Is clinician burnout linked to patient safety?

The paper most often cited on this question no longer stands. Panagioti et al. (2018) pooled 47 studies covering 42,473 physicians in JAMA Internal Medicine and reported roughly twofold odds of unsafe care, unprofessional behavior, and low patient satisfaction. In 2019 the journal published a correction headed Errors in Data Entry and Figures (179(4), 596), published letters requesting clarification followed, and the editors retracted the paper on 18 May 2020 (Bauchner and Redberg, 2020).

What is documented is the sequence: a correction to data entry and figures, then published requests for clarification, then withdrawal in a retraction notice signed by an editor in chief and an editor of the journal on 18 May 2020. A reader who wants the editors' own account should read that notice itself, listed in the references below. The documented sequence is enough to retire the 2018 numbers from any deck or business case, and not enough to conclude that the underlying association is false.

The paper that now carries the claim is Hodkinson et al. (2022) in the BMJ, and it is far larger: 170 studies against 47, and 239,246 physicians against 42,473. It is not, however, work by outside replicators. Maria Panagioti is first author of the retracted 2018 paper and senior author of the 2022 meta-analysis, and Hodkinson, Geraghty, Johnson, Panagopoulou, Chew-Graham, Peters, Riley and Esmail appear on both author lists. Read the 2022 result as the same team redoing the work at greater scale, rather than independent replication.

Hodkinson et al. (2022) screened 4,732 records, reviewed 684 in full text, and pooled 170 observational studies of 239,246 physicians: 150 cross-sectional studies covering 231,964 physicians and 20 prospective or longitudinal studies covering 7,282. Reviewer agreement was kappa 0.89 (95% CI 0.81 to 0.96), and 84 of the 170 study authors, 49%, confirmed their data, with the review team's extractions found accurate in 96% of those studies.

On patient care, burnout was associated with 2.04 times the odds of patient safety incidents (95% CI 1.69 to 2.45, I2=87%) across 35 studies and 41,059 physicians, and with 2.33 times the odds of low professionalism (1.96 to 2.70, I2=96%) across 40 studies and 32,321 physicians. Patient dissatisfaction came in at 2.22 (1.38 to 3.57, I2=75%), but on 8 studies and 1,002 physicians, the thinnest evidence base in the paper. Do not lead a case with that third number.

The largest associations were not about patients at all. Job dissatisfaction reached 3.79 times the odds (3.24 to 4.43) on 73 studies and 146,980 physicians, career development 3.77 (2.77 to 5.14) on only two studies and 3,411 physicians under a fixed-effect model, career choice regret 3.49 (2.43 to 5.00) on 16 studies and 33,871 physicians, turnover intention 3.10 (2.30 to 4.17) on 25 studies and 32,271 physicians, and reduced productivity 1.82 (1.08 to 3.07) on 7 studies and 9,581 physicians.

Several cautions travel with all of it. Heterogeneity ran from I2=75% to I2=97% in every headline comparison, and for career choice regret the 95% prediction interval was 0.90 to 13.49, meaning a future study could plausibly find no association at all. Egger's test flagged publication bias for that same outcome (P=0.004). In 31 of the 39 studies reporting patient safety incidents (79%) and 37 of the 46 reporting professionalism (80%), the outcome was the physician's own report rather than a chart audit or surveillance record, and whether burned-out physicians are also more willing to report their errors is a contested point in this literature.

Two boundary conditions matter for how you read the totals. The search ran from database inception to May 2021, so burnout data from the years after the pandemic peak sits outside it, and only English-language publications were eligible. The authors also acknowledge method bias, since exposure and outcome were often measured with the same instrument, and they note that the career engagement outcomes overlap conceptually with the personal accomplishment subscale of burnout itself. On causality they are explicit: with only 20 prospective or longitudinal studies, assessing direct causality was not feasible, and they call for work on the causal and temporal relations.

The most repeated line about this literature is that burnout tracks hardest with safety in emergency medicine and among the youngest physicians. In univariable meta-regression that is what Hodkinson et al. (2022) found: the association with patient safety incidents was strongest in physicians aged 20 to 30 (1.88, 1.07 to 3.29, P=0.03) and in emergency medicine and intensive care settings (2.10, 1.09 to 3.56, P=0.02), with trainees in the Commonwealth region at 3.03 (0.83 to 11.25, P=0.09) on an interval that already crosses one.

