Abstract
We study nonparametric robust tail coefficient estimation when the variable of interest, assumed to be of Weibull type, is observed simultaneously with a random covariate. In particular, we introduce a robust estimator for the tail coefficient, using the idea of the density power divergence, based on the relative excesses above a high threshold. The main asymptotic properties of our estimator are established under very general assumptions. The finite sample performance of the proposed procedure is evaluated by a small simulation experiment.
Similar content being viewed by others
References
Basu, A., Harris, I. R., Hjort, N. L., Jones, M. C. (1998). Robust and efficient estimation by minimizing a density power divergence. Biometrika, 85, 549–559.
Beirlant, J., Broniatowski, M., Teugels, J. L., Vynckier, P. (1995). The mean residual life function at great age: applications to tail estimation. Journal of Statistical Planning and Inference, 45, 21–48.
Billingsley, P. (1995). Probability and measure. Wiley series in probability and mathematical statistics. New York: Wiley.
Brazauskas, V., Serfling, R. (2000). Robust estimation of tail parameters for two-parameter Pareto and exponential models via generalized quantile statistics. Extremes, 3, 231–249.
Broniatowski, M. (1993). On the estimation of the Weibull tail coefficient. Journal of Statistical Planning and Inference, 35, 349–366.
Daouia, A., Gardes, L., Girard, S., Lekina, A. (2011). Kernel estimators of extreme level curves. Test, 20, 311–333.
Daouia, A., Gardes, L., Girard, S. (2013). On kernel smoothing for extremal quantile regression. Bernoulli, 19, 2557–2589.
de Haan, L., Ferreira, A. (2006). Extreme value theory: an introduction. New York: Springer.
de Wet, T., Goegebeur, Y., Guillou, A. Osmann, M. (2013). Kernel regression with Weibull-type tails. Submitted.
Diebolt, J., Gardes, L., Girard, S., Guillou, A. (2008). Bias-reduced estimators of the Weibull tail-coefficient. Test, 17, 311–331.
Dierckx, G., Beirlant, J., De Waal, D., Guillou, A. (2009). A new estimation method for Weibull-type tails based on the mean excess function. Journal of Statistical Planning and Inference, 139, 1905–1920.
Dierckx, G., Goegebeur, Y., Guillou, A. (2013). An asymptotically unbiased minimum density power divergence estimator for the Pareto-tail index. Journal of Multivariate Analysis, 121, 70–86.
Dierckx, G., Goegebeur, Y., Guillou, A. (2014). Local robust and asymptotically unbiased estimation of conditional Pareto-type tails. Test. doi:10.1007/s11749-013-0350-6
Dupuis, D., Field, C. (1998). Robust estimation of extremes. Canadian Journal of Statistics, 26, 119–215.
Gannoun, A., Girard, S., Guinot, C., Saracco, J. (2002). Reference ranges based on nonparametric quantile regression. Statistics in Medicine, 21, 3119–3135.
Gardes, L., Girard, S. (2005). Estimating extreme quantiles of Weibull tail distributions. Communications in Statistics-Theory and Methods, 34, 1065–1080.
Gardes, L., Girard, S. (2008a). A moving window approach for nonparametric estimation of the conditional tail index. Journal of Multivariate Analysis, 99, 2368–2388.
Gardes, L., Girard, S. (2008b). Estimation of the Weibull-tail coefficient with linear combination of upper order statistics. Journal of Statistical Planning and Inference, 138, 1416–1427.
Gardes, L., Stupfler, G. (2013). Estimation of the conditional tail index using a smoothed local Hill estimator. Extremes. doi:10.1007/s10687-013-0174-5
Gardes, L., Girard, S., Lekina, A. (2010). Functional nonparametric estimation of conditional extreme quantiles. Journal of Multivariate Analysis, 101, 419–433.
Geluk, J.L., de Haan, L. (1987). Regular variation, extensions and Tauberian theorems. CWI Tract 40. Amsterdam: Center for Mathematics and Computer Science.
Girard, S. (2004). A Hill type estimator of the Weibull tail coefficient. Communications in Statistics-Theory and Methods, 33, 205–234.
Goegebeur, Y., Guillou, A. (2011). A weighted mean excess function approach to the estimation of Weibull-type tails. Test, 20, 138–162.
Goegebeur, Y., Beirlant, J., de Wet, T. (2010). Generalized kernel estimators for the Weibull tail coefficient. Communications in Statistics-Theory and Methods, 39, 3695–3716.
