Lecture 11: Tests of Restrictions

In which you …

Learn how to test hypotheses about the value of more than 1

variable in a multiple regression model (more F tests)

Find out the consequences of getting the model specification

wrong
(OLS estimates may be biased)

Learn what to do about it

 Tests of Restrictions

Remember the F test for goodness of fit of the model

F = ESS/k-1 = R2/k-1 ~ F(J, N-K)

RSS/N- K (1- R2)/n-K
 Tests of Restrictions

Remember the F test for goodness of fit of the model

F = ESS/k-1 = R2/k-1 ~ F(J, N-K)

RSS/N- K (1- R2)/n-K

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero
 Tests of Restrictions

Remember the F test for goodness of fit of the model

F = ESS/k-1 = R2/k-1 ~ F(J, N-K)

RSS/N- K (1- R2)/n-K

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

Often useful if want to know whether worth adding (or removing) sets of
variables,

particularly now we know that there are potential problems with

multicolinearity and precision of estimates
Called the Chow test – after Gregory Chow of Princeton
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1= 0
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1= 0 or b2 = 0 or b3=0
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1= 0 or b2 = 0 or b3=0

Now test a subset

Y = b0 + b1X1 + b2X2 + b3X3 + u
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them (except the constant) are zero as with the
original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1= 0 or b2 = 0 or b3=0

Now test a subset

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b2= 0 & b3 = 0
 Tests of Restrictions

A variant of the test of goodness of fit of the model is instead to test a

hypothesis that a sub-set of the right hand side variables are zero

rather than test that all of them are zero as with the original F test
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1 = 0 & b2= 0 & b3 = 0

or just one of them as in the t test

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b1= 0 or b2 = 0 or b3=0

Now test a subset

Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b2= 0 & b3 = 0 (or b1 & b2=0)
This amounts to running 2 different regressions
This amounts to running 2 different regressions

1) Y = b0 + b1X1 + b2X2 + b3X3 + u

This amounts to running 2 different regressions

1) Y = b0 + b1X1 + b2X2 + b3X3 + u

2) Y = b0 + b1X1 + u
This amounts to running 2 different regressions

1) Y = b0 + b1X1 + b2X2 + b3X3 + u

2) Y = b0 + b1X1 + u

and looking to see whether the RSS is significantly different in the two
specifications
Can show that test becomes

F = RSSrestricted – RSSunrestricted /J ~ F(J, N-Kunrestricted)

RSSunrestricted /N- Kunrestricted

(compared with F = ESS/k-1 ~ F(J, N-K)

RSS/N- K
For the test of goodness of fit of the whole
model )
Can show that test becomes

F = RSSrestricted – RSSunrestricted /J ~ F(J, N-Kunrestricted)

RSSunrestricted /N- Kunrestricted

(compared with F = ESS/k-1 ~ F(J, N-K)

RSS/N- K

For the test of goodness of fit of the whole model )

Or equivalently (sub. in RSS = TSS(1-R2) )

F = R2unrestricted – R2restricted /J ~ F(J, N-Kunrestricted)

1-R2unrestricted /N- Kunrestricted
F = R2unrestricted – R2restricted /J ~ F(J, N-Kunrestricted)
1-R2unrestricted /N- Kunrestricted

Where
J = No. of variables to be tested
F = R2unrestricted – R2restricted /J ~ F(J, N-Kunrestricted)
1-R2unrestricted /N- Kunrestricted

Where

J = No. of variables to be tested

restricted = values from model with variables set to zero (ie excluded
from the regression specification) Y = b0 + b1X1 + u
F = R2unrestricted – R2restricted /J ~ F(J, N-Kunrestricted)
1-R2unrestricted /N- Kunrestricted

Where

J = No. of variables to be tested

restricted = values from model with variables set to zero (ie excluded
from the regression specification) Y = b0 + b1X1 + u

unrestricted = values from model with variables included in the

regression specification Y = b0 + b1X1 + b2X2 + b3X3 + u
F = R2unrestricted – R2restricted /J ~ F(J, N-Kunrestricted)
1-R2unrestricted /N- Kunrestricted

where

J = No. of variables to be tested

restricted = values from model with variables set to zero (ie excluded
from the regression specification) Y = b0 + b1X1 + u

unrestricted = values from model with variables included in the

regression specification Y = b0 + b1X1 + b2X2 + b3X3 + u

Intuition: Under the null that the extra variables have no explanatory
power then wouldn’t expect the RSS (or the R2 ) from the two models to
differ much

Hence reject null if estimated F > Fcritical

(because then the RSS from the two regressions are far apart)
 F-Test for restriction on a sub-set of variables

