© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 7

4.2 Tests of Structural Changes

This section focuses on Chow Test and leaves general discussion on dummy variable models to
other section.

4.2.1 Chow Test: Simple Example1

Chow Test examines whether parameters of one group of the data are equal to those of other
groups. Simply put, the test checks whether the data can be pooled. If only intercepts are
different across groups, this is a fixed effect model, which is simple to handle. Let us consider
two groups.

y = α + βx + ε for all observations

y = α1 + β1 x + ε1 for n1 observations (group 1)
y = α 2 + β2 x + ε 2 for n2 observations (group 2)

The null hypothesis is α1 = α 2 and β1 = β 2 . If the null hypothesis is rejected, two groups have
different slopes and intercepts; data are not poolable.

(e' e − e1' e1 − e2' e2 ) / J ( SSE − SSE1 − SSE2 ) J

F ( J , n1 + n2 − 2 K ) = = where
(e1e1 + e2e2 ) /(n1 + n2 − 2 K ) ( SSE1 + SSE2 ) (n1 + n2 − 2 K )
' '

e' e is the SSE of the pooled model and J is the number of restrictions (often equal to K—all
parameters).2

In order to conduct the Chow test,

1. Run pooled OLS to get SSEpooled

2. Run separate OLS to get SSE1 and SSE2
3. Apply the formula

In the following example, we assume that two groups have difference slopes of cost and
different intercepts; there are two restrictions, J=2.
. use http://www.indiana.edu/~statmath/stat/all/panel/airline.dta, clear
(Cost of U.S. Airlines (Greene 2003))

. gen d1=(airline<=2)
. gen d0=(d1==0)
. gen d2=(airline>=2 & airline<=4)
. gen d3 =(airline>=5)

. gen cost0=cost*d0
. gen cost1=cost*d1
. gen cost2=cost*d2
. gen cost3=cost*d3

First, fit the pooled model.

1
Greene 2003 (289-291), http://www.stata.com/support/faqs/stat/chow.html
2
If we want to test the null hypothesis that only intercept is different, J will be K-1 (all the slopes are equal)

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 8
. regress output cost // pooled model

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 1, 88) = 880.73
Model | 107.123089 1 107.123089 Prob > F = 0.0000
Residual | 10.7034329 88 .12162992 R-squared = 0.9092
-------------+------------------------------ Adj R-squared = 0.9081
Total | 117.826522 89 1.32389351 Root MSE = .34875

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .9691952 .032658 29.68 0.000 .9042942 1.034096
_cons | -14.12819 .4380397 -32.25 0.000 -14.99871 -13.25768
------------------------------------------------------------------------------

And then fit separate models using two subsets of data.

. regress output cost if d1==0 // if airline >= 3

Source | SS df MS Number of obs = 60

-------------+------------------------------ F( 1, 58) = 314.51
Model | 30.7715 1 30.7715 Prob > F = 0.0000
Residual | 5.6746959 58 .097839584 R-squared = 0.8443
-------------+------------------------------ Adj R-squared = 0.8416
Total | 36.4461959 59 .617732134 Root MSE = .31279

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .8454175 .047671 17.73 0.000 .7499936 .9408413
_cons | -12.64286 .610876 -20.70 0.000 -13.86566 -11.42006
------------------------------------------------------------------------------

. regress output cost if d1==1 // if airline <= 2

Source | SS df MS Number of obs = 30

-------------+------------------------------ F( 1, 28) = 166.32
Model | 2.80505683 1 2.80505683 Prob > F = 0.0000
Residual | .472224836 28 .016865173 R-squared = 0.8559
-------------+------------------------------ Adj R-squared = 0.8508
Total | 3.27728166 29 .113009712 Root MSE = .12987

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .5144647 .0398915 12.90 0.000 .4327507 .5961787
_cons | -7.328991 .5798709 -12.64 0.000 -8.516802 -6.141179
------------------------------------------------------------------------------

(e' e − e1' e1 − e2' e2 ) / J (10.7034 − 5.6747 − .4722) 2

F= = = 31.8745
(e1e1 + e2e2 ) /(n1 + n2 − 2 K ) (5.6747 + .4722) (60 + 30 − 2 * 2)
' '

. di ((10.7034329-.472224836-5.6746959)/2)/((.472224836+5.6746959)/(30+60-2*2))
31.8745

. di Ftail(2,86,31.8745)
4.393e-11

The large F 31.8745 (2, 68) rejects the null hypothesis of equal slope and intercept (p<.0000).

