# A Monte Carlo Study of Time Varying Coefficient (TVC) Estimation

- 185 Downloads

## Abstract

A number of recent papers have proposed a time-varying-coefficient (TVC) procedure that, in theory, yields consistent parameter estimates in the presence of measurement errors, omitted variables, incorrect functional forms, and simultaneity. The key element of the procedure is the selection of a set of driver variables. With an ideal driver set the procedure is both consistent and efficient. However, in practice it is not possible to know if a perfect driver set exists. We construct a number of Monte Carlo experiments to examine the performance of the methodology under (i) clearly-defined conditions and (ii) a range of model misspecifications. We also propose a new Bayesian search technique for the set of driver variables underlying the TVC methodology. Experiments are performed to allow for incorrectly specified functional form, omitted variables, measurement errors, unknown nonlinearity and endogeneity. In all cases except the last, the technique works well in reasonably small samples.

## Keywords

Time-varying coefficients Specification errors Monte Carlo## JEL Classification

C130 C190 C220## 1 Introduction

A series of papers have proposed the use of time-varying coefficient (TVC) models to uncover the bias-free estimates of a set of model coefficients in the presence of omitted variables, measurement error and an unknown true functional form.
^{1} There have also been a reasonably-large number of successful applications of the technique.^{2} However, it is difficult to establish the usefulness of a technique strictly through applications since we can never be certain of the accuracy of the results. This paper attempts to bridge the gap between the asymptotic theoretical results of the theoretical papers and the apparently good performance of the applied papers by constructing a set of Monte Carlo experiments to examine (1) how well the technique performs under clearly-defined conditions and (2) the limits on the technique’s ability to perform successfully under a broad range of model misspecifications.

The technique is motivated by an important theorem that was first proved by Swamy and Mehta (1975) and has recently been confirmed by Granger (2008) who quoted a proof that he attributed to Hal White. This theorem states that any nonlinear function may be exactly represented by a linear relationship with time-varying parameters. The importance of this theorem is that it allows us to capture an unknown true functional form in this framework. The parameters of this time-varying-coefficient model are, of course, not consistent estimates of the true functional form since they will be contaminated by the usual biases due to omitted variables, measurement error and simultaneity. The technique being investigated here allows us, in principal, to decompose the TVCs into two components; we associate the first component with the true nonlinear structure, which we interpret as the derivative of the dependent variable with respect to each of the independent variables in the unknown, nonlinear, true function; we associate the second component with the biases emanating from misspecification, and which we then remove from the TVC to give us our consistent estimates. Potentially, this technique offers an interesting way forward in dealing with model misspecification. It has generally been applied in a time series setting but it can equally well be interpreted as a cross section^{3} or panel estimation technique.

The remainder of this paper is structured as follows. Section 2 outlines the basic (TVC) theoretical framework. Section 3 discusses some computational issues associated with estimating the model. Section 4 reports on a series of Monte Carlo experiments. Section 5 concludes. An “Appendix” provides details on the computational methods used in the Monte Carlo simulations.

## 2 The Theoretical Framework

## 3 Computational Aspects

- (i)
Obtain \( {\boldsymbol{\uppi}}_{\text{r}} \) from its conditional distribution:

