–1.16)	0.448
10	Stool study: mucus	0.41	1.02 (0.90–1.14)	0.783

Note: Each variable is shown with its associated IncMSE, OR and P value. Variables with an OR > 1 are predictive of bacterial aetiology and variables with an OR < 1 are predictive of non-bacterial aetiology. Refer to Figure S7 (Supplementary data are available at JTM online) for partial dependency plots fit to characterize the marginal impacts and relationships between the top 9 most important variables and the probability of bacterial diarrhoea aetiology.

In order to identify models that would balance a high AUC with ease of use, we selected the three model variations below. Our first model included demographic and clinical data such as symptoms, stool appearance and stool grade. We initially excluded weather and stool microscopy data as these data may be challenging for travellers to obtain. Weather data were included in our second model iteration and stool microscopy included in the third given that stool studies represent the most challenging variables for a traveller to obtain. Our final three model variations included:

• (i) Model 1: patient symptoms and demographic variables only, no weather or stool microscopy variables;

• (ii) Model 2: symptoms, demographics and weather variables without stool microscopy;

• (iii) Model 3: all variables including weather and stool microscopy.

In 5-fold cross-validation, we found that the inclusion of weather and microscopy variables yielded an increase in the maximum discriminative ability among all models. Maximum discriminative ability with only symptoms and history (Model 1) occurred at 8 variables with an AUC of 0.0.63 using LR [calibration intercept −0.01 (standard deviation, SD 0.30) and slope 0.57 (SD 0.44), Figure 1A]. Maximum discriminative ability with inclusion of weather data (Model 2) occurred at 1 variable with an AUC of 0.70 using LR [calibration intercept −0.05 (SD 0.30) and slope 1.21 (SD0.65), Figure 1B]. The inclusion of both weather and stool microscopy variables (Model 3) demonstrated an AUC of 0.73 at 3 variables using LR [calibration intercept −0.01 (SD 0.31) and slope 0.95 (SD 0.36), Figure 1C]. Overall, Model 3 demonstrated the highest AUC and best calibration, with logistic methods demonstrating a higher AUC compared to RF methods. In sensitivity analysis, a prediction model discriminating Shigella/Campylobacter infections from non-bacterial infections resulted in a maximum AUC of 0.75 (Figure S5, Supplementary data are available at JTM online).

Figure 1.

Area under the ROC curve (AUC) for bacterial aetiology obtained at a range of variables using LR and RF methods. Refer to Table 4 for a list of the top 10 predictive variables ranked in descending order. (A) Model 1 includes demographic and symptom variables and excludes weather and stool microscopy variables. Maximum predictive accuracy in Model 1 occurred at 9 variables using LR [AUC 0.70, calibration intercept 0.2 (SD 0.32) and slope 0.74 (SD 0.39)]. (B) Model 2 includes demographic, symptom and weather variables while excluding stool microscopy variables. Maximum predictive accuracy in Model 2 occurred at 2 variables using LR [AUC 0.71, calibration intercept 0.04 (SD 0.29) and slope 1.05 (SD 0.57)]. (C) Model 3 includes all demographic, symptom, weather and stool microscopy variables. Maximum predictive accuracy in Model 3 occurred at 3 variables using LR [AUC 0.74, calibration intercept 0 (SD 0.31) and slope 1 (SD 0.36)].

Discussion

Diarrhea is the most common health complaint among international travellers, and one that occurs more frequently in travellers to low and middle-income countries (LMICs).^1,21 Current guidelines recommend traveller-initiated self-treatment of moderate to severe diarrhoea, with limited consideration of aetiology; thus, tools differentiating bacterial from non-bacterial aetiologies can provide practical support for antibiotic use decision-making. In this study, we used data from two studies of traveller’s diarrhoea to derive clinical predictors of bacterial aetiology.

Our model identified the top predictors of bacterial diarrhoea to be a lower location-specific environmental temperature and the presence of RBCs on stool microscopy. A recent study of 1450 patients evaluated for TD in the UK found male gender, younger age and WBC on microscopy to be predictive of bacterial disease,²¹ whereas a non-traveller-specific study found associations of bacterial diarrhoea with bloody stools, fevers, lack of vomiting and longer diarrheal duration.²² Neither of these studies considered weather variables among potential predictors, and our demonstration of the importance of a location-specific variable such as environmental temperature, suggests that this should be included in future predictive modelling efforts.

We found that a 14-day average of environmental temperature and precipitation increased model performance. In our two-site study, we found that lower environmental temperature predicted bacterial disease (Table 4), suggesting location-specific variation. A study with more site locations may show location-specific variation depending on local seasonality. We also found an association between high precipitation and bacterial TD; although this association was nonsignificant, removal of this variable decreased model accuracy.

