Cross-Validation: What Does It Estimate and How Well Does It Do It?

被引:54
|
作者
Bates, Stephen [1 ,2 ]
Hastie, Trevor [3 ]
Tibshirani, Robert [4 ]
机构
[1] Univ Calif Berkeley, Dept Stat, Berkeley, CA 94720 USA
[2] Univ Calif Berkeley, EECS, Berkeley, CA 94720 USA
[3] Stanford Univ, Dept Stat & Biomed Data Sci, Stanford, CA USA
[4] Stanford Univ, Dept Biomed Data Sci & Stat, Stanford, CA USA
基金
美国国家卫生研究院; 美国国家科学基金会;
关键词
Bootstrap/resampling; Computationally intensive methods; Cross-validation; Goodness-of-fit methods; MODEL SELECTION; PREDICTION ERROR; VARIANCE;
D O I
10.1080/01621459.2023.2197686
中图分类号
O21 [概率论与数理统计]; C8 [统计学];
学科分类号
020208 ; 070103 ; 0714 ;
摘要
Cross-validation is a widely used technique to estimate prediction error, but its behavior is complex and not fully understood. Ideally, one would like to think that cross-validation estimates the prediction error for the model at hand, fit to the training data. We prove that this is not the case for the linear model fit by ordinary least squares; rather it estimates the average prediction error of models fit on other unseen training sets drawn from the same population. We further show that this phenomenon occurs for most popular estimates of prediction error, including data splitting, bootstrapping, and Mallow's C-p. Next, the standard confidence intervals for prediction error derived from cross-validation may have coverage far below the desired level. Because each data point is used for both training and testing, there are correlations among the measured accuracies for each fold, and so the usual estimate of variance is too small. We introduce a nested cross-validation scheme to estimate this variance more accurately, and show empirically that this modification leads to intervals with approximately correct coverage in many examples where traditional cross-validation intervals fail. Lastly, our analysis also shows that when producing confidence intervals for prediction accuracy with simple data splitting, one should not refit the model on the combined data, since this invalidates the confidence intervals. for this article are available online.
引用
收藏
页码:1434 / 1445
页数:12
相关论文
共 50 条
  • [31] Does corruption hurt green innovation? Yes - Global evidence from cross-validation
    Wen, Jun
    Yin, Hua-Tang
    Jang, Chyi-Lu
    Uchida, Hideaki
    Chang, Chun-Ping
    [J]. TECHNOLOGICAL FORECASTING AND SOCIAL CHANGE, 2023, 188
  • [32] Resistance to what, does it matter? How do we study it?
    Holmes, David R., Jr.
    [J]. EUROPEAN HEART JOURNAL, 2008, 29 (08) : 957 - 958
  • [33] What Does the Reform Do? How Dungeon Became Prison
    Zaharijevic, Adriana
    [J]. PHILOSOPHY AND SOCIETY-FILOZOFIJA I DRUSTVO, 2014, 25 (03): : 247 - 266
  • [34] Inclusion - What Does It Cost and How Do We Measure This?
    Schraner, Ingrid
    Bolzan, Natalie
    [J]. ASSISTIVE TECHNOLOGY FROM ADAPTED EQUIPMENT TO INCLUSIVE ENVIRONMENTS, 2009, 25 : 777 - 782
  • [35] Uveitis: what do we know and how does it help?
    Lightman, S
    [J]. CLINICAL AND EXPERIMENTAL OPHTHALMOLOGY, 2001, 29 (02): : 48 - 51
  • [36] Comparative effectiveness research - what is it and how does one do it?
    Goss, Christopher H.
    Tefft, Nathan
    [J]. PAEDIATRIC RESPIRATORY REVIEWS, 2013, 14 (03) : 152 - 156
  • [37] Cross Validation Can Estimate How Well Prediction Variance Correlates with Error
    Viana, Felipe A. C.
    Haftka, Raphael T.
    [J]. AIAA JOURNAL, 2009, 47 (09) : 2266 - 2270
  • [38] How well does the Parkland Formula estimate actual fluid resuscitation volumes?
    Cartotto, RC
    Innes, M
    Musgrave, MA
    Gomez, M
    Cooper, AB
    [J]. JOURNAL OF BURN CARE & REHABILITATION, 2002, 23 (04): : 258 - 265
  • [39] THE UNIX SHELL ... WHAT IT DOES AND HOW IT DOES IT
    PADOVANO, M
    [J]. SYSTEMS INTEGRATION BUSINESS, 1990, 23 (05): : 25 - 25
  • [40] The Kunming-Montreal Global Biodiversity Framework: what it does and does not do, and how to improve it
    Hughes, Alice C.
    Grumbine, R. Edward
    [J]. FRONTIERS IN ENVIRONMENTAL SCIENCE, 2023, 11