Detecting failed WBC-reduction processes: computer simulations of intermittent and continuous process failure

BACKGROUND: By regulation, ongoing process control of WBC-reduced processes is performed on 1 percent of WBC-reduced components, typically four to fiv e samples per month. However, prospective study of the power of this small sample has been difficult. Using computer-generated "residual WBC" distribu tions, sample size sensitivity to continuous or intermittent WBC-reduction failure was examined. STUDY DESIGN AND METHODS: Populations of log-normally distributed values (m ean +/- SD, 4.5 +/- 0.5; n = 10(5)) were generated. Continuous failure (log -normality maintained) was simulated by incrementally increasing the popula tion mean or its SD. Intermittent failure (bimodal distributions with discr ete subpopulations of WBCs > the FDA cutoff) was simulated by admiring incr easing percentages of secondary outlier populations. Sample sizes of 4 to 6 0 were examined (500 repetitions each) for their power to detect drift or f ailure by standard control criteria. RESULTS: Normally distributed low variance failure was easily detected by c omparison of the mean of four samples to an upper control limit (95% confid ence of detecting 2% failure). However, 40 samples were required to detect > 5 percent intermittent (bimodal) failure or high variance failure with 90 -percent confidence, and only if individual WBC values were compared to cut off. CONCLUSION: Sampling error limits the detection of high variance or bimodal distributions. While the mean of a small sample is highly sensitive to shi fts in a low-variance normal distribution, the detection of a high-variance bimodal population requires a large number of individual values compared t o cutoff. Therefore, the number of samples required for confident failure d etection depends on both the nature of the underlying distribution and the interpretive criteria. Further research is necessary to determine the true distributions of WBC-reduction process failure, as well as clinically relev ant quality limits.