ECONOMETRICS II

TUTORIAL I
The first tutorial will mainly deal with model selection carried out through
correct specification tests. It will also address the issues of coefficients’ inter-
pretation, especially those of dummy variables, of (perfect) multicollinearity
and of the correction of coefficients’ variance/covariance matrix when errors
are heteroskedastic.

We will use the bwages.wf1 workfile (the example is taken from Verbeek,
chapter 3). The workfile contains 4 variables observed in 1994 for 1, 472
Belgian individuals (therefore, data are cross sectional ): 1) wage, hourly
gross wage, in euro; 2) educ, education level: 1 if primary school; 2 if lower
vocational training; 3 if high-school; 4 if higher vocational training; 5 if
degree; 3) exper, working experience, in years; 4) male, dummy variable: 1 if
the individual is a man and 0 otherwise (there are 893 men and 579 women in
the sample). The workfile also contains three additional series: the natural
logs of wage, educ and exper.

We will estimate a wage equation with gender, education level and working
experience as regressors. We will first use a linear functional form and then
a log form. We will also use interactions of the variables to enrich the spec-
ification of the model. Notice how such a dataset allows us to try to assess
whether gender discrimination in the labour market occurs.

1 DUMMY VARIABLES
Dummy variables are variables that take the value 1 if a condition is satisfied
and 0 otherwise (e.g. 1 if the individual is a man and 0 otherwise).

Example 1 Compute first descriptive statistics for male, and then for wage,
educ and exper by gender (i.e. by classifying for the male variable).

The table with descriptive statistics for the male variable is the following:

1
Suggestion: to obtain the descriptive statistics for subsamples use the com-
mand

View...
Descriptive Statistics & Tests...
Stats by Classification

The three tables with descriptive statistics by sub-samples are the following:

2
3
Example 2 After generating the dummy variable F EM ALE which takes
the value 1 if the individual is female and 0 otherwise, estimate three models
with OLS:
wage = β0 ×i + β1 ×Male + ε (1)
wage = γ0 ×i + γ1 ×Female + ε (2)
wage = δ0 ×Male + δ1 ×Female + ε (3)
Save the three equations as eq01, eq02 and eq03. Interpret the coefficients
and the variance-covariance matrices of the coefficients.

Suggestion: to generate the dummy variable female use the expression

female=(male=0) or the expression female=1-male

The outputs with the three regressions:

4
Notice how the three models are a re-parametrization the one of the other.
For instance β0 = δ1 and β1 = δ0 −δ1 so that var(β0 ) = var(δ1 ) and var(β1 ) =
var(δ0 − δ1 ) = var (δ0 ) + var (δ1 )

5
2 THE LINEAR MODEL
Example 3 Estimate the model adding as additional regressors education
and working experience:

wage = β0 ×i + β1 ×Male + β2 ×Educ + β3 ×Exper + ε (4)

Save the equation as eq04 and interpret the new coefficients.

The output of the estimated equation is:

Notice: (i) β1 is the absolute difference between the wage of a man and the
one of a woman with same education and working experience levels; (ii) For
continuous variables, the coefficients are marginal effects (e.g. the marginal
effect of Exper is β3 : increasing by 1 the number of years of experience, the
wage increases by 0.19 euro, all else equal).

Example 4 Suppose that working experience affects wage in a non linear

way; more specifically, suppose that the marginal impact of working expe-
rience on wage decreases when working experience increases. Which model

6
would you estimate and how would you interpret the coefficients? (save the
equation as eq05)

Example 5 Compute the marginal effect of experience on wage for an indi-

vidual with a working experience of 20 years.

Suggestion: The last estimated model is not linear in Exper. To obtain

the marginal effect (which is not constant any more), digit the command

scalar marg eff 20=eq05.@coef(4)+2*eq05.@coef(5)*20

in the command window.

Example 6 We suspect that the last estimated model con be heteroskedastic.

How would you try to verify this hypothesis with a graphical analysis and a
formal test?

7
Suggestion: to perform an heteroskedasticity test digit the command:
View...
Residual Diagnostics...
Heteroskedasticity tests

The graph with the residuals plotted as function of the predicted values of
the dependent variable is:

8
Example 7 How would you modify either the model to obtain the FGLS
estimator or the variance covariance matrix of OLS estimators to be able to
perform reliable hypothesis tests?

