The exact ORs (95% CIs) are given in eAppendix 2 in the Supplement. Error bars indicate 95% CIs. AKI indicates acute kidney failure; AMI, acute myocardial infarction; ARF, acute renal failure; BTE, bleeding or transfusion event; CVA, cardiovascular event; DVT, deep vein thrombosis; PE, pulmonary embolus; SSI, surgical site infection; UTI, urinary tract infection.
aA complication with P > .002, the Bonferroni-corrected significance level. The exact ORs (95% CIs) are given in eAppendix 2 in the Supplement. Error bars indicate 95% CIs. AKI indicates acute kidney failure; ARF, acute renal failure; BTE, bleeding or transfusion event; CVA, cardiovascular event; DVT, deep vein thrombosis; PE, pulmonary embolus; SSI, surgical site infection; UTI, urinary tract infection.
aA complication with P > .002, the Bonferroni-corrected significance level. The exact ORs (95% CIs) are given in eAppendix 2 in the Supplement. Error bars indicate 95% CIs. AKI indicates acute kidney failure; AMI, acute myocardial infarction; ARF, acute renal failure; BTE, bleeding or transfusion event; CVA, cardiovascular event; DVT, deep vein thrombosis; PE, pulmonary embolus; UTI, urinary tract infection.
aA complication with P > .002, the Bonferroni-corrected significance level. The exact ORs (95% CIs) are given in eAppendix 2 in the Supplement. Error bars indicate 95% CIs. AKI indicates acute kidney failure; AMI, acute myocardial infarction; ARF, acute renal failure; CVA, cardiovascular event; DVT, deep vein thrombosis; PE, pulmonary embolus; SSI, surgical site infection; UTI, urinary tract infection.
aA complication with P > .002, the Bonferroni-corrected significance level. The exact ORs (95% CIs) are given in eAppendix 2 in the Supplement. Error bars indicate 95% CIs. AKI indicates acute kidney failure; AMI, acute myocardial infarction; BTE, bleeding or transfusion event; CVA, cardiovascular event; DVT, deep vein thrombosis; PE, pulmonary embolus; SSI, surgical site infection; UTI, urinary tract infection.
eAppendix 1. Procedural Characteristics of the Cohorts After Matching
eAppendix 2. Odds Ratios (95% CIs) for Outcomes for Each Index Complication
Customize your JAMA Network experience by selecting one or more topics from the list below.
Wakeam E, Hyder JA, Jiang W, Lipsitz SA, Finlayson S. Risk and Patterns of Secondary Complications in Surgical Inpatients. JAMA Surg. 2015;150(1):65–73. doi:10.1001/jamasurg.2014.1795
Little empirical evidence exists on how a first (index) complication influences the risk of specific subsequent secondary complications. Understanding these risks is important to elucidate clinical pathways of failure to rescue or death after postoperative complication.
To understand patterns of secondary complications in the American College of Surgeons National Surgical Quality Improvement Program (NSQIP).
Design, Setting, and Participants
Matched analysis using a cohort of 890 604 patients undergoing elective inpatient surgery from January 1, 2005, through December 31, 2011, identified in the NSQIP Participant Use Data File. Five index complications were studied: pneumonia, acute myocardial infarction, deep space surgical site infection, bleeding or transfusion event, and acute renal failure. Each patient with an index complication was matched to a control patient based on propensity for the index event and the number of event-free days. Outcomes were compared using conditional logistic regression.
Main Outcomes and Measures
Rates of 30-day secondary complications and 30-day mortality.
