SAMPLE 1 — EHR data, cross-site unit mixing Input: DATA SOURCE/TYPE: EHR extract from 4 hospitals KEY VARIABLES AND SAMPLE SIZE: n=18,000 patients. Variables: creatinine, glucose, HbA1c, weight, mortality (yes/no) KNOWN OR SUSPECTED ISSUES: Some sites recorded creatinine in µmol/L and others in mg/dL, but the unit column is incomplete — not every record has a unit specified INTENDED ANALYSIS: not specified Output: Priority issues specific to this dataset: Creatinine unit mixing with incomplete unit metadata is the most severe issue here. Because the unit column is incomplete, you can't simply convert based on a known label for every record — you need a distributional/plausibility-based detection strategy. Plot the raw creatinine value distribution per hospital: µmol/L values will cluster roughly 80-90x higher numerically than mg/dL for the same physiological range (normal creatinine is roughly 0.6-1.2 mg/dL, or about 53-106 µmol/L). If one hospital's distribution clusters around 0.5-2 and another clusters around 50-150, that's a strong signal of which unit was used, even without an explicit label. Records with values falling in an ambiguous overlapping range should be flagged as indeterminate rather than guessed. Glucose and HbA1c may have the same undocumented unit-mixing risk. You only mentioned creatinine explicitly, but glucose (mg/dL vs. mmol/L) commonly has the same cross-site unit problem in multi-site EHR extracts. Don't assume only creatinine is affected — check the numeric distributions of glucose and HbA1c by site as well. Mortality as an outcome combined with unresolved unit issues creates a serious downstream risk. An unresolved creatinine unit error is one of the most well-known ways to catastrophically bias a clinical prediction or association model, since it makes the same clinical severity look wildly different across hospitals. Full checklist (key items): Before applying a single plausibility range to all creatinine values, first determine each record's likely unit via the distributional approach above. Reconstruct or verify the unit field per record where possible (e.g. cross-check against hospital-level system defaults) rather than relying solely on the incomplete unit column. Check for duplicate patient records across the 4 hospitals. What NOT to do: Don't assume a single global conversion factor applies uniformly without first confirming which records are actually in which unit. Don't drop records with ambiguous/indeterminate units without documenting how many were dropped and why. Documentation reminder: Log the number of records affected by each unit correction, by site, so the derivation is reproducible.