- (ii)where \( {\hat{\boldsymbol{\uppi }}}_{\text{r}} = \left( {{\mathbf{X}}^{\prime } {\boldsymbol{\Omega}}^{ - 1} {\mathbf{X}}} \right)^{ - 1} {\mathbf{X}}^{\prime } {\boldsymbol{\Omega}}^{ - 1} {\mathbf{y}},{\mathbf{V}} = \left( {{\mathbf{X}}^{\prime } {\boldsymbol{\Omega}}^{ - 1} {\mathbf{X}}} \right)^{ - 1} \), \( {\boldsymbol{\Omega}} = {\text{diag}}( \upsigma_{\text{u}}^{2} + {\mathbf{x}}^{\prime }_{\text{et}} {\boldsymbol{\Sigma}}_{\text{r}} {\mathbf{x}}_{\text{et}} ,{\text{t}} = 1,\ldots,{\text{T}}) \).$$ {\boldsymbol{\uppi}}_{\text{r}} |\upsigma_{\text{r}} ,\upsigma_{\text{u}} ,{\boldsymbol{\Sigma}}_{\text{r}} ,{\mathbf{Y}} \sim {\text{N}}\left( {{\hat{\boldsymbol{\uppi }}}_{\text{r}} ,{\mathbf{V}}} \right) $$(18)
- (iii)
Reparametrize \( {\boldsymbol{\Sigma}}_{\text{r}} \) using \( {\mathbf{C}} \) where \( {\boldsymbol{\Sigma}}_{\text{r}} = {\mathbf{C^{\prime}}}_{\text{r}} {\mathbf{C}}_{\text{r}} \), \( \sigma_{\text{u}} \propto \exp \left({c_{0}} \right) \) and \( \sigma_{r} \propto \exp (C_{00}) \). Assuming that different non-zero elements of \( {\mathbf{C}}_{\text{r}} \) are \( {\text{c}}_{1},\ldots,{\text{c}}_{\text{p}} \) the new parameter vector is \( {\boldsymbol{\uppi}}_{\text{r}} \) and \( {\mathbf{c}} = [{\text{c}}_{0} ,{\text{c}}_{{00}} ,{\text{c}}_{1} , \ldots ,{\text{c}}_{{\text{p}}} ]^{\prime } \in \mathbb{R}^{{{\text{p}} + 2}} \). Drawings from the conditional posterior distribution of \( {\mathbf{c}}|{\boldsymbol{\uppi}}_{\text{r}},{\mathbf{Y}} \) can be realized using the Girolami and Calderhead (2011) Metropolis Adjusted Langevin Diffusion method described in the “Appendix”.

In this form we can avoid a possibly inefficient Gibbs sampler which relies on drawing \( {\boldsymbol{\Pi}} \) and \( {\boldsymbol{\Sigma}} \) from (13), \( \{{\boldsymbol{\upbeta}}_{\text{t}},{\text{t}} = 1,\ldots,{\text{T}}\} \) from (19) and \( \upsigma \) from (12).

Selecting the drivers

^{4}The SSVS involves a specific prior of the form:

For the elements of \( {\mathbf{c}} \) we follow a similar approach. If \( {\text{c}}_{{\text{j}}} \) corresponds to a diagonal element it is always included in the model. If not, we use a mixture-of-normals SSVS approach as above.

## 4 Monte Carlo results

In all cases below \( \upgamma_{0} = \upgamma_{1} = \upgamma_{2} = .1 \). The number of Monte Carlo simulations is set to 10,000. All \( \upvarepsilon_{tj} \sim iidN(0,1) \). In case IV, we set \( \upsigma_{\upvarepsilon} = .3 \).

### 4.1 Model I: Incorrect Functional Form

The true model is \( {\text{y}}_{\text{t}} = \upgamma_{0} + \upgamma_{1} {\text{x}}_{\text{t}} + \tfrac{1}{2}\upgamma_{2} {\text{x}}_{\text{t}}^{2},\:{\text{t}} = 1, \ldots,{\text{T}} \) and we have omitted the nonlinear term. The driver is \( {{z}}_{{t}} = \alpha {{x}}_{{t}} + \upvarepsilon_{{t}} \). We have \( \upvarepsilon_{\text{t}} \sim {\text{iidN(0,1)}} \) and \( {\text{x}}_{{\text{t}}} \sim{\text{iidN(1,1)}} \). In this case the correlation between \( {\text{z}}_{\text{t}} \) and \( {\text{x}}_{{\text{t}}}^{2} \) is \( \uprho = \frac{3\alpha}{{\sqrt {3\left({3\alpha^{2} + 1} \right)}}} \).

If the correlation were equal to 1, then this would be a perfect driver as it exactly recreates the missing quadratic term. The estimation procedure would then be unbiased and efficient. If the correlation were zero, then z_{t} would contain no information about the missing nonlinearity. We are, therefore, interested in varying this correlation and seeing how low the correlation can fall before the estimator ceases to be useful.

