Fall 2013
ECON 5340 Applied Econometrics – Exam #2
True or False. Five points per question: 2 pts for a correct T/F and 3 pts for the explanation.
1. Measurement error in the dependent variable is more serious than measurement error in the
explanatory variables.
False. Measurement error in the dependent variable only adds more noise to the model,
reducing the goodness of fit, but it does not bias the coefficients. Measurement error in the
independent variables, however, biases the coefficient estimates toward zero.
2. Autocorrelation biases the coefficients towards zero.
False. Autocorrelation does not bias the coefficient estimates but OLS estimates are no longer
efficient.
3. Severe multicollinearity will bias the coefficient estimates.
False. Provided the multicollinearity (MC) is not perfect, MC does not bias the coefficients.
OLS estimates are still unbiased and efficient. However, the standard errors are inflated relative
to the case of no MC.
4. Ordinary least squares is a special case of generalized least squares.
True. Generalized least squares (GLS) involves a transformation of the variables so that the
Classical assumptions hold. If there is no heteroscedasticity or autocorrelation, then OLS is
simply a special case of GLS where no transformation is necessary.
5. The “runs test” for autocorrelation is immune from Type I errors.
False. A Type I error involves rejecting a true null hypothesis. This is possible if the sample is
not representative and suggests autocorrelation when there is none in the data-generating
process.
Fall 2013
#6. (50 pts) Consider the earnings model: 𝑊𝑎𝑔𝑒𝑖 = 𝛽1 + 𝛽2 𝐸𝑥𝑝𝑒𝑟𝑖 + 𝛽3 𝐸𝑑𝑢𝑐𝑖 + 𝑢𝑖 , where
𝑊𝑎𝑔𝑒 is the measured in dollars per hour, 𝐸𝑥𝑝𝑒𝑟 is work experience in years, and 𝐸𝑑𝑢𝑐 is the
number of years of schooling. Tables 1-3 and Figure 1 show the OLS regression results for
𝑁 = 100 males in a given year. Use the tables and figures to answer the following questions:
a) (5 pts) Using the results in Table 1, summarize the overall goodness of fit of the model. Do
the signs of the coefficients match your explanations? Explain.
The overall goodness of fit is given by the R2 value of 0.25. This implies that the 25% of the
variation in the dependent variable can be explained by variation in the independent variables.
The F test for overall goodness of fit is also significant. Both coefficient estimates are positive,
which is to be expected.
b) (10 pts) Interpret the residual pattern in Figure 1. What conclusion do you draw? And based
on that conclusion, what are the impacts on the OLS estimates in Table 1? Explain.
To my eye, I see heteroscedasticity. The variance of the errors appears to increase with
Education. If this is true, the OLS estimates are unbiased but inefficient. GLS would provide
efficient estimates.
c) (10 pts) Using the results in Table 3, perform White’s test for heteroscedasticity. Be sure to
carefully set up the null and alternative hypotheses and draw a conclusion.
White’s test involves running a regression with squared residuals on the left side of the
regression model and the levels, squares and crosses of all the explanatory variables on the right
side. The results are shown in Table 3. The test statistic is chi squared with 5 degrees of
freedom. The statistic is 𝑛𝑅2 = 100 ∗ 0.0949 = 9.49. The critical value at the 5% level is 11.07.
Therefore, we fail to reject the null of no heteroscedasticity.
Fall 2013
d) (10 pts) Based on your answers to parts (b) and (c), describe a procedure to obtain the
efficient estimators for the coefficients.
The results in parts (b) and (c) differ. However, if we assume there is some heteroscedasticity
related to, say, the square of Education (i.e., 𝜎𝑖2 = 𝜎 2 𝐸𝑑𝑢𝑐𝑖2 ), then the efficient thing to do is
GLS. We would then transform all the variables by dividing them by Education and running
OLS on the transformed variables.
e) (5 pts) Use the results in Table 2 to discuss the severity of the multicollinearity and the likely
impacts on the OLS results in Table 1.
The pairwise correlation between Education and Experience is -0.3665. This is not strong
enough to be a concern, but it will inflate the standard errors relative to the case where there is
no correlation. The OLS estimates are still unbiased and efficient.
f) (10 pts) Show how to perform a t test that the EXPER and EDUC coefficients are equal. Do
you have all the necessary information to complete the test? If so, complete the test.
The null hypothesis is 𝐻0 : 𝛽2 = 𝛽3 . The t statistic would be
𝛽̂2 − 𝛽̂3
𝑡= .
̂ ̂ ̂ ̂
�𝑣𝑎𝑟�𝛽2 � + 𝑣𝑎𝑟(𝛽3 ) − 2𝑐𝑜𝑣(𝛽2 , 𝛽3 )
The covariance between the two estimates is not provided so the test cannot be completed.
