Assignment 8 solution

You are estimating the following model with the help of OLS approach

𝑟𝑖𝑡 = 𝑎0 + 𝑎1 𝑂𝐼𝐵𝑖𝑡 + µ𝑖𝑡 . Here, 𝑟𝑖𝑡 is the return and 𝑂𝐼𝐵𝑖𝑡 is the order imbalance (excess of
buy over sell) for security ‘i’ at time ‘t’. Theoretically, when OIB rises positively, returns
increase and when it falls and becomes negative, returns decrease. However, you believe that
there is sufficient heterogeneity across individual units (e.g., size, profitability of the firm) and
also time-effects, which may affect the estimation. Answer questions 1-10 using this
information.

1. If you estimate the model with pooled OLS approach, which of the following is
incorrect?

a) There may be time-invariant terms ‘α𝑖 ’ that are specific to security and may mix with
the error term
b) There may be time-fixed effects ‘𝑣𝑡 ’ which may mix with the error term
c) If estimating with pooled OLS, the errors may be serially correlated
d) If estimating with pooled OLS, the errors may be correlated with the regressor terms
e) All of the statements are correct

Hint: The model may include time-fixed effects and security specific heterogeneity, which may
mix with the error term and result in serially correlated errors and also error terms that are
correlated with the regressors (endogeneity issue).

2. If there are 100 securities, which of the following panel data method is least desirable

a) LSDV (Dummy variable) approach

b) First Difference (FD) approach
c) Fixed effects (FE) approach
d) Random Effects (RE) approach

Hint: LSDV approach would require 100-1 =99 dummies to account for the unobserved
heterogeneity corresponding to each of the 1000 securities. This would result in an extremely
non-parsimonious model.
 3. The following is incorrect in the context of LSDV approach to account for the
unobserved heterogeneity across securities

a) In case of ‘N’ securities, ‘N-1’ dummy variables are required

b) With LSDV approach it is feasible to estimate the unobserved heterogeneity, i.e., αi
c) The intercept acts as the reference dummy for the security corresponding to which there
is no dummy
d) The coefficients of the dummy variables indicate the differences in the unobserved
heterogeneity of all the securities with respect to the reference dummy (or reference
security)
e) All the above are correct

Hint: In case of LSDV approach, N-1 dummies are needed to account for N securities. The
unobserved heterogeneity corresponding to the security which is not included in the dummy,
will appear as the reference variable on the intercept. For each of the securities, the dummy
variable coefficient reflects the differences between the security ‘i' and the reference dummy
security. That is, adding this coefficient with the intercept results in the unobserved
heterogeneity (αi) of the security ‘i'.

4. The following is incorrect in the context of First Difference (FD) approach to account
for the unobserved heterogeneity across securities

a) FD approach requires differences across the one time-lag

b) For each security one observation is consumed due to differencing procedure
c) If the original errors from the model were uncorrelated, FD approach may result in
autocorrelated errors
d) With FD approach, the time-invariant terms ‘αi’s’ are eliminated and hence can not be
estimated
e) All the above are correct

Hint: FD approach requires taking one time-lag differences across the data. This results
reduction of one observation for each of the security. For example, if there are 10 time periods
and 100 securities, that is, 1000 security-period observations; FD approach would result in
consumption of 100 observations and remaining 900 for estimation.
 5. The following is incorrect in the context of Fixed-effects (FE) and First difference (FD)
approaches to account for the unobserved heterogeneity across securities

a) FE approach requires differencing each security observations from the corresponding

time-average (time-demeaning)
b) FE approach removes any time-constant terms
c) If original error terms are uncorrelated, then FE approach results in a more efficient
(less standard errors) estimates as compared to FD
d) If original error terms are correlated, then FE approach results in a more efficient
(less standard errors) estimates as compared to FD

Hint: FE approach requires time-demeaning. That is, for each security, observations are
subtracted from their time-averages. This removes any time-constant terms. If the original
errors are uncorrelated, the FD approach produces autocorrelated errors and hence less efficient
estimates compared to FE. If the original errors are heavily correlated, the FD produces errors
that are less autocorrelated, and hence more efficient estimates compared to FE.

