Lecture on Bootstrap

Yu-Chin Hsu (許育進)

Academia Sinica

December 4, 2014

1 / 29
This lecture is based on Xiaoxia Shi’s lecture note and my
understanding of bootstrap.
This is an introductory lecture to Bootstrap Method in that I
won’t provide any proofs.
For further reading, please see
Horowitz, J.L. (2001). “The Bootstrap”, in J.J. Heckman and
E. Leamer, eds, Handbook of Econometrics, vol. 5, Elsevier
Science, B.V., p. 3159-3228,
and references within.
Xiaoxia Shi’s lecture note is available at
http://www.ssc.wisc.edu/∼ xshi/econ715/Lecture 10 bootstrap.pdf

2 / 29
Introduction

Let W = (W1 , . . . , Wn ) denote an i.i.d. sample with

distribution function F .
Let θ0 be the parameter of interest and θ̂n (W ) denote the
estimator based on W.
For example, P
θ0 can be the mean of W , E [W ], and
θ̂(W) = n−1 Wi .
To make inference or to construct confidence interval (CI) for
θ0 , in general, we need to know the exact (or limiting)
√
distribution of n(θ̂ − θ0 ).
√ d
Most of the time, n(θ̂ − θ0 ) → N(0, σ 2 ).
Then given the availability of consistent estimator σ̂n for σ,
we can make inference and construct CI.

3 / 29
Introduction (cont’d)

However, sometimes, (1) the exact form of σ is hard to obtain

or (2) it is very complicated to construct consistent estimator
σ̂ for σ.
(1) can happen when the derivatives of the objective function
of a maximum likelihood model are complicated.
(2) can happen when the σ̂ involves nonparametric
components, e.g., θ0 is the medium of of F and θ̂n (W) is the
sample medium. Then σ 2 will be 1/(4f (θ0 )) where f (·) is the
pdf of F . Then to estimate σ̂, one needs fˆ(θ̂n ) which is a
nonparametric estimator.
Therefore, it is hard to make inference and to construct CI for
θ0 .
Bootstrap can be served as an alternate method for this
purpose.
4 / 29
What is bootstrap?

Shi: “Bootstrap is an alternative to asymptotic approximation

for carrying out inference. The idea is to mimic the variation
from drawing different samples from a population by the
variation from redrawing samples from a sample.”
Horowitz: “The bootstrap is a method for estimating the
distribution of an estimator or test statistic by resampling
one’s data or a model estimated from the data.”
Shi: “The name comes from the common English phrase
“bootstrap” which alludes to “pulling oneself over the fence
by pulling on ones own bootstrap”, and means solving a
problem without external help.”

5 / 29
How does bootstrap work?

Let W ∗ = (W1∗ , . . . , Wn∗ ) denote an i.i.d. sample with

distribution function F ∗ .
Let θ0∗ be the parameter under F ∗ and θ̂n (W ∗ ) denote the
estimator based on W ∗ .
Basic Idea:
√
When F ∗ is close to F , then the distribution of n(θ̂∗ − θ0∗ )
√
should be close to n(θ̂ − θ0 ).
Therefore, if we can find an F ∗ (known to√ us) that is close to
F (unknown),
√ then we can approximate n(θ̂ − θ0 ) (unknown)
by n(θ̂∗ − θ0∗ ) (known).
A natural choice of F ∗ is the empirical cdf Fbn since we can
show that when sample size is large enough, Fbn is consistent
for F . This leads to the “nonparametric bootstrap”.

6 / 29
Confidence Interval

What is the concept of a confidence interval (from a

frequentist point of view)?
√ D p
Suppose n(θ̂ − θ0 ) → N(0, σ 2 ) and σ̂ 2 → σ 2 .
Then the two-sided 95% confidence interval of θ is:
 1.96σ̂ 1.96σ̂ 
θ̂ − √ , θ̂ + √ .
n n
Why?
√
P(−1.96 < n(θ̂ − θ0 )/σ̂ < 1.96) ≈ 95%
√
⇒P(−1.96σ̂ < n(θ̂ − θ0 ) < 1.96σ̂) ≈ 95%
√ √
⇒P(−1.96σ̂/ n < (θ̂ − θ0 ) < 1.96σ̂/ n) ≈ 95%
√ √
⇒P(−1.96σ̂/ n < (θ0 − θ̂) < 1.96σ̂/ n) ≈ 95%
√ √
⇒P(θ̂ − 1.96σ̂/ n < θ0 < θ̂ + 1.96σ̂/ n) ≈ 95%
7 / 29
Bootstrap Confidence Interval

