Copula modeling for discrete data

Transcription

1 Copula modeling for discrete data Christian Genest & Johanna G. Nešlehová in collaboration with Bruno Rémillard McGill University and HEC Montréal ROBUST, September 11, 2016

2 Main question Suppose (X 1, Y 1 ),..., (X n, Y n ) are from a discrete distribution H. More specifically, assume X, Y {0, 1,...}. Can we still do copula modeling?

3 Lack of uniqueness of the copula In the continuous case, there is a unique function C : [0, 1] 2 [0, 1] such that H(x, y) = C{F (x), G(y)}, x, y R. In the discrete case, there are several functions A : [0, 1] 2 [0, 1] such that H(x, y) = A{F (x), G(y)}, x, y R. This class of functions is denoted A, but note that not all its members are copulas!

4 Lack of uniqueness of the copula In the continuous case, C is the distribution function of the pair (U, V ) = (F (X ), G(Y )), i.e., C(u, v) = Pr(U u, V v), u, v (0, 1). In the discrete case, D(u, v) = Pr(U u, V v), u, v (0, 1). is a distribution function too, but it is not a copula! As a consolation, D A.

5 Many paradoxical results follow... As soon as F or G are discrete, the set C H of admissible copulas is infinitely large (though its bounds can be identified). Members of C H can embody completely different types of dependence. Measures of association, dependence concepts, and orderings become margin dependent. Genest & Nešlehová (2007), ASTIN Bulletin.

6 Saving the connection between copulas and dependence? 1. H defines a contingency table. 2. Spread the mass uniformly in each cell. 3. Call the resulting copula C C H the bilinear extension copula. Illustration for Bernoulli variates X and Y :

7 Our main discovery C is the best possible candidate if you want to think of the copula associated with a discrete H, because... C is an absolutely continuous copula. X Y C (u, v) = uv. For any concordance measure, κ(h) = κ(c ). If ( X, Ỹ ) is distributed as C, then DEP(X, Y ) DEP( X, Ỹ ). Here, DEP can refer to PQD, LTD, RTI, SI, LRD.

8 Are copula models for discrete data of interest? A copula model for H, viz. H(x, y) = C{F (x), G(y)} with F (F α ), G (G β ) and C (C θ ) is a perfectly valid construction, even if F and G are discrete.

9 Yes, they are! Y Clayton copula with Kendall's tau 0.5 Gumbel copula with Kendall's tau Y Y Gauss copula with Kendall's tau X X X Y t copula with 4 dof and Kendall's tau -0.5 Clayton copula with Kendall's tau Y Y Frank copula and Kendall's tau X X X X P(5) and Y G(0.6) with various copulas and positive (top) and negative (bottom) association.

10 Theoretical back-up H often inherits dependence properties from C because DEP(U, V ) DEP(X, Y ), where (U, V ) C and DEP means PQD, LTD, RTI, SI, LRD. θ can continue to govern association between X and Y, viz. C θ PQD C θ H θ PQD H θ.

11 Can we do inference? Assume (X 1, Y 1 ),..., (X n, Y n ) is an iid sample from with F and G discrete. H θ (x, y) = C θ {F (x), G(y)} How can one fit and validate such a model? The quest for the answer is the subject of our ongoing research.

12 Naïve suggestion Draw 10,000 samples (X 1, Y 1 ),..., (X n, Y n ) of size n = 100 from H θ (x, y) = C θ {F (x), G(y)}, where C θ is a Clayton copula and F, G are discrete distributions. Since τ = θ/(θ + 2), pick ˆτ {τ n, τ a,n, τ b,n } and let ˆθ = 2 ˆτ 1 ˆτ.

13 Illustration: Poisson margins Take θ = 2 and Poisson margins with E(X ) = 1 and E(Y ) = 2. ˆθ based on τ n ˆθ based on τa,n ˆθ based on τb,n

14 What is going on? It can be seen that τ n is an unbiased estimator of τ(h) = τ(c ). BUT: C C θ for most copula families except at independence. In general, τ a,n and τ b,n are biased estimators of τ(c θ ) because X i = F 1 (U i ) and Y i = G 1 (V i ) (F (X i ), G(Y i )) C θ. In short, the discretization of (U i, V i ) is irreversible.

15 Can t we just randomize? 1. Take a sample (X 1, Y 1 ),..., (X n, Y n ) from H whose margins are count distributions. 2. Add an independent noise to each component of the pair (X i, Y i ), viz. X i = X i + U i 1, Ỹ i = Y i + V i 1, where U 1,..., U n and V 1,..., V n are independent samples from the standard uniform distribution on (0, 1). 3. The randomized sample ( X 1, Ỹ 1 ),..., ( X n, Ỹ n ) then stems from a distribution whose margins are continuous.

16 Two additional estimators 4. Compute τ n ( X, Ỹ ), the sample version of Kendall s tau based on the randomized sample ( X 1, Ỹ 1 ),..., ( X n, Ỹ n ). This gives a moment-estimate of θ, viz. ˆθ = g 1 ( τ n ). 5. Alternatively, one can compute the pseudo-likelihood estimate of θ based on the randomized sample. 6. To eliminate the uncertainty induced by randomization, repeat the previous steps N times and compute the average of the values of the estimate of θ.

17 ... that do not work either. Average and st. deviation of six estimates of θ in the Illustration: Estimate of θ based on τ n(u, V ) τ n(x, Y ) τ a,n(x, Y ) τ b,n (X, Y ) τ n( X, Ỹ ) MLE( X, Ỹ ) Av S.d Randomization is bound to fail, because the copula of ( X, Ỹ ) is C. However, remember that C C θ.