Alternatively, you may regress y on two dummies and two interaction terms with the intercept
suppressed.3 Parameter estimates are identical to those of the above, while standard errors are

3
Therefore, R2 and standard errors are not reliable.

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 9

different. This estimation is handy since parameter estimates are slopes and intercepts of
individual groups; any further computation is not need at all.
. regress output cost0 d0 cost1 d1, noconstant

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 4, 86) = 824.72
Model | 235.789788 4 58.9474471 Prob > F = 0.0000
Residual | 6.14692073 86 .071475822 R-squared = 0.9746
-------------+------------------------------ Adj R-squared = 0.9734
Total | 241.936709 90 2.68818566 Root MSE = .26735

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost0 | .8454175 .0407452 20.75 0.000 .7644187 .9264162
d0 | -12.64286 .522126 -24.21 0.000 -13.68081 -11.60491
cost1 | .5144647 .0821229 6.26 0.000 .3512098 .6777197
d1 | -7.328991 1.193756 -6.14 0.000 -9.702098 -4.955883
------------------------------------------------------------------------------

. test _b[cost0]=_b[cost1]
( 1) cost0 - cost1 = 0

F( 1, 86) = 13.03
Prob > F = 0.0005

. test _b[d0]=_b[d1], accum

( 1) cost0 - cost1 = 0
( 2) d0 - d1 = 0

F( 2, 86) = 31.87
Prob > F = 0.0000

More convenient way is to regress y on a regressor of interest, cost, an interaction term, and a
dummy with the intercept included. The intercept is the intercept of the baseline group and the
dummy coefficient is the deviation from the baseline intercept. The coefficient of the regressor is
the slope of the baseline, while the coefficient of the interaction term is the deviation of slope
from the baseline slope. That is, the intercept of compared group is 7.3290 (=5.3139-12.6429)
and the slope is .5145 (=-.3310+.8454).
. regress output cost cost1 d1

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 3, 86) = 520.83
Model | 111.679602 3 37.2265339 Prob > F = 0.0000
Residual | 6.14692073 86 .071475822 R-squared = 0.9478
-------------+------------------------------ Adj R-squared = 0.9460
Total | 117.826522 89 1.32389351 Root MSE = .26735

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .8454175 .0407452 20.75 0.000 .7644187 .9264162
cost1 | -.3309528 .0916752 -3.61 0.001 -.513197 -.1487085
d1 | 5.31387 1.302946 4.08 0.000 2.723699 7.904041
_cons | -12.64286 .522126 -24.21 0.000 -13.68081 -11.60491
------------------------------------------------------------------------------

. test _b[cost1]=0
( 1) cost1 = 0

F( 1, 86) = 13.03
Prob > F = 0.0005

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 10
. test _b[d1]=0, accum
( 1) cost1 = 0
( 2) d1 = 0

F( 2, 86) = 31.87
Prob > F = 0.0000

4.2.2 Chow Test: Comparing Three Groups

What if we want to compare three groups? Let us fit two remaining models for group 2 and 3.
Restrictions here are 1) slop2 is equal to the baseline slope (slop1), 2) slope3 is equal to the
baseline slope, 3) intercept2 is equal to the baseline intercept, and 4) intercept3 is equal to the
baseline intercept. Degrees of freedom is 84=N-(group*K)=90-3*2.
. regress output cost if d2==1

Source | SS df MS Number of obs = 30

-------------+------------------------------ F( 1, 28) = 43.01
Model | 4.48338312 1 4.48338312 Prob > F = 0.0000
Residual | 2.91859218 28 .104235435 R-squared = 0.6057
-------------+------------------------------ Adj R-squared = 0.5916
Total | 7.4019753 29 .255240528 Root MSE = .32286

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .6205532 .0946202 6.56 0.000 .4267326 .8143739
_cons | -9.498564 1.255486 -7.57 0.000 -12.07031 -6.926819
------------------------------------------------------------------------------