Given a multiple regression model (using the data set ps4data.dta)

. reg lhwage age edage union public if female==0

Source | SS df MS Number of obs = 6026
---------+------------------------------ F( 4, 6021) = 343.88
Model | 375.023052 4 93.755763 Prob > F = 0.0000
Residual | 1641.59102 6021 .272644248 R-squared = 0.1860
---------+------------------------------ Adj R-squared = 0.1854
Total | 2016.61407 6025 .33470773 Root MSE = .52215
------------------------------------------------------------------------------
lhwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
age | .0129394 .0005988 21.609 0.000 .0117656 .0141133
edage | .081065 .0025468 31.830 0.000 .0760724 .0860576
union | .0952227 .0154843 6.150 0.000 .0648679 .1255776
public | -.0121846 .0181974 -0.670 0.503 -.047858 .0234889
_cons | .1504611 .0539812 2.787 0.005 .0446387 .2562835
------------------------------------------------------------------------------

To test whether the union dummy variable is significantly different from zero, look at
the estimated t value

The equivalent F test in stata is given by

test union=0

( 1) union = 0

F( 1, 6021) = 37.82
Prob > F = 0.0000

^ 0

(β i − β i )2
(which is just the square of the t value F= ^
)
Var ( β i )
To test whether the variables union and public are (jointly) insignificant – they don’t contribute to explaining the dependent
variable

So omit union and public from the model and compare RSS

(Intuitively: If RSS is significantly different between the 2 models then suggests omitted variables do contribute something
to explain behaviour of dependent variable

. reg lhwage age edage if female==0

Source | SS df MS Number of obs = 6026

---------+------------------------------ F( 2, 6023) = 663.31
Model | 364.003757 2 182.001879 Prob > F = 0.0000
Residual | 1652.61031 6023 .27438325 R-squared = 0.1805
---------+------------------------------ Adj R-squared = 0.1802
Total | 2016.61407 6025 .33470773 Root MSE = .52382
------------------------------------------------------------------------------
lhwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
age | .013403 .0005926 22.615 0.000 .0122412 .0145648
edage | .0801733 .0024976 32.100 0.000 .0752771 .0850695
_cons | .1763613 .0532182 3.314 0.001 .0720345 .2806881
------------------------------------------------------------------------------
F test of null hypothesis that coefficients on union and public are zero (variables have no explanatory power)

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

= 1652.6 – 1641.6 /2 ~ F(2, 6026 -5)

1641.6 /6026 – 5

= 20.2

From F tables, critical value at 5% level F(2, 6021) = F(2, ∞ ) = 3.00

So estimated F > Fcritical

Stata equivalent is given by

test union public

( 1) union = 0
( 2) public = 0

F( 2, 6021) = 20.21
Prob > F = 0.0000

So reject null that union and public sector variables jointly have no
explanatory power in the model

Note that the t value on the public sector dummy indicates that the effect
of this variable is statistically insignificant from zero, yet the combined F
test has rejected the null that both variables have no explanatory power.

Be careful that test results don’t conflict

(technically the F test for joint restrictions is “less powerful” test of single
restrictions than the t test
Since this test is essentially a test of (linear) restrictions

– in the above case the restriction was that the coefficients on the sub-
set of variables were restricted to zero
Y = b0 + b1X1 + b2X2 + b3X3 + u
Test: b2= 0 b3 = 0

– other important uses of this test also include

1) Testing linear hypotheses other than a null of zero

– other important uses of this test also include

1) Testing linear hypotheses other than a null of zero

Eg. We know the Cobb-Douglas production function

y = ALαKβ
– other important uses of this test also include

1) Testing linear hypotheses other than a null of zero

Eg. We know the Cobb-Douglas production function

y = ALαKβ

with α+β=1 if there is constant returns to scale

– other important uses of this test also include

1) Testing linear hypotheses other than a null of zero

Eg. We know the Cobb-Douglas production function

y = ALαKβ

with α+β=1 if there is constant returns to scale

d.r.s. means α+β<1 i.r.s. means α+β>1

– other important uses of this test also include

1) Testing linear hypotheses other than a null of zero

Eg. We know the Cobb-Douglas production function

y = ALαKβ

with α+β=1 if there is constant returns to scale

d.r.s. means α+β<1 i.r.s. means α+β>1

Taking (natural) logs of y = ALαKβ
Taking (natural) logs of y = ALαKβ

becomes

Lny = LnA + αLnL + βLnK (1)