In this case the true effect is \( \upgamma_{1} + \upgamma_{2} {\text{x}}_{\text{t}} \), that is, the derivative of \( y \) with respect to \( x \). There are no omitted variables or other misspecifications other than the nonlinearity so the set S_{2} is empty and the estimate of the derivative is given by \( \upgamma_{1} + \upgamma_{2} {\text{x}}_{\text{t}} = \beta_{1} - \varepsilon_{{t}} \).

Monte Carlo results for Model I

Corr \( \rho \) | .95 | .90 | .80 | .70 | .60 | .50 | .40 | .30 | .20 | .10 | .00 |
---|---|---|---|---|---|---|---|---|---|---|---|

T = 50 | |||||||||||

Bias | .017 | .017 | .025 | .028 | .035 | .048 | .071 | .098 | .117 | .125 | .205 |

SD | .011 | .011 | .012 | .019 | .041 | .057 | .091 | .120 | .144 | .189 | .265 |

T = 100 | |||||||||||

Bias | .007 | .007 | .009 | .017 | .022 | .035 | .077 | .114 | .135 | .181 | .235 |

SD | .008 | .008 | .007 | .014 | .035 | .052 | .128 | .140 | .192 | .272 | .301 |

T = 200 | |||||||||||

Bias | .004 | .004 | .007 | .012 | .011 | .070 | .101 | .177 | .186 | .244 | .293 |

SD | .006 | .006 | .005 | .007 | .012 | .044 | .177 | .186 | .281 | .316 | .387 |

T = 500 | |||||||||||

Bias | .003 | .003 | .005 | .008 | .011 | .079 | .136 | .218 | .222 | .271 | .332 |

SD | .004 | .004 | .006 | .007 | .032 | .055 | .190 | .277 | .334 | .389 | .415 |

T = 1000 | |||||||||||

Bias | .001 | .001 | .003 | .005 | .007 | .065 | .225 | .280 | .345 | .381 | .414 |

SD | .003 | .003 | .005 | .009 | .041 | .062 | .217 | .305 | .376 | .414 | .520 |

### 4.2 Model II: Omitted Variables

The second model focuses on omitted variables. The true model is \( {\text{y}}_{\text{t}} = \upgamma_{0} + \upgamma_{1} {\text{x}}_{{{\text{t}}1}} + \upgamma_{2} {\text{x}}_{{{\text{t}}2}} \). The \( {\text{x}}_{{{\text{t}}1}},x_{{{\text{t}}2}} \) are correlated: \( {\text{x}}_{{{\text{t}}2}} = \upgamma {\text{x}}_{{{\text{t}}1}} + \upxi_{\text{t}},\upxi_{\text{t}},{\text{x}}_{{{\text{t}}1}} \sim {\text{iidN(0,1)}} \). The squared correlation between the two variables is \( \uprho_{12}^{2} = \frac{{\upgamma^{2}}}{{\upgamma^{2} + 1}} \). We set \( \upgamma = 2 \) so that this is .80.

We estimate the TVC model \( {\text{y}}_{\text{t}} = \upbeta_{{0{\text{t}}}} + \upbeta_{{1{\text{t}}}} {\text{x}}_{{{\text{t}}1}} \) and again use a driver \( {\text{z}}_{\text{t}} = \alpha {\text{x}}_{{{\text{t}}2}} + \upvarepsilon_{\text{t}} \) and we see how well the estimator performs as the correlation between \( {\text{z}}_{\text{t}} \) and \( {\text{x}}_{{{\text{t}}2}} \) falls. The correlation between \( {\text{z}}_{\text{t}} \) and \( {\text{x}}_{{{\text{t}}2}} \) is \( \uprho = \frac{\upalpha}{{\sqrt {\upalpha^{2} + 1}}} \).

In this case, the true effect is \( \upgamma_{1} \) and the bias free estimate is \( \upbeta_{1\text{t}} - \uppi_{1} {\text{z}}_{\text{t}} - {\text{e}}_{\text{t}} \).

