Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
38,711 characters · 10 sections · 0 citation commands
Nonparametric Tests of Tail Behavior in Stochastic Frontier Models
\doublespacing
Stochastic frontier analysis (SFA) has a vast literature, both methodological and applied, and empiricists have applied the methods to myriad industries, most notably agriculture, banking, education, healthcare, and energy. A common practice in SFA is to impose parametric assumptions on the error components, but the set of statistical tools to investigate the validity of these assumptions is still limited. This paper expands this set of tools by drawing on recently developed techniques in Extreme Value (EV) theory and by developing new diagnostic tests.
In particular, the parametric stochastic frontier model for cross-sectional data (Aigner et al. 1997) is a leading case of the error component regression model but with the unique feature that one error component ($U$) is a non-negative random variable (e.g., half-normal, exponential), while the other ($W$) is a random variable of unbounded support (e.g., normal, Laplace, Student-t). A common assumption in the stochastic frontier literature is that $W$ is drawn from a normal or Laplace distribution (both thin-tailed distributions). See Aigner et al. (1977) or Horrace and Parmeter (2018), respectively.\footnote{For other parametric specifications of the model see Li (1996), Carree (2002), Tsionas (2007), Kumbhakar et al. (2013), and Almanidis et al. (2014). } However, heavy-tailed distributions are now also being considered. For example, the findings of Wheat et al. (2019) suggest that a cost inefficiency model of highway maintenance costs in England has Student-t errors.\footnote{There are semi-parametric estimators of the model that relax the distributional assumptions on one component and estimate the density of the other using kernel deconvolution techniques. See Kneip et al. (2015), Horrace and Parmeter (2011), Cai et al. (2020), Simar et al. 2017, Hall and Simar (2002), Florens et al. (2020).} These parametric distributions, such as normal and Student-t, display similar patterns in the middle of their supports but exhibit substantially different tail behaviors. This observation motivates and plays an essential role in our diagnostic tests, which we believe are a timely and appropriate contribution to the literature.
The key idea of our test is as follows. Assuming independence of the error components, the largest order statistics of the composed error term ($Z = W-U$) (approximately) arise from the tails of $W$, because $U$ is one-sided. Also, assuming that $W$ is in the domain of attraction (DOA) of extreme value distributions, the asymptotic distribution of the largest order statistics of $W$ is the EV distribution, which may be fully characterized (after location and scale normalization) by a single parameter that captures its tail heaviness.\footnote{The assumption that $W$ is in the DOA of extreme value distributions is not restrictive, as we shall see.} Then, likelihood ratio statistics for hypotheses on this single parameter can be derived based on the limiting EV distribution.
To be specific, consider the right tail of $W$. If the DOA assumption is satisfied, then tail behavior may be entirely characterized by a tail index, $\xi \in \mathbb{R}$. If $\xi = 0$, then $W$ has thin tails. If $\xi > 0$, then $W$ has thick tails. Otherwise, $W$ has bounded support. Under very weak assumptions on the error components, we derive a test that the tails of $W$ are thin ($\xi=0$). We prove that this test is valid whether $Z$ is observed or appended to a regression model (as it is in the stochastic frontier model). If we assume that $U$ is also in the DOA of extreme value distributions and that $W$ is symmetric (a common assumption), we also derive a test that the (right) tail of $U$ is thinner than the left tail of $W$. If we further assume that $W$ is a member of the normal family, then we may test the hypotheses that the tails of $U$ and $W$ are both thin. Therefore, our nonparametric tests are useful diagnostic tools to help empiricists make parametric choices on the distributions of both $U$ and $W$. This is particularly important for the stochastic frontier model for cross-sectional data, where distributional assumptions on the components are typically necessary for the identification of the model's parameters.
The paper is organized as follows. The next section presents the tests. Section 3 provides a simulation study of the power and size of the test. Section 4 applies the tests to a stochastic cost function for a panel of US banks 1998-2005, revealing that the tails of $W$ are not thin. Therefore, a normal or Laplace assumption for $W$ is not justified, and perhaps a Student-t assumption may be appropriate. Section 5 concludes.
