The Daily Differential — Wednesday, July 15, 2026

Cognitive Challenges in EBM

Emergency physicians make high-stakes decisions under time pressure, uncertainty, interruption, and emotional load. Recognizing predictable cognitive errors helps you use evidence as a safeguard rather than selectively citing it to justify an intuitive choice. The dangerous failure is rarely ignorance of the literature; it is a reasoning process that never reopens once it has closed.

By the Blue Fish Med Desk Jul 15, 2026 · Phoenix

0:00 / –:–– AI‑generated audio

It is the tenth hour of an overnight shift. A 54-year-old man returns for the second time in a week with “the same dizziness,” already carrying a discharge diagnosis of benign positional vertigo from the prior visit. The resident opens the chart, sees the earlier workup, and feels the pull to simply reload the same plan and clear the board. Nothing about a full waiting room rewards reopening a closed question—and that is exactly the moment the most expensive errors are made. The evidence to protect this patient exists; whether it gets used depends on how the clinician thinks in the next five minutes.

— What’s your move? Read on.

Before you read

How do you use evidence as a check on your reasoning rather than a justification for a decision you have already made?
When is rapid pattern recognition appropriate, and when must you deliberately slow down?

The Framework

Cognitive error can arise during problem representation, evidence retrieval, evidence interpretation, or application to an individual patient.
Common biases include anchoring, premature closure, confirmation bias, availability bias, framing effects, search satisficing, and therapeutic misconception.
Dual-process thinking is useful: rapid pattern recognition is efficient, while deliberate analytic review is needed when the case is high-risk, atypical, or discordant with the initial impression.
Evidence itself does not eliminate bias; clinicians may preferentially remember studies that support a preferred treatment or overvalue dramatic relative effects.
Debiasing is most effective when built into workflow: explicit alternatives, diagnostic timeouts, checklists, forced reconsideration, and calibrated consultation.

How to Apply It

1. State the problem representation in one sentence, including acuity, key positives, key negatives, and degree of uncertainty. 2. Ask, “What diagnosis or treatment am I anchoring on, and what finding would make it wrong?” 3. Generate at least one dangerous alternative and one common benign alternative. 4. Separate the patient’s values and clinical facts from your emotional reaction, such as fear of missing a catastrophic diagnosis. 5. Search or review evidence using a prewritten question rather than a confirmatory phrase. 6. Before disposition, perform a brief diagnostic timeout: What changed? What remains unexplained? What follow-up failure would be dangerous?

Worked Example

A 42-year-old patient presents with pleuritic chest pain after a long flight. The clinician anchors on pulmonary embolism and orders CT pulmonary angiography despite a low-risk examination and a negative high-sensitivity D-dimer. To counter anchoring and availability bias, the clinician explicitly asks what else could explain the pain, confirms the patient’s pretest probability, and reviews the evidence supporting a rule-out strategy in low-risk patients. The final decision is to avoid imaging, provide return precautions, and evaluate alternative causes such as chest-wall pain, while documenting the residual uncertainty and follow-up plan.

Common Pitfalls

Treating “slow down” as a universal solution when structured prompts and better data are needed.
Calling a case atypical only after the test result conflicts with the initial hypothesis.
Confusing a familiar diagnosis with a likely diagnosis.
Using a checklist mechanically without changing the underlying question or decision.

Back to Our Patient

Back to our 54-year-old man with recurrent “vertigo”: instead of reloading the prior label, the resident runs a thirty-second diagnostic timeout—What is unexplained? What dangerous alternative have I not excluded? A focused reassessment adds a HINTS examination and a gait check, revealing direction-changing nystagmus and truncal instability that do not fit a peripheral cause. Recognizing anchoring on the prior discharge diagnosis, the team reframes the presentation as possible posterior-circulation stroke, obtains appropriate imaging and neurology involvement, and admits the patient. The safeguard was not a new fact in the literature; it was a structured prompt that forced the first impression to be re-examined.

Study Directive

Write down the four points where cognitive error enters—representation, retrieval, interpretation, application—and name one debiasing move for each.
For your next five shifts, force one explicit “dangerous alternative” per undifferentiated patient and note whether it changed your workup.
Practice a 30-second diagnostic-timeout script you can say out loud before disposition: What changed? What is unexplained? What miss would be catastrophic?
Review a case where a test result “surprised” you and identify which bias delayed recognition.
Build one workflow-level safeguard—an order-set prompt, checklist, or forced reconsideration—rather than relying on effort alone.

3 of 4

Critical Appraisal 2025

Modern evidence appraisal must account for complex trial designs, selective reporting, adaptive methods, estimands, pragmatic implementation, and rapidly...

0:00 / –:–– AI‑generated audio

The Case

At morning report, a new sepsis-bundle trial is presented as definitive: mortality fell in the treatment arm, and the slide shows a bold p-value. But the presenter mentions, almost in passing, that the benefit came from a per-protocol analysis and that many usual-care patients received the same antibiotics once clinicians recognized sepsis on their own. The room is ready to change practice. You have thirty seconds to decide whether this trial actually estimated what everyone thinks it did—and whether the effect belongs to the intervention or to something else entirely.

