1

Dummy Variable Models

1.1 Introduction
This chapter presents alternative equation specifications (ES) of the dummy variables
models (DVMs) presented in Chapter 6, of the main book, starting with the simplest equation
specification, which can be used to conduct the analysis based on alternative models by using
different type of endogenous (impact, down-stream, or dependent) variable, such as the
numerical endogenous variable for cell-mean models and quantile regression models, (where
some of them have advantages for testing selected hypotheses, because the hypotheses could be
tested directly by using the test statistics, such as the t-statistic and the Z-statistic, presented in
the output of the statistical results. Refer to some examples presented in the main book.

1.2 The Simplest Equation Specification of DVMs

1.2.1 Two-Way Interaction DVMs
The following equation specification (ES) is representing all IxK factorial DVMs having
only two-way interaction dummy variables, namely [A=i]*[TP=k], for all i, and k, where the
[A=i] is the dummy variable of the i-th group of the of the firms or research objects, for
i=1,…,I; and [TP=k] is the dummy variable of the k-th time period, for k=1,…,K..
Y @Expand(A,TP) (1.1)
Note that this two-way interaction DVM, in fact is a two-way ANOVA model, with the
main objectives are to test the hypotheses on the parameters or statistics differences of the
endogenous variable Y between the cells or groups generated by the factor A, and the time-period
TP. However, the parameters differences could be representing the main-effect of A, conditional

1
for TP=k, the main-effect of TP, conditional for A=i, and the interaction effect A*TP, on the
endogenous variable Y.

1.2.1.1 A 2x2 Factorial DVM

Table 1.1 presents the parameters of a 2x2 factorial DVM in (1.1). Based on the
parameters in this table, the following statistics and hypotheses can be considered.

Table 1.1 Parameters of a 2x2 factorial DVM in (1.1)

A TP=1 TP=2 TP(1-2)
1 C(1) C(2) C(1)-C(2)
2 C(3) C(4) C(3)-C(4)
A(1-2) C(1)-C(3) C(2)-C(4) C(1)-C(2)-C(3)+C(4)

(1) Three basic statistics are as follows:

1.1 The main effect of A on Y, conditional for TP=k, such as follows:
Ef(A|TP=1) = C(1) – C(3), and Ef(A|TP=2) = C(2) – C(4),
1.2 The main effect of TP on Y, conditional for A=i, such as follows:
Ef(TP|A=1) = C(1) - C(2), and Ef(TP|A=2) = C(3) - C(4),
1.3 The interaction effect of A*TP on Y, or the difference-in-differences (DID), is defined as
follows:
Ef(A*TP) = Ef(A|TP=1) - Ef(A|TP=2) = Ef(TP|A=1) - Ef(TP|A=2)
Ef(A*TP) = C(1) - C(2) – C(3) + C(4)
(2) Hypotheses
2.1 For each of the conditional main effects, either two-sided or one sided hypotheses can easily
be tested using the Wald test.
2.2 For the interaction effect, namely Ef(A*TP), in general, only two-sided hypotheses would be
tested using the Wald test, with the statistical hypothesis as follows
H0 : C(1)-C(2)=C(3)-C(4); versus H1 : Otherwise
or
. H0 : C(1)-C(3)=C(2)-C(4); versus H1 : Otherwise

2
2.3 In addition, the hypothesis on the joint effects of A and TP on Y, also can easily be tested
using the Wald test, with the statistical hypothesis as follows:
H0 : C(1)=C(2)=C(3)=C(4); versus H1 : Otherwise

1.2.1.2 A 2x4 Factorial DVM

As an extension of the2x2 factorial DVM in Table 1.1, Table 1.2 presents the parameters
of a 2x4 factorial DVM in (1.1). Based on the parameters in this table, similar conditional main-
effects statistics and hypotheses to the 2x2 DVM could easily be defined. For this reason, this
section would present only new additional hypotheses, such as follows:

Table 1.2 Parameters of a 2x4 factorial DVM in (1.1)

A TP=1 TP=2 TP=3 TP=4
1 C(1) C(2) C(3) C(4)
2 C(5) C(6) C(7) C(8)
A(1-2) C(1)-C(5) C(2 )-C(6) C(3)-C(7) C(4)-C(8)

(1) The interaction effect of A*TP on Y, with the statistical hypothesis

H0 : C(1)-C(5) = C(2)-C(6) = C(3)-C(7) = C(4)-C(8); versus H1 : Otherwise
(2) The effect of the time-period TP on Y, conditional for A=1, with the statistical hypothesis:
H0 : C(1)=C(2)=C(3)=C(4); versus H1 : Otherwise
(3) The effect of the time-period TP on Y, conditional for A=2, with the statistical hypothesis:
H0 : C(5)=C(6)=C(7)=C(8); versus H1 : Otherwise
(4) The joint effects of A and TP on Y, with the statistical hypothesis:
H0 : C(1)=C(2)=C(3)=C(4)=C(5)=C(6)=C(7)=C(8); versus H1 : Otherwise

1.2.1.3 A 3x4 Factorial DVM

As an extension of the 2x4 factorial DVM in Table 1.2, Table 1.3 presents the parameters
of a 3x4 factorial DVM in (1.1). Based on the parameters in this table, similar conditional main-
effects statistics and hypotheses to the 2x2 and 2x4 DVMs could easily be defined. For this
reason, this section would present only a new additional hypothesis, such as follows:

3
 Table 1.3 Parameters of a 3x4 factorial DVM in (1.1)
A TP=1 TP=2 TP=3 TP=4
1 C(1) C(2) C(3) C(4)
2 C(5) C(6) C(7) C(8)
3 C(9) C(10) C(11) C(12)
A(1-3) C(1)-C(9) C(2 )-C(10) C(3)-C(11) C(4)-C(12)
A(2-3) C(5)-C(9) C(6 )-C(10) C(7)-C(11) C(8)-C(12)