. regress output cost if d3==1

Source | SS df MS Number of obs = 30

-------------+------------------------------ F( 1, 28) = 817.74
Model | 9.93465282 1 9.93465282 Prob > F = 0.0000
Residual | .340168734 28 .012148883 R-squared = 0.9669
-------------+------------------------------ Adj R-squared = 0.9657
Total | 10.2748216 29 .354304191 Root MSE = .11022

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .7370425 .0257742 28.60 0.000 .6842466 .7898385
_cons | -11.47175 .3181417 -36.06 0.000 -12.12344 -10.82007
------------------------------------------------------------------------------

(e' e − e1' e1 − e2' e2 − e3' e3 ) / J (10.703 − .472 − 2.919 − .340) 4

F= = = 39.245
(e1e1 + e2e2 + e3e3 ) /(n1 + n2 + n3 − gK ) (.472 + 2.919 + .340) (30 + 30 + 30 − 3 * 2)
' ' '

. di ((10.7034329-.472224836-2.91859218-.340168734)/4)/ ///
((.472224836+2.91859218+.340168734)/(30+30+30-3*2))
39.244693

. di Ftail(4,84,39.244693)
1.696e-18

Now, include all interaction terms and dummies without the regressor of interest. The model
report all parameter estimates, which are identical to the above.
. regress output cost1 d1 cost2 d2 cost3 d3, noconstant

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 6, 84) = 893.83
Model | 238.205723 6 39.7009539 Prob > F = 0.0000

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 11
Residual | 3.73098575 84 .044416497 R-squared = 0.9846
-------------+------------------------------ Adj R-squared = 0.9835
Total | 241.936709 90 2.68818566 Root MSE = .21075

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost1 | .5144647 .0647376 7.95 0.000 .3857268 .6432026
d1 | -7.328991 .9410399 -7.79 0.000 -9.200352 -5.457629
cost2 | .6205532 .0617658 10.05 0.000 .4977251 .7433813
d2 | -9.498564 .8195514 -11.59 0.000 -11.12833 -7.868796
cost3 | .7370425 .049282 14.96 0.000 .6390399 .8350452
d3 | -11.47175 .6083095 -18.86 0.000 -12.68144 -10.26206
------------------------------------------------------------------------------

. test _b[cost1]=_b[cost2]=_b[cost3], notest

. test _b[d1]=_b[d2]=_b[d3], accum
( 1) cost1 - cost2 = 0
( 2) cost1 - cost3 = 0
( 3) d1 - d2 = 0
( 4) d1 - d3 = 0

F( 4, 84) = 39.24
Prob > F = 0.0000

Finally, include the regresor, two interaction terms, and two dummies excluding baseline
interaction and dummy. The coefficient of the regressor is the baseline slope and the intercept is
the baseline intercept. Coefficients of interaction terms are deviations from the baseline slope. As
a result, the slope of group 1 is .5144647 (=-.2225778+.7370425) and the intercept is -7.328991
(=4.142762-11.47175).
. regress output cost cost1 d1 cost2 d2

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 5, 84) = 513.75
Model | 114.095537 5 22.8191073 Prob > F = 0.0000
Residual | 3.73098575 84 .044416497 R-squared = 0.9683
-------------+------------------------------ Adj R-squared = 0.9665
Total | 117.826522 89 1.32389351 Root MSE = .21075

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | .7370425 .049282 14.96 0.000 .6390399 .8350452
cost1 | -.2225778 .0813614 -2.74 0.008 -.3843739 -.0607817
d1 | 4.142762 1.120534 3.70 0.000 1.914458 6.371067
cost2 | -.1164893 .0790173 -1.47 0.144 -.2736239 .0406453
d2 | 1.973189 1.020639 1.93 0.057 -.0564649 4.002842
_cons | -11.47175 .6083095 -18.86 0.000 -12.68144 -10.26206
------------------------------------------------------------------------------

. test _b[cost1]=0, notest

. test _b[cost2]=0, accum notest
. test _b[d1]=0, accum notest
. test _b[d2]=0, accum
( 1) cost1 = 0
( 2) cost2 = 0
( 3) d1 = 0
( 4) d2 = 0

F( 4, 84) = 39.24
Prob > F = 0.0000

Hypothesis testing is identical to the above.