Before You Read

What exact treatment effect did this analysis estimate—for whom, compared with what, and over what time?
Does the reported benefit come from the intention-to-treat analysis, or from a population defined after randomization?
Did contamination, crossover, or adherence quietly dilute or distort the comparison?

Why It Matters

Modern evidence appraisal must account for complex trial designs, selective reporting, adaptive methods, estimands, pragmatic implementation, and rapidly changing standards of care. A contemporary reader must evaluate not only whether a result is valid, but exactly which treatment effect was estimated and whether it remains decision-relevant.

The Framework

Begin with the estimand: the precise treatment effect of interest, in a defined population, over a defined period, including how intercurrent events such as treatment switching or rescue therapy are handled.
Distinguish intention-to-treat, per-protocol, as-treated, and modified analyses; each answers a different question and may introduce different biases.
Examine whether the control group reflects current practice and whether co-interventions, crossover, and protocol adherence dilute or distort the contrast.
Look for multiplicity, subgroup fishing, outcome switching, missing-data assumptions, and selective emphasis on secondary or surrogate outcomes.
Contemporary applicability also requires attention to equity, representation, resource intensity, patient-centered outcomes, and implementation burden.
A “real-world” or pragmatic label does not guarantee low bias; pragmatic trials can sacrifice explanatory control for operational relevance.

How to Apply It

1. Write the treatment effect you actually need: for whom, compared with what, over what time, and on which patient-important outcome. 2. Identify the primary analysis population and determine how deviations, crossover, death, and rescue treatment were handled. 3. Check whether the comparator and background care match current local practice. 4. Separate prespecified primary outcomes from exploratory subgroups and post hoc findings. 5. Review missing-data methods, absolute effects, precision, and harms. 6. Ask whether the intervention is feasible, affordable, equitable, and acceptable in your ED. 7. State the conclusion with calibrated language: “supports,” “suggests,” “does not establish,” or “is not applicable.”

Worked Example

A contemporary randomized trial evaluates a new sepsis alert plus pharmacist intervention versus usual care. The study reports lower mortality in a per-protocol analysis, but many control patients received the same antibiotic and fluid recommendations after clinician recognition, and the intention-to-treat analysis is neutral. The correct interpretation is that the bundled strategy may help when implemented as intended, but the trial does not establish a mortality benefit from the alert itself; contamination and adherence affect the estimand, while workflow burden and false alerts affect applicability.

Common Pitfalls

Treating a prespecified subgroup as reliable merely because it has a plausible biologic explanation.
Interpreting a per-protocol result as if it preserved randomization.
Assuming newer, larger, or more technologically sophisticated means more clinically useful.
Ignoring the difference between a patient-centered outcome and a surrogate or process measure.

Back to Our Patient

Back to the morning-report sepsis trial: you name the estimand out loud—effect of being assigned the bundle, in the enrolled population, on 30-day mortality. The intention-to-treat analysis is neutral; the per-protocol benefit is vulnerable to confounding because adherent patients differ systematically from nonadherent ones, and heavy contamination in the control arm dilutes the true contrast. Your calibrated verdict: the strategy is promising and worth piloting as a workflow, but the trial does not establish that the alert reduces mortality. The room wanted “proven”; the honest verdict is “suggested, not established.”

Study Directive

For one recent trial, write the estimand in a single sentence: population, comparator, outcome, time horizon, and handling of intercurrent events.
Compare an intention-to-treat and a per-protocol result in the same paper and explain why they differ.
Identify one trial where control-arm contamination or crossover likely diluted the effect.
Flag a headline subgroup finding and decide whether it was prespecified or exploratory.
Rewrite an overconfident conclusion using calibrated language: supports, suggests, does not establish, or is not applicable.

High-Yield Pearls

Always ask, “What exact treatment effect did this analysis estimate?”
A neutral trial may reflect no effect, treatment crossover, poor adherence, inadequate power, or a diluted comparison.
Modern appraisal includes implementation, equity, and opportunity cost—not just internal validity.

Board Question

In a randomized trial, the intention-to-treat analysis shows no difference in mortality, but a per-protocol analysis shows benefit among participants who received at least 80% of the assigned intervention. Which conclusion is most defensible?

AThe intervention definitely reduces mortality when adherence is good
BThe per-protocol result is unbiased because nonadherent patients were excluded after randomization
CThe intention-to-treat result estimates the effect of assignment, while the per-protocol result may be confounded by factors associated with adherence
DThe study proves that adherence is unnecessary because the intention-to-treat analysis was negative

Reveal answer

Correct: C

Intention-to-treat preserves the benefit of randomization and estimates the effect of being assigned the intervention. Adherence-based analyses can be clinically informative but are vulnerable to confounding because adherent participants may differ systematically from nonadherent participants.

TheDaily Differential

Cognitive Challenges in EBM

The Framework

How to Apply It

Common Pitfalls

Study Directive

More in Today's Issue

Critical Appraisal

Critical Appraisal 2025

Strategies for Reading Papers

Yesterday’s Differential

Journal Watch

Trustworthy AI in Cardiology

Can’t Intubate Can’t Oxygenate (CICO)

Podcast Picks

High-Flow or Standard O2 for Hypoxemic Respiratory Failure?

Finger Pointers

Pharmacology Corner

Trimethoprim-Sulfamethoxazole (TMP-SMX)

Fomepizole

ECG of the Day

Case of the Day