The hypothesis on the interaction effect of A*TP on Y, with the statistical hypothesis as
follows:
H0 : C(1)-C(9)=C(2)-C(10)=C(3)-C(11)=C(4)-C(12),
C(5)-C(9)=C(6)-C(10)=C(7)-C(11)=C(8)-C(12),
H1 : Otherwise

1.2.2 Three-Way Interaction DVMs

As the extension of the two-way interaction DVMs in (1.1), The following equation
specification (ES) is representing all IxJxK factorial DVMs having only three-way interaction
dummy variables, namely [A=i]*[B=j]*TP=k], for all i, j, and k, where the [A=i] is the
dummy variable of the i-th level of factor A, [B=j] is the dummy variable of the j-th level of
factor B, and [TP=k] is the dummy variable of the k-th time period.
Y @Expand(A,B,TP) (1.2)

1.2.2.1 A 2x2x2 Factorial DVM

As an extension of the 2x2 factorial DVM in Table 1.1, Table 1.4 presents the parameters
of a 2x2x2 factorial DVM in (1.2). Based on this table, the following notes and comments are
presented.
(1) Note that for each level of B, we have a 2x2 factorial DVM of Y by the factors A and TP,
which is exactly the same as the 2x2 factorial DVM in Table 1.1. Thence, we would have
similar main-effects, interaction effects, and hypotheses, but conditional for each B=j, as
presented in Table 1.5, such as follows:

4
 Table 1.4 Parameters of a 2x2x2 factorial DVM in (1.3)
B=1 B=2
A TP=1 TP=2 TP(1-2) TP=1 TP=2 TP(1-2)
1 C(1) C(2) C(1)-C(2) C(3) C(4) C(3)-C(4)
2 C(5) C(6) C(5)-C(6) C(7) C(8) C(7)-C(8)
A(1-2) C(1)-C(5) C(2)-C(6) (a) C(3)-C(7) C(4)-C(8) (b)

Table 1.5 Selected Statistics and hypotheses based on the 2x2x2 factorial DVM in Table 1.2
Statistics / Effects Parameters Hypothesis
No. (1) (2) (3)
1 Ef(A|B=1,TP=1) C(1) - C(5) One/two-Sided
2 Ef(A|B=1,TP=2) C(2) - C(6) One/two-Sided
3 Ef(A|B=2,TP=1) C(3 ) - C(7) One/two-Sided
4 Ef(A|B=2,TP=2) C(4) – C(8) One/two-Sided

5 Ef(A*TP|B=1)=(a) (C(1) - C(5)) – (C(2) - C(6)) Two-Sided

6 Ef(A*TP|B=2)=(b) (C(3) - C(7)) – (C(4) - C(8)) Two-Sided
7 Ef(A*B|TP=1) (C(1) - C(5)) – (C(3) - C(7)) Two-Sided
8 Ef(A*B|TP=2) (C(2) - C(6)) – (C(4) - C(8)) Two-Sided

9 Ef(A*B*TP) Ef(A*TP|B=1) - Ef(A*TP|B=2) Two-Sided

1.1 The main effects of A, conditional for B=j and TP=k, namely Ef(A|B=j,TP=k).
1.2 The two-way interaction effect of A*TP, conditional for B=j, namely Ef(A*TP|B=j).
1.3 The three-way interaction effect of A*B*TP, defined as Ef(A*TP|B=1) - Ef(A*TP|B=2).
(2) The others statistics: Ef(B|A=i,TP=k), Ef(TP|A=i,B=j), Ef(A*B|TP=k), and Ef(B*TP|A=i),
are left for exercises.
(3) In addition, we have the hypotheses on the joint effects of (A,TP) on Y, conditional for B=j,
such as follows:

5
 H0 : Ef(A,TP|B=1) = 0, or C(1)=C(2)=C(5)=C(6); versus H1: Otherwise
and
H0 : Ef(A,TP|B=2) = 0, or C(3)=C(4)=C(7)=C(8); versus H1: Otherwise

(4) The other hypotheses on the joint effects of (A,B) on Y, conditional for TP=k and the joint
effects (B,TP) on Y, conditional for A=i are left for the exercises.
(5) Finally, we also have the hypothesis on the joint effects of (A,B,TP) on Y. However, for
testing this hypothesis it is recommended to apply the following ES. Moreover, for the
factors A, B, and TP having more levels.

Y C @Expand(A,B,TP,@Droplast) (1.3)

1.2.2.2 Special Notes and Comments

Note that the 2x2x2 factorial MDV in Table 1.4 can be viewed as a 4x2 factorial MDV of
a cell-factor, CF, generated by the factor A, and B; and the time-periods or time-points TP.
Similarly, for all MDVs in (1.2), and (1.3).

1.3 Special Recommended Equation Specifications

Note that that by using the simplest equation specification of the interaction DVMs,
presented above, all hypotheses should be tested using the Wald test. It is recognized that by
using special equation specifications, many of the hypotheses can easily be tested directly based
on the output of the statistical results.

1.3.1 Bi-Factorial DVMs having a Main Factor and Two-way Interactions

Referring to the MDVs in (1.1) the following special equation specification (ES) is
representing all IxK factorial DVMs of Y on a factor A and the time-period TP.

Y C @Expand(TP,@Droplast) @Expand(A,@Droplast)*@Expand(TP) (1.4)

6
 Table 1.6 Parameters of a 2x4 factorial DVM in (1.4)
A TP=1 TP=2 TP=3 TP=4 TP(1-4) TP(2-4) TP(3-4)

1 C(1)+C(2)+C(5) C(1)+C(3)+C(6) C(1)+C(4)+C(7) C(1)+C(8)

2 C(1)+C(2) C(1)+C(3) C(1)+C(4) C(1) C(2) C(3) C(4)

A(1-2) C(5) C(6) C(7) C(8)

For an illustration, Table 1.6 presents the model parameters for a 2x4 factorial DVM in
(1.4). Based on this table the following and comments are presented.
(1) Referring the statistics and hypotheses defined based on the parameters of the 2x4 DVM in
(1.4), but they have different forms of functions.
(2) Note that the hypotheses, either one-sided or two sided, based on each of the parameters C(2)
to C(8), can easily be tested using the test statistic presented in the output of the statistical
result.
(3) In addition, the hypotheses on the joint effects A and TP on the endogenous variable Y,
namely H0 : C(2)=C(3)=…=C(8)=0; versus H1 : otherwise, also can be tested using the test
statistic presented in the output.
(4) The hypothesis on the interaction effect of A*TP on Y, namely H0: C(5)=C(6)=C(7)=C(8),
versus H1 : Otherwise, should be tested using the Wald test.