Monte Carlo results for Model II

Corr \( \rho \) | .95 | .90 | .80 | .70 | .60 | .50 | .40 | .30 | .20 | .10 | .00 |
---|---|---|---|---|---|---|---|---|---|---|---|

T = 50 | |||||||||||

Bias | .021 | .021 | .029 | .032 | .038 | .041 | .058 | .069 | .082 | .098 | .125 |

SD | .013 | .013 | .015 | .019 | .022 | .033 | .066 | .083 | .102 | .128 | .155 |

T = 100 | |||||||||||

Bias | .014 | .014 | .021 | .028 | .032 | .055 | .067 | .091 | .122 | .144 | .171 |

SD | .009 | .009 | .012 | .017 | .019 | .028 | .077 | .107 | .135 | .176 | .193 |

T = 200 | |||||||||||

Bias | .009 | .009 | .017 | .022 | .027 | .051 | .083 | .129 | .142 | .185 | .215 |

SD | .007 | .008 | .010 | .015 | .017 | .020 | .154 | .196 | .226 | .287 | .303 |

T = 500 | |||||||||||

Bias | .007 | .008 | .011 | .018 | .022 | .049 | .124 | .155 | .189 | .212 | .288 |

SD | .006 | .007 | .008 | .013 | .016 | .022 | .187 | .234 | .288 | .317 | .355 |

T = 1000 | |||||||||||

Bias | .005 | .006 | .008 | .010 | .017 | .047 | .171 | .222 | .287 | .334 | .345 |

SD | .004 | .005 | .007 | .011 | .014 | .020 | .199 | .276 | .302 | .344 | .381 |

### 4.3 Model III: Measurement Error

The third model deals with measurement error, so we generate data from \( {\text{y}}_{\text{t}} = \upgamma_{0} + \upgamma_{1} {\text{x}}_{\text{t}} \) then create \( {\text{y}}_{\text{t}}^{*} = {\text{y}}_{\text{t}} + \upvarepsilon_{t1} \) and \( {\text{x}}_{\text{t}}^{*} = {\text{x}}_{\text{t}} + \upvarepsilon_{\text{t2}} \) then we estimate the TVC model \( {\text{y}}_{{\text{t}}}^{*} = \upbeta _{0} {\text{t}} + \upbeta _{{1{\text{t}}}} {\text{x}}_{{\text{t}}}^{*} \) and use two *z*’s as drivers \( {\text{z}}_{\text{t1}} = \upalpha_{1} \upvarepsilon_{\text{t1}} + \upvarepsilon_{\text{t3}} \) and \( {\text{z}}_{{{\text{t2}}}} = {\upalpha }_{2} {\upvarepsilon }_{{{\text{t2}}}} + {\upvarepsilon }_{{{\text{t4}}}} \) and again see how things change as \( \upalpha \) gets bigger.

Monte Carlo results for Model III, ρ_{ε1,z1} = .50

Corr \( \rho \) | .95 | .90 | .80 | .70 | .60 | .50 | .40 | .30 | .20 | .10 | .00 |
---|---|---|---|---|---|---|---|---|---|---|---|

T = 50 | |||||||||||

Bias | .022 | .024 | .031 | .036 | .041 | .055 | .062 | .077 | .085 | .097 | .105 |

SD | .011 | .011 | .015 | .019 | .022 | .037 | .055 | .071 | .080 | .092 | .101 |

T = 100 | |||||||||||

Bias | .018 | .019 | .022 | .029 | .035 | .050 | .071 | .082 | .091 | .108 | .117 |

SD | .009 | .009 | .015 | .021 | .030 | .047 | .077 | .093 | .105 | .120 | .146 |

T = 200 | |||||||||||

Bias | .009 | .009 | .015 | .021 | .031 | .047 | .087 | .095 | .119 | .126 | .141 |

SD | .007 | .008 | .012 | .019 | .027 | .045 | .090 | .101 | .138 | .155 | .188 |

T = 500 | |||||||||||

Bias | .007 | .008 | .011 | .016 | .027 | .040 | .090 | .122 | .139 | .155 | .196 |

SD | .004 | .005 | .009 | .017 | .020 | .039 | .115 | .137 | .152 | .188 | .217 |

T = 1000 | |||||||||||

Bias | .005 | .005 | .007 | .009 | .016 | .032 | .117 | .144 | .162 | .196 | .213 |

