Analyzing counterintuitive data

Purpose: To explore the issue of counterintuitive data via analysis of a representative case and further discussion of those situations in which the data appear to be inconsistent with current knowledge. Case: 844 postoperative CABG patients, who were extubated within 24 hours of surgery were identified in a critical care database (MIMIC-III). Nurse elicited pain scores were documented throughout their hospital stay on a scale of 0 to 10. Levels were tracked as mean, median, and maximum values, and categorized as no (0/10), mild (1-3), moderate (4-6) and severe pain (7-10). Regression analysis was employed to analyze the relationship between pain scores and outcomes of interest (mortality and hospital LOS). After covariate adjustment, increased levels of pain were found to be associated with lower mortality rates and reduced hospital LOS. Conclusion: These counterintuitive results for post-CABG pain related outcomes have not been previously reported. While not representing strong enough evidence to alter clinical practice, confirmed and reliable results such as these should serve as a research trigger and prompt further studies into unexpected associations between pain and patient outcomes. With the advent of frequent secondary analysis of electronic health records, such counterintuitive data results are likely to become more frequent. We discuss the issue of counterintuitive data in extended fashion, including possible reasons for, and approaches to, this phenomenon.

💡 Research Summary

The authors set out to illustrate how “counterintuitive data” can emerge from large‑scale secondary analyses of electronic health records (EHRs). Using the publicly available MIMIC‑III intensive‑care database, they identified a cohort of 844 adult patients who underwent isolated coronary artery bypass grafting (CABG) and were extubated within 24 hours of ICU admission—a “fast‑track” subgroup that represents relatively uncomplicated postoperative courses. Throughout each patient’s ICU stay, bedside nurses recorded self‑reported pain scores on a 0‑10 numeric rating scale. For each individual, the authors derived three summary metrics (mean, median, and maximum pain) and also categorized pain into four ordinal groups (none = 0, mild = 1‑3, moderate = 4‑6, severe = 7‑10) in accordance with NIH recommendations.

The primary outcome was 30‑day mortality; secondary outcomes included 1‑year mortality and hospital length of stay (LOS). To adjust for confounding, multivariable logistic regression (for mortality) and linear regression (for LOS) incorporated age, sex, the Elixhauser comorbidity index, and the Oxford Acute Severity of Illness Score (OASIS). Additional ordinal regression examined the categorical pain variable, while ANOVA tested whether concurrent vital signs (heart rate, respiratory rate, systolic blood pressure) varied across pain levels. Two sensitivity analyses were performed: (1) expanding the sample to all CABG patients regardless of ventilation duration (n = 1,889) and (2) excluding patients who died during the index hospitalization. As a falsification test, the authors examined nausea—a symptom presumed unrelated to outcomes—to ensure that spurious associations would not appear.

Across all primary models, higher pain metrics were associated with lower odds of death. For example, each unit increase in mean pain reduced the odds of 30‑day mortality by roughly 30 % (adjusted OR ≈ 0.70) and similarly lowered 1‑year mortality (adjusted OR ≈ 0.75). Linear models showed that higher mean or median pain corresponded to shorter LOS, whereas the model using maximum pain showed a paradoxical increase in LOS. The sensitivity analyses largely reproduced these patterns; the expanded cohort retained the protective association for 30‑day mortality and LOS, but the effect on 1‑year mortality lost statistical significance, likely due to limited events. The falsification test confirmed that nausea had no relationship with outcomes, while delirium—known to worsen prognosis—behaved as expected, supporting the validity of the analytic approach.

The authors acknowledge that these findings run counter to the prevailing clinical dogma that better postoperative pain control improves recovery. They propose two speculative mechanisms. First, pain may be a surrogate for a robust inflammatory response; cytokines such as IL‑1β, IL‑6, and TNF‑α both amplify nociception and promote wound healing, so patients who mount a stronger inflammatory reaction might experience more pain yet heal faster. Second, patients with superior physiologic reserve may metabolize analgesics more rapidly, resulting in higher perceived pain despite adequate analgesic dosing; such patients also tend to have better cardiac output, cerebral perfusion, and overall organ function, which could independently drive better outcomes.

Methodological limitations are extensively discussed. Selection bias is inherent because only fast‑track, uncomplicated CABG cases were examined, limiting generalizability. Pain scores are subjective and susceptible to reporting bias; cultural, psychological, or opioid tolerance factors were not captured. Crucially, the database lacks detailed analgesic dosing information, preventing adjustment for medication exposure—a major potential confounder. The low absolute mortality rates in this cohort reduce statistical power for detecting modest effects, especially at 1 year. Finally, as an observational study, causality cannot be inferred; the observed associations may reflect unmeasured confounding or reverse causation.

In conclusion, the paper serves as a methodological case study highlighting that counterintuitive associations can arise from well‑executed EHR analyses and should not be dismissed outright. Instead, such signals can generate new hypotheses that merit prospective validation. The authors call for replication in larger, more diverse datasets, incorporation of medication data, and ideally randomized trials to determine whether the observed link between higher postoperative pain and improved outcomes is genuine or an artifact of residual confounding.