1.3.2 Three-Factorial DVMs having only Two- and Three-way Interactions

The following special equation specification (ES) is representing all IxJxK factorial
DVMs having two-way interactions [A=i]*[B=j], and three-way interactions
[A=i]*[B=j]*[TP=k].. For an illustration, Table 1.7 presents the model parameters of a 2x2x2
factorial DVM, using the following equation specification (ES).

Y C @Expand(B,TP,@Droplast) @Expand(A,@Droplast)*@Expand(B,TP) (1.5)

7
 Table 1.7 Parameters of a 2x2x2 factorial DVM in (1.5)
B=1 B=2
A TP=1 TP=2 TP=1 TP=2
1 C(1)+C(2)+C(5) C(1)+C(3)+C(6) C(1)+C(4)+C(7) C(1)+C(8)
2 C(1)+C(2) C(1)+C(3) C(1)+C(4) C(4)
TP(1-2) C(5) C(6) C(7) C(8)

Table 1.8 Various Statistics and hypotheses based on the parameters in Table 1.6
No. Statistics /Effects Parameters H0 Hypothesis
1 Ef(A|B=1,TP=1) C(5) C(5) = 0 One/two Sided

2 Ef(A|B=1,TP=2) C(6) C(6) = 0 One/two Sided

3 Ef(A|B=2,TP=1) C(7) C(7) = 0 One/two Sided

4 Ef(A|B=2,TP=2) C(8) C(8) = 0 One/two Sided

5 Ef(A*TP|B=1)=(a) C(5)-C(7) C(5)=C(7) One/two Sided

6 Ef(A*TP|B=2)=(b) C(6)-C(8) C(6)=C(8) One/two Sided

7 Ef(A*B|TP=1) C(5)-C(6) C(5) = C(6) One/two Sided

8 Ef(A*B|TP=2) C(7)-C(8) C(7) = C(8) One/two Sided

9 Ef(A*B*TP) C(5)-C(6)-(7)+C(8) C(5)-C(6) = C(7)-C(8) Two Sided

Referring to the 9 statistics in Table 1.5, Table 1.8 presents exactly the same statistics,
but using different functions of the parameters.
The other DVMs having only two- and three-way interactions independent variables are
as follows:
Y C @Expand(A,B,@droplast) @Expand(A,B)*@Expand,TP,@Droplast)) (1.6)
Y C @Expand(A,TP,@Droplast) @Expand(A,TP)*@Expand(B,@Droplast) (1.7)

8
1.4 Alternative Equation Specifications
1.4.1 Bi-Factorial Hierarchical DVMs
Y C @Expand(A,@Droplast) @Expand(TP,@Droplast)
@Expand(A,@Droplast)*@Expand(TP,@Droplast) (1.8)

1.4.2 Three-way Hierarchical DVMS

By using the following general equation specification (ES), it looks too long and difficult,
but by using Microsoft Word, you can save this ES in a file, and then you can use the block-
copy-paste method to conduct the data analysis. Note that this ES presents all DVMs having the
intercept parameter, indicated by the symbol “C”.
Y C @Expand(A,@Droplast) @Expand(B,@Droplast) @Expand(TP,@Droplast)
@Expand(A,@Droplast)*@Expand(B,@Droplast)
@Expand(A,@Droplast)*@Expand(TP,@Droplast)
@Expand(B,@Droplast)*@Expand(TP,@Droplast)
@Expand(A,@Droplast)*@Expand(B,@Droplast)*@Expand(TP,@Droplast) (1.9)

1.5 Application of Alternative DVMs

This section would present the applications of alternative 2x2x2 factorial DVMs of a
numerical, a zero-one, or an ordinal dependent variables. Since, all of the IxJxK factorial
equation specifications presented above in fact are presenting exactly the same DVM, then this
section only presents the empirical statistical results based on the equation specification (1.5),
where the dichotomous TP indicates two time periods, two years before and after a new tax
policy.

1.5.1 Applications of DVMs of a Numerical Dependent Variable

1.5.1.1 Application Cell-Means or ANOVA Models
Example 1.1 (An ANOVA Model and its AR(1) Model) Figure 1.1 presents the statistical
results of an ANOVA Model in ES (1.9), and its AR(1) Model. Based on these results, the
following notes and comments are presented.

9
 Figure 1.1 Statistical results of the ANOVA model in (1.9), and its AR(1) Model

(1) The model in Figure 1.1a has a very small DW Statistic of 0.517773, which indicates that the
model has autoregressive problem. For this reason, it is presented its AR(1) in Figure 1.1b,
having the equation as follows:
lnSALE = C(1)+C(2)[TP=1,B=1] +C(3)[TP=1,B=1]+C(4)[TP=1,B=1]
+C(5)[A=1,TP=1,B=1] +C(6)[A=1,TP=1,B=1]+C(7)[A=1,TP=1,B=1]
+C(8)[A=1,TP=1,B=1]+[AR(1)=C(9)]+ε (1.10)

(2) Note that the differences of estimated parameters C(1) up to C(8), based on the two models
presented in Figure 1.1, because the unpredictable impact of the term AR(1). Refer to the
Example presented in Chapter 1 of the main book. Furthermore, note that the two statistical
results also present difference values of other statistics.
(3) For testing various hypotheses refer to the parameters of the model presented in Table 1.8.
Note that corresponding to each of the parameters C(5) up to C(8), the hypothesis, either one
or two sided hypotheses, can be tested using the t-test statistic in Figure 1.1. Do it for an
exercise.
(3) Based on Figure 1.1b, it can be concluded that the joint effects of the independent variables
has a significant effect based on the F-statistic of F0 = 622.4516 with a p-value = 0.0000.
(4) In addition, the following hypotheses can be tested using the Wald test.
4.1 For the two-sided hypothesis H0: C(5)=C(6), versus H1: C(5)≠C(6), the null hypothesis is
accepted based on the F-statistic of F0 = 0.008156 with df = (1,1222) and a p-value = 0,9281.

