Statistical Testing of Covariate Effects in Conditional Copula Models

In conditional copula models, the copula parameter is deter-ministically linked to a covariate via the calibration function. The latter is of central interest for inference and is usually estimated nonparametrically. However, in many applications it is scientifically important to test whether the calibration function is constant or not. Moreover, a correct model of a constant relationship results in significant gains of statistical efficiency. We develop methodology for testing a parametric formulation of the calibration function against a general alternative and propose a generalized likelihood ratio-type test that enables conditional copula model diagnostics. We derive the asymptotic null distribution of the proposed test and study its finite sample performance using simulations. The method is applied to two data examples.


Introduction
Copulas are an important tool for modeling dependence.The recent development of conditional copulas by Patton (2006) widely expands the range of possible applications, as it allows covariate adjustment in copula structures and thus enables their use in regression settings.Specifically, if X is a covariate that affects the dependence between the continuous random variables Y 1 and Y 2 , then the conditional joint distribution H x of Y 1 and Y 2 given X = x can be written as , where F i|x is the conditional marginal distribution of Y i given X = x for i = 1, 2 and C x is the conditional copula, i.e. the joint distribution of When the conditional dependence structure is in the inferential focus, one needs to specify a functional connection between the covariate X and the copula C x as in Jondeau and Rockinger (2006); Patton (2006); Bartram et al. (2007); Rodriguez (2007).Facing the problem of specifying a model for this functional relationship, one cannot usually rely on graphical or empirical pointers and must therefore use flexible models (e.g.semi-or non-parametric) that can potentially capture a large variety of patterns.There is now a rich body of work on modelling and estimating conditional copula models, e.g.Gijbels et al. (2011); Abegaz et al. (2012); Veraverbeke et al. (2011); Craiu and Sabeti (2012).In the context of a parametric copula family, Acar et al. (2011) have studied a nonparametric estimator of the calibration function η(X) in where g : Θ → R is a known link function that allows unrestricted estimation for η.In this model, the copula family, i.e. the form of dependence between uniform random variables U 1 and U 2 remains the same for each value of the covariate X = x, but the strength of the dependence between U 1 and U 2 , measured by the copula parameter θ, is allowed to vary with X according to a smooth function η.Hence, in order to assess the covariate effect on the strength of dependence, one needs to infer the functional form of η(X).
It is known that if a parametric model for η(X) is suitable, then fitting a nonparametric model leads to an unnecessary loss of efficiency.For instance, in Table 1 in Acar et al. (2011) this loss is illustrated in the case of an underlying linear calibration function.It is thus of great practical importance to construct rigorous hypothesis tests for the specification of calibration functions.
Our development focuses on the hypotheses of the form H 0 : "η(•) is linear in X" versus H 1 : "η(•) is not linear in X" under the conditional copula model in (1).This class of hypotheses includes the important special case of H (c) 0 : "η(•) is constant" versus H (c) 1 : "η(•) is not constant".In most applications we have encountered, the constant calibration hypothesis is most relevant scientifically.In comparison, the precise specification of a non-constant calibration function (linear, quadratic, cubic) is of relatively much smaller importance.Therefore, the paper focuses on the constant/nonconstant dichotomy and uses the linear/nonlinear one to illustrate the possible generalizations of the testing procedure proposed here.From a statistical viewpoint, it is important to establish the validity of a constant calibration because then one can rely on inferential methods developed for the classical copula model.Canonical approaches based on likelihood ratio tests are possible when the calibration function is specified parametrically (Jondeau and Rockinger, 2006).Within a Bayesian approach in which regression splines are used to model η, Craiu and Sabeti (2012) suggest that novel criteria for testing H (c) 0 are needed.Hypotheses like H 0 cannot be tested using the canonical likelihood ratio test (LRT) because estimation under the alternative hypothesis is performed nonparametrically.Exploration of the asymptotic distribution of the ratio test falls within the scope of the generalized likelihood ratio test (GLRT) developed by Fan et al. (2001) for testing a parametric null hypothesis versus a nonparametric alternative hypothesis.Since nonparametric maximum likelihood estimators are difficult to obtain and may not even exist, Fan et al. (2001) suggested using any reasonable nonparametric estimator under the alternative model.In particular, using a local polynomial estimator to specify the alternative model of a number of hypothesis testing problems, Fan et al. (2001) showed that the null distribution of the GLRT statistic follows asymptotically a chi-square distribution with the number of degrees of freedom independent of the nuisance parameters.This result, referred to as Wilks phenomenon, holds for Gaussian white-noise model (Fan et al., 2001), varying-coefficient models, which include the regression model as a special case (Fan et al., 2001), spectral density (Fan and Zhang, 2004), additive models (Fan and Jiang, 2005) and single-index models (Zhang et al., 2010).
We expand the GLRT-based approach to testing the calibration function in conditional copula models.The test procedure employs the nonparametric estimator proposed by Acar et al. (2011) when evaluating the local likelihood under the alternative hypothesis.The major contribution of this work is the construction of a rigorous framework for such GLRTs in the conditional copula context.It is worth mentioning that the proposal can easily accommodate the test for an arbitrary parametric form specified under the null hypothesis.The description of the test, the derivation of its asymptotic null distribution and the discussion of practical implementation are included in Section 2. The finite sample performance of the test is illustrated using simulations and two data examples in Section 3 and 4, respectively.The paper ends with concluding remarks.