SD | .003 | .003 | .006 | .010 | .016 | .030 | .141 | .166 | .195 | .225 | .255 |

### 4.4 Detecting Irrelevant Drivers

^{5}procedure, which we have not applied so far, can correctly identify the drivers \( {\text{z}}_{\text{t1}},{\text{z}}_{\text{t2}} \). To this end, we construct ten other drivers, say \( {\text{z}}_{\text{t2}},\ldots,{\text{z}}_{t,12} \) from a multivariate normal distribution with zero means and equal correlations of .70. In Table 4 we report the equivalent of Table 3 plus the proportion of cases, say \( \Pi^{*} \), in which SSVS has correctly excluded \( {\text{z}}_{\text{t2}},\ldots,{\text{z}}_{\text{t,12}} \) from the set of possible drivers.

^{6}

Monte Carlo results for Model III, ρ_{ε1,z1} = .50, SSVS

Corr \( \rho \) | .95 | .90 | .80 | .70 | .60 | .50 | .40 | .30 | .20 | .10 | .00 |
---|---|---|---|---|---|---|---|---|---|---|---|

T = 50 | |||||||||||

Bias | .025 | .026 | .033 | .038 | .044 | .059 | .067 | .079 | .091 | .099 | .109 |

SD | .012 | .012 | .016 | .020 | .023 | .039 | .057 | .075 | .086 | .095 | .114 |

\( \Pi * \) | 60.5% | 60.0% | 58.2% | 57.3% | 51.3% | 33.3% | 12.2% | 8.3% | 4.5% | .0% | .0% |

T = 100 | |||||||||||

Bias | .019 | .020 | .024 | .031 | .038 | .053 | .075 | .087 | .096 | .112 | .120 |

SD | .009 | .011 | .017 | .023 | .034 | .049 | .079 | .098 | .114 | .126 | .151 |

\( \Pi * \) | 71.2% | 71.0% | 62.3% | 64.8% | 58.2% | 55.4% | 9.3% | 7.5% | 3.3% | .0% | .0% |

T = 200 | |||||||||||

Bias | .012 | .012 | .019 | .027 | .035 | .049 | .089 | .099 | .121 | .127 | .148 |

SD | .008 | .008 | .015 | .021 | .029 | .047 | .082 | .103 | .140 | .159 | .192 |

\( \Pi * \) | 87.3% | 87.0% | 77.3% | 71.2% | 61.5% | 59.2% | 8.2% | 3.7% | .0% | .0% | .0% |

T = 500 | |||||||||||

Bias | .009 | .009 | .014 | .019 | .029 | .043 | .094 | .128 | .140 | .162 | .200 |

SD | .005 | .006 | .012 | .019 | .023 | .040 | .119 | .141 | .158 | .193 | .232 |

\( \Pi * \) | 97.3% | 96.5% | 93.4% | 85.5% | 79.3% | 62.7% | 4.4% | 1.0% | .0% | .0% | .0% |

T = 1000 | |||||||||||

Bias | .006 | .006 | .009 | .015 | .019 | .035 | .121 | .147 | .168 | .201 | .217 |

SD | .004 | .004 | .007 | .015 | .018 | .034 | .144 | .169 | .198 | .230 | .266 |

\( \Pi * \) | 99.5% | 98.3% | 97.7% | 91.2% | 85.2% | 77.7% | 2.1% | .0% | .0% | .0% | .0% |

There is again a remarkable cut off at the correlation level of .5. Above this level the true driver set is correctly identified in around 60% of cases and for the largest sample in over 90% of cases, even for small samples. Once the correlation falls below .5, however, the proportion of correct identifications falls dramatically. An obvious conclusion here is that when we have drivers that are effective enough so that we will get reasonably good parameter estimates, the SSVS algorithm is very effective at detecting them.

