Regression Models with Dummy Independent Variables

 Dummy variables classify data into mutually exclusive categories

 Regression model with only dummy or qualitative variables – Analysis of Variance (ANOVA) –
compares the mean of two or more categories
 Regression model with mix of qualitative and quantitative variables – Analysis of Covariance
(ANCOVA) – examines the main and interaction effects of categorical variables on a continuous
dependent variable, controlling the effects of other continuous variables
 In addition to examining impact of qualitative aspects or attributes, dummy (independent) variables
are also used for seasonality analysis and examining structural breaks/differences

ANOVA: Example 1

Model Specification Interpretation

To examine if average monthly per Three Alternatives:
capita consumption expenditure (1) If D1i = 0, D2i = 0 => E Yi   
(MPCE) varies depending on whether
(2) If D1i = 1, D2i = 0 => E Yi      1
the households belong to rural, urban
and semi-urban areas, i.e., (3) If D1i = 0, D2i = 1 => E Yi      2

Yi     1 D1i   2 D2 i  ui Possibilities
(a) Comparison between rural and urban households
Yi = Monthly per capita consumption
expenditure (1) Β1 is not statistically significant => average MPCE of urban households is not
Household from three different types significantly different from that of rural households
of locations (2) Β1 is statistically significant and positive => average MPCE of urban
D1i = 1 if the household is in urban households is significantly higher than that of rural households
area (3) Β1 is statistically significant and negative => average MPCE of urban
D1i = 0 otherwise households is significantly lower than that of rural households
D2i = 1 if the household is in semi-
urban area (b) Comparison between rural and semi-urban households
D2i = 0 otherwise
Base category = Households from (1) Β2 is not statistically significant => average MPCE of semi-urban households
rural area is not significantly different from that of rural households
(2) Β2 is statistically significant and positive => average MPCE of semi-urban
households is significantly higher than that of rural households
(3) Β2 is statistically significant and negative => average MPCE of semi-urban
households is significantly lower than that of rural households

(c) Comparison between urban and semi-urban households

(1) Β1 is not significantly different from Β2 => average MPCE of urban

households is not significantly different from that of semi-urban households
(2) Β1 is significantly higher that Β2 => average MPCE of urban households is
significantly higher than that of semi-urban households
(3) Β1 is significantly lower than Β2 => average MPCE of urban households is
significantly lower than that of semi-urban households

ANOVA: Example 2

Model Specification Interpretation

To examine if monthly per capita
consumption expenditure varies Four Alternatives:
depending on (i) if the households belong (1) If D1i = 0, D2i = 0 => E Yi    (Rural, Female Head)
to rural or urban areas, and (ii) if the (2) If D1i = 1, D2i = 0 => E Yi      1 (Urban, Female Head)
household head is male or female, i.e.,
(3) If D1i = 0, D2i = 1 => E Yi      2 (Rural, Male Head)

Yi    1 D1i   2 D2i  3 ( D1i * D2i )  ui (4) If D1i = 1, D2i = 1 => E Yi      1   2   3 (Urban, Male Head)
Some Possibilities
Yi = Monthly per capita consumption (a) Differential impact of being in urban area (with female head)
expenditure (1) Β1 is not statistically significant => average MPCE of urban households
D1i = 1 if the household is in urban area female head is not significantly different from that of rural households
D1i = 0 otherwise (2) Β1 is statistically significant and positive => average MPCE of urban
D2i = 1 if the household head is male households with female head is significantly higher than that of others
D2i = 0 otherwise (3) Β1 is statistically significant and positive => average MPCE of urban
Base category = Households from rural households with female head is significantly lower than that of others
area with female head
(b) Differential impact of having male household head (in rural areas)
(1) Β2 is not significant => average MPCE of rural households with male head
is not significantly different from those with female head
(2) Β2 is significant and positive => average MPCE of rural households with
male head is significantly higher than those with female head
(3) Β2 is significant and negative => average MPCE of rural households with
male head significantly lower than those with female head

(c) Differential impact of having male head in urban area

(1) Β3 is not statistically significant => average MPCE of urban households
with male head is not significantly different from others
(2) Β3 is statistically significant and positive => average MPCE of urban
households with male head is significantly higher than that of others
(3) Β3 is statistically significant and negative => average MPCE of urban
households with male head is significantly lower than that of others