4.2.3 Chow Test: Including Covariates

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 12

Now, suppose we need to include some covariates for control, load and fuel. We may regress
the dependent variable on all interactions, dummies, and covariates with the intercept suppressed.
Again R2 is not reliable here. A coefficient of an interaction term is the slope of cost of the group,
just as the dummy coefficient is the intercept of the group.
. regress output cost1 d1 cost2 d2 cost3 d3 load fuel, noconstant

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 8, 82) = 1526.59
Model | 240.323102 8 30.0403877 Prob > F = 0.0000
Residual | 1.61360768 82 .019678142 R-squared = 0.9933
-------------+------------------------------ Adj R-squared = 0.9927
Total | 241.936709 90 2.68818566 Root MSE = .14028

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost1 | 1.017854 .0884737 11.50 0.000 .8418518 1.193856
d1 | -9.777424 .7217463 -13.55 0.000 -11.21321 -8.34164
cost2 | 1.070322 .0839545 12.75 0.000 .9033093 1.237334
d2 | -10.56474 .5832504 -18.11 0.000 -11.72501 -9.404465
cost3 | 1.10826 .0672781 16.47 0.000 .9744229 1.242098
d3 | -11.10194 .4070026 -27.28 0.000 -11.9116 -10.29229
load | 2.041787 .398699 5.12 0.000 1.248648 2.834927
fuel | -.4733269 .0556113 -8.51 0.000 -.5839555 -.3626983
------------------------------------------------------------------------------

This model is written as,

output= -9.7774+1.0179*cost+2.0418*load-.4733*fuel (group 1)
output=-10.5647+1.0703*cost+2.0418*load-.4733*fuel (group 2)
output=-11.1019+1.1083*cost+2.0418*load-.4733*fuel (group 3)

Let us conduct the hypothesis test for the four restrictions.

. test _b[cost1]=_b[cost2]=_b[cost3], notest
. test _b[d1]=_b[d2]=_b[d3], accum
( 1) cost1 - cost2 = 0
( 2) cost1 - cost3 = 0
( 3) d1 - d2 = 0
( 4) d1 - d3 = 0

F( 4, 82) = 0.86
Prob > F = 0.4923

We may also fit the same model including the regressor of interest and excluding baseline
interaction term and its intercept. Covariates remain unchanged. Note that slope2 is 1.0703 (=-
.0379+1.1083) and the intercept of group 2 is -10.5647 (=.5372079-11.10194).
. regress output cost cost1 d1 cost2 d2 load fuel

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 7, 82) = 843.67
Model | 116.212915 7 16.6018449 Prob > F = 0.0000
Residual | 1.61360768 82 .019678142 R-squared = 0.9863
-------------+------------------------------ Adj R-squared = 0.9851
Total | 117.826522 89 1.32389351 Root MSE = .14028

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | 1.10826 .0672781 16.47 0.000 .9744229 1.242098
cost1 | -.0904063 .0572606 -1.58 0.118 -.2043159 .0235034
d1 | 1.32452 .8454028 1.57 0.121 -.3572557 3.006295

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 13
cost2 | -.0379388 .0545639 -0.70 0.489 -.1464838 .0706061
d2 | .5372079 .7215669 0.74 0.459 -.8982187 1.972634
load | 2.041787 .398699 5.12 0.000 1.248648 2.834927
fuel | -.4733269 .0556113 -8.51 0.000 -.5839555 -.3626983
_cons | -11.10194 .4070026 -27.28 0.000 -11.9116 -10.29229
------------------------------------------------------------------------------

. test _b[cost1]=0, notest

. test _b[cost2]=0, accum notest
. test _b[d1]=0, accum notest
. test _b[d2]=0, accum

( 1) cost1 = 0
( 2) cost2 = 0
( 3) d1 = 0
( 4) d2 = 0

F( 4, 82) = 0.86
Prob > F = 0.4923

4.2.4 Chow Test: Different Slopes of Multiple Regressors4

If we assume that more than one regressor have different slopes across group, model will be
complicated. But the underlying logic remains same. Let us create a set of interaction terms for
the regressor load.
. gen load1=load*d1
. gen load2=load*d2
. gen load3=load*d3