Goegebeur, Y., Guillou, A., Schorgen, A. (2013a). Nonparametric regression estimation of conditional tails–the random covariate case. Statistics. doi:10.1080/02331888.2013.800064
Goegebeur, Y., Guillou, A., Stupfler, G. (2013b). Uniform asymptotic properties of a nonparametric regression estimator of conditional tails. Submitted. http://hal.archives-ouvertes.fr/hal-00794724
Hall, P. (1982). On some simple estimates of an exponent of regular variation. Journal of the Royal Statistical Society, Series B: Statistical Methodology, 44, 37–42.
Juárez, S., Schucany, W. (2004). Robust and efficient estimation for the generalized Pareto distribution. Extremes, 7, 237–251.
Kim, M., Lee, S. (2008). Estimation of a tail index based on minimum density power divergence. Journal of Multivariate Analysis, 99, 2453–2471.
Klüppelberg, C., Villaseñor, J. A. (1993). Estimation of distribution tails–a semiparametric approach. Blätter der Deutschen Gesellschaft für Versicherungsmathematik, 21, 213–235.
Lehmann, E. L., Casella, G. (1998). Theory of point estimation. New York: Springer.
Parzen, E. (1962). On estimation of a probability density function and mode. Annals of Mathematical Statistics, 33, 1065–1076.
Peng, L., Welsh, A. (2001). Robust estimation of the generalized Pareto distribution. Extremes, 4, 53–65.
Severini, T. (2005). Elements of distribution theory. Cambridge series in statistical and probabilistic mathematics. New York: Cambridge University Press.
Vandewalle, B., Beirlant, J., Christmann, A., Hubert, M. (2007). A robust estimator for the tail index of Pareto-type distributions. Computational Statistics and Data Analysis, 51, 6252–6268.
Wang, H., Tsai, C. L. (2009). Tail index regression. Journal of the American Statistical Association, 104, 1233–1240.
Yao, Q. (1999). Conditional predictive regions for stochastic processes. Technical report, University of Kent at Canterbury.
Author information
Authors and Affiliations
Corresponding author
Additional information
The authors are very grateful to the referee for her/his very constructive comments on the paper. The suggestions have definitely improved the presentation of the material.
Appendix
Appendix
1.1 Proof of Lemma 1
The case \(\alpha =\beta =r=0\) is trivial, so we only consider case (ii). Let \(p_n:=F(u_n;x)\). We remark that
where \(U\) is a uniform \([0,1]\) random variable and \(\widetilde{p}_n := 1-{1-p_n \over \ln {e \over 1-p_n}}\).
We will study the two terms separately. First, we remark that
Thus by the change of variable \(z={1-u \over 1-p_n}\), Assumption \(({\mathcal {R}})\) and the bound \({\rho (x) - 1 \over 2} z^2 \le D_{\rho (x)}(1+z)-z\le 0\), for \(z \ge 0\), we deduce that
Now, we remark that
Thus,
Now, concerning the \(m^{(2)}(u_n,\alpha , \beta ,r;x)\) term, using the monotonicity of \(Q\) and of the exponential function leads to the inequality
Clearly, using (9), Assumption \(({\mathcal {R}})\) and the bound for \(D_{\rho (x)}(1+.)\), we have
This implies that
since \(\rho (x)<0\).
Now, concerning the term \(T_2\), using the tail quantile function \(U(y; x) := Q\left( 1-{1 \over y}; x\right) ,\,y >1\), combined with the change of variables \(z={1-p_n \over 1-u}\), we deduce that
where \(a\) is the positive function that appears in the max-domain of attraction condition
We have to study two cases depending on the sign of \(\beta \).
First case: \(\beta \le 0\). Using the fact that \(U(.)\) is an increasing function combined with Corollary B.2.10 in de Haan and Ferreira (2006, p. 376), we deduce that for \(p_n\) sufficiently large and \(\varepsilon \) sufficiently small that
where we have also used that
see, e.g., the proof of Lemma 1 in de Wet et al. (2013).
Second case: \(\beta >0\). Using again Corollary B.2.10 in de Haan and Ferreira (2006, p. 376), we have for \(p_n\) sufficiently large, \(\delta \) and \(\widetilde{\delta }\) positive constants, and \(\varepsilon \) and \(\widetilde{\varepsilon }\) sufficiently small that
Finally
Combining all these results together leads to Lemma 1. \(\square \)
1.2 Proof of Lemma 2
From the rule of repeated expectations, we have that
Straightforward operations give
We now analyze each of the terms separately. By \(({\mathcal {F}})\) and \(({\mathcal {K}}),\) we have that
and, by \(({\mathcal {M}})\) and \(({\mathcal {K}})\),
Using similar arguments, one obtains \(T_5=O(m(u_n,\alpha ,\beta ,r;x)h^{\eta _f}\Phi _n(x))\). This proves the statement about the unconditional expectation.