To fix ideas, we begin a review of the DOA assumption and present the test in the case where $Z$ is directly observed in Section (ref). Then in Section (ref), we move to the case where $Z$ is appended to a regression model and has to be estimated, which covers the linear regression stochastic frontier model. Additional tests under different sets of weak assumptions are also presented.\footnote{While the analyses that follow are for cross-sectional data, they can easily be applied to panel data, as long as one is willing to assume independence in both the time and cross-sectional dimensions.}
Consider a random sample of $Z_{i}=W_{i}-U_{i}$ for $i=1,\ldots ,n$, where $U_{i}\geq 0$ represents inefficiency, and $W_{i}\in \mathbb{R}$ is noise with unbounded support. We start with testing the shape of the right tail of $W_{i}$ in a nonparametric way.
The key assumption is that the distribution of $W_{i}$ is within the domain of attraction of EV distributions. In particular, a cumulative distribution function $F$ is in the domain of attraction of $G_{\xi }$, denoted as $F\in \mathcal{D} \left( G_{\xi }\right) $, if there exist constants $a_{n}>0$ and $b_{n}$ such that
where $G_{\xi }$ is the generalized EV distribution,
and $\xi $ is the tail index, measuring the decay rate of the tail.
The domain of attraction condition is satisfied by a large range of commonly used distributions. If $\xi $ is positive, this condition is equivalent to regularly varying at infinity, i.e.,
This covers Pareto, Student-t\footnote{The tail index of the Student-t distribution with $\nu$ degrees of freedom is $\xi =1/\nu$.}, and F distributions, for example. The case with $\xi =0$ covers the normal family, and the case with $\xi <0$ corresponds to distributions with a bounded support.\footnote{ The uniform distribution has $\xi =-1$, and the triangular distribution has $ \xi =-1/2$.} See de Haan and Ferreira (2007), Ch.1 for a complete review.
Note that the above notation is for the right tail of $W$, which can be easily adapted to the left tail by considering $-W$. For expositional simplicity, we denote $\xi_{W_{-}}$ and $\xi_{W_{+}}$ as the tail indices for the left and right tails of $W$, respectively. The same notation applies to other variables (e.g., $U$ and $Z$) introduced later.
Returning to SFA, a common assumption is that $W_{i}$ is normal or Laplace, which implies that $\xi_{W_{+}} =0$. So our hypothesis testing problem is as follows:
If the null hypothesis is rejected, we would then argue that some heavy-tailed distribution should be used to model the noise and maybe the inefficiency as well.
To obtain a feasible test, we argue that, since $U_{i}$ is bounded from below at zero, the largest order statistics of $ Z_{i} $ are approximately stemming from the right tail of $W_{i}$. This is formalized in Proposition (ref), which requires the following conditions. Let $Z_{n:n}\geq ,\ldots ,\geq Z_{n:1}$ be the order statistics of $\{Z_{i}\}_{i=1}^{n}$ by descend sorting. Denote
as the $k$ largest observations. From now on, we use bold letters to denote vectors. Denote $F_{W}$ and $Q_{W}(p)=\inf \{y\in \mathbb{R} \left. :\right. p\leq F_{W}(y)\}$ as the CDF and the quantile function of $ W_{i}$, respectively. Write $Q_{W}(1)$ as the right end-point of the support of $W_{i}$. For a generic column vector $\mathbf{X}$ and scalar $c$, the notation $ \mathbf{X}-c $ means $\mathbf{X}-(c,\ldots ,c)^{\intercal }$.
\paragraph{Assumption 1}
Assumptions 1(i)-(iii) are common in the SFA literature (see Horrace and Parmeter (2018) and the references therein). Assumption 1(iv) requires the tail of $F_{W}$ to be within the domain of attraction of EV distributions with an infinite upper bound. Moreover, it requires that the density derivative monotonically increases to zero. This is a mild assumption and is satisfied by many commonly used distributions. For example, the normal distribution is covered as seen by
and the Pareto distribution is covered as seen by
Under Assumption 1, the following proposition derives the asymptotic distribution of $\mathbf{Z}_{+}$.