Generalized likelihood ratio test for copula functions
The construction of the GLRT is detailed under the assumption that the conditional marginal distributions A discussion of the impact of estimating the conditional marginal distributions is provided at the end of this section.
Suppose that {(U 11 , U 21 , X 1 ), . . ., (U 1n , U 2n , X n )} is a random sample from the conditional copula model (1).The null hypothesis of interest restricts the space of calibration functions to a subspace f that is fully specified parametrically.Without loss of generality, we assume is the set of all linear functions on X .Then we are interested in testing In what follows, we assume that the density c x of C x exists and for simplicity we use the notation ℓ(t, u 1 , u 2 ) = ln c x {u 1 , u 2 ; g −1 (t)}.Furthermore, the first and second partial derivatives of ℓ with respect to t are assumed to exist and are denoted by ℓ j (t, u 1 , u 2 ) = ∂ j ℓ(t, u 1 , u 2 )/∂t j , for j = 1, 2.

Proposed GLRT for the conditional copula model
A natural way to approach (2) is through the likelihood ratio of the restricted (i.e., conditional copula with a linear calibration function) and the full (i.e., conditional copula with an arbitrary calibration function) models, or equivalently, through the difference sup where The supremum of the log-likelihood function under the null hypothesis is given by where η(X) = ã0 + ã1 X, with ã = (ã 0 , ã1 ) denoting the maximum likelihood estimator of the parameter a = (a 0 , a 1 ).
Under the alternative, the general unknown form of η(•) adds significant complexity to the calculation of the supremum.We use the nonparametric estimator of η(•) proposed by Acar et al. (2011) to define the log-likelihood under the full model.Specifically, for each observation X i in a neighbourhood of an interior point x, we approximate η(X i ) linearly by provided that η(x) is twice continuously differentiable.Estimates of β = (β 0 , β 1 ) and of η(x) = β 0 , are then obtained by maximizing a kernel-weighted local likelihood function where h > 0 is a bandwidth parameter controlling the size of the neighbourhood around x, K is a symmetric kernel density function and weighs the contribution of each data point based on their proximity to x.Similarly, if one uses a pth order local polynomial estimator, the local linear approximation in (3) will be replaced by p ℓ=0 β ℓ (X i −x) ℓ and the resulting estimator is given by ηh (x) = β0 .The bandwidth h is chosen to maximize the cross-validation criterion ) is the estimate of the copula parameter θ at X i when the ith sample (U 1i , U 2i ) is left out.
Then we evaluate the log-likelihood function under the alternative hypothesis of (2) as The difference between the two log-likelihoods allows us to evaluate the evidence in the data in favor of (or against) the null model.Hence, the generalized likelihood ratio statistic is given by While large values of λ n (h) suggest the rejection of the null hypothesis, we need to determine the rejection region for the test.In order to inform the decision in finite samples we investigate the asymptotic distribution of the GLRT statistic under the null hypothesis.

Asymptotic distributions of proposed GLRT statistic
To facilitate our presentation we introduce the following notation.Let f (x) > 0 be the density function of X with support X and denote by |X | the range of the covariate X.Also, denote by K * K the convolution of the kernel K and define The following result states that the GLRT statistic follows asymptotically a normal or equivalently a chi-square distribution in the case of negligible bias, where the mean and variance are related to the quantities µ n and ν n , respectively.The technical conditions and proofs are deferred to Appendix I.
Theorem 1. Assume that the conditions (A1)-(A7) in Appendix I hold and the GLRT statistic λ n (h) is constructed from (4) with a local linear estimator.