3 THE LOGLINEAR MODEL

Example 8 An alternative approach to address the heteroskedasticity prob-
lem (and/or to introduce non linearity in the model) is to estimate the model
in the log form:
ln wage = α0 ×i + α1 ×Male + α2 × ln Educ + α3 × ln Exper + α4 × (ln Exper)2 + υ
(5)
Estimate this model (call it eq06) and graphically check how heteroskedasticity
is now very much reduced.

9
10
Example 9 Interpret the coefficients, notably the one of the dummy variable
Male.

α1 = E (ln wage | X, Male = 1) − E (ln wage | X, Male = 0) =

= E (ln wage | X, Male = 1 − ln wage | X, Male = 0) =
  
wage | X, Male = 1
= E ln (6)
wage | X, Male = 0

    
wage | X, Male = 1 wage | X, Male = 1
exp (α1 ) = exp E ln ≈E
wage | X, Male = 0 wage | X, Male = 0
(7)

 
wage | X, Male = 1 − wage | X, Male = 0
[exp (α1 ) − 1] ×100 ≈ E ×100
wage | X, Male = 0
(8)

[exp (α1 ) − 1] ×100 ≈ α1 ×100 (9)

Therefore, notice how α1 ×100 is the relative difference (%) between a

man’s and a woman’s wage holding education and working experience con-
stant.

For continuous variables, the estimated coefficients in log-log specifications

are not marginal effects but elasticities. E.g. the elasticity of wage with
respect to Educ, the education level, is α2 : if the education level increases
by 1% , the wage will increase by 0.44%, all else equal. The elasticity of wage
with respect to Exper is not constant and it is given by: α3 +2×α4 ×(Exper)

Example 10 Test the joint hypothesis

H0 : α3 = 0; α4 = 0

11
To perform a joint significance test digit the command:
View...
Coefficient Diagnostics...
Wald Test - Coefficient restrictions
How you would reconcile the result of the test with the p-values of the single
hypotheses?

Example 11 Estimate the model without the (ln Exper)2 variable, save as
eq07 and consider this model as the reference one from this point onwards.
Which restriction on the impact of education on wage are we implicitly im-
posing in eq07? How could you relax this assumption? How could you test
it?

Suggestion: create 5 dummy variables (deduc1, deduc2, deduc3, deduc4,

deduc5), one for each education level. Then estimate a new model (save
as eq08), where you use the set of education dummy variables instead of
lnEduc. When including the full set of dummy variables, be aware of multi-
collinearity.
Finally, use the R2 (or the sum of squared residuals) of the two models
to perform the test on the restrictions (eq07 is the constrained model, while
eq08 is the unconstrained model).
Suggestion:
Digit
scalar f=((eq07.@ssr-eq08.@ssr)/3)/(eq08.@ssr/(eq08.@regobs-eq08.@ncoef))
or
scalar f1=((eq08.@r2-eq07.@r2)/3)/((1-eq08.@r2)/(eq08.@regobs-eq08.@ncoef))
in the command window. To obtain the pvalue digit:
scalar pvalue=@fdist(f1,3,eq08.@regobs-eq08.@ncoef)

4 STRUCTURAL BREAK TEST

Example 12 Test the hypothesis that the following model is the same for
men and women:
ln wage = α0 ×i + α1 × ln Exper + α2 ×dEduc2 + α3 ×dEduc3 + α4 ×dEduc4 +
+α5 ×dEduc5 + υ (10)

12
Which conclusion would you draw?

Suggestion: Estimate a new model (eq09) on the whole sample, by including

all original regressors and their interaction with the dummy variable male.
What is the interpretation of the estimated coefficients for this model?
Then use an F test for the null hypothesis that all the interacted terms are
jointly equal to zero. If the null hypothesis is rejected, the coefficients for
the two groups are actually different.

5 THE RESET TEST

Example 13 Test the hypothesis that the general model (i.e. split by gender,
eq09) just estimated is correctly specified by performing a RESET TEST.

Suggestion: to perform a RESET test the command to use is:

View...
Stability Diagnostics...
Ramsey RESET Test...

digiting “Q” in the the window that opens, i.e. the number of powers of the
fitted values to insert in the auxiliary equation (NB: Q in EViews is to be
interpreted as Q+1 of the notes and of Verbeek).

13