### 4.5 Model IV: A More Complex Nonlinearity

The true model is \( {\text{y}}_{\text{t}} ={\upgamma}_{ 0} +{\upgamma}_{ 1} {\text{x}}_{\text{t}} + {\text{ exp}}\left({- {\updelta}{\text{x}}_{\text{t}}^{2}} \right) +{\upvarepsilon}_{\text{t}} \), t = 1,…,T and we have omitted the nonlinear term. The drivers form a Fourier basis \( \left\{ {\cos ({\text{jx}}_{\text{t}} ),\sin ({\text{jx}}_{\text{t}} ),{\text{j = 1,}} \ldots {\text{J}}} \right\} \) after transforming all series to lie in (− π, π). We have \( {\upvarepsilon}_{\text{t}} \sim {\text{iidN(0,1)}} \) and \( {\text{x}}_{\text{t}} \sim {\text{iidN(0,1)}} \) ordered from smallest to largest. The drivers, that is powers of \( {\text{x}}_{\text{t}} \) are selected through the SSVS procedure. We set the maximum value of J to 10.

We again estimate the TVC model \( {\text{y}}_{\text{t}} = \upbeta_ 0 {\text{t}} + \upbeta_{\text{1t}} x_{t1} \) and this time the derivative of y with respect to x is \( \upgamma_{1} - 2\delta {\text{x}}_{\text{t}} \exp (- \delta {\text{x}}_{\text{t}}^{2}) \). Our estimate of this is again given by \( \beta_{{1{t}}} - {\text{e}}_{\text{t}} \).

*δ*in the range .1–5 the bias remains very small, as does the standard deviation. There is also a noticeable reduction in both bias and standard deviation as the sample size increases.

Monte Carlo results for Model IV, nonlinearity, SSVS/Fourier basis

\( \delta \) | .1 | .3 | .5 | 1.00 | 5.00 |
---|---|---|---|---|---|

T = 50 | |||||

Bias | .022 | .025 | .028 | .031 | .035 |

SD | .014 | .015 | .016 | .019 | .023 |

T = 100 | |||||

Bias | .017 | .022 | .025 | .029 | .032 |

SD | .011 | .012 | .014 | .017 | .020 |

T = 200 | |||||

Bias | .013 | .015 | .017 | .019 | .021 |

SD | .008 | .009 | .011 | .012 | .017 |

T = 500 | |||||

Bias | .009 | .010 | .012 | .014 | .018 |

SD | .005 | .007 | .009 | .010 | .014 |

T = 1000 | |||||

Bias | .005 | .006 | .007 | .011 | .013 |

SD | .003 | .004 | .006 | .009 | .011 |

### 4.6 Model V: An Endogeneity Experiment

In this experiment we have: \( {\text{y}}_{\text{t}} = {\upgamma}_{1} + {\upgamma}_{2} {\text{x}}_{\text{t1}} + {\upgamma}_{3} {\text{x}}_{\text{t2}} + {\text{u}}_{\text{t}} \). The correlation between \( {\text{u}}_{\text{t}} \) and \( {\text{x}}_{\text{tj}} \) is .80 (*j *= 1, 2) so that endogeneity is quite strong in this model. Our drivers are four variables \( {\text{z}}_{{{\text{t}}5}}, \ldots,{\text{z}}_{{{\text{t}}8}} \) *orthogonal* to the error \( {\text{u}}_{\text{t}} \) and four drivers \( {\text{z}}_{{{\text{t}}1}}, \ldots,{\text{z}}_{{{\text{t}}4}} \) which are *correlated* with the error \( {\text{u}}_{\text{t}} \) but they are orthogonal to each other as well as orthogonal to the other four drivers. The degree of correlation between the drivers and \( {\text{u}}_{\text{t}} \) is \( {\uprho} \). We are interested in \( \Pi^{*} \), the proportion of cases where *all* the drivers \( {\text{z}}_{{{\text{t}}1}}, \ldots,{\text{z}}_{{{\text{t}}4}} \) are included in the model *and* the drivers \( {\text{z}}_{{{\text{t}}5}}, \ldots,{\text{z}}_{{{\text{t}}8}} \) are *all* excluded. Of course, we do not force the correct drivers in final estimation.