The proof is in Appendix A. This proposition implies that the distributions of $Z_{i}$ and $W_{i}$ share the same (right) tail shape, which is entirely characterized by the tail index $\xi_{W_{+}} $. Such tail equivalence does not hold, however, for the left tails due to the existence of $U$. This is studied in Section (ref) under the additional assumption that $W$ is symmetric.
If the constants $a_{n}$ and $b_{n}$ were known, $\mathbf{Z_{+}}$ is then approximately distributed as $\mathbf{V_{+}}$, and the limiting problem is reduced to the well-defined finite sample problem: constructing some inference method based on one draw $\mathbf{V_{+}}$ whose density $f_{\mathbf{V_{+}}|\xi_{W_{+}} }$ is known up to $\xi_{W_{+}} $. However, $a_{n}$ and $b_{n}$ depend on $F_{W}$ and hence are unknown a priori.
To avoid the need for knowledge of $a_{n}$ and $b_{n}$, we consider the following self-normalized statistic
It is easy to establish that $\mathbf{Z}^{\ast }$ is maximally invariant with respect to the group of location and scale transformations (cf., Lehmann and Romano (2005), Ch.6). In words, the estimator constructed as a function of $\mathbf{Z}^{\ast }$ remains unchanged if data are shifted and multiplied by any non-zero constant. This makes senses since the tail shape should be preserved no matter how data are linearly transformed. This invariance property allows us to construct nonparametric tests for a stochastic frontier model that is otherwise not identified without parametric assumptions on $U$ and $W$.\footnote{In particular the non-zero expectation of $U$ precludes identification of unknown parameter $\delta$ in the model $Z_i=\delta+W_i-U_i$. } As such, our tests do not reveal anything about the location or the scale of the error components.
The continuous mapping theorem and Proposition (ref) imply that for any fixed $k$, as $n\rightarrow \infty $,
The CDF\ of $\mathbf{V}_{+}^{\ast }$ can be calculated via change of variables as
where $\mathbf{v}_{+}^{\ast }=(v_{1}^{\ast },\ldots ,v_{k}^{\ast })$, $ b_{0}\left( \xi \right) =\infty $ if $\xi \geq 0$ and $-1/\xi $ otherwise, and $\Gamma \left( k\right) $ is the gamma function. Note that the invariance restriction costs two degrees of freedom since the first and last elements of $\mathbf{V}_{+}^{\ast }$ are always 1 and 0, respectively. We calculate this density by numerical quadrature.
Given $f_{\mathbf{V}_{+}^{\ast }\mathbf{|}\xi_{W_{+}} }$, we can construct the generalized likelihood-ratio test for problem ((ref)). Since the alternative hypothesis is composite, we follow Andrews and Ploberger (1994) and Elliott et al. (2015) to consider the weighted average alternative
where $w(\cdot )$ is a weighting function that reflects the importance of rejecting different alternative values. Then our test is constructed as \footnote{In later sections, we set $w(\cdot )$ to be the standard uniform distribution over $(0,1)$ for simplicity.}
where the critical value $\mathrm{cv}(k,\alpha )$ depends on $k$ and the level of significance $\alpha $. We can obtain it by simulation. By Proposition (ref) and the continuous mapping theorem, this test controls size asymptotically as $\lim_{n\rightarrow \infty }\varphi (\mathbf{Z}_{+}^{\ast })=\alpha$.
We end this subsection by briefly discussing the choice of $k$, that is, the number of the largest order statistics used to approximate the EV distribution. On the one hand, larger $k$ means including more mid-sample observations, which induces a larger finite sample bias in the EV approximation. On the other hand, smaller $k$ provides a better asymptotic approximation but uses less sample information, leading to a lower power test. This trade-off leads to difficulty in theoretical justification of an optimal $k$ in standard EV theory literature (cf., M\"{u}ller and Wang (2017)). It is even more difficult, if at all possible, in our case, since we only observe $Z$, and not $W$. Nonetheless, our asymptotic arguments show that the test ((ref)) controls size for any fixed $k$, as long as $n$ is sufficiently large. Figure (ref) depicts the asymptotic power of the test ((ref)) with $\mathbf{V}_{+}^{\ast}$ generated from the density ((ref)) based on 10,000 simulation draws. The test controls size for all values of $k$ by construction and has reasonably large power when $k$ exceeds 20.