6. The following is incorrect in the context of Random-effects (RE) and Fixed-effects (FE)
approach to account for the unobserved heterogeneity across securities

a) If the issue of endogeneity (correlation between unobserved heterogeneity and

exogenous regressor) is less, RE approach is more suitable
b) The essence of RE approach is captured by a parameter ‘λ’ which varies from 0-1,
which is multiplied to the time-demeaning process
c) If λ=0, then RE approach is similar to pooled OLS
d) If λ=1, then RE approach is similar to FE
e) All of the statements are correct

Hint: RE approach requires time-demeaning with a multiple of λ. λ varies from 0-1. If λ is 0,

the model approaches pooled OLS estimates. If λ=1, the model approaches FE estimation.

7. The following is incorrect in the context of Random-effects (RE) and Fixed-effects (FE)
approach to account for the unobserved heterogeneity across securities

a) In RE approach λ calculation, if the size of unobserved heterogeneity is too small

relative to the model errors, λ approaches 0
 b) In RE approach λ calculation, if the size of unobserved heterogeneity is too large
relative to the model errors, λ approaches 1
c) If the correlation between unobserved heterogeneity and regressors is large
(endogeneity issue) then FE is preferred over RE
d) Given the heavy nature of transformation, FE estimates are less efficient (higher
standard error) compared to RE
e) All of the statements are correct

Hint: If unobserved heterogeneity is too small, λ=0, RE estimation approaches pooled OLS.
Conversely, if λ=1, RE estimation approaches FE. If the correlation between unobserved
heterogeneity and regressors is high, the model suffers endogeneity. Thus, pooled OLS and RE
estimates are not consistent, only FE is consistent. However, given the heavy nature of
transformation, RE estimates are more efficient (lower standard error) compared to FE.

8. The following is incorrect in the context of Hausmann test statistic regarding the
selection of Random-effects (RE) or Fixed-effects (FE) approach to account for the
unobserved heterogeneity across securities

a) Hausman test statistics reflects the trade-off between model consistency and efficiency
b) The numerator reflects how consistent the estimates are and the denominator reflects
efficiency of the estimates
c) If both the RE and FE estimates are consistent, the efficiency of the estimates is
considered and RE is chosen
d) If only FE is consistent, FE is chosen
e) All of the statements are correct

Hint: In general, the Hausman test statistics has coefficient differences in the numerator, and
the standard error squares in the denominator. If only FE is consistent, then numerator is large,
and Hausman stat is high, null is rejected and FE is selected. If both the estimates are consistent,
then numerator is close the zero. We fail to reject the null and RE is selected.

9. Which one of the following commands estimates the random effect model in R

a) Model= lm(rit ~ OIBit, data =panel, model=”random”)

b) Model= plm(rit ~ OIBit, data =panel, model=”random”)
c) Model= plm(rit ~ OIBit, data =panel, model=”pooled”)
d) Model= plm(rit ~ OIBit, data =panel, model=”within”)
 e) None of the above

Hint: RE estimation in R, requires PLM “plm” function with model=”random” argument.

10. When compared between Random Effects (RE) and Fixed Effects (FE), the following
is an incorrect statement.
a. If the null “H0: Both models are consistent” is rejected that means FE is the correct
model. Hint: FE is always consistent, while RE is consistent depending on the nature
of unobserved.
b. If we fail to reject the null “H0: Both models are consistent” that means RE is the correct
model. Hint: RE is always more efficient than FE. So if both the models are consistent
then RE is better.
c. Selection of RE vs FE is essentially a trade-off between the relative gains in efficiency
and consistency of the model. Hint: If gains from consistency (numerator) are more
then FE is selected; if gains from efficiency (denominator) are more then RE is selected.
d. If the covariance between the unobserved heterogeneity and independent variable
is not equal to zero then RE is appropriate. Hint: If the covariance Between the
unobserved heterogeneity and independent variable is not equal to zero, then
model faces potential endogeneity, and only FE is consistent.
e. None of the above