The PCC is not a robust statistic for measuring reproducibility of ChIP-chip data. (A) A scatterplot between probe signals of two simulated ChIP-chip replicates. The total number of probes is 300,000; 5% of the probes are assumed to have binding signal, with signal-to-noise ratio (SNR) of 3. Gaussian noise with standard deviation s.d.= 1 was added to all the probes for each simulated replicate. The PCC is 0.3. (B) Same as (A) but with 20% of the probes displaying binding signal. The PCC nearly doubles to 0.59. (C) The PCC at the probe level depends on the amount of binding for range of SNRs. Each point is an average over 100 simulated replicate pairs; error bars represent a 95% confidence interval. Although the PCC correctly reflects the data quality as defined by SNR, it underestimates substantially when the amount of binding signal is low for any fixed SNR.