When looking for signs of climate change in data, or comparing satellite measurements to ground validation data, it is often tempting to examine data at many locations, to test whether any differences are statistically significant. Unless one is careful about how one goes about this, though, serious overestimates of the statistical significance of a result are possible. We discuss some pitfalls often encountered in multi-hypothesis testing. We describe examples where this might have occurred in some recent papers about detecting weekly cycles in meteorological data. We conclude with a brief discussion of our efforts to avoid these pitfalls in some early research on the weekly cycle seen in TRMM satellite rainfall data, and how a physical explanation based on weekly variability in pollution has been reinforced by investigations of other and more recent datasets.
Branch Seminar Series Coordinators:Lazaros.Oraiopoulos@nasa.govCharles.K.Gatebe@nasa.gov
Contact: Lazaros Oraiopoulos