ANOVA: Example 3

Model Specification Interpretation

To examine if monthly MPCE varies depending on (i) Eight Alternatives:
if the households belong to rural or urban areas, (ii) (1) If D1i = 0, D2i = 0, D3i=0 =>
if the household head is male or female, and (iii) if E Yi    (Rural, Female Head, BPL)
the household is of APL or BPL category, i.e., (2) If D1i = 1, D2i = 0, D3i=0 =>
E Yi      1 (Urban, Female Head, BPL)
Yi    1D1i  2 D2i  3 D3i  4 ( D1i * D2i ) 
(3) If D1i = 0, D2i = 1, D3i=0 =>,
5 ( D1i * D3i )  6 ( D2i * D3i )  7 ( D1i * D2i * D3i )  ui E Yi      2 (Rural, Male Head, BPL)
(4) If D1i = 0, D2i = 0, D3i=1 =>,
Yi = Monthly per capita consumption expenditure E Yi      3 (Rural, Female Head, APL)
D1i = 1 if the household is in urban area
(5) If D1i = 1, D2i = 1, D3i=0 =>
D1i = 0 otherwise
D2i = 1 if the household head is male E Yi      1   2   4 (Urban, Male Head, BPL)
D2i = 0 otherwise (6) If D1i = 1, D2i = 0, D3i=1 =>
D3i = 1 if the household is APL E Yi      1   3   5 (Urban, Female Head, APL)
D3i = 0 otherwise (7) If D1i = 0, D2i = 1, D3i=1 =>
Base category = BPL households from rural area with E Yi      2   3   6 (Rural, Male Head, APL)
female head
(8) If D1i = 1, D2i = 1, D3i=1 =>
E Yi      1   2   3   4   5   6  7 (Urban, Male
Head, APL)
 ANCOVA: Example 1

Model Specification Interpretation

To examine if monthly per capita Two Alternatives:
consumption expenditure varies depending (1) Given Xi and D1i = 0, E Yi      2 X i
on income and whether the households
(2) Given Xi and D1i = 1, E Yi   (    1 )  (  2   3 ) X i
belong to rural or urban areas, i.e.,
The PRFs will differ depending on statistical significance and sign of β 1 and β3
Yi     1 D1i   2 X i   3 ( D1i * X i )  ui
Possibilities

Yi = Monthly per capita consumption (1) If both β1 and β3 are not significant, the two PRFs will coincide
expenditure
Xi = Monthly income of the household (2) If β1 is significant and β3 not, the two PRFs will be parallel (difference
Household from two different types of will be only in respect of intercept – autonomous consumption)
location – rural and urban (a) If β1 is positive, PRF for urban households will have higher intercept
D1i = 1 if the household is in urban area (b) If β1 is negative, PRF for urban households will have lower intercept
D1i = 0 otherwise
Base category = Households from rural area (3) If β3 is significant and β1 not, the two PRFs will be concurrent
(difference will be only in respect of slope – induced consumption)
(a) If β3 is positive, PRF for urban households will be steeper
(b) If β3 is negative, PRF for urban households will be flatter

(4) If both β1 and β3 are significant, the two PRFs will be dissimilar
(a) If both β1 and β3 positive, PRF for urban households will be steeper
with a higher intercept
(b) If both β1 and β3 negative, PRF for urban households will be flatter
with a lower intercept
(c) If β1 is positive but and β3 is negative, PRF for urban households will
be flatter with a higher intercept
(d) If β1 is negative but and β3 is positive, PRF for urban households will
be steeper with a lower intercept
ANCOVA: Example 2

Model Specification Interpretation

To examine if MPCE varies depending on income, Four Alternatives:
the households belong to rural or urban areas, and if (1) Given Xi and D1i = 0, D2i=0: E Yi      4 X i
the household head is male or female, i.e.,
(2) Given Xi and D1i = 1, D2i=0: E Yi   (    1 )  (  4   5 ) X i
Yi     1 D1i   2 D2 i   3 ( D1i * D2 i ) 
(3) Given Xi and D1i = 0, D2i=1: E Yi   (    2 )  (  4   6 ) X i
 4 X i   5 ( D1i * X i )   6 ( D2 i * X i )  ui
(4) Given Xi and D1i = 1, D2i=1:
i E Y   (        )  (      ) X
1 2 3 4 5 6 i
Yi = Monthly per capita consumption expenditure
Xi = Monthly income of the household The PRFs will differ depending on statistical significance and sign of β 1,
Household from two different types of location – β2, β3, β4, β5, and β6
rural and urban
D1i = 1 if the household is in urban area Possibilities
D1i = 0 otherwise
D2i = 1 if the household head is male How will you explain the coefficients of alternative PRFs?
D2i = 0 otherwise
Base category = Households from rural area with
female head
Two Alternative Forms: (i) Yi     1 D1i   2 D2 i  ui , and (ii) Yi   1 D1i   2 D2 i   3 D3i  vi