Practical implications and further aspects
In order to use the asymptotic result proven in Theorem 1 in a practical setting, one needs to choose values for the bandwidth parameter and the order of the polynomial fitting.We give below guidelines for these choices and discuss other aspects relevant to the implementation of the GLRT.
Choice of bandwidth parameter.It should be noted that when η is linear, the asymptotic bias d n becomes exactly zero, as shown in (8) in the Appendix I, and thus the condition nh 9/2 → 0 is not needed (the optimal bandwidth for estimation is of the order n −1/5 , see Acar et al., 2011).More importantly, this facilitates the calculation of the GLRT statistic λ n (h) in practice, since one can use directly the bandwidth used for estimation, chosen by the leave-one-out cross-validated likelihood (Acar et al., 2011).Our simulation study in Section 3 provides empirical support for this suggestion.
Order of local polynomial fitting and testing polynomial functions.The asymptotic results in Theorem 1 can be easily extended to the case where λ n (h) is based on a pth order local polynomial estimator, by substituting the kernel function K with its equivalent kernel K * in c K and r K (see Fan and Gijbels, 1996, page 64, for the expression of K * ) induced by the local polynomial fitting (Fan et al., 2001).The asymptotic chi-square distribution (6) continues to hold if either η is a polynomial of degree p or nh (4p+5)/2 → 0, as the asymptotic bias ).The practical implication of such an extension is that, if the interest is to test a null hypothesis of a polynomial form η(x) = p ℓ=0 β ℓ x ℓ , it is recommended to calculate λ n (h) using the local polynomial estimator with the corresponding degree p.This avoids the possible necessity of undersmoothing in order to have the asymptotic bias negligible.
Testing constancy of the copula parameter.As pointed out earlier, the hypothesis of η being constant is a special case of the linearity constraint and leads to the classical copula model (i.e., no covariate adjustment is required).If this hypothesis is of interest, using a local constant estimator, i.e., p = 0, to calculate λ n (h) may be more appealing (as confirmed by the simulations in Section 3) than using a local linear estimator.The latter tends to overfit even with large bandwidth when H 0 indeed holds, thus resulting in an inflated type I error.
Testing independence.While the test is quite general and can be used, in principle, to test any preset copula parameter value, extra caution is needed when testing independence.Under the null hypothesis of independence, the copula parameter is at the boundary of the parameter space (e.g., θ = 0 for the Frank and Clayton families, and θ = 1 for the Gumbel family), and the asymptotic result in Theorem 1 does not hold.For the canonical likelihood ratio test, the asymptotic null distribution is known to be a mixture of chi-squares when some parameters lie on boundary of the parameter space (Self and Liang, 1987).A similar result is expected to hold for the GLRT as the two tests are fairly similar in nature (see the next paragraph).However, an investigation in this direction may not be too much of practical value, considering that there are a number of independence tests in the copula literature (see, for instance, Genest and Rémillard, 2004).
Relationship to the canonical likelihood ratio test.One can conclude from Theorem 1 that the GLRT is fairly similar to the classical likelihood ratio test.The tabulated value of the scaling constant r K is close to 2 for commonly used kernels.For instance, r K = 2.115 for the commonly used Epanechnikov kernel the asymptotic null distribution of the GLRT tends to infinity when h → 0, due to the nonparametric nature of the alternative hypothesis.One can interpret the quantity |X |/h as the number of nonintersecting intervals on X , and thus r K c K |X |/h approximates the effective number of parameters in the nonparametric estimation.For the Epanechnikov kernel with c K = 0.45, the degrees of freedom is given by 0.968 |X |/h.Impact of estimating conditional marginal distributions.An important aspect in (conditional) copula model implementation is the estimation of unknown (conditional) marginal distributions.Although the proposed GLRT procedure was presented assuming that the conditional marginal distributions are known, Theorem 1 provides a basis for more general approach where estimation is performed jointly or using a two-step method (i.e., parameters for the marginal distributions are estimated first and only subsequently the inference for the conditional copula parameters is performed).
In a conditional copula model with parametric conditional marginal distributions, one can easily accommodate joint estimation under the null hypothesis since the model for the calibration function is then parametric.On the other hand, joint analysis under the alternative requires iterative estimation of parametric conditional marginal distributions and nonparametric calibration function model.Although, such iterative procedures have been long studied for nonparametric regression problems, for instance in partially linear models, the problem of joint estimation has not been addressed yet in the conditional copula setting.
As shown in Appendix I, the asymptotic distribution of the GLR statistic λ n (h) is governed by the nonparametric part λ 1n (h) since the parametric part λ 2n , which corresponds to the canonical likelihood ratio statistic, vanishes compared to λ 1n (h).Hence, even one employs joint maximum likelihood estimation or a less efficient inference for margins approach (Joe, 2005) under the null hypothesis, the result in Theorem 1 will not change.Generally speaking, the proposed test remains valid as long as the conditional marginal distributions are estimated with the usual parametric rate √ n, both under the null and under the alternative (see Remark 1 at the end of Appendix I).A two-step approach can be safely employed in the testing procedure provided that the conditional marginal distributions are estimated parametrically.This observation, in fact, motivated the use of parametric models to specify the conditional marginal distributions in the data examples of Section 4. Note that an alternative two-step approach is to use kernel-based smoothing methods as in Abegaz et al. (2012) in the first stage.However, this case requires a careful treatment as both the null and the alternative models will have nonpararametric convergence rates due to nonparametric specification of the conditional marginal distributions.
Choice of copula family.The proposed GLRT approach assumes that the true copula family is used in the test procedure.If the copula is misspecified, the asymptotic result in Theorem 1 would not hold.This phenomenon is similar to the departures from the chi-squared asymptotic limit exhibited by the canonical likelihood ratio tests under misspecified models.Furthermore, copula misspecification can lead to serious bias in the estimation results.Nonetheless, in our simulations we have observed a good performance of the cross-validated prediction error criterion of Acar et al. (2011) in choosing the true copula family.