10
 So that it can be concluded that the interaction A*B, conditional for TP=1, has insignificant
effect on lnSale.
4.2 For the two-sided hypothesis H0: C(7)=C(8), versus H1: C(7)≠C(8), the null hypothesis is
accepted based on the F-statistic of F0 = 2.050326 with df = (1,1222) and a p-value = 0.1524,
at the 10% level of significance. So that it can be concluded that the interaction A*B,
conditional for TP=2, has insignificant effect on lnSale. In this case, I would present the
illustration to test the two possible one-sided hypotheses, at the α = 0.10 level of significance.
 For testing the one-sided hypothesis using the t-test statistic, it is important to see the sign
of the estimated value of C(7) - C(8) = - 0.214836 < 0. So that t0 < 0, then we have the
conclusion as follows:
 For the left-sided hypothesis H0: C(7)-C(8) ≥ 0, versus H1: C(7)-C(8) < 0, the null
hypothesis is rejected based on the t-statistic of t0 = - SQRT(F0) = - SQRT(2.050326),
with df = 1222, and a p-value = 0.1524/2 = 0.0762 < 0.10. Note that by using EViews 6,
the value of the t-statistic should be computed, however by using EViews 7, the results of
the Wald test would present the value of the t-test statistic.
 For the right-sided hypothesis H0: C(7)-C(8) ≤ 0, versus H1: C(7)-C(8)> 0, the null
hypothesis is accepted, since t0 < 0, which has the same sign as the null hypothesis.
4.3 The null hypothesis H0: C(5)-C(6)=C(7)-C(8) is accepted based on the F-statistic of F0 =
0.771370 with df = (1,1222) and a p-value = 0,3800. Then it can e concluded that the three-
way interaction A*B*TP has an insignificant effect on lnSale.

1.5.1.2 Application Quantile Regressions

Example 1.2 (Application of Quantile Regression Models) Figure 1.2 presents the statistical
results of two quantile regressions using the ES in (1.9). Based on each of these statistical results
the same statistical hypotheses on the median or first-quartile of LNSALE as in the Example 1.1
can easily be tested. Do it an exercise. The equation of the models are as follows:

Med(lnSALE) = C(1)+C(2)[TP=1,B=1] +C(3)[TP=1,B=1]+C(4)[TP=1,B=1]

+C(5)[A=1,TP=1,B=1] +C(6)[A=1,TP=1,B=1]+C(7)[A=1,TP=1,B=1]
+C(8)[A=1,TP=1,B=1]+ε (1.11)

11
 Figure 1.2 Statistical results of two Quantile Regression model in (1.5)

Q1(lnSALE) = C(1)+C(2)[TP=1,B=1] +C(3)[TP=1,B=1]+C(4)[TP=1,B=1]

+C(5)[A=1,TP=1,B=1] +C(6)[A=1,TP=1,B=1]+C(7)[A=1,TP=1,B=1]
+C(8)[A=1,TP=1,B=1]+ε (1.12)

1.5.1.3 Application Instrumental Variables Models

In has been found an error message as presented on the right would be obtained, which
indicates that we should increase the number of the instrumental variables. It is very easy to
increase the number of the instrumental variables. But it is not an easy task to select the best
possible set of instruments. In fact, based on the
author’s experiment, the statistical results can be
obtained by using any set of instruments. For this
reason, Agung (2009a) recommended not to apply an
instrumental variables model, since the other models
can be applied with confidence. The following
examples present the statistical results of the models using alternative instruments.

Example 1.3 (Application of TSLS estimation method) Figure 1.3 presents equation
specification (ES) of the model in Figure 1.1, and the statistical results based on an instrumental

12
variables model with the set of instruments, which is exactly the same set of the regressors of the
ES. For a comparison, Figure 1.4 presents the statistical results of two instrumental variables
model, with different sets of instruments. Based on these statistical results, the following notes
and comments are presented.
(1) The ES of the instrumental variables models has eight regressors or parameters, and each of
instrumental variables model (IVM) also has a list of eight variables, such as follows:
1.1 The list of instruments of the first IVM is exactly the same as the regressors of the ES,
namely the list : C @Expand(TP,B,@Droplast) @Expand(A,@Droplast)*@Expand(TP,B),
which contains (1+3+1x4) = 8 variables.

The

Figure 1.3 The equation specification of the model in Figure 1.1a with a set of instrumentals

1.2 The list of instruments of the second IVM is @Expand(A,B,TP), which represent a list of
2x2x2=8 dummy variables. Note that its statistical results are exactly the same as the
statistical results of the first IVM, because both lists of instruments are representing the same
set of 8 dummy variables, generated by the dichotomous variables A, B and TP.
1.3 The list of instruments of the third IVM is the list of eight variables: C, X1 up to X7. Note
that the statistical results of this model has a negative R-squared. So that this model is not an
acceptable IVM, in a statistical sense.

13
 Figure 1.4 Statistical results of the model in Figure 1.1, with two alternative sets of instruments

Figure 1.5 Statistical results of the model in Figure 1.1b, with an arbitrary list of instruments

1.4 Since the first two IVMs have small DW statistic, Figure 1.5 presents the statistical results of
an AR(1) IVM, which is an acceptable model, using unintentional or arbitrary list of
instruments, which are selected using the trial and error method.
(2) By inserting one or more instrumental variables, including to choose the option “Include
lagged regressors for linear equations with ARMA terms”, as presented in Figure 1.5,
acceptable statistical results could be obtained. Do it for exercises.