√ ∗
Suppose that the limiting distribution of n(θ̂ − θ̂) is close
√
to n(θ̂ − θ0 ).
Let’s pretend that we know the 2.5% and 97.5% quantiles of
√ ∗ ∗ and q ∗ .
n(θ̂ − θ̂) for now and they are denoted as q2.5 97.5
∗ √ ∗ √
Then the 95% CI is (θ̂−q97.5 / n, θ̂−q2.5 / n).
Note that
∗ √
P(q2.5 < n(θ̂ ∗ − θ̂) < q97.5
∗
) = 95%
∗ √ ∗
⇒P(q2.5 < n(θ̂ − θ0 ) < q97.5 ) ≈ 95%
∗ √ ∗ √
⇒P(q2.5 / n < (θ̂ − θ0 ) < q97.5 / n) ≈ 95%
∗ √ ∗ √
⇒P(−q97.5 / n < (θ0 − θ̂) < −q2.5 / n) ≈ 95%
∗ √ ∗ √
⇒P(θ̂−q97.5 / n < θ0 < θ̂−q2.5 / n) ≈ 95%

8 / 29
Remarks:

In this example, we know that the limiting distribution of

√
n(θ̂ − θ0 ) is symmetric. Let α95 such that
√
P( n|(θ̂ ∗ − θ̂)| < α95 ) = 95%.
√ √
Then P(θ̂−α∗95 / n < θ0 < θ̂−α∗95 / n) = 95%.
Both CI’s are asymptotically valid.
In general, the second one can have higher-order improvement
in that the converge rate of this CI converge to 95% at a
faster rate than the first one. (Why?)
√
In general, if the finite sample distribution of n(θ̂ − θ0 ) is
known to be skewed, then the first one might be a better one
to use.

9 / 29
How to obtain those quantiles?

∗ , q∗
So far, we pretend that q2.5 ∗
97.5 and α95 are known.
√ ∗
n(θ̂ − θ̂) is known as we pointed out, because we know Wi∗
are drawn from Fb , the empirical CDF.
√
Of course, the close form of the CDF of n(θ̂ ∗ − θ̂) is still
hard to get!
Then this is where the well-known bootstrap simulations come
into play.

10 / 29
How to obtain those quantiles? (Cont’d)

We know that Wi∗ ’s are drawn from Fb which is equivalent to

randomly draw Wi∗ from {W1 , . . . , Wn } with prob 1/n.
Therefore, a bootstrap sample {W1∗ , . . . , Wn∗ } is formed from
n random sample with replacement.
This step can be done by computer.
Generate U[0, 1] random variables. Let u be a realization and
we can have the index be k if (k − 1)/n < u ≤ k/n.

11 / 29
Bootstrap simulations:

1. We can use computer to draw {W1,b ∗ , . . . , W ∗ } for

n,b
b = 1, . . . , B and obtain θ̂b∗ .
√
2. Then the n(θ̂ ∗ − θ̂) can be further approximated by the
√
empirical distribution of n(θ̂b∗ − θ̂) from b = 1, . . . , B.
√ ∗
3. Rank n(θ̂(b) − θ̂) in an ascending order such that
√ ∗ √ ∗ √ ∗
n(θ̂(1) − θ̂) ≤ n(θ̂(2) − θ̂) ≤ . . . ≤ n(θ̂(B) − θ̂).
∗ and q ∗ ∗ √ ∗
4. q2.5 97.5 can be approximated by q̂2.5 = n(θ̂(⌊2.5∗B⌋) − θ̂)
∗
√ ∗
and q̂97.5 = n(θ̂(⌊97.5∗B⌋) − θ̂), respectively, where ⌊c⌋ denote
the largest integer a such that a ≤ c.
5. That is, if B = 1000, then q̂2.5 ∗ = √n(θ̂ ∗
(25) − θ̂) and
∗ √ ∗
q̂97.5 = n(θ̂(975) − θ̂), respectively.
6. α̂∗95 is defined similarly except that the ranking is based on
√ ∗ − θ̂)|.
n|(θ̂(b)
12 / 29
How to obtain those quantiles? (Cont’d)

Note that this approximation can be as accurate as you please

by setting B large enough.
When B is too large, it might take too much time to compute.
Therefore, there is a trade-off between accuracy and time.
In general, setting B = 700 ∼ 1000, the approximation can be
good.
∗ √ ∗
Note that q̂97.5 = n(θ̂(975) − θ̂). Therefore, the lower bound
of the CI can be rewritten as
√ ∗
∗
q̂97.5 n(θ̂(975) − θ̂) ∗
θ̂ − √ = θ̂ − √ = θ̂ − (θ̂(975) − θ̂).
n n
∗
Similarly, the upper bound can be rewritten as θ̂ − (θ̂(25) − θ̂).