In the multivariable regression, only the younger-physician association remained, and it weakened to 1.55 (0.94 to 2.56, P=0.08), which crosses one and is not statistically significant. The emergency medicine and intensive care result did not survive. For job satisfaction the authors say plainly that the subgroup associations did not remain significant in the multivariable regressions, which covers hospital settings (1.88, 0.91 to 3.86, P=0.09), physicians aged 31 to 50 (2.41, 1.02 to 5.64, P=0.04), emergency medicine and intensive care (2.16, 0.98 to 4.76, P=0.06), and the lowest value of all, general practitioners at 0.16 (0.03 to 0.88, P=0.04).

The professionalism subgroups are reported as univariable results and should be quoted that way: smallest in physicians older than 50 (0.36, 0.19 to 0.69, P=0.003), greatest in those still in training or residency (2.27, 1.45 to 3.60, P=0.001), higher in hospital-based practice (2.16, 1.46 to 3.19, P<0.001) and in emergency medicine (1.48, 1.01 to 2.34, P=0.042), with low to middle income countries at 1.68 (0.94 to 2.97, P=0.08). Presenting any of these as a settled ranking of specialties or age groups overstates what the paper supports.

The physician evidence says nothing about nurses, and the nurse evidence takes a different shape. Cimiotti et al. (2012) merged a 2006 survey of 7,076 registered nurses in 161 Pennsylvania hospitals with the state's hospital infection report and the American Hospital Association annual survey. Mean staffing was 5.7 patients per nurse (SD 1.1), and 2,544 nurses, 36.5% of those who completed the emotional exhaustion subscale of the Maslach Burnout Inventory Human Services Survey, scored at or above the high-burnout threshold of 27. Across those hospitals there were 30,213 cases with infection, 19.2 per 1,000 patients overall, including 13,567 urinary tract infections (8.6 per 1,000) and 1,668 surgical site infections (4.2 per 1,000).

The coefficients here are ordinary least squares betas on infections per 1,000 patients, not odds ratios, and they must not be restated as percentages. Modeled alone, the patient-to-nurse ratio carried 0.86 (P=.02) for urinary tract infection and 0.93 (P=.04) for surgical site infection, while burnout carried 0.85 (P=.02) and 1.58 (P<.01). In the fully adjusted model, controlling for patient severity and nurse and hospital characteristics, the staffing coefficients fell to 0.21 (P=.54) and 0.78 (P=.09), both non-significant, while the burnout coefficients held at 0.82 (P=.03) and 1.56 (P<.01). The authors' plain-language reading is that a 10% increase in a hospital's share of high-burnout nurses was associated with nearly one more urinary tract infection and two more surgical site infections per 1,000 patients.

That result cuts against a staffing-first reading of this literature, and it should be held loosely. The design is cross-sectional, confined to one state, and built on 2006 data that is now two decades old and predates substantial change in infection prevention practice. Nurses could not be linked to specific patients, and the authors twice decline to claim causation. One cross-sectional model in which a staffing coefficient lost significance does not overturn the staffing literature; it is a reason to measure burnout as its own variable rather than treating it as a stand-in for workload.

The money in this paper is a model, not a result, and it is one straight line rather than a set of alternatives. Cimiotti et al. (2012) model reductions of 10, 20 and 30 percentage points in a hospital's share of high-burnout nurses, preventing 2,079, 4,160 and 6,239 infections. The 6,239 figure quoted in the abstract, 4,006 urinary tract and 2,233 surgical site, is the 30-point case, which means a hospital with no high-burnout nurses left at all; taking a 30% share down to 10% is the 20-point case, 4,160 infections and about $41 million. The dollars come from attributable per-patient costs of $749 to $832 and $11,087 to $29,443 in 2007 dollars, and the paper's own totals are inconsistent: the abstract gives an annual saving of up to $68 million, the discussion gives nearly $28 million to more than $69 million, and Table 4 lists $3,000,661 to $3,333,178 for urinary tract infections and $24,755,692 to $65,742,026 for surgical site infections. No hospital in the study changed its burnout level; every dollar here is extrapolated from the fitted coefficients.