14
(3) The illustrative statistical results above clearly show that we would never know, weather we
have been using a relevant and good list of instruments. Furthermore, we would never know
the true IVM for the corresponding population (Agung, 2009a, Section 2.14).

1.5.1.4 Application of GMM Estimation Method

Example 1.4 (An application of GMM method) As another type of the IVMs, Figure 1.6
presents two statistical results based on an AR(1) IVM of the AR(1) IVM in Figure 1.5, and a
first-order lagged (LV(1)) IVM, using the GMM method.
(1) In addition to using the option “Include lagged regressors for linear equations with ARMA
terms”, the lists of instruments used are exactly the same sets as the regressors of the
corresponding models, the equation specifications in particular.
(2) It is found that by inserting one or more instrumental variables, acceptable statistical results
also would be obtained. Refer to the notes presented in Example 1.7, pointer (3).

Figure 1.6 Statistical results of an AR(1) IVM, and LV(1) IVM, using the GMM method

Figure 1.6 Statistical results of an AR(1) , and LV(1) IVM, using the GMM method
(3) Both models, as well as other possible IVMs by using different list of instruments could be
acceptable model, in a statistical sense, but we would never know which one is the best fit
model, because we will not applied all possible models.

1.5.2 Application of DVMs of a Zero One Dependent Variable

15
 For the illustrative example, it is generated a dummy variable, namely DLNSALE =
1*(LNSALE<@Quantile(SALE,0.20). In this section, two DVMs would be considered, namely
the cell-proportions model, and the binary choice (probit, logit, and extreme value) models

1.5.2.1 Application of Cell-proportion Models

Example 1.5 (Application of cell-proportion models) Figure 1.7 presents the statistical results
of a cell-mean model, using two types of equation specifications. Based on these statistical
results are the following notes and comments.

(1) Both statistical results are presenting exactly the same cell-mean model. Note that the
coefficients of the second models are in fact the sampled proportions of (DLNSALE=1)
within the eight cells generated by the factors A, B, and TP. These proportions are important
to be noted, since they are also the same as the predicted probabilities of the corresponding
binary choice models.

Figure 1.7 Statistical results of a cell-proportion model, using two types of equation specifications

(1) Since the model has a very small DW-statistic, then it is recommended to apply its auto-
regressive model. However, it is left for exercise.

16
1.5.2.2 Application of Binary Choice Models
Example 1.6 (Application of a binary logit model) Figure 1.8 presents the statistical results of
a binary logit model using the same ES as the cell-proportion model in Figure 1.7b. Based on
these statistical results the following, findings, notes and comments are presented.
(1) Its predicted probabilities obtained by clicking Quick/Generate Series…, then inserts the
following equation, by using the block-copy-paste of the equation “Substituted Coefficient”.
However, remember to replace the dependent variable DLNSALE by any other symbol, such
as PP_1. Then PP_2, and PP_3 can be used for the binary probit, and extreme value
models, respectively, which are left for exercises.
PP_1 = 1-@CLOGISTIC(-(0.0512932943876*(A=1 AND B=1 AND TP=1)
+ 0.0893451003358*(A=1 AND B=1 AND TP=2)
- 1.24171313231*(A=1 AND B=2 AND TP=1)
- 0.998528830111*(A=1 AND B=2 AND TP=2)
- 2.75153531302*(A=2 AND B=1 AND TP=1)
- 2.57515352762*(A=2 AND B=1 AND TP=2)
- 3.7841896338*(A=2 AND B=2 AND TP=1)
- 3.86523084322*(A=2 AND B=2 AND TP=2))) (1.13)

Figure 1.8 Statistical results of a binary logit model,using the ES for the model in Figure 1.7b

17
(2) Figure 1.9 presents a part of the statistical results, namely the means of PP_1, by the factors
A, B, and TP, using the object View/Descriptive Statistics & Tests/Equality Test by
Classification…, and entering “ TP B A” as the “Series/Group for classify” … OK.

Category Statistics
Std. Err.
A B TP Count Mean Std. Dev. of Mean
1 1 1 78 0.512821 0 0
1 1 2 229 0.522321 0 0
1 2 1 58 0.224138 0 0
1 2 2 185 0.269231 0 0
2 1 1 50 0.060000 0 0
2 1 2 213 0.070755 0 0
2 2 1 90 0.022222 0 0
2 2 2 341 0.020528 0 0
All 1244 0.200555 0.201982 0.005727
Figure 1.9 Statistical results of an Equality Test by Classification of PP_1

(3) Note that the means of PP_1 in Figure 1.9 are exactly the same as the coefficient of the cell-
proportions model in Figure 1.7b. However, it is found that by using the ES for the model Figure
1.7a, the means obtained have small differences, because of using other stages in the estimation
process.

1.3.2.3 Application of Instrumental Variables Models

It is found that, the instrumental variables DVMs also could be applied using a zero-one
problem indicator. Find the following example.

Example 1.7 (Application of TSLS and GMM methods) For an illustration, Figure 1.10
presents the statistical results of an AR(1) instrumental variables DVM of DLNSALE, using the
simplest ES, namely DLNSALE @Expand(A,B,TP) AR(1), using the TSLS estimation method,
with its predicted probabilities. In addition, Figure 1.11 presents the statistical results, using the
GMM estimation method.

18
 Figure 1.10 Statistical results of an AR(1) Instrumental Variables DVM, using TSLS method

Figure 1.11Statistical results of an AR(1) Instrumental Variables DVM, using GMM method

(1) Both models give acceptable predicted probabilities, but they are not the same as the sampled
proportions of DLNSALE, by the factors A, B, and TP.
(2) Note that many other lists of instrumental variables also could give acceptable predicted
probabilities, which have been demonstrated in previous examples of IVMs. On the other
hand, poor models could be obtained, such as the statistical results having negative R-
squared, as presented in Figure 1.4b..

19
1.5.3 Application of Ordered Choice DVMs
Example 1.8 (Application of an ordered extreme value models) For an illustration, Figure
1.12 presents the statistical results of an ordered extreme value model of O3lnSale, having three
levels, by dichotomous factors A, B, and TP, using the ES in (1.5). Based on these results, the
following findings, notes and comments are presented.