13 / 29
Hypothesis testing

Let Wi ∼ N(µ, 1). We want to test H0 : µ = 1 v.s.

H0 : µ 6= 1 at 5% significance level.
√
Test statistic: n(µ̂n − 1) where µ̂n is the sample average.
√
We would reject H0 when | n(µ̂n − 1)| > 1.96
Under H0 : µ = 1, we will falsely reject the null hypothesis
5% of the time.
Under H1 : µ 6= 1, we will reject the null hypothesis with
probability 1 asymptotically. (when n → ∞)

14 / 29
A wrong bootstrap procedure!

The following procedure is WRONG!

1 Generate bootstrap samples: {W1,b , . . . , Wn,b } for
b = 1, . . . ,√B, say B = 1000. √
2 Calculate n(µ̂∗b − 1) and obtain q̃2.5
∗
= n(µ̂∗(25) − 1) and
∗
√
q̃97.5 = n(µ̂∗(975) − 1).
√ ∗
√ ∗
3 Reject H0 when n(µ̂n − 1) < q̃(25) or n(µ̂n − 1) > q̃(975) .
To see why?

15 / 29
Note that
√ ∗
P( n(µ̂n − 1) < q̃(25) )
√ √
=P( n(µ̂n − 1) < n(µ̂∗(25) − 1))
√
=P(0 < n(µ̂∗(25) − µ̂n )) → 0.
Similarly,
√ ∗
P( n(µ̂n − 1) > q̂(975) )
√ √
=P( n(µ̂n − 1) > n(µ̃∗(975) − 1))
√
=P(0 > n(µ̂∗(975) − µ̂n )) → 0.
Note that the previous two results hold no matter the true
parameters are.
Therefore, no matter under the null or under the alternative,
the size or the power of such test is zero.
16 / 29
A Right way to do!

The following procedure is correct!

1 Generate bootstrap samples: {W1,b , . . . , Wn,b } for
b = 1, . . . ,√B, say B = 1000. √
2 Calculate n(µ̂∗b − µ̂n ) and obtain q̂2.5
∗
= n(µ̂∗(25) − µ̂n ) and
∗
√
q̂97.5 = n(µ̂∗(975) − µ̂n ).
√ ∗
√ ∗
3 Reject H0 when n(µ̂n − 1) < q̂(25) or n(µ̂n − 1) > q̂(975) .
Why this is a valid procedure?

17 / 29
Under the null hypothesis H0 : µ = 1,
√ ∗ √ ∗
P( n(µ̂n − 1) < q̂(25) or n(µ̂n − 1) > q̂(975) )
∗ √ ∗
=1 − P(q̂(25) < n(µ̂n − 1) < q̂(975) )
∗ √ ∗
=1 − P(q̂(25) < n(µ̂n − µ) < q̂(975) ) ≈ 0.05.

Under the alternative H1 : µ 6= 1, we have

√ ∗ ∗
n(µ̂n − 1) → ±∞. Also, q̂(25) and q̂(975) are bounded in
probability. Therefore,
√ ∗ √ ∗
P( n(µ̂n − 1) < q̂(25) or n(µ̂n − 1) > q̂(975) ) → 1.

18 / 29
Remarks

The key is to approximate the “null distribution” no matter

we are under the null or under the alternative.
√
In this case, the null distribution is n(µ̂n − µ) no matter the
value of true parameter is.
Therefore, we cannot just plug in the value that we want to
test in the bootstrap repetitions.

