1. 2008 Chingchun 1 Bootstrap Chingchun Huang ( 黃敬群 ) Vision Lab, NCTU

2. 2008 Chingchun 2 Introduction A data-based simulation method For statistical inference –finding estimators of the parameter in interest –Confidence of the parameter in interest

3. 2008 Chingchun 3 An example Two statistics definition for a random variable Average: sample mean Standard error: The standard deviation of the sample means Calculation of two statistics Carry out measurement “ many ” times Observations from these two statistics Standard error decreases as N increases Sample mean becomes more reliable as N increases

4. 2008 Chingchun 4 Central limit theorem Averages taken from any distribution Averages taken from any distribution (your experimental data) will have a normal (your experimental data) will have a normal distribution distribution The error for such an statistic will The error for such an statistic will decrease slowly as the number of decrease slowly as the number of observations increase observations increase

5. 2008 Chingchun 5 Normal distribution Averages of N.D.    distribution Averages of    distribution

6. 2008 Chingchun 6 Uniform distribution Averages of U.D.

7. 2008 Chingchun 7 Consequences of central limit theorem Bootstrap --- the technique to the rescue But nobody tells you how big the sample has to be.. But nobody tells you how big the sample has to be.. Should we believe a measurement of “Average”? Should we believe a measurement of “Average”? How about other objects rather than “Average”How about other objects rather than “Average”

8. 2008 Chingchun 8 Basic idea of bootstrap Originally, from some list of data, one computes an object (e.g.: statistic). Create an artificial list by randomly drawing elements from that list. Some elements will be picked more than once. Nonparametric mode (later) Parametric mode (later) Compute a new object. Repeat 100-1000 times and look at the distribution of these objects.

9. 2008 Chingchun 9 A simple example Data available comparing grades before and after leaving graduate school Some linear correlation between grades  =0.776 But how reliable is this result (  =0.776)?

10. 2008 Chingchun 10 A simple example

11. 2008 Chingchun 11

12. 2008 Chingchun 12

13. 2008 Chingchun 13 A simple example

14. 2008 Chingchun 14 Confidence intervals Consider the similar situation as before The parameter of interest is  (e.g. Mean) is an estimator of  based on the sample. We are interested in finding the confidence interval for the parameter.

15. 2008 Chingchun 15 The percentile algorithm Input the level=2 for the confidence interval. Generate B number of bootstrap samples. Compute for b = 1,…, B Arrange the new data set with ‘s in order. Compute and percentile for the new data. C.I. is given by ( th, th ) Percentile5%10%16%50%84%90%95% Percentile 49.756.462.786.9112.3118.7126.7 of

16. 2008 Chingchun 16 How many bootstraps ? No clear answer to this. Rule of thumb : try it 100 times, then 1000 times, and see if your answers have changed by much.

17. 2008 Chingchun 17 How many bootstraps ? B50100200500100020003000 Std. Error 0.08690.08040.07900.07450.07590.07560.0755

18. 2008 Chingchun 18 Convergence This histogram is showing the distribution of the correlation coefficient for the bootstrap sample. Here B=200, B=500

19. 2008 Chingchun 19 Contd… B=1000, B=2000

20. 2008 Chingchun 20 Contd….. B=3000 B=4000 Now it can be seen the sampling distributions of correlation coefficient are more or less identical.

21. 2008 Chingchun 21 Contd….. The above graph is showing the similarity in the distribution of the bootstrap distribution and the direct enumeration from random samples from the empirical distribution

22. 2008 Chingchun 22 Is it reliable ? Observations Good agreement for Normal (Gaussian) distributions Skewed distributions tend to more problematic, particularly for the tails A tip: For now nobody is going to shoot you down for using it.

23. 2008 Chingchun 23 Schematic representation of bootstrap procedure

24. 2008 Chingchun 24 Bootstrap The bootstrap can be used either non- parametrically or parametrically In nonparametric mode, it avoids restrictive and sometimes dangerous parametric assumptions about the form of the underlying population. In parametric mode it can provide more accurate estimates of errors than traditional methods.