Simulation study
We conduct simulations to evaluate the finite sample performance of the proposed test for the linear hypothesis given in (2).We consider four simulation scenarios corresponding to four calibration functions, The copula used belongs to the Frank family and has the form Since the range of θ is R \{0} for the Frank copula, an identity link is used, i.e., θ k (X) = η k (X) for k = 0, 1, 2, 3. Similar findings, summarized in Appendix II, were obtained in simulations produced using the Clayton and Gumbel copulas in the true generating models.
As mentioned earlier, the copula family describes the overall shape of the dependence.For the Frank family, this shape is symmetric and shows no tail dependence as shown in the left panel of Figure 1.The four calibration models are displayed in the middle panel of Figure 1 over the covariate range (2, 5).For a scale-free interpretation of the strength of dependence, we also provide a graphical summary of the variation in Kendall's tau for each η (see the right panel of Figure 1) obtained using the conversion where k , we first simulate the covariate values X i from Uniform(2, 5).Then, for each i = 1, . . ., n, we calculate the copula parameter θ k (X i ) = η k (X i ) induced by the calibration model M (F ) k , and simulate the uniform pairs (U 1i , U 2i ) conditional on X i from the Frank family with copula parameter θ k (X i ).Throughout the simulations we have used the Epanechnikov kernel.For each Monte Carlo sample, the leave-one-out cross-validated likelihood method of Acar et al. (2011) is employed to select, out of 12 pilot values ranging from 0.33 to 2.96 and equally spaced in logarithmic scale, the optimum bandwidth h for the local polynomial estimation of the calibration function.
We have followed the suggestion made in Section 2 and have calculated the nonparametric estimator for η using a local polynomial of the same degree as specified by the null hypothesis.For instance, in Table 1, when testing H 0 : η = c, we consider a local constant estimator (with p = 0) for η under the alternative model.Subsequently, the GLRT statistic λ n (h) is computed using the same bandwidth h that is used for estimation.We also assume that in practice one would first test for constant calibration function and, conditional on rejection, would test for linear calibration.For this reason, in Table 1 we do not report the results of testing H 0 : η(x) = a 0 + a 1 x when the generating model is 1.00 1.00 1.00 1.00 1.00 1.00 One can notice from Table 1 that the rejection rates under the null are very close to the target values of the type I error probabilities α ∈ {0.1, 0.05, 0.01}, for both linear and constant nulls (models M (F ) 0 and M (F ) 1 ), except for the case with n = 50, which is slightly conservative.Since nonparametric methods often require relatively large samples, the latter observation confirms that caution is recommended when working with small samples.Overall, our approach leads to high power in detecting departures from the null, as one can see from the results generated under models and M (F ) 3 .For clearer visualization, the entries in the table that correspond to power are shown in bold face.As expected, the rejection rates increase with the sample size and the nonlinearity of the underlying calibration function (see the middle or right panel of Figure 1 for a visual comparison).For instance, when testing the linear calibration hypothesis, we observe lower rejection rates for M (F ) 2 , which is quadratic, than for M (F ) 3 , which exhibits more variation.