Five cohorts were developed, each with 1:1 matching to controls, which were well balanced. Index pneumonia (n = 7947) was associated with increased odds of 30-day reintubation (odds ratio [OR], 17.1; 95% CI, 13.8-21.3; P < .001), ventilatory failure (OR, 15.9; 95% CI, 12.8-19.8; P < .001), sepsis (OR, 7.3; 95% CI, 6.2-8.6; P < .001), and shock (OR, 13.0; 95% CI, 10.4-16.2; P < .001). Index acute myocardial infarction was associated with increased rates of secondary bleeding or transfusion events (OR, 4.3; 95% CI, 3.3-5.8; P < .001), pneumonia (OR, 5.1; 95% CI, 2.6-10.2; P < .001), cardiac arrest (OR, 12.0; 95% CI, 7.5-19.2; P < .001), and reintubation (OR, 11.7; 95% CI, 8.4-16.3; P < .001). Deep space surgical site infection was associated with dehiscence (OR, 30.4; 95% CI, 19.9-46.5; P < .001), sepsis (OR, 13.1; 95% CI, 10.2-16.7; P < .001), shock (OR, 10.6; 95% CI, 6.4-17.7; P < .001), kidney injury (OR, 8.6; 95% CI, 3.9-18.8; P < .001), and acute renal failure (OR, 10.5; 95% CI, 3.8-29.3; P < .001). Index acute renal failure was associated with increased odds of cardiac arrest (OR, 25.3; 95% CI, 9.3-68.6; P < .001), reintubation (OR, 11.3; 95% CI, 7.4-17.1; P < .001), ventilatory failure (OR, 12.4; 95% CI, 8.2-18.8; P < .001), bleeding or transfusion events (OR, 11.3; 95% CI, 6.3-20.5; P < .001), and shock (OR, 11.2; 95% CI, 7.2-17.3; P < .001).
Conclusions and Relevance
This investigation quantified the effect of index complications on patient risk of specific secondary complications. The defined pathways merit investigation as unique targets for quality improvement and benchmarking.
Quality improvement in relation to surgical complications has traditionally focused on their incidence in the postoperative period and on the provision of perioperative processes of care. However, more recently, literature has begun to refocus efforts to improve surgical quality on the consequences of complications, not simply their incidence. The concept of failure to rescue (FTR) or death after a postoperative complication was introduced more than 20 years ago and has become an endorsed performance measure for surgery.1-3 The FTR literature suggests that a hospital’s surgical mortality rate may be determined largely by how well hospitals and surgical care teams rescue patients after complications have occurred.3,4 Structural characteristics, such as teaching status, hospital size, and measures of technology, affect FTR.5,6 In the FTR literature, patients’ pathways to failure or rescue have not been well characterized, and hence improvement strategies are also largely undefined.
To improve rates of rescue for patients with postoperative complications, it is necessary to understand the epidemiology of complications in the postoperative patient. An understanding of the patterns and outcomes of complications would help surgeons and hospitals design strategies to interrupt the progression of complications toward eventual death. First (index) complications are known to affect the risk of death in complication-specific patterns, but the pathway through which deteriorating patients transition from index event to death is unclear.7,8 Prior research and clinical practice suggest that index complications lead to secondary complications,9 but these investigations have not examined specific sequences or associations among complications that might constitute such pathways. More specific knowledge of these sequences, based on clinical data, would provide useful information for those seeking to interrupt them because it would clarify where quality improvement and safety resources should be focused and would aid in real-time risk prediction for postoperative patients with complicated disease courses.10-13
Hence, the focus of our study was to investigate whether we could identify predictable patterns of secondary complication occurrences in the postoperative period after an index complication. This epidemiologic information would be useful for health care professionals attempting to understand the quantitative risks to patients after an index event. We hypothesized that index complications would substantially alter a patient’s risk profile for subsequent complications within 30 days in patterns that were specific to each index complication.
We used data from the National Surgical Quality Improvement Program (NSQIP) Participant Use Data File from January 1, 2005, through December 31, 2011. The NSQIP Participant Use Data File is a clinical registry developed by the American College of Surgeons. The details of the NSQIP have been outlined previously.14,15
In brief, this clinical registry uses trained surgical clinical reviewers to examine medical records and classify perioperative clinical data and outcomes using standard definitions. Reliability of the data is ensured through structured training and data audits. At the time of this study, the NSQIP included the records of approximately 1.7 million surgical patients.
From this initial population, we identified 964 561 patients undergoing inpatient and elective general, vascular, cardiothoracic, orthopedic, urologic, and gynecologic operations. Only patients with complete comorbidity and operative data were included to allow subsequent propensity scoring (5565 patients excluded). We also excluded patients with acute medical conditions, such as preoperative sepsis, pneumonia at the time of operation, and preoperative ventilator dependence, and patients receiving transfusions of more than 4 U of blood in the 72 hours before operation (68 392 patients excluded). These exclusions resulted in a final cohort of 890 604 patients. This study was exempted from review by the institutional review board of the Brigham and Women’s Hospital.
