Methods and reagent-lot comparisons by regression analysis: False positives

Abstract

Background

The significance level specified for a statistical analysis indicates the expected false positive rate. Failure to confine false positives to their expected rates can lead to misleading conclusions and especially in studies investigating required sample sizes. A simulation study was used to investigate false positives associated with regression analysis.

Methods

Randomly generated X, Y pairs with zero underlying bias were subjected to Deming regression analysis. The design incorporated equal X, Y errors, unequal but parallel errors and non-parallel errors, each in combination with small (2:1), moderate (10:1) and large (2667:1) maximum:minimum range ratios. Bias (false positives in this context) was signified by failure of 95% confidence intervals (CIs) or, alternatively, the joint 95% slope, intercept confidence region (CR) to enclose target values.

Results

False positive rates assessed by CIs ranged from 6 to 10% and were clearly range ratio dependent, while those assessed by CRs were stable and very close to the expected 5% throughout.

Conclusions

CRs are distinctly more reliable than CIs in controlling false positive rates and should be the preferred option when using regression analysis in methods and reagent-lot comparisons. The computer program used to perform the study is freely available.