With ideas fixed, we now turn to the regression version of the test, with application to SFA.
Now consider the linear regression with
where $Z_{i}=-U_{i}+W_{i}$ is as in the previous section, and $\beta _{0}$ is some pseudo-true parameter in some compact parameter space. This could be a Cobb-Douglas production function (in logarithms), where $Y$ is productive output and $U$ is now called technical efficiency, which measures distance ($U_i$) from a stochastic frontier ($\mathbf{X}_{i}^{\intercal }\beta _{0}+W_{i}$). The slopes ($\beta_0$) are marginal products of the productive inputs, $\mathbf{X}_{i}$. It could also be a stochastic cost function if we multiply $U$ by $-1$. Suppose we have some estimator, $\hat\beta$ of $\beta_0$. The following assumption is imposed to construct our diagnostic test.
\paragraph{Assumption 2}
Assumption 2 is similar to Assumption 1 with additional restrictions on the covariate $\mathbf{X}$. In particular, Assumption 2(v) bounds the norm of $\hat{\beta}$ and $\left\vert \left\vert \mathbf{X}_{i}\right\vert \right\vert $. A sufficient condition when $\xi_{W_{+}} $ is positive is that $\left\vert \left\vert \hat{\beta}-\beta _{0}\right\vert \right\vert =O_{p}(n^{-1/2})$ and $\sup_{i}\left\vert \left\vert \mathbf{X} _{i}\right\vert \right\vert =o_{p}(n^{1/2})$, which is easily satisfied in many applications.\footnote{Even though $\mathbb{E}\left[ \left\vert U_{i}\right\vert \right] \ne 0$, ordinary least squares (OLS) will typically suffice for $\hat\beta$, because our test is invariant to relocation. } When $\xi $ is zero, we need slightly stronger bounds. Straightforward calculations show that the normal distribution satisfies Assumption 2(v) for the $\xi_{W_{+}} = 0$ case, if $\left\vert \left\vert \hat{\beta}-\beta _{0}\right\vert \right\vert =O_{p}\left( n^{-1/2}\right) $ and $ \sup_{i}\left\vert \left\vert \mathbf{X}_{i}\right\vert \right\vert \left. =\right. O_{p}(n^{1/2-\varepsilon })$ for some $\varepsilon >0$. This is seen by $1/f_{W}\left( Q_{W}\left( 1\left. -\right. 1/n\right) \right) \left. \leq \right. O(\log (n))$ (cf.\ Example 1.1.7 in de Haan and Ferreira (2007)).
Denote $\hat{Z}_i$ as the OLS residuals and
the largest $k$ order statistics. Then given Assumption 2, the following proposition derives the asymptotic distribution of $\mathbf{\hat{Z}}_{+}$
The proof is in Appendix A. Proposition (ref) implies that the largest order statistics of the regression residuals satisfy the same convergence as the no-covariate case. In other words, the estimation error from the OLS becomes negligible so that the largest order statistics are stemming from the right tail of $W$ asymptotically. This validates the construction of the test ((ref)) by replacing $\mathbf{Z}_{+}^{\ast }$ with $\mathbf{\hat{Z}}_{+}^{\ast }$, where
Proposition (ref) and the continuous mapping theorem, we similarly have $\lim_{n\rightarrow \infty }\varphi (\mathbf{\hat{Z}_{+}}^{\ast })=\alpha$.
The previous analysis studies the right tail of $W$ (and equivalently $Z$). Suppose we assume $W$ has a symmetric distribution, then the tail indices of both tails of $W$ become equivalent, and hence we can learn about the tail of $U$ using the left tail index of $Z$. To this end, we make the following additional assumption.