Using the aforementioned cohort of patients, we sought to define groups of patients experiencing index complications. We chose 5 index complications based on their frequency, significant associated mortality, importance to quality improvement efforts, and commonality to all operations (eg, graft failure was excluded because it only applies to those operations with a graft). The 5 index complications selected were pneumonia, acute myocardial infarction (AMI), acute renal failure (ARF), deep space surgical site infection (SSI), and bleeding or transfusion events (BTEs; defined as ≥4 U of blood transfused within 72 hours postoperatively).
To assess the risk of subsequent complications after an index event, we created a reference cohort for each index complication. To assess the effect of an index complication on a patient’s risk of secondary complications, the patient’s outcomes should be compared with the outcomes for an identical patient, on the same postoperative day, in the absence of the index event. To create such a comparison group, we devised a strategy to mimic this clinical scenario, taking into account 2 main considerations. The first consideration was that index complications may be more likely to occur in patient populations with specific clinical characteristics, and these characteristics may also confound the association with subsequent complications. The second consideration was timing. Specifically, our methods needed to take into account the fact that an index event on postoperative day 2 is not equivalent to the same event on day 28 with respect to the days at risk for secondary complications because the NSQIP data are censored at 30 days. Furthermore, one cannot compare a patient with an index complication on day 2 with a patient who has not had any complications using standard proportional hazards techniques because these methods do not allow comparison of outcomes in the group that has not had the event. We therefore devised a propensity scoring and matching strategy to address these 2 issues.
For each patient in the original cohort of 890 604 patients, 5 propensity scores were calculated for the propensity to experience each index complication. Variables used in the propensity score (logistic regression) models were based on the most recent NSQIP general morbidity risk adjustment model available from January 2013. In all, 29 variables (24 patient comorbidity and demographic variables and 5 procedural variables) were used for propensity scoring (Table 1 and Table 2). Because several of the index complications and secondary complications are rare, we attempted to maximize our statistical power by choosing an inclusive sample with respect to procedure type. We adjusted for procedure by incorporating 5 procedural variables (Current Procedural Terminology group, wound class, operative time, work relative value units, and surgical specialty) into the propensity scores in addition to the 24 variables used to characterize patient comorbidity and health status. We used Current Procedural Terminology groupings previously used by the NSQIP.16
Next we developed a unique comparison cohort for each index complication matched on event-free days and propensity to have the index complication. We matched each patient with an index event in a 1:1 ratio to another patient in the cohort based on number of event-free days (exact match) and propensity score (nearest neighbor). To assess success of matching, we compared comorbidity and operative characteristics of the matched cohorts.
The outcomes used in this analysis were the occurrence of secondary complications after an index complication. We analyzed 18 complications occurring on or after the day of the index complication as outcomes: AMI, pneumonia, urinary tract infection, pulmonary embolus, deep vein thrombosis, superficial SSI, deep space SSI, organ space SSI, BTE, cerebrovascular accident, ARF, acute kidney injury, dehiscence, sepsis, shock, cardiac arrest, ventilatory failure, reintubation, and death. Among other candidate complications collected by the NSQIP, we eliminated graft failure because it is unique to certain operations and coma because its rarity (0.08% overall) prevented meaningful statistical comparisons.
We compared the absolute rates of complications between the 2 matched cohorts for each index complication. Unadjusted marginal, absolute risk differences, 95% CIs, and P values between the matched cohorts were calculated using unadjusted generalized estimating equations that were clustered by matched pair. Absolute risk differences with their 95% CIs are presented in the Results section when these differences were large (>5%) because these differences were more likely to represent clinically significant alterations in risk. Conditional logistic regression was used to generate conditional odds ratios (ORs) accounting for the effect of the matched pairs. The ORs were displayed graphically using Forrest plots. A Bonferroni-corrected P < .002 was used for statistical significance (corrected for 19 outcomes). All statistical programming was performed using SAS statistical software, version 9.3 (SAS Institute Inc).
We conducted several sensitivity analyses to confirm the robustness of the findings. The first was to assess the effect of procedural heterogeneity on the results. We did this by limiting the cohort to a subgroup of gastrointestinal surgery patients and repeating the analyses. The second used multivariable conditional logistic regression models to assess the effect of any persistently unmatched comorbidity variables between the matched cohorts.