How will the results and the interpretation of the coefficients differ?
Empirical Example:

 Cross-sectional data set of 100 households

 Dependent Variable: (i) Monthly per capita consumption expenditure (LNMPCE)*
Independent Variables: (ii) Age of the household head (LNHHAGE)*, (iii) Per capita landholding
(PCLAND), (iii) Household size (LNHHSIZE)*, (iv) Location of the households (RURAL, URBAN and
SEMIURBAN), (vi) If the household is below the poverty line (BPL)
 Two interaction terms: (i) Rural households below the poverty line (RURBPL=RURAL*BPL), and (ii)
Per capita landholding of rural households (RURLAND=RURAL*PCLAND)
 Estimations of alternative models with URBAN as the base (reference) category
 Shortlisting of the following two models initially based on R 2:

*These three variables are measured in natural logarithmic scale

Model I

Number of Observation 100

F( 8, 91) 59.11
Prob. > F-Stat < 0.001
R-squared 0.363
Root MSE 0.50841
Variable Coefficient Robust Std. Err. t-Stat P>t
SEMIURBAN 0.17379 0.14038 1.24 0.219
RURAL 0.24295 0.15306 1.59 0.116
LNHHSIZE -0.61185 0.10184 -6.01 <0.001
LNHHAGE 0.22987 0.19642 1.17 0.245
BPL -0.18597 0.13794 -1.35 0.181
RURBPL 0.09078 0.19716 0.46 0.646
RURLAND -0.42537 0.84882 -0.5 0.617
PCLAND 2.17516 0.75790 2.87 0.005
Intercept 7.64819 0.72726 10.52 <0.001

Model II

Number of Observation 100

F(6, 93) 70.53
Prob. > F-Stat <0.001
R-squared 0.3519
Root MSE 0.50728
Variable Coefficient Robust Std. Err. t-Stat P>t
SEMIURBAN 0.1379 0.1394 0.99 0.325
RURAL 0.2690 0.1110 2.42 0.017
LNHHSIZE -0.6099 0.0970 -6.28 0.001
BPL -0.1585 0.1098 -1.44 0.152
RURLAND -0.4974 0.6963 -0.71 0.477
PCLAND 2.2442 0.6875 3.26 0.002
Intercept 8.5250 0.1487 57.32 0.001

Comparison between the two models:

 Not much difference in R2 between the two models (to be tested statistically)
 More variables turnout to be statistically significant in Model II
 No change in sign of the statistically significant coefficients
 Finally, selection of Model II for further discussions
Interpretation of the Results

 Coefficient of RURAL is significant and positive => Average MPCE of rural households is higher
than that of urban households.
 Coefficient of SEMIURBAN is not significant => Average MPCE of semi-urban households is not
significantly different from that of the urban households.
 Coefficient of LNHHSIZE is statistically significant and negative => Households with more
members in the family have lower average MPCE.
 Coefficient of PCLAND is statistically significant and positive => Households with more
landholding per member in the family have higher average MPCE.
 Coefficient of BPL is not statistically significant=>Average MPCE does not differ depending on
whether households below to the BPL category

Use of Dummy Variables for Seasonality Analysis

 Assumption: The components of the time-series is additive, i.e.,

TS = Trend (T) + Seasonal (S) + Cyclical (C) + Randomness (U)

 Steps to be followed:

 Use of dummy for every quarter (without intercept) or for three quarters (with
intercept) treating the omitted quarter as the base or reference
 Estimation of the residuals – deseasonalized values of the time-series

 Important Questions:

 Will the results differ depending on selection of the base/reference quarter?

Answer: NO – The deseasonalized time-series will be the same irrespective of selection of the
base/reference quarter

 Consider the function: GDP = f(GFCF). How to account for seasonality in GFCF, if any?