Figure 1.12 Statistical results of ordered extreme value models, using the ES in (1.9)

Table 1.13 Parameters of a 2x2x2 factorial ordered extreme value model in (1.9)
TP=1 TP=2
A B=1 B=2 B(1-2) B=1 B=2 B(1-2)
1 C(1)+C(4) C(2)+C(5) (*) C(3)+C(6) C(7) (**)
2 C(1) C(2) C(1)-C(2) C(3) --- C(3)
A(1-2) C(4) C(5) C(5)-C(6) C(6) C(7) C(6)-(7)

(1) Note that the ES in (1.5) contains the intercept parameter “C”, as presented in the Estimation
Command, but the model in the Estimation Equation does not have an intercept parameter.
Similarly, for the ordered probit, and logit models. For this reason, compare to the
parameters of the other models in (1.5) – as presented in Table 1.10, then the ordered choice

20
 model, using the ES in (1.5), would have different form of the table parameters, such as
presented in Table 1.13, which has only 7 parameters, with an empty cell (A=2,B=2,TP=2).
(2) Various hypotheses on the cell differences can easily be tested using either the Z-Statistic in
the output, or the Wald test. Do it for exercises.
(3) Table 1.14 presents the summary of the predicted probabilities of the model. Which is
obtained using the steps as follows:

Table 1.14 Summary of the predicted probabilities of the ordered extreme value model in Figure 1.12
Category Statistics Predicted Probabilities
A B TP Count O3lnsale_1 O3lnsale_2 O3lnsale_3 Total
1 1 1 78 0.0122 0.027863 0.959937 1.000
1 1 2 229 0.783687 0.210216 0.006098 1.000
1 2 1 58 0.982969 0.01703 1.28E-06 1.000
1 2 2 185 0.76806 0.224247 0.007693 1.000
2 1 1 50 0.003903 0.009038 0.987060 1.000
2 1 2 213 0.181812 0.305664 0.512524 1.000
2 2 1 90 0.238989 0.358374 0.402637 1.000
2 2 2 341 0.12311 0.231309 0.645581 1.000
All 1244 0.387403 0.216619 0.395977 1.000

Table 1.15 Two-way tabulation of O3LNSALE and the cell-factor CF, generated by (A,B,TP)
Count CF = Cell (A=i,B=j,TP=k) = Cell(ijk)
% Col. 111 112 121 122 211 212 221 222 Total
47 148 17 69 3 19 4 10 317
1 60.26 64.63 29.31 37.3 6 8.92 4.44 2.93 25.48
29 79 38 95 23 70 28 98 460
2 37.18 34.5 65.52 51.35 46 32.86 31.11 28.74 36.98
2 2 3 21 24 124 58 233 467
3 2.56 0.87 5.17 11.35 48 58.22 64.44 68.33 37.54
78 229 58 185 50 213 90 341 1244
O3LNSale Total 100 100 100 100 100 100 100 100 100

3.1 By using the equation in “Substituted Coefficients” a new variable I_O3SALE can easily be
generated.
3.2 Then using each of the last three equations, namely O3SALE_1, O3SALE_2, and O3SALE_1,
the variables of the predicted probabilities can be generated, as presented in Table 1.14.

21
3.3 Finally, the summary table in (1.13) could be developed, based on three statistical results, by
using the object View/Descriptive Statistics & Tests/Equality Test by Classification…, and
entering “ TP B A” as the “Series/Group for classify” … OK, for the three predicted
probabilities. In fact, it is sufficient to conduct the data analysis twice only, because the
three predicted probabilities have a total of 1.000.
(4) For a comparison, Table 1.15 presents a tabulation of O3LNSALE and a cell-factor, CF,
generated by the factors A, B, and TP. Specifically, this table is presenting the distribution of
O3LNSALE by CF(A,B,TP), indicated by the total of “% Col” equal to 100.

1.6 Under-identified and the Worst DVMs

It is recognized that many DVMs having less number of parameters than the number of
cells generated by the categorical independent variables have been presented by the researchers,
because they have the statistical results, which could be considered as acceptable in a statistical
sense. Those models are defined as the under-identified models, which have hidden assumptions
which are not mentioned by the researchers. For instance, based on three dichotomous
categorical independent variables, A, B and TP, the DVMs only have three up to seven
parameters, which could be classified using the general equation specifications presented in the
following subsections.

1.6.1 The Worst DVMs

The following equation specification is representing the DVMs having the categorical
independent variables A, B, and TP with I, J and K levels, but the modesl only have . (I+J+K-3)
parameters, or additive dummy independent variables. Note that the best fit DVM should have
IxJxK parameters.
Y @Expand(A,@Droplast) @Expand(B,@Droplast)
@Expand(TP,@Droplast) (1.14)

22
 Table 1.16 The parameters of a 2x2x3 factorial DVM in (1.14)
A B TP=1 TP=2 TP=3 TP(1-3) TP(2-3)
1 C(1)+C(2)+C(3) C(1)+C(2)+C(4) C(1)+C(2) C(3) C(4)
1 2 C(1) +C(3) C(1) +C(4) C(1) C(3) C(4)
B(1-2) C(2) C(2) C(2) 0 0
1 C(2) +C(3) C(2) +C(4) C(2) C(3) C(4)
2 2 C(3) C(4) 0 C(3) C(4)
B(1-2) C(2) C(2) C(2) 0 0
A(1-2) B=1 C(1) C(1) C(1) 0 0
B=2 C(1) C(1) C(1) 0 0

For an illustration, Table 1.16 presents the parameters of a 2x2x3 factorial DVM, or
ANOVA model in (1.16). Note that the table has 12 cells, but the model has only (1+1+2)= 4
parameters. So that the cells differences have a specific pattern, which would never be observed
in reality.

1.6.2 Other Under-identified DVMs

The following equation specifications are representing only some of many DVMs having
the categorical independent variables A, B, and TP with I, J and K levels, but the models have
less than IxJxK parameters. So that each of the models have a specific hidden assumption.
However, they are acceptable models in a statistical sense.