Five matched cohorts were successfully created, and all the patients with index complications were matched to a control in a 1:1 fashion. Patient characteristics are given in Table 1 and operative characteristics in eAppendix 1 in the Supplement. Cohorts were well matched for comorbidity and procedure, and the absolute magnitude of any residual difference was small.
Index pneumonia was associated with an absolute difference in the rates and odds of 30-day reintubation (19.3% vs 1.5%; +∆17.8%; 95% CI, 17.0%-18.8%; P < .001; OR, 17.1; 95% CI, 13.8-21.3; P < .001), ventilatory failure (17.8% vs 1.4%; +∆16.4%; 95% CI, 15.5%-17.3%; P < .001; OR, 15.9; 95% CI, 12.8-19.8; P < .001), sepsis (15.8% vs 2.5%; +∆13.3%; 95% CI, 12.4%-14.1%; P < .001; OR, 7.3; 95% CI, 6.2-8.6; P < .001), and shock (13.9% vs 1.2%; +∆12.7%; 95% CI, 11.8%-13.5%; P < .001; OR, 13.0; 95% CI, 10.4-16.2; P < .001). Examination of the ORs for index pneumonia (Figure 1) also revealed increased odds of secondary AMI (OR, 6.6; 95% CI, 4.2-10.3; P < .001), pulmonary embolus (OR, 6.1; 95% CI, 4.0-9.4; P < .001), and ARF (OR, 11.8; 95% CI, 7.7-18.3; P < .001). Odds of mortality were 6.2 times greater for those with index pneumonia (8.6% vs 1.5%; +∆7.1%; 95% CI, 6.4%-7.8%; P < .001).
Index AMI was associated with increased rates of secondary BTEs (10.6% vs 2.9%; +∆7.7%; 95% CI, 6.4%-9.1%; P < .001; OR, 4.3; 95% CI, 3.3-5.8; P < .001), pneumonia (10.6% vs 2.4%; +∆8.2%; 95% CI, 6.9%-9.5%; P < .001; OR, 5.1; 95% CI, 2.6-10.2; P < .001), cardiac arrest (8.9% vs 0.8%; +∆8.1%; 95% CI, 7.0%-9.3%; P < .001; OR, 12.0; 95% CI, 7.5-19.2; P < .001), reintubation (17.7% vs 1.9%; +∆15.8%; 95% CI, 14.2%-17.3%; P < .001; OR, 11.7; 95% CI, 8.4-16.3; P < .001), ventilatory failure (13.3% vs 1.8%; +∆11.5%; 95% CI, 10.2%-13.0%; P < .001; OR, 9.3; 95% CI, 6.6-13.1; P < .001), and death (14.2% vs 2.0%; +∆12.2%; 95% CI, 10.7%-13.6%; P < .001; OR, 7.7; 95% CI, 5.7-10.4; P < .001). Examination of the ORs also revealed that the index AMI (Figure 2) was associated with large increases in the odds of pulmonary embolus, deep vein thrombosis, ARF, acute kidney injury, and cerebrovascular accident (ORs, 5.1, 4.3, 8.7, 4.7, and 5.2, respectively; all P < .001). Figure 2 plots the ORs for index AMI.
Deep wound SSI was associated with a +∆11.2% rate of dehiscence (11.6% vs 0.4%; 95% CI, 10.3%-12.0%; P < .001; OR, 30.4; 95% CI, 19.9-46.5; P < .001) and sepsis (15.8% vs 1.5%; +∆14.3%; 95% CI, 13.3%-15.3%; P < .001; OR, 13.1; 95% CI, 10.2-16.7; P < .001). In addition to dehiscence and sepsis, deep space SSI was associated with an increase in the odds of shock (OR, 10.6; 95% CI, 6.4-17.7; P < .001), kidney injury (OR, 8.6; 95% CI, 3.9-18.8; P < .001), and ARF (OR, 10.5; 95% CI, 3.8-29.3; P < .001). The ORs are plotted in Figure 3. Absolute rates of secondary complications are listed in Table 2.