Fall 2013
#7. (25 pts) Provide brief answers to the following three questions.
a) (5 pts) Write down an AR(1) process for the error terms and describe how to perform a
Durbin-Watson test for autocorrelation.
An AR(1) process is 𝑢𝑡 = 𝜌 𝑢𝑡−1 + 𝜖𝑡 . The Durbin Watson statistic is 𝐷𝑊 ≅ 2(1 − 𝜌�). The
null hypothesis is no autocorrelation. The DW statistic is calculated and compared to the low
and high critical values from the DW table.
b) (10 pts) Re-specify the earnings model in question #6, to test the hypothesis that men in a
labor union earn more on average than men that are not members of a labor union. How
would you test this hypothesis? Draw an appropriate figure of the regression model to
support your test.
The new model would be: 𝑊𝑎𝑔𝑒𝑖 = 𝛽1 + 𝛽2 𝐸𝑥𝑝𝑒𝑟𝑖 + 𝛽3 𝐸𝑑𝑢𝑐𝑖 + 𝛽4 𝑈𝑛𝑖𝑜𝑛𝑖 + 𝑢𝑖 , where Union
is a dummy variable for union membership. The model results in two parallel regression lines, a
higher one for union men and a lower one for non-union men. The hypothesis can be tested with
a t test for 𝛽4 = 0.
c) (10 pts) Re-specify the earnings model in question #6, to test the hypothesis that men in a
labor union earn more for each additional year of experience than men that are not members
of a labor union. How would you test this hypothesis? Draw an appropriate figure of the
regression model to support your test.
The new model is
𝑊𝑎𝑔𝑒𝑖 = 𝛽1 + 𝛽2 𝐸𝑥𝑝𝑒𝑟𝑖 + 𝛽3 𝐸𝑑𝑢𝑐𝑖 + 𝛽4 𝑈𝑛𝑖𝑜𝑛𝑖 + 𝛽5 (𝐸𝑥𝑝𝑒𝑟𝑖 ∗ 𝑈𝑛𝑖𝑜𝑛𝑖 )+𝑢𝑖 ,
where Union is a dummy variable for union membership. The model results in two regression
lines with different intercepts and slopes. If the theory is correct, the union membership
regression line has a steeper slope. The hypothesis can be tested with a t test for 𝛽5 = 0.
Fall 2013
BONUS QUESTION. (5 pts) Describe an economic situation (other than the one in my lecture
notes) where panel data will allow you tease out two impacts that would not be possible using
only cross-section or only time-series data.
Table 1. STATA results from OLS estimation of the earnings model
Source | SS df MS Number of obs = 100
-------------+------------------------------ F( 2, 97) = 16.47
Model | 2057.5037 2 1028.75185 Prob > F = 0.0000
Residual | 6059.71269 97 62.4712648 R-squared = 0.2535
-------------+------------------------------ Adj R-squared = 0.2381
Total | 8117.21639 99 81.9920847 Root MSE = 7.9039
------------------------------------------------------------------------------
wage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Educ | 1.435782 .321546 4.47 0.000 .7976026 2.073962
Exper | .328525 .0658247 4.99 0.000 .1978813 .4591687
_cons | -11.91922 4.750254 -2.51 0.014 -21.34716 -2.491275
------------------------------------------------------------------------------
Figure 1. Wage residuals versus education
Fall 2013
Table 2. Pairwise correlations
| grade exper wage
-------------+---------------------------
grade | 1.0000
exper | -0.3665 1.0000
wage | 0.2485 0.3163 1.0000
Table 3. STATA results with squared OLS residuals as the dependent variable
Source | SS df MS Number of obs = 100
-------------+------------------------------ F( 5, 94) = 1.97
Model | 498933.661 5 99786.7323 Prob > F = 0.0901
Residual | 4759291.93 94 50630.7652 R-squared = 0.0949
-------------+------------------------------ Adj R-squared = 0.0467
Total | 5258225.59 99 53113.3898 Root MSE = 225.01
------------------------------------------------------------------------------
res2 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Grade | -7.357599 79.35932 -0.09 0.926 -164.9274 150.2122
Exper | -23.67913 16.87954 -1.40 0.164 -57.19386 9.835591
Grade^2 | -1.048003 2.223082 -0.47 0.638 -5.461984 3.365978
Exper^2 | .270444 .162453 1.66 0.099 -.0521102 .5929982
Exper*Grade | .5788711 .7165818 0.81 0.421 -.8439188 2.001661
_cons | 108.2517 582.867 0.19 0.853 -1049.044 1265.548
------------------------------------------------------------------------------
