A new Slop University research poster re-analyses nine seasons of finish times at a free weekly community 5K, finding the recorded median falls three minutes as volunteer rosters thin while the censoring-adjusted median does not fall at all.
Nine seasons of finish times show a falling median, and no improvement at all once the hour the timing tent packs up is treated as censoring.
Nine seasons of a free weekly community 5K show a field getting steadily quicker. A research poster released today by the School of Continuous Improvement finds the improvement belongs to the timing tent rather than to the runners.
The work sits within the School’s long-running interest in what a record leaves out, and the University regards it as a useful demonstration of that interest applied somewhere entirely ordinary.
Nobody in this study did anything wrong. The volunteers stayed as long as they could, and the results file recorded exactly what it was built to record. What none of us had done was ask the file what time it stopped.
— Dr Bram Ntuli, Lecturer and Convenor, Improvement Grand Rounds
Working across eleven sites and 84,610 recorded finishes, the team reconstructed each week’s closing time from volunteer sign-off sheets rather than from the results themselves, then re-estimated every season’s median with post-closure finishers carried as censored rather than dropped. Recorded medians fall by three minutes across the period; adjusted medians do not fall at all. A season’s volunteer numbers track the published time and not the estimated one.
“We came to this from our own habit of showing the finished version,” said Dr Thandiwe Solberg, Senior Lecturer and Convenor of Demo Quarter. “Every school here has a point where the record stops and the work carries on without it, and I would not claim ours is the exception.”
The poster proposes publishing a cut-off-aware median beside the recorded one from next season, and a two-clock trial at three sites.
The full poster is available from the University’s research repository under an open licence, doi:10.5555/slop.bfcpph.