Index ARF was associated with an elevated rate of postoperative BTEs (9.9% vs 1.4%; +∆8.5%; 95% CI, 7.0%-10.2%; P < .001), a +∆5.4% difference in rate of urinary tract infection (7.6% vs 2.2%; 95% CI, 3.8%-6.9%; P < .001), and a +∆9.9% difference in rate of pneumonia (12.3% vs 2.4%; 95% CI, 8.1%-11.8%; P < .001; OR, 6.0; 95% CI, 4.0-8.8; P < .001). Index ARF (Figure 3) was also associated with increased odds of cardiac arrest (7.0 vs 0.4%; +∆6.7%; 95% CI, 5.4%-8.0%; P < .001; OR, 25.3; 95% CI, 9.3-68.6; P < .001), reintubation (18.9% vs 1.8%; +∆17.1%; 95% CI, 14.9%-19.2%; P < .001; OR, 11.3; 95% CI, 7.4-17.1; P < .001), ventilatory failure (21.7% vs 2.8%; +∆18.9%; 95% CI, 16.6%-21.1%; P < .001; OR, 12.4; 95% CI, 8.2-18.8; P < .001), BTEs (9.9% vs 1.4%; +∆8.6%; 95% CI, 7.0%-10.2%; P < .001; OR, 11.3; 95% CI, 6.3-20.5; P < .001), and shock (17.3% vs 1.9%; +∆15.5%; 95% CI, 13.4%-17.5%; P < .001; OR, 11.2; 95% CI, 7.2-17.3; P < .001). The ORs are plotted in Figure 4.
Index BTEs were not associated with large absolute differences in the rates of secondary complications. The BTEs had the weakest association with subsequent events of all the index complications studied. The ORs for all the subsequent complications were elevated, but no OR exceeded 3. The strongest association was seen with ventilatory failure (6.1% vs 2.7%; +∆3.4%; 95% CI, 3.1%-3.7%; OR, 2.4; 95% CI, 2.2-2.6; P < .001). The ORs for secondary complications after index BTEs are plotted in Figure 5.
Using the matched cohort, we ran conditional logistic regression models restricted to patients undergoing gastrointestinal surgery only and found no material differences in the patterns observed. In restricted analyses, rare outcomes (such as cerebrovascular accident) could not be analyzed for smaller cohorts because of lack of events. Control for unmatched covariates did not materially alter the results. We present the analysis of the matched cohorts without further adjustment for simplicity and because the patterns of secondary complications were similar in both sensitivity analyses.
In this study, we quantified the elevated risk of secondary complications after the occurrence of 5 index complications compared with cohorts matched on propensity for the index event and number of event-free days. We quantified the extent to which a patient’s risk profile is altered after the occurrence of an index complication and found that the patterns are complication specific. The constellations of secondary complications that we describe likely represent pathways to patient failure. Although many of the pathways follow expected patterns (eg, it is not surprising that pneumonia increases the risk of reintubation), our analysis is important for 2 reasons. First, the novel propensity- and timing-matched method allows for quantification of the magnitudes of these associations in reliable, contemporary clinical data, thus producing meaningful clinical comparisons. Second, it provides specific information that can be used to devise hospital-level strategies to address rescue, as well as a possible benchmarking approach that uses intermediate outcomes for FTR.
Complications and their outcomes are a vital and growing area of interest within the surgical quality improvement literature. Several complication outcomes, such as death,3 unplanned readmission,17,18 and prolonged length of stay,19 have been analyzed previously. Our work expands these investigations and suggests that key secondary complications are another important outcome after index events. Prior analyses, including those pertaining to FTR, have treated complications as independent events. Our analysis, however, indicates that in fact complications are not independent; rather, index complications greatly alter a patient’s risk profile for subsequent complications within 30 days. Although we did not specifically address how secondary complications contribute to mortality risk, given prior literature on mortality risk in patients with multiple complications,9,20,21 these complication sequences likely represent pathways through which patients transition before eventual death.
The results of our analysis could also provide the basis for several strategies to prevent or interrupt the pathway to patient demise. Index complications may represent a pause point at which interventions, such as communication protocols,22 checklists,23 and rapid response involvement, can be initiated. These interventions will need to be targeted at the epidemiologic patterns that we describe in our analysis. For example, patients with index pneumonias may need strategies aimed at preventing disease progression to decrease the need for reintubation or to prevent the development of sepsis, shock, or renal failure.