Monte Carlo results for Model V, Endogeneity, SSVS

Corr \( \rho \) | .1 | .3 | .5 | .7 | .9 |
---|---|---|---|---|---|

T = 50 | |||||

Bias | .491 | .401 | .387 | .266 | .212 |

SD | .716 | .710 | .688 | .644 | .601 |

\( \Pi^{*} \) | .044 | .081 | .225 | .447 | .617 |

T = 100 | |||||

Bias | .386 | .316 | .303 | .201 | .138 |

SD | .355 | .350 | .281 | .277 | .252 |

\( \Pi^{*} \) | .051 | .128 | .315 | .517 | .645 |

T = 200 | |||||

Bias | .300 | .216 | .287 | .181 | .101 |

SD | .314 | .310 | .277 | .201 | .196 |

\( \Pi^{*} \) | .081 | .201 | .403 | .615 | .717 |

T = 500 | |||||

Bias | .295 | .290 | .101 | .087 | .071 |

SD | .201 | .200 | .096 | .061 | .055 |

\( \Pi^{*} \) | .091 | .261 | .462 | .687 | .775 |

T = 1000 | |||||

Bias | .282 | .280 | .047 | .031 | .028 |

SD | .195 | .193 | .032 | .027 | .016 |

\( \Pi^{*} \) | .101 | .316 | .518 | .784 | .801 |

T = 10,000 | |||||

Bias | .190 | .190 | .039 | .001 | .001 |

SD | .182 | .182 | .030 | .011 | .008 |

\( \Pi^{*} \) | .115 | .320 | .615 | .813 | .917 |

## 5 Conclusions

This paper has investigated the performance of the TVC estimation procedure in a Monte Carlo setting. The key element of TVC estimation is the identification and selection of a set of driver variables. With an ideal driver set, it is straightforward to show that the procedure is both consistent and efficient. However, in practice it is not possible to know if we have a perfect driver set. Therefore, we need to know how the procedure performs when the driver set is less than perfect. In this paper, we dealt with this issue in a Monte Carlo setting.

We construct a number of Monte Carlo experiments to examine the performance of the methodology under (i) clearly-defined conditions and (ii) a range of model misspecifications. We also propose a new Bayesian search technique for the set of driver variables underlying the TVC methodology. Experiments are performed to allow for incorrectly specified functional form, omitted variables, measurement errors, unknown nonlinearity and endogeneity. Our broad conclusion is that, even for relatively small samples, the technique works well so long as the correlation between the driver set and the misspecification in the model is greater than about .5. Both the bias and the efficiency of the estimators also improve as the sample size grows, but again a correlation of over .5 seems to be required. The only caveat to this result is that if we are considering strong simultaneity bias; in that case the sample size needs to be quite large (over 500) before the technique works reasonably well. Finally, we find that the SSVS technique also seems to perform well in finding an appropriate driver set from a much larger set of possible drivers.

## Footnotes

- 1.
- 2.
- 3.
Time varying coefficients are meaningless in a cross section setting; in such a setting the coefficients vary across the individual units in the cross section. We simply re-interpret the t-subscript as i-subscripts.

- 4.
See also Jochmann et al (2010).

- 5.
An alternative to using the SSVS procedure would be the LASSO prior. The procedures are similar in terms of timing and purpose. There is some evidence that both perform well (Pavlou et al. 2016) and in a similar manner but further work is needed in this area.

- 6.
There is an issue here as to whether we need to start from a superset of drivers which includes the true ones. Clearly, if we do this then this is an ideal situation and the Monte Carlo tells us how well the procedure performs. However, from a theoretical point of view what we need is that the superset includes variables be highly correlated with the true drivers. In a data rich environment this would not be a strong restriction. Bai and Ng (2010) prove that common factors which drive all the variables in a system are valid instrumental variables. By the same reasoning, we could construct a set of factors from a large set of variables which would work well as drivers.

- 7.
This guarantees the existence of moments up to order four.