1,6.2.1 Additive DVMs having (I+J+K-2) parameters

One of these models has the following equation specification.
Y C @Expand(A,@Droplast) @Expand(B,@Droplast)
@Expand(TP,@Droplast) (1.15a)

Y @Expand(A,@Droplast) @Expand(B,@Droplast)
@Expand(TP) (1.15b)

23
1.6.2.2 Two-way Interaction DVMs
Two out of many possible two-way interaction DVMs have the equations specifications
as follows:
(i) DVMs having (IxJ + K-2) parameters
Y @Expand(A,B,@Droplast) @Expand(TP,@Droplast) (1.16)

(ii) DVMs having (IxJ + K-1) parameters

Y C @Expand(A,B,@Droplast) @Expand(TP,@Droplast) (1.17a)

Y @Expand(A,B) @Expand(TP,@Droplast) (1.17b)

1.6.2.3 Three-way Interaction DVMs (IxJxK-1) parameters

One out of several possible three-way interaction DVMs has the following equation
specification.
Y @Expand(A,B,TP,@Droplast)) (1.18)

1.7 Illustrative Fixed-Effects Dummy Variables Models

1.7.1 DVMs from International Journals
1.7.1.1 DVMs from Puri, and Zarutskie (2012)
Puri and Zarutskie (2012) presents several of the simplest fixed-effects DVMs, namely
the industry-year fixed effects models, having only a single variable, namely the Venture Capital
(VC), which is a dummy variable equal to one for VC-financed firms (= 1 in all firm-years for
the firm-financed firms), and zero for non-VC-financed firms, based on the data having 101,936;
and 17,885 observations. The models can be represented using the following equation
specification.
G(Y) VC @Expand(Industry) @Expand(Year,@Dropfirst) (1.19a)
or
G(Y) VC @Expand(Industry,@Dropfirst) @Expand(Year) (1.19b)

24
where G(Y) is representing alternative dependent variables, such as Log(Employment),
Log(Sales), and Log(Sales-Payroll)/Sales. They also present the simplest industry-year fixed
effects multinomial logits models, having only the independent dummy variable VC, based on
105,031 observations. In addition, they also present several more advanced industry-year fixed
effects models having interactions independent variables between VC and other variables, which
will be consider in the following chapter.
Note that those DVMs would have thousands or hundred-thousands of additive dummy
independent variables, generated by the three categorical variables, namely VC, Industry, and
Year These models are similar to the additive DVM in (1.14) having a special hidden assumption
between its parameters, as illustrated by Table 1.16 based on a 2x2x3 additive ANOVA model.
So that I would say that the models are acceptable in a statistical sense only, but they are not
acceptable in a theoretical sense, more over for an evaluation study – refer to Agung (2011).

1.7.1.2 DVMs from Jotikasthira, et.al (2012)

It happens I have found 22 DVMs of G-7 Betas, presented by Jotikasthira, et.al (2012).
Based on their models, the following notes and comments are presented.
(1) Four of the DVMs are indicated as the country fixed effects models, where the country fixed
effects are presented using a set of 26 dummy variables. Two of the FEMs have 10 dummy
independent variables, where 8 are interaction dummies, and 26 country dummy variables.
So that they can be presented using the general ES as follows:
G7_Betas DV1 DV2 … DV10 @Expand(Country) (1.20)

where DVk is a dummy of a dichotomous variable Vk. For this general ES, the following
notes need to be considered.
1.1 All possible DVMs in (1,21) would have 36 =(10+26) parameters, which is less than the
number of cells generated by the variables Vk’s and COUNTRY. The number of cells could
be a minimum of 11x26 = 286, if the 10 dummies are the dummies of a categorical variable
having 11 levels, and a maximum of 210x26 = 26,624, if the 10 dummies are the dummies of
10 dichotomous variables. Refer to the under-identified DVMs presented above.
1.2 The Country-Fixed-Effects Additive DVMs of G7_Betas are in fact the DVMs, because the
function @Expand(Country) is presenting 26 dummies. Referring to the additive DVMs in

25
 (1.16), and in Table 1.16, then these models would be acceptable only in a statistical sense,
but they would never be observed in practice. Compare to DVMs in the following examples,
specifically the firm-year FEM presented in Figure 1.15.
1.3 Both models do not have the intercept parameters, so that the statistical results do not present
the F-statistic for testing the joint effects of the independent variables. Furthermore, the
statistical results also do not present the Durbin Watson statistic, so that the autoregressive
problem can not be identified.
(2) The 22 DVMs presented have very small R-Squared, with a minimum of 0.000, and a
maximum of 0.060. These very small R-Squared are indicating that the scores of G7_Betas
have very large variances or ranges within the 26 countries, as well as within the 10 Vk’s. In
other words, the independent dummy variables are not good predictors for G7_Betas. Find
the following examples for a comparison.

1.7.2 Illustrative Empirical Statistical Results of DVMs

Dependent Variable : LNSALE

Method: Panel Least Squares
Date: 04/03/13 Time: 16:35
Sample (adjusted): 2 8
Periods included: 7
Cross-sections included: 224
Total panel (unbalanced) observations: 1561
Variable Coeff. s.e. t-Stat. Prob.
C 8.8082 0.2555 34.4694 0.0000
A=1 -0.8090 0.0742 -10.9083 0.0000
B=1 -0.1763 0.0530 -3.3273 0.0009
TP=1 -0.5373 0.0348 -15.4464 0.0000
# Firm dummies (224-1)
R-squared 0.9032 Mean dependent var 6.1778
Adjusted R-squared 0.8868 S.D. dependent var 1.9618
S.E. of regression 0.6601 Akaike info criterion 2.1407
Sum squared resid 581.21 Schwarz criterion 2.9192
Log likelihood -1443.85 Hannan-Quinn criter. 2.4302
F-statistic 55.0707 Durbin-Watson stat 1.3103
Prob(F-statistic) 0.0000
Figure 1.13 Statistical results of a firm-fixed-effects additive DVM in (1.20)

26
Example 1.9 ( A Firm-Fixed-Effects Additive DVM) For a comparison with the country-fixed-
effects additive DVMs presented by Jotikasthira, et.al (2012), Figure 1.13 presents the statistical
results of a firm-fixed-effects additive model of LNSALE on dummies of three dichotomous
variables A, B, and TP, and (224 -1) firm dummies. Based on these statistical results, the
following findings and notes are presented.
(1) In order to reduce the space, the 223 firm dummies are not presented. The statistical results
are obtained using the following ES, with the intercept parameter “C”. So that based on the
F-statistic of F0 = 55.0707 with df = (226,1334) and a p-value= 0.0000, it can be concluded
that the 226 dummies independent variables have a significant joint effects on LNSALE.