Data application
In this section, we apply the GLRT to the two data examples studied in Acar et al. (2011).Our aim is to check whether a constant copula model or a conditional copula model with a linear calibration function fits these examples reasonably well, i.e. whether the nonparametric calibration estimates of Acar et al. (2011) are in fact necessary.

Twin birth data
This data set contains information on 450 pairs of twins from the Matched Multiple Birth Data Set (MMB) of the National Center for Health Statistics.We consider the birth weights (BW 1 , BW 2 ) of the first-and second-born twins who survived their first year, and whose mothers were between 18 and 40 years old.The gestational age GA is an important factor for fetal growth and is therefore included in the analysis as a covariate.The question of interest here is whether the dependence between the weights BW 1 and BW 2 varies with GA.Subsequent interrogations regarding the exact parametric forms of η are much less relevant.We transform the data on the uniform scale, as shown in the left panel of Figure 2, after fitting parametric marginal regression models.We use the Frank family of copulas to model the dependence structure, as this was the family chosen by Acar et al. (2011) according to their cross-validated prediction error criterion.The middle panel of Figure 2 shows the maximum likelihood estimates obtained under the constant calibration assumption (solid line), linear calibration assumption (long-dash line), the nonparametric estimates with p = 0 (dot-dashed line), p = 1 (dashed line) and 90% pointwise confidence intervals for the local linear estimates (dotted lines), obtained as in Acar et al. (2011).These results are also displayed in terms of Kendall's tau in the right panel of Figure 2. As seen in Figure 2, the maximum likelihood estimates under constant and linear calibration assumptions are not within the pointwise confidence intervals of the local linear estimates, suggesting that these simple parametric formulations may not be appropriate.This empirical observation is confirmed by the GLRT tests, which yielded p-values smaller than 10 −5 for both tests (test statistics are 13.58 on 3.92 df and 12.95 on 3.36 df for the constant and linear hypothesis, respectively).
Thus, we conclude that the variation in the strength of dependence between the twin birth weights at different gestational ages, as represented by the non- parametric estimates in the right panel of Figure 2 is statistically significant.The estimated nonlinear pattern indicates a relatively stronger dependence between the birth weights of the preterm (28-32 weeks) and post-term (38-42 weeks) twins compared to the twins delivered at term (33-37 weeks).While the factors affecting the twin fetal growth patterns are not fully known, the relatively stronger intra-twin dependence for preterm twins may be due to the fact that fat accumulation only begins in the third trimester of gestation (around 28 weeks).Hence, during the weeks 28-32, twins are expected to be still very similar in their growth.The increase in the strength of dependence after week 38, on the other hand, is more puzzling, but perhaps can be explained by intrauterine growth restriction factors.

Framingham heart study data
This data set comes from the Framingham Heart Study (FHS) and contains the log-pulse pressures of 348 subjects at the first two examination periods (1956 and 1962), denoted by log(P P 1 ) and log(P P 2 ), respectively, as well as the change in body mass index ∆BMI between these periods.Pulse pressure, defined as the difference between systolic and diastolic blood pressure, reflects arterial stiffness and is associated with an increased risk in stroke incidence.
For the 348 subjects, who experienced a stroke during the rest of follow-up period, of interest is to investigate the dependence structure between log(P P 1 ) and log(P P 2 ) conditional on ∆BMI.The left panel of Figure 3 displays the conditional marginal distributions of the log-pulse pressures given ∆BMI, which are obtained parametrically as in Acar et al. (2011).
In conditional copula selection, Acar et al. (2011) used the cross-validated prediction errors only for the second log-pulse pressure as the selection criterion, and chose the Frank copula family.Proceeding with their choice, we obtained the calibration function estimates using the maximum likelihood estimation with constant and linear calibration forms and the nonparametric estimation with p = 0 and p = 1.The results are shown in the middle panel of Figure 3, and converted to the Kendall's tau scale in the right panel.
Based on the Figure 3 we suspect that a constant copula model may be appropriate, even though it is not fully contained in the pointwise confidence intervals.Note that the latter provides only partial guidance, and simultaneous confidence intervals are needed for a decisive visual conclusion.To decide whether the fitted constant copula model is appropriate, we perform the GLRT using the local constant estimates at the bandwidth value h = 3.45.This bandwidth choice leads to 2.66 df of the chi-square distribution.The difference between the loglikelihoods of the alternative and null conditional copula models is 0.91 and consequently the p-value is 0.514.Thus, we conclude that the change in body mass index does not have any significant effect on the strength of dependence between the two log-pulse pressures.