\paragraph*{Assumption 3}
Assumption 3(i) implies that $\xi_{W_{-}}=\xi_{W_{+}}$, and the condition that $U>0$ implies its left tail index is negative. Therefore, in this subsection only, we simply denote $\xi_{U}$ and $\xi_{W}$ as the right tail indices of $U$ and $W$, respectively. Now we can test if $U$ has a thinner or equal right tail than $W$ by specifying the following hypothesis testing problem,
Moreover, if $W$ is in the normal or Laplace family ($\xi _{W}=0$), since we limit the tail indices to be non-negative, the null hypothesis then reduces to $\xi _{U}=\xi _{W}=0$.
Under the null hypothesis of ((ref)), $W$ is the leading term in $Z$ in both the left and right tails. Then the DOA assumption for both $W$ and $U$ implies that $\xi _{Z_{-}}=\max \{\xi _{U},\xi _{W}\}$, and Proposition (ref) entails $\xi _{Z_{+}}=\xi _{W}$. Therefore, the above testing problem becomes equivalent to
We now construct a test for ((ref)). Define $\mathbf{\hat{Z}} _{-}$ as the smallest $k$ order statistics of the estimation residuals, that is,
and its self-normalized analogue as
The following proposition establishes that $\mathbf{\hat{Z}}_{-}^{\ast }$ asymptotically has the EV distribution with tail index $\xi _{Z_{-}}$ and is independent from $\mathbf{\hat{Z}}_{+}^{\ast }$.
The proof is in Appendix A. Given the above proposition, we aim to construct a generalized likelihood ratio test for ((ref)) as follows,
where $\Xi$ denotes the parameter space of the tail indices, and $w\left( \cdot ,\cdot \right) $ is the weighting function for the alternative hypothesis as in ((ref)). We set $\Xi$ to be $[0,1)$ to cover all distributions with a finite mean and $w(\cdot)$ to be uniform over the alternative space. The weight $\Lambda \left( \cdot \right) $ can be considered as the least favorable distribution, which we discuss more now.
Note that the null hypothesis of ((ref)) is composite. We need to control size uniformly over all $\xi _{Z_{-}}=\xi _{Z_{+}}\in \Xi $. To that end, we can transform the composite null into a simple one by considering the weighted average density with respect to the weight $\Lambda $. Together with a suitably chosen the critical value, this test ((ref)) maintains the uniform size control. Now the problem reduces to determining an appropriate weight $\Lambda$. Elliott et al. (2015) study the generic hypothesis testing problem where a nuisance parameter exists in the null hypothesis. We tailor their argument for our test ((ref)) and adopt their computational algorithm for implementation. In particular, $\Lambda \left( \cdot \right) $ and $\mathrm{cv}(k,\alpha )$ are numerically calculated only once by the authors instead of the empiricists who use our test. They only need to construct the order statistics $\mathbf{\hat{Z}}_{-}^{\ast }$ and $\mathbf{\hat{Z}}_{+}^{\ast }$ and numerically evaluate the density. We provide more computational details in the Appendix and the corresponding MATLAB code in the supplemental materials. By the continuous mapping theorem and Proposition (ref), for any fixed $k$ , $\lim \sup_{n\rightarrow \infty }\mathbb{E}\left[ \varphi _{\pm }\left( \mathbf{\hat{Z}}_{-}^{\ast },\mathbf{\hat{Z}}_{+}^{\ast }\right) \right] \leq \alpha $ under the null hypothesis of ((ref)).
As we discussed above, the hypothesis testing problem ((ref)) simplifies to
if $W$ is assumed to be in the normal family ($\xi_{W}=0$). Proposition (ref) implies $\mathbf{\hat{Z}}_{-}^{\ast }$ and $\mathbf{\hat{Z}}_{+}^{\ast }$ are asymptotically independent and both of them are EV distributed. Then accordingly, our test ((ref) ) reduces to
which is identical to ((ref)). This suggests that we can simply substitute $ \mathbf{\hat{Z}}_{-}^{\ast }$ into ((ref)) for implementation.
