Bandwidth selection for kernel distribution function estimation

Naomi Altman, Christian Léger

Research output: Contribution to journalArticlepeer-review

136 Scopus citations


Leave-one-out cross-validation is a popular and readily implemented heuristic for bandwidth selection in nonparametric smoothing problems. In this note we elucidate the role of leave-one-out selection criteria by discussing a criterion introduced by Sarda (J. Statist. Plann. Inference 35 (1993) 65-75) for bandwidth selection for kernel distribution function estimators (KDFEs). We show that for this problem, use of the leave-one-out KDFE in the selection procedure is asymptotically equivalent to leaving none out. This contrasts with kernel density estimation, where use of the leave-one-out density estimator in the selection procedure is critical. Unfortunately, simulations show that neither method works in practice, even for samples of size as large as 1000. In fact, we show that for any fixed bandwidth, the expected value of the derivative of the leave-none-out criterion is asymptotically positive. This result and our simulations suggest that the criteria are increasing and that for sufficiently large samples (e.g., n = 100), the smallest available bandwidth will always be selected, thus contradicting the optimality result of Sarda for this estimator. As an alternative to minimizing a selection criterion, we propose a plug-in estimator of the asymptotically optimal bandwidth. Simulations suggest that the plug-in is a good estimator of the asymptotically optimal bandwidth even for samples as small as 10 observations and is not too far from the finite sample bandwidth.

Original languageEnglish (US)
Pages (from-to)195-214
Number of pages20
JournalJournal of Statistical Planning and Inference
Issue number2
StatePublished - Aug 1 1995

All Science Journal Classification (ASJC) codes

  • Statistics and Probability
  • Statistics, Probability and Uncertainty
  • Applied Mathematics


Dive into the research topics of 'Bandwidth selection for kernel distribution function estimation'. Together they form a unique fingerprint.

Cite this