For what concerns the convergence in probability, we already have from the first part of the proof that
Also, again by using the result from the first part of the proof
Thus,
under the assumptions of the lemma and the convergence in probability follows. \(\square \)
1.3 Proof of Corollary 1
First, note that
is a classical kernel density estimator for \(f\). As shown in Parzen (1962), if \(nh^p \rightarrow \infty \), then for all \(x \in {\mathbb {R}}^p\) where \(f(x)>0\) one has that \(\widehat{f}_n(x) \mathop {\rightarrow }\limits ^{{\mathbb {P}}} f(x)\). The result follows then by noting that
1.4 Proof of Theorem 1
To prove the theorem, we will adjust the arguments used to prove the existence and consistency of solutions of the likelihood estimating equation, see e.g., Theorem 3.7 and Theorem 5.1 in Chapter 6 of Lehmann and Casella (1998), to the MDPD framework. We rescale the objective function \({\widehat{\Delta }}_\alpha (\theta ;{\widehat{c}}_n)\) as
First, we will show that
as \(n \rightarrow \infty \), for any \(\theta \) sufficiently close to \(\theta _0(x)\).
By Taylor’s theorem
where \({\widetilde{\theta }}\) is a value between \(\theta \) and \(\theta _0(x)\). The term \({\widetilde{\Delta }}_\alpha ^\prime (\theta _0(x);{\widehat{c}}_n)\) can be obtained from (4). Write \(\widetilde{\Delta }_\alpha ^\prime (\theta _0(x);{\widehat{c}}_n)=: R_1+R_2+R_3-R_4\). For analyzing the term \(R_1\), we use the recursive relationships
Lemma 2, and the consistency of \(\widehat{\overline{F}}(u_n;x)\), giving
For \(R_2\) we rearrange the terms to obtain
By Lemma 2, we have that \(R_{2,1} \mathop {\rightarrow }\limits ^{{\mathbb {P}}} (1+\alpha )^{-1}\). For the term \(R_{2,2}\), we use the mean value theorem to obtain, with \(\widetilde{c}_n\) being a random value between \(c_n\) and \({\widehat{c}}_n\),
which can be easily bounded as follows:
and, therefore, by the consistency of \(\widehat{{\overline{F}}}(u_n;x)\), the convergence \(R_{2,2} \mathop {\rightarrow }\limits ^{{\mathbb {P}}}0\) follows. Combining all results gives
The terms \(R_3\) and \(R_4\) can be analyzed in an analogous way and yield
Thus, \(\widetilde{\Delta }_\alpha ^\prime (\theta _0(x);{\widehat{c}}_n) \mathop {\rightarrow }\limits ^{{\mathbb {P}}} 0\). Let \(|\theta -\theta _0(x)|=r,\,r>0\). With probability tending to 1, we have that
We now turn to the analysis of \(\widetilde{\Delta }_\alpha ^{\prime \prime }(\theta _0(x);{\widehat{c}}_n)\). Let
and
Note that the function \(\phi (a,b)\) satisfies the recursive relationship
After tedious calculations, one obtains the following expression for \(\widetilde{\Delta }_\alpha ^{\prime \prime }(\theta _0(x);{\widehat{c}}_n)\):
By a line of argumentation similar to that used for \(\widetilde{\Delta }_\alpha ^\prime (\theta _0(x);{\widehat{c}}_n)\) and also using (11), one obtains that under the conditions of the theorem
Write
The random part in the right-hand side of the above display is in absolute value less than \(r^3\) with probability tending to 1. There exist thus a \(\delta _1 >0\) and an \(r_0>0\) such that for \(r<r_0\)
with probability tending to 1.
For the third-order derivative, one can show that \(|\widetilde{\Delta }_\alpha ^{\prime \prime \prime }(\theta ;{\widehat{c}}_n)| \le M(\varvec{V})\), where \(\varvec{V} :=[(X_1,Y_1),\ldots , (X_n,Y_n)]\), for \(\theta \in (\theta _0(x)-r,\theta _0(x)+r)\), with \(M(\varvec{V}) \mathop {\rightarrow }\limits ^{{\mathbb {P}}} M\), which is bounded. The derivation is straightforward but lengthy and is therefore omitted from the paper. We can thus conclude that with probability tending to 1,
Overall, we have that with probability tending to 1,
which is positive if \(r < \delta _1/(1+M/3)\) and thus (10) follows.