We set $w(\cdot )$ to be the uniform weight on $\left[ 0,0.99\right] $ to include all distributions with a finite mean and the level of significance to be $0.05$. In Table 1, we report the small sample rejection probabilities of the test ((ref)). We generate $U_{i}$ from the right half-standard normal and the right half-Laplace(0,1) distributions and $W_{i}$ from four distributions: standard normal, Laplace(0,1) (denoted La(0,1)) Student-t(2), Pareto(0.5) and F(4,4). The normal and Laplace distributions correspond to the null hypothesis, and the other three are alternative hypotheses. The results suggest that the test ((ref)) has an excellent performance in size and power. Note that when $k=50$ and $n=100$, we essentially include too many mid-sample observations so that the EV approximation is poor.
Now we consider the linear regression model that $Y_{i}=\mathbf{X} _{i}^{\intercal }\beta _{0}+Z_{i}$ with $\mathbf{X}_{i}=\left( 1,X_{2i}\right) ^{\intercal }$ and $\beta _{0}=\left( 1,1\right) ^{\intercal }$. We assume $X_{2i}\sim \mathcal{N}\left( 0,1\right) $ and independent from $Z_{i}$. Table 2 reports the rejection probabilities of our test ((ref)). Findings are similar to those in Table 1.
Consider the hypothesis testing problem ((ref)). We implement the test ((ref)) with the same setup as above. Table 3 reports the rejection probabilities under the null and alternative hypotheses. We make the following observations. First, the test controls size well unless $k$ is too large relative to $n$, as seen in the column with $n=100$ and $k=50$. This is again because we are using too many mid-sample observations to approximate the tail so that the EV convergence in Propositions 1-3 provides poor approximations. Second, the test has good power properties as seen from the last five rows. In particular, using only the largest 50 order statistics from 1000 observations leads to the power of 0.94. Finally, the power decreases as the alternative hypothesis becomes closer to the null, as we move down along rows.
Now we consider the special case where $W$ is in the normal family. Then we implement ((ref)) with $\mathbf{\hat{Z}}_{-}^{\ast }$ as the input. Table 4 contains the rejection probabilities under the null and alternative hypotheses. The rows with $F_{U}$ being half-normal or Laplace correspond to the size under the null hypothesis, while other rows the power under the alternative hypothesis. The new test has excellent size and power properties.
We illustrate the new method using the US bank data collected by Feng and Serletis (2009). The data are a sample of US banks covering the period from 1998 to 2005 (inclusive). After deleting banks with negative or zero input prices, we are left with a balanced panel of 6,010 banks observed annually over the 8-year period. A more detailed description of the data may be found in Feng and Serletis (2009). Here we specify a stochastic cost function, letting $Z=W+U$, so $U \geq 0$ is cost inefficiency, and more inefficient banks have higher total costs, $Y$. Since our tests are designed for cross-sectional data, we divide the original panel data into cross-sections (one for each year) and regress the logarithm of total bank cost on a constant and the logarithms of six control variables, including the wage rate for labor, the interest rate for borrowed funds, the price of physical capital, and the amounts of consumer loans, non-consumer loans, and securities. Since the object of interest is the cost function, we multiply the OLS residuals by $-1$ and take the smallest and the largest $k\in\{25, 50, 75, 100\}$ order statistics, respectively, to implement the test ((ref)). The p-values are reported in Table 5. Under the assumption that $W$ is symmetric\footnote{The symmetry assumption is reasonable here and is imposed in Feng and Serletis (2009).}, these small p-values suggest that $W$ has heavy tails on both sides, so a Student-t assumption (e.g., Wheat, Stead, and Greene, 2019) is more appropriate.
We derive several nonparametric tests of the tail behavior of the error components in the stochastic frontier model. The tests are easy to implement in MATLAB and are useful diagnostic tools for empiricists.
Often a first-step diagnostic tool for SFA is to calculate the skewness of the OLS residuals to see if they are properly skewed. See Waldman (1982), Simar and Wilson (2010), and Horrace and Wright (2020). If they are positively skewed, the maximum likelihood estimator of the variance of inefficiency is zero, and OLS is the maximum likelihood estimator of $\beta_0$. If they are negatively skewed, then OLS is not a stationary point in the parameter space of the likelihood, and the stochastic frontier model is well-posed. After calculating negatively skewed OLS residuals, a useful second-step diagnostic tool is to implement our nonparametric tests to understand the tail behaviors of the error component distributions and to guide parametric choices subsequently .