This interesting discussion was sparked by a post arguing that the statistical QC of the 1970s just can’t handle the complexities of today’s modern lab, suggesting it’s time for a major paradigm shift! (also on Linkedin)
A great question came up from Sanford Moos: “Hypothetically, if controls were run every other sampling…patient…control…patient…control…would the Westgard Rules still work?”
My short answer is: It depends on what you mean by ‘work.’
If your definition of “work” is guaranteeing zero patient errors are reported, my answer is generally No.
BUT… there’s a big exception! I’d lean toward “Probably Yes” if two things are true:
1. You verify a clean QC flag both immediately before and after each patient sample is run.
2. Your method’s Sigma metric (calculated from units, not percent) is greater than 5.8.
Why the Skepticism? Four Key Risks
Even with alternating QC, there are still a few low-probability but critical risks:
- Tiny Analytical Window: It’s theoretically possible for an error to pop up in the brief time between a patient sample and its bracketing QC. Think of it: you might change a reagent lot or briefly recalibrate in that small gap.
- Missing the Full Picture: What if a problem only shows up at the low end of your assay range, but you only ran the high-level control (or vice versa)? Running just one QC sample per “batch” (even if that batch is one patient) means you risk missing concentration-specific errors.
- The Single-Run Problem: Only the 1-3s Westgard Rule can catch a failure in a single control measurement (and the older 1-2s rule is no longer recommended). If your method has a lower Sigma value that requires multi-rules (like 2-2s or R-4s), these rules inherently need more than one QC run to detect a problem.
- The Sigma Sweet Spot (or not): Consider a method with a respectable, but not elite, Sigma of 5.7. Westgard recommends a simple 1-3s rule here. However, this single rule is incapable of ensuring 100% error detection in a single run.
The Real-World Data
Look at the numbers when a method has a Sigma of 5.7. If you use the recommended 1-3s rule with a simulated failure rate of 1 error per 100 samples, the rule only caught 13 out of 20 errors (65% detection).
If you loosen the rule to a 1-2.5s, the detection rate jumps to 85%.
The other multi-rules (2-2s, R-4s, 4-1s, or 10x) simply cannot help detect failure when you are only measuring one QC sample at a time.
I suppose one could argue for an “acceptable” error rate of 100% (meaning you accept that all errors will pass unnoticed). But if that were the case, why would we bother running QC at all?
Thanks for the great question, Sanford! This is exactly the kind of discussion we need to have.
Professor Zoe.