Evidence Relating Health Care Provider Burnout and Quality of Care

Health care providers face a rapidly changing landscape of technology, care delivery methods, and regulations that increase the risk for professional burnout. Studies suggest that nearly half of health care providers may have burnout symptoms at any given time (1). Burnout has been linked to adverse effects, including suicidality, broken relationships, decreased productivity, unprofessional behavior, and employee turnover, at both the provider and organizational levels (26). Recent attention has been focused on the relation between health care provider burnout and reduced quality of care, with a growing body of primary literature and systematic reviews reporting associations between burnout and adherence to practice guidelines, communication, medical errors, patient outcomes, and safety metrics (711). Most studies in this field use retrospective observational designs and apply a wide range of burnout assessments and analytic tools to evaluate myriad outcomes among diverse patient populations (12). This lack of a standardized approach to measurement and analysis increases risk of bias, potentially undermining scientific progress in a rapidly expanding field of research by hampering the ability to decipher which of the apparent clinically significant results represent true effects (13). The present analysis sought to appraise this body of primary and review literature, developing an understanding of true effects within the field by using a detailed evaluation for reporting biases. Reporting biases take many forms, each contributing to overrepresentation of positive findings in the published literature. Publication bias occurs when studies with negative results are published less frequently or less rapidly than those with positive results (14). Selective outcome reporting occurs when several outcomes of potential interest are evaluated, but only those with positive results are presented or emphasized (13). Selective analysis reporting occurs when several analytic strategies are used, but those that produce the largest effects are presented. Overall, these biases result in an excess of statistically significant results in the published literature, threatening reproducibility of findings, promoting misappropriation of resources, and skewing the design of studies assessing interventions to reduce burnout or improve quality (13). Methods We conducted a systematic literature review and meta-analysis to provide summary estimations of the relation between provider burnout and quality of care, estimate study heterogeneity, and explore the potential of reporting bias in the field. We followed the PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) and MOOSE (Meta-analysis of Observational Studies in Epidemiology) guidelines for methodology and reporting (15, 16). Data Sources and Searches We searched MEDLINE, PsycINFO, Health and Psychosocial Instruments (EBSCO), Mental Measurements Yearbook (EBSCO), EMBASE (Elsevier), and Web of Science (Clarivate Analytics) from inception through 28 May 2019, with no language restrictions. We used search terms for burnout and its subdomains (emotional exhaustion, depersonalization, and reduced personal accomplishment), health care providers, and quality-of-care markers, as shown in Supplement Tables 1 to 3. Supplement. Appendix Material Study Selection We included all peer-reviewed publications reporting original investigations of health care provider burnout in relation to an assessment of patient care quality. Providers included all paid professionals delivering outpatient, prehospital, emergency, or inpatient care, including medical, surgical, and psychiatric care, to patients of any age. We chose an inclusive method of identifying burnout studies, considering assessments to be related to burnout if the authors defined them as such and used any inventory intended to identify burnout, either in part or in full. Likewise, we chose an inclusive approach to identify quality-of-care metrics, including any assessment of processes or outcomes indicative of care quality. We included objectively measured and subjectively reported quality metrics originating from the provider, other sources within the health care system, or patients and their surrogates. We considered medical malpractice allegations a subjective patient-reported quality metric. Although patient satisfaction is an important outcome, it is not consistently indicative of care quality or improved medical outcomes, suggesting that it may be related to factors outside the provider's immediate control, such as facility amenities and access to care (1720). Thus, for the purposes of this review, we excluded metrics solely indicative of patient satisfaction to reduce bias from these nonprovider-related factors that may affect satisfaction. We included peer-reviewed, indexed abstracts if they reported a study population not previously or subsequently reported in a full-length article. For study populations described in more than 1 full-length article, we included the primary result from the paper with the earliest publication date as the primary outcome, with any unique outcomes from subsequent articles as secondary outcomes. We supplemented the database searches with manual bibliography reviews from included studies and related literature reviews (79, 2124). In line with our aim to look for reporting bias, we did not expand our search beyond peer-reviewed publications and did not contact authors for unpublished data. If an article presented insufficient data to calculate an effect size, we supplemented the information with data from subsequent peer-reviewed publications when available; however, we still attributed these effect sizes to the initial report. We excluded any studies that were purely qualitative. All investigators contributed to the development of study inclusion and exclusion criteria. The literature review and study selection were conducted by 2 independent reviewers in parallel (D.S.T. and either A.S. or K.C.A.), with ambiguities and discrepancies resolved by consensus. Data Extraction and Quality Assessment We extracted data into a standard template reflecting publication characteristics, methods of assessing burnout and quality metrics, and strength of the reported relationship. Data were extracted by 2 independent reviewers (D.S.T. and A.S.), with discrepancies resolved by consensus. We estimated effect sizes and precision using the Hedges g and SEs, respectively. The Hedges g estimates effect size similarly to the Cohen d, but with a bias correction factor for small samples. In general, 0.2 indicates small effect; 0.5, medium effect; and 0.8, large effect. We classified each assessment of burnout as overall burnout, emotional exhaustion, depersonalization, or low personal accomplishment. We also identified burnout assessments as standard if defined as an emotional exhaustion score of 27 or greater or a depersonalization score of 10 or greater on the Maslach Burnout Inventory, or as the midpoint and higher on validated single-item scales. We categorized quality metrics within 5 groupsbest practices, communication, medical errors, patient outcomes, and quality and safetyand reverse coded any high-quality metrics such that positive effect sizes indicate burnout's relation to poor-quality care. For publications with several distinct (nonoverlapping) study populations reported separately, we considered each population separately for analytic purposes. For publications with more than 1 outcome for the same study population, we decided to perform analyses using only 1 outcome per study, ideally the specified primary outcome. If no primary outcome was clear, we chose the first-listed outcome, consistent with reporting conventions of presenting the primary outcome first. We considered other outcomes secondary, excluding them from the primary analyses to avoid bias from intercorrelation but including them in selected descriptive statistics and stratified analyses when appropriate. Data Synthesis and Analysis We calculated the Hedges g from odds ratios (dichotomized data) by using the transformation or from correlation coefficients (unscaled continuous data) by using the transformation both multiplied by a bias correction factor consistent with published norms (25, 26). Further details are provided in the Supplement. Most studies reported burnout as a dichotomous variable or with unscaled effect size estimates, facilitating the aforementioned transformations. We scaled effect sizes accordingly for the 6 studies reporting burnout only as a continuous variable in order to maintain comparability, adapting our methods from published guidelines (27, 28). On the basis of known distributions of burnout scores among providers (2931), we calculated the difference between the mean scores of providers with and without burnout to average 47.6% of the span of the particular burnout scale used. We thus converted effect sizes from continuous scales to the corresponding effect size reflecting a 47.6% change in scale score when needed to extrapolate to dichotomized burnout. We also performed sensitivity analyses excluding these few scaled effect sizes. Details of this process are presented in the Supplement. Initially, we intended to primarily perform a random-effects meta-analysis including all primary (or first-listed) effect sizes, with secondary meta-analyses stratified by quality metric category and by each unique burnoutquality metric combination. However, because of high heterogeneity in the pooled meta-analyses, we report only summary effects from the unique burnoutquality metric combinations. We also performed sensitivity analyses limited to studies with standard burnout assessments and those with independently observed or objectively measured quality-of-care markers. We used the empirical Bayes method with KnappHartung modification to estimate the between-study variance 2 (32). We evaluated study heterogeneity