In addition, secondary complication patterns could be used to inform real-time risk prediction models that alert surgical teams to the changing risks of associated events for individual patients. Currently, risk prediction strategies typically rely on preoperative data, although some have begun to formally incorporate intraoperative and postoperative data from electronic health records as well.10,24 Our analysis implies that complication prediction and prevention may meaningfully extend to the postoperative period. Strategies based on real-time analysis of risk incorporating events occurring after operation may greatly improve our ability to predict, and hence initiate strategies to reduce, secondary complications and FTR.
Most important, the associations presented in this analysis not only elucidate patterns of adverse postoperative events but also could represent a new method to help hospitals identify areas of the rescue process in which they are deficient and to benchmark hospitals on their rescue capability. Currently, the FTR metric that is designed by the Agency for Healthcare Research and Quality and reported by Medicare may give a global picture of overall quality but may not be adequately granular for quality improvement efforts because the end point (mortality) may be too complex, rare, or influenced by numerous factors.25 Use of secondary complications as an outcome may avoid these issues because the progression among events is more clearly linked. For example, if a hospital with poor performance with respect to FTR were given data identifying specific deficiencies (eg, benchmarked rates of sepsis, shock, or ARF after an index pneumonia event), that hospital could more easily identify specific process or surveillance targets for quality improvement. Secondary complication patterns, if they can be shown to vary by hospital, could be a useful quality metric for hospitals that seek to improve their rates of rescue and surgical mortality.
There are limitations to our analysis. The time data within the NSQIP are recorded by day (per 24 hours), which limits our ability to sequence complications that occur on the same day. However, our analysis was focused on risk of associated secondary complications within 30 days and hence was less focused on precise delineation on the day of occurrence because this is often difficult or impossible in retrospective medical record review.8 Furthermore, the associations among the complications are not symmetric (ie, odds of a secondary pneumonia after index AMI and a secondary AMI after index pneumonia are not equal). If the complications all occur on the same day, one would expect these associations to be equivalent.
Because the NSQIP lacks data on severity of some complications, we are unable to determine how severity of the initial complication affects the risk of subsequent complication events. In addition, it is possible that more severe complications result in more rapid mortality, and hence death could potentially appear protective with respect to subsequent complications. However, the mortality rates are uniformly higher in the cohort with index complications compared with the matched cohort; thus, estimates presented in this article may in fact be conservative. We also note that some imbalances remain in a small proportion of the covariates after matching, although the absolute differences in the cohorts are very small. Because sensitivity analyses adjusting for unmatched covariates did not reveal alterations to the observed patterns of secondary complications, it is unlikely that this represents a threat to the validity of the findings. Residual confounding is another possible limitation in retrospective studies; however, the use of the robust NSQIP data to risk adjust the result, coupled with the strength of the reported associations, makes residual confounding unlikely to explain the patterns we observe.
Index complications significantly alter the risk of specific secondary complications. This analysis should inform real-time risk estimation for clinical decision making and quality improvement strategies for targeted interventions aimed at preventing secondary complications. Most important, our analysis draws attention to the need to expand the definition of patient rescue to include intermediate outcomes, such as secondary complications, which can potentially be used for FTR measurement. Rescue interventions that target patients with index complications are needed to arrest their progression to secondary complications and death.
Accepted for Publication: April 21, 2014.
Corresponding Author: Elliot Wakeam, MD, Center for Surgery and Public Health, Department of Surgery, Brigham and Women’s Hospital, 1 Brigham Circle, Boston, MA 02130 (firstname.lastname@example.org).
Published Online: December 3, 2014. doi:10.1001/jamasurg.2014.1795.
Author Contributions: Drs Wakeam and Finlayson had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: Wakeam, Hyder, Finlayson.
Acquisition, analysis, or interpretation of data: Wakeam, Jiang, Lipsitz, Finlayson.
Drafting of the manuscript: Wakeam, Hyder, Lipsitz, Finlayson.
Critical revision of the manuscript for important intellectual content: All authors.
Statistical analysis: Wakeam, Jiang, Lipsitz.
Administrative, technical, or material support: Hyder, Finlayson.
Study supervision: Hyder, Finlayson.
Conflict of Interest Disclosures: None reported.
Funding/Support: This study is supported by an Arthur Tracy Cabot Fellowship at the Center for Surgery and Public Health (Dr Wakeam).
Role of the Funder/Sponsor: The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and the decision to submit the manuscript for publication.
Correction: This article was corrected on January 22, 2015, to fix errors in the titles for Figures 4 and 5.