18 The empirical bilinear extension copula Compute the empirical cdf H n corresponding to the sample (X 1, Y 1 ),..., (X n, Y n ). Denote its bilinear extension copula by C n. C n is explicit; its density is given by n ij cn (u, v) = n n i n j for all u (F n (i 1), F n (i)], v (G n (j 1), G n (j)], i, j N. Observe that C n is rank-based.

19 Wait a minute... A sample (X 1, Y 1 ),..., (X n, Y n ) defines a contingency table. Pearson s chi-squared statistic for testing independence χ 2 = I J i=1 j=1 (n ij n i n j /n) 2 n i n j /n Spearman s mid-rank coefficient for testing monotone trend { n } ρ n = 12 (R n 3 i R)(S i S) i=1 Kendall s coefficient for testing monotone trend τ n = 2 n 2 {#(concordant pairs) #(discordant pairs)}

20 Surprise! It can be seen that χ 2 = n ρ n = τ n = {c n (u, v) 1} 2 du dv, {C n (u, v) uv}du dv, C n (u, v) dc n (u, v).

21 Here s an idea In the continuous case, many inferential procedures derive from the limiting behavior of the empirical copula process C n = n (C n C). In the discrete case, one can investigate the asymptotic behavior of the empirical Maltese copula process C n = n (C n C ), hoping that it would be as useful as in the continuous case.

22 Known margins in the continuous case When F and G are known and continuous, C can be estimated by the empirical distribution function B n of the sample (F (X i ), G(Y i )), 1 i n. It is well-known that in this case, B n = n (B n C) converges weakly in C[0, 1] 2 to a C-Brownian sheet B C, i.e., to a centered Gaussian process with covariance function cov{b C (u, v), B C (w, z)} = C(u w, v z) C(u, v)c(w, z).

23 Known margins in the discrete case When F and G are known and supported on N, C can be estimated by bilinear interpolation Bn of the empirical distribution function of the sample (F (X i ), G(Y i )), 1 i n. Theorem As n, the process B n = n (Bn C ) converges weakly in C[0, 1] 2 to a centered Gaussian process B C. Here, B C is no longer a C -Brownian sheet, but a bilinear interpolation thereof.

24 Illustration The limiting process B C is illustrated below in the univariate case when F is binomial with p = 0.4 and N = 3, 10 and 100. Displayed are ten realizations of B n when n = process C_n process C_n process C_n u u u

25 Unknown margins in the continuous case Under suitable regularity conditions, the process C n = n (C n C) converges weakly in C[0, 1] 2 to a centered Gaussian process C, C(u, v) = B C (u, v) u C(u, v) B C (u, 1) v C(u, v) B C (1, v).

26 Bad news in the discrete case Suppose that H is a bivariate Bernoulli distribution with F (0) = p, G(0) = q, H(0, 0) = r, where p, q (0, 1) and r = C(p, q) for some copula C. It can be established that the finite-dimensional margins of C n converge in law, although the limit may not be Gaussian. However, the sequence C n probability unless r = pq. is not asymptotically equicontinuous in In other words, C n does not converge in C[0, 1] 2 unless r = pq.

27 What went wrong? To illustrate, consider a similar process in the univariate case. Take F Bernoulli with F (0) (0, 1) and let F n be its empirical counterpart based on a sample of size n from F. Set F (0) = p and F n (0) = p n and consider { { u, u [0, pn ] u, u [0, p] E n (u) = (1 u) 1 p n p n, u [p n, 1], E(u) = (1 u). 1 p p, u [p, 1] Then E n = n (E n E) does not converge in law in C[0, 1] even though its finite-dimensional margins converge.

28 Illustration n= Frequency Frequency u p= p=0.5 Ten realizations of the process E n for the Bernoulli distribution with p = 0.4 and sample size 1000 (left). Histograms of 5, 000 realizations of E n (u) when u = 0.4 (middle) and u = 0.5 (right) based on samples of size n = 10, 000.

29 All s well that ends well! Consider the set O = ( ) ( ) F (k 1), F (k) G(l 1), G(l). (k,l) N 2 Theorem Let K be an arbitrary compact subset of O. Then C n converges weakly on C(K) as n to C given for every u, v O by B C (u, v) u C (u, v) B C (u, 1) v C (u, v) B C (1, v), where B C is the weak limit of B n.

30 Example: Spearman s rho Consider the non-normalized version of Spearman s rho, viz. ρ = ρ(h) = ρ(c ). Its consistent estimator is given by ρ n = 12 n 3 n (R i R)(S i S) = ρ(cn ), i=1 where R i and S i are the componentwise mid-ranks. Consequently, n {ρ n ρ(h)} = C n (u, v) du dv.

31 It works! Because [0, 1] 2 \ O has Lebesgue measure zero, C n (u, v) du dv = 12 C n (u, v) du dv. O Furthermore, O can be approximated arbitrarily closely by compact sets. This lies at the heart of the following result: Theorem As n, n {ρ n ρ(h)} 12 C (u, v) du dv. O

32 References C. Genest & J. Nešlehová (2007). A primer on copulas for count data. The ASTIN Bulletin, 37, C. Genest & J.G. Nešlehová & B. Rémillard (2014). On the empirical multilinear copula process for count data. Bernoulli, 20, C. Genest & J.G. Nešlehová & B. Rémillard (2014). Asymptotics of the empirical multilinear copula process. Submitted.

33 Research funded by