 Taking (natural) logs of y = ALαKβ

becomes

Lny = LnA + αLnL + βLnK (1)

and can test the null H0: by imposing the restriction that α+β=1 in (1)
against an unrestricted version that does not impose the constraint.
Example: Using the data set prodfn.dta containing information on the output,
labour input and capital stock of 27 firms

The unrestricted regression (ie not constraining the coefficients to sum to one) is
. reg logo logl logk
Source | SS df MS Number of obs = 27
-------------+------------------------------ F( 2, 24) = 200.25
Model | 14.2115637 2 7.10578187 Prob > F = 0.0000
Residual | .85163374 24 .035484739 R-squared = 0.9435
-------------+------------------------------ Adj R-squared = 0.9388
Total | 15.0631975 26 .57935375 Root MSE = .18837
------------------------------------------------------------------------------
logo | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
logl | .6029994 .125954 4.79 0.000 .3430432 .8629556
logk | .3757102 .085346 4.40 0.000 .1995648 .5518556
_cons | 1.170644 .326782 3.58 0.002 .4961988 1.845089

and test for the restriction using the command

constraint define 1 logl=1-logk

This produces a restricted OLS regression with the coefficients on logL and logk
constrained to add to one
. cnsreg logo logl logk, constraint(1)

Constrained linear regression Number of obs = 27

Root MSE = .18501
( 1) logl + logk = 1.0
------------------------------------------------------------------------------
logo | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
logl | .6369701 .075408 8.45 0.000 .4816644 .7922758
logk | .3630299 .075408 4.81 0.000 .2077242 .5183356
_cons | 1.069265 .1317588 8.12 0.000 .7979026 1.340627
(note that the coefficients on logl & logk do add to one)

Using the formula

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict
Stata produces the following output
. test _b[logl]+_b[logk]=1

( 1) logl + logk = 1.0

F( 1, 24) = 0.12
Prob > F = 0.7366
So estimated F < Fcritical at 5% level
So accept null that H0: α+β=1

So production function is Cobb-Douglas constant returns to scale

2) Testing Stability of Coefficients Across Sample Splits
2) Testing Stability of Coefficients Across Sample Splits

Often the estimated relationship varies over time or across easily

characterised sub-groups of your data (eg by gender)

In this case test the restricted model

Y = β0 + β1X1 + β2X2 + u (1)
2) Testing Stability of Coefficients Across Sample Splits

Often the estimated relationship varies over time or across easily

characterised sub-groups of your data (eg by gender)

In this case test the restricted model

Y = β0 + β1X1 + β2X2 + u (1)
coefficients same in both periods/both sub-groups
2) Testing Stability of Coefficients Across Sample Splits

Often the estimated relationship varies over time or across easily

characterised sub-groups of your data (eg by gender)

In this case test the restricted model

Y = β0 + β1X1 + β2X2 + u (1)
coefficients same in both periods/both sub-groups
Against an unrestricted model which allows the coefficients to vary
across the two-subgroups/time periods
Period/Group 1 Y = β01 + β11 X1 + β21 X2 + u1 (2)
2) Testing Stability of Coefficients Across Sample Splits

Often the estimated relationship varies over time or across easily

characterised sub-groups of your data (eg by gender)

In this case test the restricted model

Y = β0 + β1X1 + β2X2 + u (1)
(ie coefficients same in both periods/both sub-groups)
Against an unrestricted model which allows the coefficients to vary
across the two-subgroups/time periods
Period/Group 1 Y = β01 + β11 X1 + β21 X2 + u1 (2)
Period/Group 2 Y = β02 + β12 X1 + β22 X2 + u2 (3)
2) Testing Stability of Coefficients Across Sample Splits

Often the estimated relationship varies over time or across easily

characterised sub-groups of your data (eg by gender)

In this case test the restricted model

Y = β0 + β1X1 + β2X2 + u (1)
(ie coefficients same in both periods/both sub-groups)
Against an unrestricted model which allows the coefficients to vary
across the two-subgroups/time periods
Period/Group 1 Y = β01 + β11 X1 + β21 X2 + u1 (2)
Period/Group 2 Y = β02 + β12 X1 + β22 X2 + u2 (3)

Question: How get the unrestricted RSS if the regression is split across
two sub-samples?
Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)
Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)

So that

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict
Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)

So that

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

becomes

F = RSSrestrict – (RSSgroup1+ RSSgroup2) /J

(RSSgroup1+ RSSgroup2) /N- Kunrestrict ~ F(J, N-Kunrestrict)

where J is again the number of variables restricted (in this case the
entire set of rhs variables including the constant Y = β0 + β1X1 + β2X2 + u)
Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)