Conclusion
Adjusting statistical dependence for covariates via conditional copulas is an active area of research where model fitting and validation are currently in early development.This paper takes a first step towards establishing conditional copula model diagnostics by presenting a formal test of hypothesis for the calibration function.Inspired by the generalized likelihood ratio idea of Fan et al. (2001), the proposed test uses the local likelihood estimator of Acar et al. (2011) to specify the model under the alternative when testing a parametric calibration function hypothesis.The asymptotic null distribution of the test statistic, shown to be a chi-squared distribution with the number of degrees of freedom determined by the estimation-optimal bandwidth, is used to determine the rejection region in finite samples.Simulations suggest that the method has high power of detecting departures from the null model and yields the targeted type I error probability.
The GLRT procedure presented here can be easily adapted to test an arbitrary parametric calibration function.Furthermore, the approach can be extended to employ other nonparametric estimators, such as smoothing splines, although with additional effort of deriving the asymptotic null distribution.Nevertheless, the asymptotic null distribution may not always be appropriate for determining the rejection region in finite samples.While conditional bootstrap is usually used to assess the null distribution of the GLRT in regression-based problems, defining a similar bootstrap procedure in the conditional copula setting is not straightforward and requires further study.
Although the focus of the paper is on bivariate copulas, the GLRT approach is potentially applicable to multivariate copulas.However, complications arise in the latter case since the number of copula parameters is likely to increase and simultaneous testing is therefore necessary.This represents an interesting direction for future research.
In addition, for some constants ξ > 2 and k 0 > 0, j = 1, 2, 3, Before proving Theorem 1, we shall introduce additional notation.Let γ n = 1/ √ nh and define and set Lemma 1-3 are used in our derivations, and their proofs are given at the end of this appendix.
Lemma 2. Under conditions (A1)-(A7), as h → 0 and nh 3/2 → ∞ To introduce Lemma 3, we first restate a proposition in de Jong (1987), where the notation is adapted to ours.Let X 1 , X 2 , . . .be independent variables, and w ijn (•, •) Borel functions such that W (n) = 1≤i≤n 1≤j≤n w ijn (X i , X j ), and W ij = w ijn (X i , X j ) + w jin (X j , X i ), where the index n is suppressed in W ij .Following de Jong (1987, Definition 2.1), W n is called clean if the conditional expectations of W ij vanish: E[W ij |X i ] = 0 a.s.for all i, j ≤ n.Proposition 3.2 (de Jong, 1987) Let W (n) be clean with variance ν * n , if G I , G II and G IV be of lower order than ν * 2 n , then where We now define the following U-statistic, Lemma 3.Under conditions (A1)-(A7), W n defined in (7) is clean and, as h → 0 and Proof of Theorem 1.To provide a general framework, we use η(X k ) and η(X k ) to denote the true value under the null hypothesis and its maximum likelihood estimator, respectively.Then, the GLRT statistic can be written as Here λ 2n corresponds to the canonical likelihood ratio statistic and it is λ 1n (h) that governs the asymptotic distribution of λ n (h).
To derive the asymptotic distribution of λ 1n (h), first approximate ℓ Applying Lemma 1 and 2 yields By calculating of the leading terms R n1 , R n2 and R n3 , one can show that where ω 0 = t 2 (s + t) 2 K(t)K(s + t) ds dt.Thus, This results in where W n is as defined in (7).Applying Lemma 3, we arrive at W (n) where ν n = (4h) −1 ν * .For the asymptotic null distribution of λ n (h), this result can be re-written as Since λ 2n = O p (1), it vanishes compared to λ 1n (h) = O p (h −1 ) and we obtain For the second result, note that the distribution N (a n , 2a n ) is approximately same as the chi-square distribution with degrees of freedom a n , for a sequence a n → ∞.Letting a n = 2µ 2 n /ν n and r K = 2µ n /ν n , we have so that each component has the same rate of convergence.Then, we have where z i,x = (1, (X i −x)/h) T .The local log-likelihood function can be re-written in terms of b, which can be written as In the following, we will show that ), with µ i = t i K(t)dt, and o p (1) is uniform in x ∈ X and ||b|| < m 0 , for some fixed constant m 0 > 0. To show this, we need the following smoothness result.Let Then, under the conditions (A1)-(A6), we can show that for some integrable function k(X, U 1 , U 2 ).Thus, using the triangle inequality, for η in a compact set and x ∈ X .The first sum goes to zero by the previous argument and the Dominated Convergence theorem.Similarly, the second sum converges to zero provided that hn (ξ−2)/ξ = O(1) and ||b|| < m 0 , for some fixed constant m 0 > 0. The first part in the last term goes to zero with probability one by the uniform weak law of large numbers and the second part vanishes by direct calculation.We thus obtain uniformly for x ∈ X , where Using the quadratic approximation lemma (Fan and Gijbels, 1996, p. 210) provided that W n is a stochastically bounded sequence of random vectors.The first entry of b directly yields the result, i.e.
Note that, when η is linear, then the second sum directly becomes zero as for each This is clearly also the case when η is constant.
Proof of Lemma 2. Note that The approximation of the first term yields the first result.We can decompose T n2 = T n21 + T n22 , where We deal with T n21 and T n22 separately.For T n21 , note that The first sum can be shown to be Therefore, let and the second sum becomes (V n + o(1))/2 + O p n −3/2 h −2 + o p (h −1/2 ).The decomposition theorem for U-statistics (Hoeffding, 1948) allows us to show that V ar(V n ) = O(n −1 h −2 ) as follows.First note that the leading term of V n is −h −1 E f −1 (X) K 2 (t)dt.Hence, as nh → ∞ and h → 0, we obtain Similarly, we can decompose T n22 = T n221 + T n222 with For k = i, j, define It can be easily shown that V ar(n −1 k =i,j Q ijk,h ) = O(n −1 h −2 ).Then, where E(Qijk,h|Xi, X j ) = − {h σ 2 (X i )f (X i )} −1 K(t) K((X j − X i )/h)dt.
Combining T n21 , T n221 and T n222 yields Proof of Lemma 3. Recall that We shall show that W n satisfies conditions in Proposition 3.2.Let