25. 2008 Chingchun 25 Parametric Bootstrap Real World Statistic of interest Bootstrap World Estimated Bootstrap probability sample model Bootstrap Replication ( distribution) P x (samples)

26. 2008 Chingchun 26 Bootstrap The technique was extended, modified and refined to handle a wide variety of problems including: –(1) confidence intervals and hypothesis tests, –(2) linear and nonlinear regression, –(3) time series analysis and other problems 26

27. 2008 Chingchun 27 Fit a cubic spline: (N=50 training data) Example: one-dimensional smoothing

28. 2008 Chingchun 28 Least squares where  where  The bootstrap and maximum likelihood method

29. 2008 Chingchun 29 Nonparametric bootstrap Repeat B=200 times: - draw a dataset of N=50 with replacement from the training data z i =(x i,y i ) - fit a cubic spline Construct a 95% pointwise confidence interval: At each x i compute the mean and find the 2,5% and 97,5% percentiles The bootstrap and maximum likelihood method

30. 2008 Chingchun 30 Parametric bootstrap We assume that the model errors are Gaussian: Repeat B=200 times: - draw a dataset of N=50 with replacement from the training data z i =(x i,y i ) - fit a cubic spline on z i : and estimate - simulate new responses : z i * =(x i,y i * ) - fit a cubic spline on z i * : Construct a 95 pointwise confidence interval: At each x i compute the mean and find the 2,5% and 97,5% percentiles The bootstrap and maximum likelihood method

31. 2008 Chingchun 31 Parametric bootstrap Conclusion: least squares = parametric bootstrap as B   (only because of Gaussian errors) The bootstrap and maximum likelihood method

32. 2008 Chingchun 32 Some notations The Bootstrap is –A computer-based method for assigning measures of accuracy to statistical estimates. –The basic idea behind bootstrap is very simple, and goes back at least two centuries. – The bootstrap method is not a way of reducing the error ! It only tries to estimate it. –Bootstrap methods depend only on the Bootstrap samples. It does not depend on the underlying distribution.

33. 2008 Chingchun 33 A general data set-up We have dealt with –The standard error –The confidence interval –With the assumption that distribution is either unknown or very complicated. The situation can be more general –Like regression, –Sometimes using maximum likelihood estimation.

34. 2008 Chingchun 34 Conclusion The bootstrap allow the data analyst to –Asses the statistical accuracy of complicated procedures, by exploiting the power of the computer. The use of the bootstrap either –Relief the analyst from having to do complex mathematical derivation or –Provide an answer where no analytical answer can be obtained.

35. 2008 Chingchun 35 Addendum : The Jack-knife Jack-knife is a special kind of bootstrap. Each bootstrap subsample has all but one of the original elements of the list. For example, if original list has 10 elements, then there are 10 jack-knife subsamples.

36. 2008 Chingchun 36 Introduction (continued) Definition of Efron ’ s nonparametric bootstrap. Given a sample of n independent identically distributed (i.i.d.) observations X 1, X 2, …, X n from a distribution F and a parameter  of the distribution F with a real valued estimator  (X 1, X 2, …, X n ), the bootstrap estimates the accuracy of the estimator by replacing F with F n, the empirical distribution, where F n places probability mass 1/n at each observation X i. 36

37. 2008 Chingchun 37 Introduction (continued) Let X 1 *, X 2 *, …, X n * be a bootstrap sample, that is a sample of size n taken with replacement from F n. The bootstrap, estimates the variance of  (X 1, X 2, …, X n ) by computing or approximating the variance of  * =  (X 1 *, X 2 *, …, X n * ). 37

38. 2008 Chingchun 38 Introduction (continued) The bootstrap is similar to earlier techniques which are also called resampling methods: –(1) jackknife, –(2) cross-validation, –(3) delta method, –(4) permutation methods, and –(5) subsampling.. 38

39. 2008 Chingchun 39 Bootstrap Remedies In the past decade many of the problems where the bootstrap is inconsistent remedies have been found by researchers to give good modified bootstrap solutions that are consistent. For both problems describe thus far a simple procedure called the m-out-n bootstrap has been shown to lead to consistent estimates. 39

40. 2008 Chingchun 40 The m-out-of-n Bootstrap This idea was proposed by Bickel and Ren (1996) for handling doubly censored data. Instead of sampling n times with replacement from a sample of size n they suggest to do it only m times where m is much less than n. To get the consistency results both m and n need to get large but at different rates. We need m=o(n). That is m/n→0 as m and n both → ∞. This method leads to consistent bootstrap estimates in many cases where the ordinary bootstrap has problems, particularly (1) mean with infinite variance and (2) extreme value distributions. 40 Don ’ t know why.

41. 2008 Chingchun 41 Examples where the bootstrap fails Athreya (1987) shows that the bootstrap estimate of the sample mean is inconsistent when the population distribution has an infinite variance. Angus (1993) provides similar inconsistency results for the maximum and minimum of a sequence of independent identically distributed observations. 41