So that

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

becomes

F = RSSrestrict – (RSSgroup1+ RSSgroup2) /J

(RSSgroup1+ RSSgroup2) /N- Kunrestrict ~ F(J, N-Kunrestrict)

where J is again the number of variables restricted (in this case the
entire set of rhs variables including the constant Y = β0 + β1X1 + β2X2 + u)
Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)

So that

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

becomes

F = RSSrestrict – (RSSgroup1+ RSSgroup2) /J

(RSSgroup1+ RSSgroup2) /N- Kunrestrict ~ F(J, N-Kunrestrict)

where J is again the number of variables restricted (in this case the
entire set of rhs variables including the constant Y = β0 + β1X1 + β2X2 + u)

and Kunrestrict =2k

Can show that the unrestricted RSS in this case equals the sum of the
RSS from the two sub-regressions (2) & (3)

So that

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

becomes

F = RSSrestrict – (RSSgroup1+ RSSgroup2) /J

(RSSgroup1+ RSSgroup2) /N- Kunrestrict ~ F(J, N-Kunrestrict)

where J is again the number of variables restricted (in this case the
entire set of rhs variables including the constant Y = β0 + β1X1 + β2X2 + u)

and Kunrestrict =2k

- with this form of the test there are twice as many coefficients in the unrestricted
regressions compared to the restricted
Group 1: Y = β01 + β11 X1 + β21 X2 + u1 Group 2: Y = β02 + β12 X1 + β22 X2 + u2
As before
reject null
As before
reject null

(that coefficients in the two sub-samples are the same)

As before
reject null

(that coefficients in the two sub-samples are the same)

if estimated F > Fcritical

(because then the RSS from the unrestricted two regressions are
significantly different from the RSS in the pooled (restricted) sample
Example 1: Chow Test for Structural Break in Time Series Data

u cons /* read in consumption function data for years 1955-99 */

twoway (scatter cons year, msymbol(none) mlabel(year)), xlabel(55(5)100) xline(90)

600000

99
500000

98
97
96
9495
90 93
400000

89 9192
88
cons

87
86
85
8384
300000

79808182
78
7374757677
72
7071
69
6768
200000

646566
63
62
596061
58
555657

55 60 65 70 75 80 85 90 95 100
year

Graph suggests relationship between consumption and income changes over the
sample period.
(slope is steeper in 2nd period)
Try sample split before and after 1990

. reg cons income if year<90

Source | SS df MS Number of obs = 35
---------+------------------------------ F( 1, 33) = 3190.74
Model | 1.5750e+11 1 1.5750e+11 Prob > F = 0.0000
Residual | 1.6289e+09 33 49361749.6 R-squared = 0.9898
---------+------------------------------ Adj R-squared = 0.9895
Total | 1.5913e+11 34 4.6803e+09 Root MSE = 7025.8
------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
income | .9467359 .0167604 56.487 0.000 .9126367 .9808351
_cons | 6366.214 4704.141 1.353 0.185 -3204.433 15936.86

. reg cons income if year>=90

Source | SS df MS Number of obs = 10
---------+------------------------------ F( 1, 8) = 75.58
Model | 1.2302e+10 1 1.2302e+10 Prob > F = 0.0000
Residual | 1.3020e+09 8 162754768 R-squared = 0.9043
---------+------------------------------ Adj R-squared = 0.8923
Total | 1.3604e+10 9 1.5115e+09 Root MSE = 12758
------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
income | 1.047171 .1204489 8.694 0.000 .7694152 1.324927
_cons | -53227.48 59208.12 -0.899 0.395 -189761.6 83306.68

Looks like coefficients are different across periods, but standard error for second period
estimate is much larger. (why?)

Compare with regression pooled over both periods (restricting coefficients to be the same
in both periods).
. reg cons income
 Source | SS df MS Number of obs = 45
---------+------------------------------ F( 1, 43) = 5969.79
Model | 4.7072e+11 1 4.7072e+11 Prob > F = 0.0000
Residual | 3.3905e+09 43 78849774.6 R-squared = 0.9928
---------+------------------------------ Adj R-squared = 0.9927
Total | 4.7411e+11 44 1.0775e+10 Root MSE = 8879.7
------------------------------------------------------------------------------
cons | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
income | .9172948 .0118722 77.264 0.000 .8933523 .9412372
_cons | 13496.16 4025.456 3.353 0.002 5378.05 21614.26

Chow Test for sample split

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict

= 3.39 - (1.63+1.30)/2 ~ F(2, 45-2(2))

