Prediction analysis of microarrays-derived age signature validated against independent estrogen receptor-positive breast cancer datasets. (b) University of California San Francisco (UCSF) RNA sample set 2 (n = 66, younger women and older women) was used to train prediction analysis of microarrays (PAM) and to derive a 145-probe (128-gene) age cohort classifying signature, arranged in ascending order of the PAM score for cases in the older cohort. (b) Actual and signature-predicted age cohort designations for the validating UCSF RNA sample set 1 (n = 35) and two external validating datasets restricted to estrogen receptor-positive cases with identical age cohort characteristics: Sotiriou and colleagues  and Miller and colleagues . Prediction accuracies are indicated, with Fisher's exact test P values presented for significance. (c) Age-signature-defined subsets from all four sample datasets show similar differences in log2 expression levels (mean ± standard deviation) of AREG and ESR1.