Abstract
Median absolute deviation (hereafter MAD) is known as a robust alternative to the ordinary variance. It has been widely utilized to induce robust statistical inferential procedures. In this paper, we investigate the strong and weak Bahadur representations of its bootstrap counterpart. As a useful application, we utilize the results to derive the weak Bahadur representation of the bootstrap sample projection depth weighted mean—a quite important location estimator depending on MAD.
Similar content being viewed by others
References
Bahadur R (1966) A note on quantiles in large samples. Ann Stat 37:577–580
Donoho DL (1982) Breakdown properties of multivariate location estimators. Statistics, Harvard University, Ph.D. Qualifying Paper. Dept
Dutta S, Ghosh AK (2012) On robust classification using projection depth. Ann Inst Stat Math 64(3):657–676
Falk M (1997) On MAD and comedians. Ann Inst Stat Math 49(4):615–644
He X, Shao Q (1996) A general Bahadur representation of \(M\)-estimators and its application to linear regression with nonstochastic designs. Ann Stat 24:2608–2630
Kiefer J (1967) On Bahadur’s representation of sample quantiles. Ann Math Stat 38(5):1323–1342
Liu RY (1992) Data depth and multivariate rank tests. In: Dodge Y (ed) L1-Statistical analysis and related methods. North-Holland, Amsterdam, pp 279–294
Mazumder S, Serfling R (2009) Bahadur representations for the median absolute deviation and its modifications. Statist Probab Lett 79:1774–1783
Small G (1990) A survey of multidimensional medians. Int Stat Rev 58:263–277
Serfling RJ (1980) Approximation theorems of mathematical statistics, vol 162. Wiley
Serfling R, Mazumder S (2009) Exponential probability inequality and convergence results for the median absolute deviation and its modifications. Statist Probab Lett 79:1767–1773
Stahel WA (1981) Breakdown of covariance estimators. Res Rep 31. Fachgruppe für Statistik. ETH, Zürich
Wang D, Chen S (2009) Empirical likelihood for estimating equations with missing values. Ann Stat 37:490–517
Wendler M (2011) Bahadur representation for U-quantiles of dependent data. J Multivar Anal 102(6):1064–1079
Wu WB (2005) On the Bahadur representation of sample quantiles for dependent sequences. Ann Stat 33:1934–1963
Zuo YJ (2003) Projection based depth functions and associated medians. Ann Stat 31:1460–1490
Zuo YJ (2006) Multidimensional Trimming Based on Projection Depth. Ann Stat 34:2211–2251
Zuo YJ (2010) Is the \(t\) confidence interval \(X\pm t_\alpha (n - 1)s / \sqrt{n}\) optimal? Am Stat 64:170–173
Zuo Y (2015) Bahadur representations for bootstrap quantiles. Metrika 78(5):597–610
Zuo YJ, Cui HJ, He XM (2004) On the Stahel-Donoho estimators and depth-weighted means for multivariate data. Ann Stat 32:189–218
Acknowledgements
The authors thank one anonymous referee and the editors for their valuable comments, which have led to many improvements in this paper. Qing Liu’s research was supported by the Key Research Base Project of Humanities and Social Sciences in Jiangxi Province Universities (Grant No. JD20021), the NSF project of Jiangxi provincial education department (Grant No. GJJ190261), the China Postdoctoral Science Foundation funded project (Grant No. 2020M671961), and Postdoctoral Program of Jiangxi Province (Grant No. 2019KY47). Xiaohui Liu’s research is supported by the NSF of China (Grant No. 11971208), the National Social Science Foundation of China (Grant No. 21 &ZD152), and the Outstanding Youth Fund Project of the Science and Technology Department of Jiangxi Province (Grant No. 20224ACB211003).
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Proofs of the lemmas
Appendix: Proofs of the lemmas
In this appendix, we provide the detailed proofs for all lemmas given above.
Proof of Lemma 1
By Hoeffding’s inequality in Serfling (1980), we have
By the Glivenko–Cantelli theorem, \(F_{n}(v+\frac{\epsilon }{2})\rightarrow F(v+\frac{\epsilon }{2}), a.s.\) Thus for sufficiently large n,
by noting that \(F(v+\frac{\epsilon }{2})>\frac{1}{2}\) for any \(\epsilon >0\). Hence
Similar discussion leads to
Then the conclusion follows. \(\square \)
Proof of Lemma 2
Denote by \(G^{*}_{n}(y)\) the empirical distribution function of \(W^{*}_{1,l},W^{*}_{2,l},\ldots , W^{*}_{n,l}\). Set \(\alpha _{n}=(\left\lfloor \frac{n+m}{2}\right\rfloor -1)/n\), then we have
where the last inequality follows from
For the first part, by Hoeffding’s inequality, we have
where \(Y_{i}=1-I\left( v-\xi -\frac{\epsilon }{2} \le X^{*}_{i}\le v+\xi +\frac{\epsilon }{2}\right) \), \(i=1,\ldots ,n\) and \(p_{n1}=F_{n}\left( v+\xi +\frac{\epsilon }{2}\right) -F_{n}\left( v-\xi -\frac{\epsilon }{2}-\right) \). Note that \(G(\xi +\frac{\epsilon }{2})=F\left( v+\xi +\frac{\epsilon }{2}\right) -F\left( v-\xi -\frac{\epsilon }{2}-\right) \), the Glivenko–Cantelli theorem implies that