Frisch-Waugh Theorem: Use of dummy variables in GDP = f(GFCF) will deseasonalize both GDP
and GFCF

 Running regression of GDP on the dummy variables and estimating the residuals (U1)
 Running regression of GFCF on the dummy variables and estimating the residuals (U2)
 Regressing U1 on U2 – Same coefficient of GFCF as one gets when GDP is regressed on
GFCF and the dummy variables

 What to do when the components of a time-series are multiplicative? – To be discussed in time-

series econometrics
Seasonality Analysis: Example
SUMMARY OUTPUT (Model I) SUMMARY OUTPUT (Model II)
Multiple R 0.261 Multiple R 0.261
R Square 0.068 R Square 0.068
Adjusted R Square -0.049 Adjusted R Square -0.049
Standard Error 0.155 Standard Error 0.155
Observations 28 Observations 28
df SS MS F df SS MS F
Regression 3 0.042161 0.014054 0.582704 Regression 3 0.042161 0.014054 0.582704
Residual 24 0.578828 0.024118 Residual 24 0.578828 0.024118
Total 27 0.620988 Total 27 0.620988
Coefficients Standard Error t Stat P-value Coefficients Standard Error t Stat P-value
Intercept 14.923 0.058698 254.231 1.1E-42 Intercept 14.824 0.058698 252.5 1.29E-42
D1 -0.0989 0.083011 -1.19149 0.24511 D2 0.0085 0.083011 0.1 0.919106
D2 -0.0904 0.083011 -1.08886 0.287029 D3 0.0402 0.083011 0.5 0.63239
D3 -0.0587 0.083011 -0.70695 0.486408 D4 0.0989 0.083011 1.2 0.24511

Model I Model II
Obs. Predicted GDP Residuals Obs. Predicted GDP Residuals Obs. Predicted GDP Residuals Obs. Predicted GDP Residuals
1 14.824 -0.21485 15 14.864 -0.00601 1 14.824 -0.21485 15 14.864 -0.00601
2 14.832 -0.22893 16 14.923 0.003556 2 14.832 -0.22893 16 14.923 0.003556
3 14.864 -0.19649 17 14.824 0.084883 3 14.864 -0.19649 17 14.824 0.084883
4 14.923 -0.20196 18 14.832 0.086258 4 14.923 -0.20196 18 14.832 0.086258
5 14.824 -0.15017 19 14.864 0.066738 5 14.824 -0.15017 19 14.864 0.066738
6 14.832 -0.15386 20 14.923 0.07161 6 14.832 -0.15386 20 14.923 0.07161
7 14.864 -0.13404 21 14.824 0.143054 7 14.864 -0.13404 21 14.824 0.143054
8 14.923 -0.146 22 14.832 0.151753 8 14.923 -0.146 22 14.832 0.151753
9 14.824 -0.07783 23 14.864 0.140815 9 14.824 -0.07783 23 14.864 0.140815
10 14.832 -0.07596 24 14.923 0.149779 10 14.832 -0.07596 24 14.923 0.149779
11 14.864 -0.07551 25 14.824 0.219563 11 14.864 -0.07551 25 14.824 0.219563
12 14.923 -0.08343 26 14.832 0.219427 12 14.923 -0.08343 26 14.832 0.219427
13 14.824 -0.00465 27 14.864 0.204511 13 14.824 -0.00465 27 14.864 0.204511
14 14.832 0.001316 28 14.923 0.206451 14 14.832 0.001316 28 14.923 0.206451
Testing for Structural Break

(1) Two sub-periods, i.e., one possible structural break

(2) Regression equation for the first sub-period: Yi   1   1 X i   1 Z i  u1i
(3) Regression equation for the second sub-period: Yi   2   2 X i   2 Z i  u 2 i
(4) Combined model: Yi   1  (  2   1 )D1i   1 X i  (  2   1 )D2 i   1 Z i  (  2   1 )D3 i  u i
(Unrestricted model)

(i) D1 = 1 for the second sub-period, and D1=0 for the first sub-period
(ii) D2 = X for observations corresponding to the second sub-period, and D 2=0 for observations
corresponding to the first sub-period
(iii) D3 = Z for observations corresponding to the second sub-period, and D3=0 for observations
corresponding to the first sub-period

Restrictions:

Elimination of Dummy
All coefficients are the same  1   2 ;  1   2 ;  1   2 D1, D2, D3
Only intercepts change:  1   2 ;  1   2 D2, D3
Only coefficients of X change:  1   2 ;  1   2 D1, D3
Only coefficients of Z change:  1   2 ;  1   2 D1, D2
Only slopes change:  1   2 D1
Only intercepts and coefficients of X change:  1   2 D3
Only intercepts and coefficients of Z change:  1   2 D2
All the slope coefficients and the intercept change None (Full Model)