I have a SUR model with 22 equations, where each equation has the same 7 factors. I want to test if a coefficient (b3 in equation 1) is significantly different from another coefficient in another equation (for example, b3 in equation 11) of the SUR model.

  • 2
    $\begingroup$ Can you clarify your question? Eg, I don't see a "?" anywhere. Are you asking how this can be done? $\endgroup$ – gung - Reinstate Monica Jan 27 '16 at 0:14
  • $\begingroup$ acronyms should also be spelled out the first time used. I'm guessing SUR is seemingly unrelated regressions? $\endgroup$ – StatsStudent Jan 27 '16 at 0:23
  • $\begingroup$ i'm sorry yes SUR is Seemingly Unrelates Regression and yes i would know how could i do the test $\endgroup$ – a.paolini Jan 27 '16 at 10:44

Conceptually, there is not all that much of a difference to any other $t$ or Wald type test. Think of the SUR estimates of the 22 equations being stacked on top of each other in one long $22\cdot 7$ (plus possibly estimates for the constant term) vector $\hat\delta$.

As you can read up for example very nicely in Hayashi, Econometrics, $\hat\delta$ will be asymptotically normal provided the SUR assumptions (the most important of which is that your regressors are actually predetermined) are met,

$$ \sqrt{n}(\hat\delta-\delta)\to_dN(0,Avar(\hat\delta)) $$ Then, you may test hypotheses like $$ H_{0}:\delta_{\ell}=\bar{\delta}_{\ell} $$ with the $t$-statistic $$ t_{\ell}=\frac{\sqrt{n}(\widehat{\delta}_{\ell}-\bar{\delta}_{\ell})}{\sqrt{\widehat{\mathrm{Avar}} (\widehat{\delta}_{\ell})}} $$ where $\widehat{\mathrm{Avar}}(\widehat{\delta}_{\ell})$ is the $(\ell,\ell)$-element of the estimated version of $Avar(\hat\delta)$. Similarly, test $$ H_{0}:R\delta =r $$ (with $J$ rows for the $J$ hypotheses) with the Wald statistic $$ W=n(R\widehat{\delta}-r)'\left[R\widehat{\mathrm{Avar}}(\widehat{\delta})R'\right]^{-1}(R\widehat{\delta}-r). $$ These test statistics then follow their usual asymptotic $N(0,1)$ and $\chi^2_J$ null distributions.

The SUR estimator for $Avar(\hat\delta)$ is given by

$$ \widehat{Avar}(\widehat{\delta})=\left(% \begin{array}{ccc} \widehat{\sigma}^{11}\frac{1}{n}\sum_{i=1}^nz_{i1}z_{i1}' & \cdots & \widehat{\sigma}^{1M}\frac{1}{n}\sum_{i=1}^nz_{i1}z_{iM}' \\ \vdots & \ddots &\vdots \\ \widehat{\sigma}^{M1}\frac{1}{n}\sum_{i=1}^nz_{iM}z_{i1}' & \cdots & \widehat{\sigma}^{MM}\frac{1}{n}\sum_{i=1}^nz_{iM}z_{iM}' \\ \end{array} \right)^{-1}, $$ with $z_{im}$ the regressors of equation $m$ and $\widehat{\sigma}^{mk}$ the estimate of the covariance of the error terms of equations $m$ and $k$, using preliminary (OLS) residuals to estimate these.

So for your problem, you would just need to specify a row vector $R$ that has a $1$ and a $-1$ in the appropriate places corresponding to your coefficients of interest, as well as $r=0$, to test coefficient equality.

Here is an example slightly adapted from the systemfit package that should adaptable to your problem:

eqDemand <- consump ~ price + income
eqSupply <- consump ~ price + farmPrice + trend
system <- list(demand = eqDemand, supply = eqSupply)
fitsur <- systemfit(system, "SUR", data = Kmenta)

Rmat <- matrix(0, nrow = 1, ncol = 7)
Rmat[1, 2] <- 1
Rmat[1, 5] <- -1
qvec <- 0
linearHypothesis(fitsur, Rmat, test = "Chisq")


> linearHypothesis(fitsur, Rmat, test = "Chisq")
Linear hypothesis test (Chi^2 statistic of a Wald test)

demand_price - supply_price = 0

Model 1: restricted model
Model 2: fitsur

  Res.Df Df  Chisq Pr(>Chisq)    
1     34                         
2     33  1 31.712  1.788e-08 ***
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

As an aside, if each of your equations has the same regressors, SUR just boils down to equation-by-equation OLS.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.