clinical-research-pitfalls

Clinical Research Pitfalls

This skill documents common methodological mistakes in ICU database research and how to avoid them. These errors can invalidate study conclusions.

When to Use This Skill

• Designing research studies
• Reviewing analysis plans
• Debugging unexpected results
• Peer review of methods

1. Immortal Time Bias

Definition

Time during which the outcome cannot occur, often because the exposure has not yet been assigned or identified.

Common Mistake

-- WRONG: Patients who "received Drug X during ICU stay"
-- Survival bias: Must survive long enough to receive the drug
SELECT stay_id
FROM mimiciv_derived.antibiotic
WHERE antibiotic LIKE '%vancomycin%';

Correct Approach

-- CORRECT: Define exposure at a fixed time point (e.g., first 24h)
SELECT DISTINCT stay_id
FROM mimiciv_derived.antibiotic ab
INNER JOIN mimiciv_icu.icustays ie ON ab.stay_id = ie.stay_id
WHERE ab.starttime <= DATETIME_ADD(ie.intime, INTERVAL 24 HOUR);

Key Principle

• Define exposure status at a fixed time point (e.g., ICU admission, 24 hours, 48 hours)
• Time zero should be the same for exposed and unexposed groups
• Consider landmark analysis or time-varying covariates

2. Information Leakage (Future Data)

Definition

Using information that would not be available at the time of prediction/decision.

Common Mistake

-- WRONG: Using diagnosis codes for prediction at admission
-- ICD codes are assigned at discharge!
SELECT hadm_id, icd_code
FROM mimiciv_hosp.diagnoses_icd
WHERE icd_code LIKE 'I21%';  -- MI diagnosis

Correct Approach

-- CORRECT: Use chief complaint or admission diagnosis
-- Or clearly acknowledge this is retrospective phenotyping
SELECT hadm_id
FROM mimiciv_hosp.admissions
WHERE LOWER(admission_type) LIKE '%emergency%';

Common Sources of Leakage

• Diagnosis codes: Assigned at discharge
• Procedure codes: May be coded after completion
• Length of stay: Only known at discharge
• Discharge disposition: Future information
• Labs ordered later: Not available at admission