Fig 1 .
Fig 1. Contour plot of the density of Frank copula under M (F ) 0 , illustrated with standard normal marginal distributions (left panel), and graphical summary of the four calibration models (middle panel) and the corresponding Kendall's tau values (right panel).
1 dt is the Debye function.Our Monte Carlo experiment consists of 200 replicated samples of sizes n = 50, 100, 200 and 500 generated from each calibration model M

Fig 2 .
Fig 2. Scatterplot of transformed birth weights (left panel), and the plots of calibration function estimates (middle panel) and the corresponding Kendall's tau values (right panel) under the Frank copula: maximum likelihood estimate of the constant calibration function (solid line), maximum likelihood estimate of the linear calibration function (long-dashed line), local constant estimates (dot-dashed line), local linear estimates (dashed line), 90% pointwise confidence intervals for the local linear estimates (dotted lines).

Fig 3 .
Fig 3. Scatterplot of the transformed log-pulse pressures (left panel), and the plots of calibration function estimates (middle panel) and the corresponding Kendall's tau values (right panel) under the Frank copula: maximum likelihood estimate of the constant calibration function (solid line), maximum likelihood estimate of the linear calibration function (long-dashed line), local constant estimates (dot-dashed line), local linear estimates (dashed line), 90% pointwise confidence intervals for the local linear estimates (dotted lines).

Table 1
Demonstration of the proposed GLRT for testing the linear/constant null hypothesis H 0 at α = 0.10, 0.05 and 0.01, respectively, under the Frank copula.Shown are the rejection frequencies assessed from 200 Monte Carlo replicates.The sample sizes are n = 50, 100, 200 and 500; the generating calibration models are shown in the "True Model" column.Those entries in the table reflecting the power of the testing procedure are shown in bold face