Then for sufficiently large n
It follows from the discussion above and Lemma 1 that
Set \(\beta _{n}=\left\lfloor \frac{n+m}{2}\right\rfloor /n\), a similar argument leads to
where \(p_{n2}=F_{n}(v+\xi -\frac{\epsilon }{2}) -F_{n}(v-\xi +\frac{\epsilon }{2}-)\), and \(\beta _{n}-p_{n2}>0\) for sufficiently large n. Then we have
The conclusion has been proved. \(\square \)
Proof of Lemma 3
Put \(\epsilon _n=D \frac{(\log n)^{1/2}}{n^{1/2}}\). It follows from Lemmas 1 and 2 that, for any fixed \(l \ge 1\) and \(m \ge 1\),
Since \(F(v)=1/2\), we have
Similarly
By similar arguments using \(F(v+\xi )-F(v-\xi )=1/2\), we also obtain
and
The conclusion follows from the inequalities above and the Borel-Cantelli lemma. \(\square \)
Proof of Lemma 4
Denote by \(\theta _p\) the p-th quantile of F for \(p\in (0, 1)\). Let \(a_n =\frac{c\log n}{n^{1/2}}\) for some positive constant c, and define
It follows from Lemma 3.7 of Zuo (2015) that
Let we express \(v-\xi \) as the p-th quantile of F: \(v-\xi =F^{-1}(p)=\theta _p\), and put \(x_n=\hat{v}^{*}_{n,l} - \hat{\xi }^{*}_{n,m,l}\) for any fixed \(l,m \ge 1\), then Lemma 3 implies
Now we have
Similarly, we can obtain
The proof has been completed. \(\square \)
Proof of Lemma 5
For convenience, we set \(l=m=1\). Recall that \(G^{*}_{n}(y)\) is the empirical distribution function of \(W^{*}_{1,l},W^{*}_{2,l},\ldots , W^{*}_{n,l}\). Since \(\hat{\xi }^{*}_{n,1,1} = W^{*}_{\left\lfloor \frac{n+1}{2}\right\rfloor :n, 1}\), we have \(G_{n}^{*}(\hat{\xi }^{*}_{n,1,1})=\left\lfloor \frac{n+1}{2}\right\rfloor /n\) unless there is a tie. If such a tie exists, we have some \(X_{i}^{*}=\hat{v}^{*}_{n,1} \pm \hat{\xi }^{*}_{n,1,1}\). It follows from Zuo (2015) that for large n
That is, we have for large n
Then almost surely
This completes the proof of this lemma. \(\square \)
Proof of Lemma 6
For any \(\epsilon >0\), let \(M > \sqrt{\log (1/\epsilon )}/\root 4 \of {2}\). Put \(\epsilon _n= D\frac{M}{n^{1/2}}\), where the constant D is defined in Lemma 3. It can be seen from Lemmas 1–2 that
Similar to Lemma 3, we have
The same results hold for \(b_0(\frac{\epsilon _n}{2}, l),c_0(\frac{\epsilon _n}{2}, m)\) and \(d_0(\frac{\epsilon _n}{2},m)\). Now we have
whence for n large enough
which implies
The rest parts can be proved by using the same steps. \(\square \)
Proof of Lemma 7
Let
By Taylor expansion and Lemma 6,
Thus \(V_n\) satisfies (a) of Lemma 13 of Serfling and Mazumder (2009).
Consider the case \(t>0\). Define the right limit as
Since \(F^{-1}\) maybe not continuous at \(F(v+\xi )\), there are two cases to consider. When \(\beta =v+\xi \), using \(F(x) < p\) if and only if \(x<F^{-1}(p)\), we have
where
By (21) and the expressions of \(U_n\) and \(V_n\), we have
Since F is continuous at \(v+\xi \), which implies that \(F(\eta _n(t))-F(v+\xi )=\frac{t+\epsilon /2}{\sqrt{n}} >0\). Then for all n sufficiently large
Then for a sufficiently large n, given \(X_1,X_2,\ldots ,X_n\), we have
By using the Chebyshev inequality, and noting that \(E(p_n)=\frac{t+\epsilon /2}{\sqrt{n}}\), we have
Returning to (22), the first condition in (b) of Lemma 13 in Mazumder and Serfling (2009) is established for \(t>0\) and \(\beta =v+\xi \).
When \(t>0\) and \(\beta >v+\xi \), let \(\theta \) be any point in the open interval \((v+\xi , \beta )\). As has been proved in Sect. 2 that \(\hat{v}^{*}_{n, l}+\hat{\xi }^{*}_{n, m,l}\rightarrow v+\xi ,\, a.s.\) which implies \( \textsf{P}(\hat{v}^{*}_{n, l}+\hat{\xi }^{*}_{n, m,l}>\theta )\rightarrow 0\) and
Since \(\eta _n(t)\rightarrow \beta >\theta \), then for sufficiently large n
Then similar to (22), we have
Note that by the definition of \(\beta \) and \(\theta \), almost surely there are no sample in the interval \([v+\xi , \theta ]\), hence no bootstrap sample in the same interval. So \(F_n^*(\theta )-F_n^*(v+\xi )=0,\, a.s.\) Hence
Thus we establish the first condition in (b) of Lemma 13 in Mazumder and Serfling (2009) for \(t>0\). The case \(t\le 0\) and the second condition of (b) can be proved similarly. That is, we obtain \(H_{2n}=o_p(n^{-1/2})\).
The proof of \(H_{1n}=o_p(n^{-1/2})\) follows a similar fashion. We omit the details. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Liu, Q., Liu, X. & Hu, Z. Bahadur representations for the bootstrap median absolute deviation and the application to projection depth weighted mean. Metrika (2024). https://doi.org/10.1007/s00184-024-00958-0
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00184-024-00958-0