(1.63+1.30)/45-2(2)

Important: With this form of the test there are twice as many coefficients in the
unrestricted regressions (income and the constant for the period 1955-89, and a
different estimate for income and the constant for the period 1990-99,

so the unrestricted degrees of freedom are

N = N55-89 + N90-99 = 35+10 = 45

and k = 2*2
 )

= 3.22 ~ F(2, 41)

From table F critical at 5% level is 3.00. Therefore reject null that coefficients are
the same in both time periods. Hence mpc is not constant over time.
Eample 2: Chow Test of Structural Break – Cross Section Data

Suppose wish to test whether estimated OLS coefficients were the same for men
and women in ps4data.dta

This is like asking whether the estimated intercepts and the slopes are different
across the 2 sub-samples
Restricted regression is obtained by pooling all observations on men & women
and running a single OLS regression
. reg lhwage age edage union public

Source | SS df MS Number of obs = 12098

---------+------------------------------ F( 4, 12093) = 724.06
Model | 763.038968 4 190.759742 Prob > F = 0.0000
Residual | 3186.01014 12093 .263459038 R-squared = 0.1932
---------+------------------------------ Adj R-squared = 0.1930
Total | 3949.04911 12097 .326448633 Root MSE = .51328
------------------------------------------------------------------------------
lhwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
age | .0100706 .0004335 23.231 0.000 .0092209 .0109204
edage | .0869484 .0018669 46.574 0.000 .083289 .0906078
union | .1780204 .0109133 16.312 0.000 .1566285 .1994123
public | -.0250529 .0114298 -2.192 0.028 -.0474571 -.0026487
_cons | .0177325 .0393914 0.450 0.653 -.059481 .094946
------------------------------------------------------------------------------

Unrestricted regression obtained by running separate estimates for men and

women
(effectively allowing separate estimates of the constant and all the slope
variables) and then adding the residual sums of squares together
Men
. reg lhwage age edage union public if female==0

Source | SS df MS Number of obs = 6026

---------+------------------------------ F( 4, 6021) = 343.88
Model | 375.023052 4 93.755763 Prob > F = 0.0000
Residual | 1641.59102 6021 .272644248 R-squared = 0.1860
---------+------------------------------ Adj R-squared = 0.1854
Total | 2016.61407 6025 .33470773 Root MSE = .52215
------------------------------------------------------------------------------
lhwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
age | .0129394 .0005988 21.609 0.000 .0117656 .0141133
edage | .081065 .0025468 31.830 0.000 .0760724 .0860576
union | .0952227 .0154843 6.150 0.000 .0648679 .1255776
public | -.0121846 .0181974 -0.670 0.503 -.047858 .0234889
_cons | .1504611 .0539812 2.787 0.005 .0446387 .2562835

Women
. reg lhwage age edage union public if female==1

Source | SS df MS Number of obs = 6072

---------+------------------------------ F( 4, 6067) = 464.55
Model | 407.028301 4 101.757075 Prob > F = 0.0000
Residual | 1328.94152 6067 .21904426 R-squared = 0.2345
---------+------------------------------ Adj R-squared = 0.2340
Total | 1735.96982 6071 .285944626 Root MSE = .46802
------------------------------------------------------------------------------
lhwage | Coef. Std. Err. t P>|t| [95% Conf. Interval]
---------+--------------------------------------------------------------------
age | .0051823 .0005881 8.811 0.000 .0040293 .0063353
edage | .0854792 .0025947 32.944 0.000 .0803927 .0905657
union | .2086894 .0145456 14.347 0.000 .1801748 .2372039
public | .0784844 .0141914 5.530 0.000 .0506642 .1063045
_cons | .066159 .0545192 1.213 0.225 -.040718 .1730359

F = RSSrestrict – RSSunrestrict /J ~ F(J, N-Kunrestrict)

RSSunrestrict /N- Kunrestrict
becomes

F = RSSpooled – (RSSmen +RSSwomen)/No. Restricted (RSSmen

+RSSwomen)/N- Kunrestrict

= 3186 – (1641.6 + 1328.9) /5 ~ F(5, 12098 -10)

(1641.6 + 1328.9) /12098 – 2(5)

= 175.4
Note

1. J= 5 because 5 values are restricted – (constant, age, edage, union, public)

2. N-Kunrestricted = 12098 –2(5)

because in the unrestricted regression there are 2*5 estimated parameters (5 for
men and 5 for women)

From F tables, critical value at 5% level F(5, 12088) = F(5, ∞ ) = 2.21

So estimated F > Fcritical

Reject null that coefficients are the same for men and women