First, include all interactions, dummies, and covariate fuel. Do not forget the suppress the
intercept.
. regress output cost1 load1 d1 cost2 load2 d2 cost3 load3 d3 fuel, noconstant

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 10, 80) = 1207.56
Model | 240.344451 10 24.0344451 Prob > F = 0.0000
Residual | 1.59225866 80 .019903233 R-squared = 0.9934
-------------+------------------------------ Adj R-squared = 0.9926
Total | 241.936709 90 2.68818566 Root MSE = .14108

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost1 | 1.014764 .0975898 10.40 0.000 .8205541 1.208974
load1 | 1.817979 .6811907 2.67 0.009 .4623659 3.173591
d1 | -9.754334 .7455903 -13.08 0.000 -11.23811 -8.270562
cost2 | 1.019311 .0981769 10.38 0.000 .8239324 1.214689
load2 | 2.661021 .7210198 3.69 0.000 1.226146 4.095896
d2 | -10.38952 .6108743 -17.01 0.000 -11.6052 -9.173845
cost3 | 1.111908 .0687766 16.17 0.000 .9750386 1.248778
load3 | 1.703393 .7009252 2.43 0.017 .3085071 3.098278
d3 | -11.11345 .4096207 -27.13 0.000 -11.92862 -10.29827
fuel | -.4615669 .0578651 -7.98 0.000 -.5767222 -.3464117
------------------------------------------------------------------------------

. test _b[cost1]=_b[cost2]=_b[cost3], notest

. test _b[load1]=_b[load2]=_b[load3], accum notest
. test _b[d1]=_b[d2]=_b[d3], accum

( 1) cost1 - cost2 = 0
( 2) cost1 - cost3 = 0
( 3) load1 - load2 = 0

4
http://www.stata.com/support/faqs/stat/chow3.html

http://www.masil.org
© Jeeshim and KUCC625 (3/22/2008) Statistical Inferences in Linear Regression: 14
( 4) load1 - load3 = 0
( 5) d1 - d2 = 0
( 6) d1 - d3 = 0

F( 6, 80) = 0.74
Prob > F = 0.6152

Now, include two regressors of interest, two sets of interactions, two dummies, and a covariate,
excluding baseline interaction terms and its dummy. Make sure you include two regresors of
interest, cost and load.
. regress output cost load cost1 load1 d1 cost2 load2 d2 fuel

Source | SS df MS Number of obs = 90

-------------+------------------------------ F( 9, 80) = 648.89
Model | 116.234264 9 12.9149182 Prob > F = 0.0000
Residual | 1.59225866 80 .019903233 R-squared = 0.9865
-------------+------------------------------ Adj R-squared = 0.9850
Total | 117.826522 89 1.32389351 Root MSE = .14108

------------------------------------------------------------------------------
output | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cost | 1.111908 .0687766 16.17 0.000 .9750386 1.248778
load | 1.703393 .7009252 2.43 0.017 .3085071 3.098278
cost1 | -.0971444 .0777234 -1.25 0.215 -.2518189 .0575301
load1 | .1145858 .9857632 0.12 0.908 -1.847146 2.076317
d1 | 1.359111 .8696813 1.56 0.122 -.3716095 3.089832
cost2 | -.0925978 .0789714 -1.17 0.244 -.2497558 .0645603
load2 | .9576281 1.020969 0.94 0.351 -1.074165 2.989421
d2 | .723922 .7484569 0.97 0.336 -.7655546 2.213399
fuel | -.4615669 .0578651 -7.98 0.000 -.5767222 -.3464117
_cons | -11.11345 .4096207 -27.13 0.000 -11.92862 -10.29827
------------------------------------------------------------------------------

. test _b[cost1]=0, notest

. test _b[cost2]=0, accum notest
. test _b[load1]=0, accum notest
. test _b[load2]=0, accum notest
. test _b[d1]=0, accum notest
. test _b[d2]=0, accum

( 1) cost1 = 0
( 2) cost2 = 0
( 3) load1 = 0
( 4) load2 = 0
( 5) d1 = 0
( 6) d2 = 0

F( 6, 80) = 0.74
Prob > F = 0.6152

http://www.masil.org