198 breast cancer patients, 62 distant metastases, 3,014 genes and one weather station. Pick a predictor, split the patients at its median, and see how convincing the survival curves look. Then shuffle the patients and do it again.
Search for the "optimal" cutpoint (10th–90th percentile) instead of the median
Pick something.
log-rank p
–
hazard ratio
–
high / low
–
outcomes
real
You have looked at 0 predictors. At p < 0.05, chance alone hands you a "significant" result about once every twenty tries.
Data: GSE7390 (TRANSBIG, Desmedt et al. 2007), 198 lymph-node-negative, untreated breast cancers on Affymetrix HG-U133A; endpoint is distant-metastasis-free survival. Genes shown are the 3,000 most variable probes plus a few famous ones. Dublin temperatures are ERA5 reanalysis for Dublin Airport via Open-Meteo. Nothing here is a clinical claim about any gene; that is rather the point. Companion post: Everything correlates with survival. Built by Carson Stacy.