To complete the proof, we adjust the line of argumentation of Theorem 3.7 in Chapter 6 of Lehmann and Casella (1998). Let \(\delta >0\) be such that \(\theta _0(x)-\delta >0\) and define
For any \({\varvec{v}} \in S_n(\delta )\), since \(\widetilde{\Delta }_\alpha (\theta ; {\widehat{c}}_n)\) is differentiable with respect to \(\theta \), we have that there exists a \(\widehat{\theta }_{n,\delta }(x) \in (\theta _0(x)-\delta ,\theta _0(x)+\delta ) \) where \(\widetilde{\Delta }_\alpha (\theta ; {\widehat{c}}_n)\) achieves a local minimum, so \(\widetilde{\Delta }_\alpha ^\prime (\widehat{\theta }_{n,\delta }(x); {\widehat{c}}_n)=0\). By the first part of the proof of the theorem, \({\mathbb {P}}_{\theta _0(x)}(S_n(\delta )) \rightarrow 1\) for any \(\delta \) small enough, and hence there exists a sequence \(\delta _n \downarrow 0\) such that \({\mathbb {P}}_{\theta _0(x)}(S_n(\delta _n)) \rightarrow 1\) as \(n \rightarrow \infty \). Now, let \(\widehat{\theta }_n^*(x) = \widehat{\theta }_{n,\delta _n}(x)\) if \({\varvec{v}} \in S_n(\delta _n)\) and arbitrary otherwise. Since \({\varvec{v}} \in S_n(\delta _n)\) implies \(\widetilde{\Delta }_\alpha ^\prime (\widehat{\theta }_n^*(x); {\widehat{c}}_n)=0,\) we have that
as \(n \rightarrow \infty \), which establishes the existence part. For the consistency of the solution sequence, note that for any fixed \(\delta > 0\) and \(n\) large enough
as \(n\rightarrow \infty \), whence the consistency of the estimator sequence. \(\square \)
1.5 Proof of Theorem 2
Let \(r_n:=\sqrt{nh^p{\overline{F}}(u_n;x)}\). To prove the theorem we will make use of the Cramér-Wold device (see e.g., Severini 2005, p. 337) according to which it is sufficient to show that
for all \(\xi \in {\mathbb {R}}^J\).
Take an arbitrary \(\xi \in {\mathbb {R}}^J\). A straightforward rearrangement of terms leads to
By the model assumptions, \(W_1,\ldots ,W_n\) are i.i.d. random variables, and therefore \({\mathbb {V}} \text{ ar }(\Lambda _n)=n{\mathbb {V}} \text{ ar }(W_1)\). We have
with
By using the results of Lemmas 1 and 2, we have then
which gives that \({\mathbb {V}} \text{ ar }(\Lambda _n) = 1/f(x) \xi ^\prime \Sigma \xi (1+o(1))\). To establish the convergence in distribution to a normal random variable, we have to verify the Lyapunov condition for triangular arrays of random variables (Billingsley 1995, p. 362). In the present context, this simplifies to verifying that \(n {\mathbb {E}}|W_1|^3 \rightarrow 0\). We have
Again, by using Lemmas 1 and 2 we obtain that
and hence \(n \text{ E } |W_1|^3 \rightarrow 0\). \(\square \)
1.6 Proof of Theorem 3
Apply a Taylor series expansion to the estimating equation \({\widetilde{\Delta }}_\alpha ^\prime (\widehat{\theta }_n(x);{\widehat{c}}_n)=0\) around \(\theta _0(x)\). This gives
where \(\widetilde{\theta }_n(x)\) is a random value between \(\widehat{\theta }_n(x)\) and \(\theta _0(x)\). A straightforward rearrangement of the terms leads then to
by (12), the consistency of \(\widehat{\theta }_n(x)\) and the boundedness of the third derivative. Another application of Taylor’s theorem gives
with \(\widetilde{c}_n\) being a random value between \({\widehat{c}}_n\) and \(c_n\). Direct computations allow us to prove that, under our assumptions, and using the second part of Lemma 2 and arguments similar to those used in the proof of Theorem 1, we have
In addition, by Theorem 2 in de Wet et al. (2013), we deduce that
Finally, combining (13) and (14) with Theorem 2 and the delta-method, Theorem 3 follows. \(\square \)
About this article
Cite this article
Goegebeur, Y., Guillou, A. & Rietsch, T. Robust conditional Weibull-type estimation. Ann Inst Stat Math 67, 479–514 (2015). https://doi.org/10.1007/s10463-014-0458-9
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-014-0458-9