Although prior studies have documented complex seasonal patterns among diarrheal diseases in tropical climates,^2,8,23 studies examining weather data for individual-level clinical prediction are lacking. Climate is thought to aid pathogen transmission through mechanisms such as contamination of food or water supplies, pathogen survival on fomite surfaces or facilitating vector life cycles. Among data collected for the Global Enteric Multicenter Study (GEMS), bacterial pathogens demonstrated higher prevalence during the hot and/or rainy seasons,² whereas viral aetiologies demonstrated higher prevalence during the dry/cold season.^8,12 Although these multi-site studies found associations between seasonality and pathogen prevalence, these findings may not be generalizable globally. Our findings suggest that site-specific weather variables aid in site-specific models to predict diarrheal aetiology. For translation of prediction models into decision support tools, the most recent daily precipitation and temperature could be gathered from online weather sources, based on smartphone-based detection of GPS location. Our analysis is predictive and not intended for causal inference to inform risk mitigation.

We showed that top predictors differed by site, which is not surprising and may suggest a need for location-specific data in generating clinical prediction models. Prior analyses of data from the GeoSentinel travel clinics demonstrate the significant role of travel location on the risk of TD development.^24,25 Overall, models including weather variables (Model 2) or both weather and stool microscopy variables (Model 3) demonstrated higher discriminatory performance based on cross-validation compared to models with only symptomatic and demographic variables (Model 1). When included, RBCs on stool microscopy appear as the second most important predictive variable. Blood in stool is known to be a specific yet insensitive predictor of bacterial diarrhoea.^26–28 Unfortunately, not all travellers have access to stool microscopy testing, though the use of a point of care stool RBC diagnostic such as used for colon cancer screening deserves further consideration. Although stool RBCs may improve model accuracy, it is unclear whether the gain in performance (increase in AUC of 0.03) is worth the resources needed for stool microscopy.

Our study has several limitations. First, we acknowledge that our inclusion of only two hospital sites in Asia limit interpretation and generalizability. Exclusion of mixed-organism infections further reduces generalizability and led to the exclusion of 127 cases. We acknowledge that a focus on patients self-referred for care skews our analysis towards illness of higher severity, and that our findings’ applications may be limited to travellers who present for care at a healthcare facility. Second, not all bacterial pathogens included in analysis would necessarily require antibiotics, though studies are lacking regarding the aetiologies that would benefit most. Although infection with Salmonella, Shigella and Campylobacter spp. are thought to benefit from antibiotics,²⁹ there is also evidence from randomized controlled trials in travellers that treatment of diarrheagenic E. coli reduces symptom duration by up to 2 days. Even among diarrheagenic E. coli, certain strains such as EAEC or EPEC may be less clinically meaningful. Future models may further differentiate between treatable and non-treatable pathogens, particularly following improvements in diagnostic testing. Third, our use of a set cycle threshold (<35) for determination of causative pathogen may be overly sensitive and contributed to a higher number of mixed-organism infections. Although detection of molecular targets by PCR does not necessarily imply pathogenicity, it may be helpful among a patient population such as ours, a proportion of which received antibiotics before obtaining stool workup. We acknowledge that pre-treatment with antibiotics may have affected our ability to interpret stool cultures in the absence of molecular testing. Lastly, our results were only cross-validated and future work is required to externally validate and implement a CPR based on the predictors identified in this study.

We recognize that differing climates may have varying effects on pathogens and patterns of disease. In the future, region-specific models incorporating local weather may provide the most accurate prediction of site-specific diarrheal aetiology. Nevertheless, we demonstrate that location-specific weather variables influence diarrheal aetiology and should be considered in future work on clinical prediction and decision support tools for TD.

Authors’ contributions

LB, DTL designed the study. PP, SKS, SA, SD, SS, JAP, LB were involved in data collection. MAP, TS, BJB, JAP, DTL were involved in analysis. MAP, TS, BJB, PP, SKS, SA, SD, SS, JAP, LB, DTL undertook the responsibility of data interpretation.

Disclaimer

Material has been reviewed by the Walter Reed Army Institute of Research (WRAIR). There is no objection to its presentation and/or publication. The opinions or assertions contained herein are the private views of the author, and are not to be construed as official, or as reflecting true views of the Department of the Army or the Department of Defense. The investigators have adhered to the policies for protection of human subjects as prescribed in AR 70–25.

Conflicts of interest

The authors have declared no conflicts of interest.

Funding

This work was supported in part by the National Institute of Allergy and Infectious Diseases (NIAID) of the National Institutes of Health (NIH) under award number R01AI135114 (to D.T.L.), and by the University of Utah Study Design and Biostatistics Center, with funding in part from the National Center for Research Resources and the National Center for Advancing Translational Sciences, National Institutes of Health, through Grant UL1TR002538 (formerly 5UL1TR001067-05, 8UL1TR000105 and UL1RR025764).