LNSALE C @Expand(A,@Droplast) @Expand(B,@Droplast)ave

@Expand(TP,@Droplast) @Expand(Firm_Code, @Dropfirst) (1.21)

(2) Compare to the models of Jotikasthira, et.al (2012), this model has a much greater R2 of
0.9032, which indicates that the independent variables could explain 90.32% of the total
variation of LNSALE. However, this additive DVM also is the worst DVM.

Example 1.10 ( A Firm-Fixed-Effects Interaction DVM) As an extension of the DVM in

(1.22), Figure 1.14 presents the statistical results of a firm-fixed-effects three-way interaction
DVM. Based on these statistical results, the following findings and notes are presented.

(1) The statistical results clearly show there are 7 three-way interaction dummies, indicated by
(A=i and B=j and TP=k), which are presenting interactions (A=i)*(B=j)*(TP=k).
(2) The statistical results obtained by using the following ES.

LNSALE C @Expand(A,B,TP,@Droplast) @Expand(Firm_Code, @Dropfirst) (1.22)

(3) This model is an interaction DVM having only (1+ 8 + 223)=222 parameters, which is much
smaller than the number of cells generated by the independent variables, that is
23x224=1,792 cells. So that this interaction DVM can be considered as a better model than
the additive DVM in (1.20).

27
 Dependent Variable: LNSALE
Method: Panel Least Squares
Date: 04/03/13 Time: 16:22
Sample (adjusted): 2 8
Periods included: 7
Cross-sections included: 224
Total panel (unbalanced) observations: 1561
Variable Coeff. s.e. t-Stat. Prob.
C 8.85522 0.26171 33.83624 0.00000
A=1 AND B=1 AND TP=1 -1.53018 0.09015 -16.97468 0.00000
A=1 AND B=1 AND TP=2 -1.11986 0.10500 -10.66561 0.00000
A=1 AND B=2 AND TP=1 -1.36953 0.09167 -14.93917 0.00000
A=1 AND B=2 AND TP=2 -0.87559 0.10335 -8.47227 0.00000
A=2 AND B=1 AND TP=1 -0.79490 0.08674 -9.16399 0.00000
A=2 AND B=1 AND TP=2 -0.17785 0.09064 -1.96227 0.04990
A=2 AND B=2 AND TP=1 -0.62118 0.06550 -9.48413 0.00000
#Firm dummies (224-1)
R-squared 0.9037 Mean dependent var 6.1778
Adjusted R-squared 0.8870 S.D. dependent var 1.9618
S.E. of regression 0.6595 Akaike info criterion 2.1410
Sum squared resid 578.39 Schwarz criterion 2.9332
Log likelihood -1440.06 Hannan-Quinn criter. 2.4355
F-statistic 54.2415 Durbin-Watson stat 1.3167
Prob(F-statistic) 0.0000
Figure 1.14 Statistical results of a firm-fixed-effects interaction DVM in (1.22)

Example 1.11 ( A Firm-Year Fixed-Effects DVM) As an extension of the DVMs in (1.21), and
(1.22), Figure 1.15 presents the statistical results of a firm-year fixed-effects DVM, or a two-
way fixed-effects DVM of LNSALE , using the following ES, based on a small sub-sample of
{Code < 4, t(Year) < 5}, and its fitted values tabulation. Based on these statistical results the
following notes and comments are presented.
LNSALE C @Expand(Code,@Dropfirst) @Expand(T,@Dropfirst) (1.23)

(1) The small sub-sample is selected in order to show perfectly the limitation of the firm-year
fixed effects model of LNSALE, by comparing its statistical results with its fitted values by
the Firm and T(Year).

28
 Figure 1.15 Statistical results of a firm-year FEM in (1.23), and its fitted values

Table 1.17 Parameters of the Firm-Year Fixed-Effects Model in (1.22)

Firm T=1 T=2 T=3 T=4
1 C(1) C(1)+C(4) C(1)+C(5) C(1)+C(6)
2 C(1)+C(2) C(1)+C(2)+C(4) C(1)+C(2)+C(5) C(1)+C(2)+C(6)
3 C(1)+C(3) C(1)+C(3)+C(4) C(1)+C(3)+C(5) C(1)+C(3)+C(6)
Firm(2-1) C(2) C(2) C(2) C(2)
Firm(3-1) C(3) C(3) C(3) C(3)

(2) Note that the model has only 6 = (1+2+3) parameters, but the 3-firms and 4 time-points
generate 12 cells. So that the model is under-identified, and fitted values within the 12 cells
have special pattern of differences, as shown in Figure 1.15, and Table 1.17, which would
never be observed in practice.
(3) So that it can be said that all firm-year (time-point) FEMs are acceptable only in a statistical
sense.

29
References
Jotikasthira, K., Lundblad, C., and Ramadorai, T. 2012. Asset Fire Sales and Purchases and the
International Transmission of Funding Shocks. The Journal of Finance. Vol. LXVII,
No. 6; 2015 – 2050.
Puri, M. and Zarutskie, R. 2012. On the Life Cycle Dynamics of Venture-Capital- and Non-
Venture-Capital-Financed Firms. The Journal of Finance. Vol. LXVII, No. 6; 2247
– 2294.

30