19 / 29
Other uses of Bootstrap

Standard Error
We can use bootstrap method to approximate the asymptotic
standard error of an estimator, σ.
As we mentioned, when constructing CI’s or conducting
hypothesis testing, we need a consistent estimator for σ.
We can use bootstrap to obtain an consistent estimator:
B
1X ∗ ∗
2
σ̂n∗ = θ̂b − θ ,
B
i =1
∗
where θ is the sample average of θ̂b∗ ’s.
Then we can replace σ̂ with σ̂n∗ in the previous cases.
Shi’s remark: To use bootstrap for standard error, the
estimator under consideration must be asymptotically normal.
Otherwise, the use of standard error itself is misguided.
20 / 29
Other uses of Bootstrap

Bias Correction
We can use bootstrap to correct the bias of an estimator.
The exact bias is Bias(θ̂n , θ) = E [θ̂n ] − θ and is unknown.
The bootstrap estimator of the bias is:
B
X
d ∗ (θ̂n , θ) = 1
Bias θ̂b∗ − θ̂n .
B
i =1

Then the bootstrap bias-corrected estimator for θ is

B
X
d ∗ (θ̂n , θ) = 2θ̂n − 1
θ̂BC ,n = θ̂n − Bias θ̂b∗ .
B
i =1

Shi’s remark: Bias correction usually increases the variance

because the bias is estimated. (This causes a trade-off
between bias and variance.) Therefore it should not be used
indiscriminately.
21 / 29
Higher-order improvements of the Bootstrap

This part is rather theoretical, so we will skip it. Please see

Shi’s note for more discussions.

22 / 29
Bootstrap for Regression Models

The regression model we consider is

Yi = Xi β + Ui , for i = 1, . . . , n,

where Wi = (Yi , Xi′ ) is iid with F .

Let β̂n denote the OLS estimator for β such that
1 X
n −1 1 X
n
β̂n = Xi Xi′ Xi Yi .
n n
i =1 i =1

√ D
Under regularity conditions, n(β̂n − β) → N(0, V ) where

V = E [XX ′ ]−1 E [U 2 XX ′ ]E [XX ′ ]−1 .

23 / 29
Bootstrap for Regression Models (Cont’d)

The nonparametric bootstrap works here.

Bootstrap sample are form by the pairs of Wi = (Yi , Xi′ )′ .
That is, a bootstrap sample {(Yi∗ , Xi∗ )}ni=1 is a random
sample with replacement from {(Yi , Xi )}ni=1 .
β̂n∗ is calculated in the same way as β̂n :
1 X
n −1 1 X
n
β̂n∗ = Xi∗ Xi∗ ′ Xi∗ Yi∗ .
n n
i =1 i =1

Then, results similar to what we discussed before would hold

in this case under suitable conditions.

24 / 29
Wild Bootstrap for Regression Models

In OLS, we have Yi = Xi β̂n + êi where êi ’s are the residuals.

Let Uib ’s denote iid pseudo random variables with mean 0 and
variance 1.
Let the b-th bootstrap sample be
∗
Yb,i = Xi β̂n + êi · Uib ,
and regressors are Xi ’s.
Then the β̂b∗ is
1 X
n −1 1 X
n 1 X
n −1 1 X
n
β̂b∗ = Xi Xi′ ∗
Xi Yb,i = β̂n + Xi Xi′ Xi êi · Uib .
n n n n
i =1 i =1 i =1 i =1

√
Then we can show that n(β̂n∗ − β̂n ) can approximate
√
n(β̂n − β̂) well.
This is a residual-based bootstrap and this only works for OLS.
25 / 29
Bootstrap method for weakly dependent data

The bootstrap method we discuss above only works in iid

framework.
For weakly dependent data, the dependence among
observations plays an important role in the asymptotics.
Doing the nonparametric bootstrap above will not work
because it will break down the dependence.
We need a method that can mimic the dependence structure.

26 / 29
Blockwise Bootstrap

Instead of resample an observation, we resample a bunch of

observations together.
To be specific, let the block size be k and the sample size be
T . Then we have T − k + 1 blocks:
(W1 , W2 , . . . , Wk )
(W2 , W3 , . . . , Wk+1 )
..
.
(WT −k+1 , . . . , WT ).
To form a bootstrap sample,
1. we randomly select m blocks (with replacement) such that
m · k ≥ T and (m − 1) · k < T .
2. laying them end-to-end in the order sampled.
3. Drop the last m · k − T observations from the last sampled
block so that the sample size of the bootstrap sample is equal
to T . 27 / 29
Blockwise Bootstrap (Cont’d)

For this method to work asymptotically, we require the block

size k → ∞, but k/T → 0 at a suitable rate.
Why this method would work?
Why k → ∞, but k/T → 0??

28 / 29
Conclusion!

29 / 29