Pseudo-precision in gene expression values can reduce efficiency

When estimating the expression of genes based on the scanned images from microarrays various algorithms are applied in a so-called low-level analysis which can calculate expression values with an arbitrary number of digits beyond the decimal point. However, too many digits (decimal places) are usually not justified because they do not represent the precision of the measured expression. Thus, there is pseudo-precision and, as a result, there are no tied values.

We suggest avoiding, or omitting, the pseudo-precision: ties can remain, or be created by rounding the computed expression values. Then, average ranks can be used in order to apply nonparametric tests when ties occur. We use two actual data sets and the Wilcoxon rank sum test.

We demonstrate that rounding gives a more efficient test, i.e. the average p-value is decreased and the number of p-values smaller than 0.05 is increased.

The random noise of pseudo-precision can reduce the efficiency of statistical tests applied to detect differentially expressed genes. This result is, obviously, relevant in many other areas of our digitalized world.