Methods
Data Sources
This study was conducted among an insured population using the MarketScan Commercial Claims database (Truven Health Analytics; http://truvenhealth.com/your-healthcare-focus/Life-Sciences/MarketScan-Databases-and-Online-Tools). MarketScan contains claims data from 2005 for 95 million unique patients enrolled in > 100 nongovernmental health insurance plans in all states. All data is deidentified and each enrollee is assigned a unique identifier. The databases contain demographic information such as age, sex, dates of services, International Classification of Disease, Revision 9 (ICD-9) codes, and Current Procedural Terminology (CPT) codes (American Medical Association), but not race. Data of the enrollee's geographic location are available at the 3-digit ZIP code and metropolitan statistical area (MSA) levels. The present study was exempt from institutional review board review per Department of Health and Human Services regulation 45 CFR 46.101, category 4.
Weather data were obtained from the National Weather Service United States Air Force–Navy weather stations (National Climatic Data Center; http://www.ncdc.noaa.gov/). We determined the mean, minimum, and maximum daily (24-hr) temperatures and the mean daily relative humidity by averaging the hourly recordings for all weather stations within each city. Stations with missing hourly data were excluded from that day's values.
Study Population
The eligible population comprised adults and children living in the MSAs of the U.S. cities of Atlanta (Georgia), Chicago (Illinois), Dallas (Texas), Los Angeles (California), and Philadelphia (Pennsylvania) between 2005 and 2011. These cities represent climate zones in which 30% of the world population lives (Mellinger et al. 1999). Atlanta, Dallas, and Philadelphia have humid subtropical climates with hot summers and mild-to-cool winters. Chicago has a continental climate with hot summers and cold winters. Los Angeles has a Mediterranean climate with mild temperatures year round (Peel et al. 2007).
Case Ascertainment
The outcome was kidney stone presentation, defined as a surgical procedure, hospital admission, and/or at least two emergency room or outpatient clinic visits < 180 days apart for a primary diagnosis of nephrolithiasis using ICD-9 and CPT codes as defined by the Urologic Disease in America Project (Litwin and Saigal 2012). The date of stone presentation was the earliest date of service associated with the nephrolithiasis claim(s) as defined above. Outcomes among unique individuals with more than one presentation for kidney stones during the study period were limited to the earliest occurrence. Individual data were aggregated at the MSA level into daily series of kidney stones counts for the period 2005 to 2011.
Statistical Analysis
We performed a time-series study using distributed lag nonlinear models (DLNMs) to estimate the relationship between mean daily temperature and kidney stone presentation. Originally developed to evaluate the relationship between temperature and mortality, DLNMs are statistical models that describe associations between exposures and outcomes with potentially nonlinear and delayed effects in time-series data (Armstrong 2006; Gasparrini et al. 2010). We evaluated two aspects of the association between temperature and kidney stone presentation. First, we estimated the relative risk (RR) of kidney stone presentation in association with daily mean temperatures for each day during a 20-day period after the temperature exposure (lag–response). RRs were estimated over the distribution of mean daily temperatures for each MSA relative to a mean daily temperature of 10°C, a moderate temperature that occurred in each of the study locations. Second, we summed the estimated risks for each lag day to estimate the cumulative RR for kidney stone presentation in association with daily mean temperatures during the 20-day period after the temperature exposure (cumulative exposure–response relationship). We used a 20-day lag period based on recent evidence suggesting a short lag time between high temperatures and presentation for kidney stones (Boscolo-Berto et al. 2008; Fletcher et al. 2012).
We built Poisson regression models, allowing for overdispersion for each city as follows:
where t represents the day of observation; Yt, the observed stone counts on t; α, the intercept; l, the lag days; Tt,l, the cross-basis matrix of temperature and lag; S(RHt), the cubic spline of relative humidity on day t; and DOWt, the indicator variable for day of the week at day t to control for daily fluctuations in outdoor activities. Month and year are indicator variables to control for season, temperature trends, and differences in the annual at-risk population. We included relative humidity because of its possible independent association with nephrolithiasis as has been reported in previous studies (Boscolo-Berto et al. 2008). For any given temperature, when humidity is low and the air is dry, more water is lost through the skin, thus decreasing urine volume and increasing the supersaturation of calcium and uric acid in the urine.
We used natural cubic splines to smooth the relationships and capture nonlinear associations between temperature and kidney stone diagnoses, and fit the same model for all five study MSAs to avoid overfitting for any particular city. We evaluated one to six knots placed at equal intervals over the range of temperatures and lag days, with the latter natural log-transformed to increase sensitivity for shorter lags. Our final model included the fewest knots needed to capture inflections in the associations and minimize the Akaike information criterion (AIC), specifically two knots for temperature and four for lag (at 2, 3, 4, and 7 days). Locations of temperature knots were Atlanta (6.7°C, 18.9°C), Chicago (–8.9°C, 6.1°C), Dallas (6.5°C, 21.4°C), Los Angeles (13.0°C, 20.6°C), and Philadelphia (3.7°C, 18.4°C). We assessed for differences in MSA mean annual temperature using two-sided analysis of variance tests. Statistical significance was defined as p < 0.05. Analyses were performed with R (version 3.0.1; R Project for Statistical Computing; http://www.r-project.org/) using the dlnmpackage (Gasparrini 2011).
Sensitivity Analyses
We performed several sensitivity analyses to evaluate the robustness of our results given the sensitivity of DLNMs to model choice. First, we used quadratic splines to capture nonlinear effects at temperature extremes. Second, we defined temperature as minimum and maximum daily temperatures to assess whether alternative definitions of temperature exposure changed the estimated associations between temperature and kidney stone presentation. Finally, we increased the lag window to 30 days given the suggestion of longer lag times in previous reports (Boscolo-Berto et al. 2008; Evans and Costabile 2005). The number and location of spline knots for the temperature range and lag period in the models used for the sensitivity analyses with a 20-day lag window were the same as those used for the primary analysis. For models that assessed a 30-day lag, spline knots were placed at 2, 3, 5, and 10 days.