## References

- Bai, J., & Ng, S. (2010). Instrumental variable estimation in a data rich environment.
*Econometric Theory,**26,*1577–1606.CrossRefGoogle Scholar - George, E., Sun, D., & Ni, S. (2008). Bayesian stochastic search for VAR model restrictions.
*Journal of Econometrics,**142,*553–580.CrossRefGoogle Scholar - Girolami, M., & Calderhead, B. (2011). Riemann manifold Langevin and Hamiltonian Monte Carlo methods.
*Journal of the Royal Statistical Society: Series B (Statistical Methodology),**73*(2), 123–214.CrossRefGoogle Scholar - Granger, C. W. J. (2008). Nonlinear models: Where do we go next—Time-varying parameter models?
*Studies in Nonlinear Dynamics and Econometrics,**12*(3), 1–9.Google Scholar - Hall, S. G., Hondroyiannis, G., Swamy, P. A. V. B., & Tavlas, G. S. (2009). Where has all the money gone? Wealth and the demand for money in South Africa.
*Journal of African Economies,**18*(1), 84–112.CrossRefGoogle Scholar - Hall, S. G., Hondroyiannis, G., Swamy, P. A. V. B., & Tavlas, G. S. (2010). The fisher effect puzzle: A case of non-linear relationship.
*Open Economies Review*. https://doi.org/10.1007/s11079-009-9157-1.CrossRefGoogle Scholar - Hall, S. G., Swamy, P. A. V. B., & Tavlas, G. S. (2017). Time-varying coefficient models: A proposal for selecting the coefficient driver sets.
*Macroeconomic Dynamics,**21*(5), 1158–1174.CrossRefGoogle Scholar - Hondroyiannis, G., Kenjegaliev, A., Hall, S. G., Swamy, P. A. V. B., & Tavlas, G. S. (2013). Is the relationship between prices and exchange rates homogeneous.
*Journal of International Money and Finance,**37,*411–436. https://doi.org/10.1016/j.jimonfin.2013.06.014.CrossRefGoogle Scholar - Jochmann, M., Koop, G., & Strachan, R. W. (2010). Bayesian forecasting using stochastic search variable selection in a VAR subject to breaks.
*International Journal of Forecasting,**26*(2), 326–347.CrossRefGoogle Scholar - Kenjegaliev, A., Hall, S. G., Tavlas, G. S., & Swamy, P. A. V. B. (2013). The forward rate premium puzzle: A case of misspecification?
*Studies in Nonlinear Dynamics and Econometrics,**3,*265–280. https://doi.org/10.1515/snde-2013-0009.CrossRefGoogle Scholar - Pavlou, M., Ambler, G., Seaman, S., De Lorio, M., & Omar, Z. (2016). Review and evaluation of penalised regression methods for risk prediction in low-dimensional data with few events.
*Statistics in Medicine,**37*(7), 1159–1177.CrossRefGoogle Scholar - Swamy, P. A. V. B., Chang, I.-L., Mehta, J. S., & Tavlas, G. S. (2003). Correcting for omitted-variable and measurement-error bias in autoregressive model estimation with panel data.
*Computational Economics,**22,*225–253.CrossRefGoogle Scholar - Swamy, P. A. V. B., Hall, S. G., & Tavlas, G. S. (2012). Generalized cointegration: A new concept with an application to health expenditure and health outcomes.
*Empirical Economics,**42,*603–618. https://doi.org/10.1007/s00181-011-0483-y.CrossRefGoogle Scholar - Swamy, P. A. V. B., Hall, S. G., & Tavlas, G. S. (2015). A note on generalizing the concept of cointegration.
*Macroeconomic Dynamics,**19*(7), 1633–1646. https://doi.org/10.1017/S1365100513000928.CrossRefGoogle Scholar - Swamy, P. A. V. B., Hall, S. G., Tavlas, G. S., & Hondroyiannis, G. (2010). Estimation of parameters in the presence of model misspecification and measurement error.
*Studies in Nonlinear Dynamics & Econometrics,**14,*1–35.CrossRefGoogle Scholar - Swamy, P. A. V. B., & Mehta, J. S. (1975). Bayesian and non-Bayesian analysis of switching regressions and a random coefficient regression model.
*Journal of the American Statistical Association,**70,*593–602.Google Scholar - Tavlas, G. S., Swamy, P. A. V. B., Hall, S. G., & Kenjegaliev, A. (2013). Measuring currency pressures: The cases of the Japanese Yen, the Chinese Yuan, and the U.K. pound.
*Journal of the Japanese and International Economies,**29,*1–20. https://doi.org/10.1016/j.jjie.2013.04.001.CrossRefGoogle Scholar

## Copyright information

**Open Access**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.