CAPÍTULO II: MARCO TEÓRICO
2.2. Bases Teóricas
2.2.1. Gestión de Procesos
GMM estimators are frequently use in autoregressive linear regression models with
small T (time horizon) and large N (units) panels with unobserved firm level
heterogeneity (i.e. fixed effects) . Not considering unobserved heterogeneity faultily
assumes zero correlation between the observed variables and unobserved firm-
specifics. The used of lagged dependent variable as a regressor in a model necessitates
an application of dynamic models because the standard OLS estimators are biased and
inconsistent (i.e. dynamic panel bias). Similarly, LSDV and Within Group estimators
do not eliminate this dynamic panel bias (Nickell 1981). Often in the literature,
standard panel estimators like the fixed effect estimators are employed; however, this
technique is biased in a dynamic setting, where lagged dependent variables are
included in the model26. A dynamic capital structure model with firm-specific effects posed at least two challenges; first, the error term, υit, tends to be correlated with the
lagged dependent variable, M_Leverageit-1.; second, other regressors may be
correlated with firm-specific effect that is time-invariant. A traditional estimator for a
dynamic panel model with firm fixed effects involves first-differencing the model to
26Econometrically, an explanatory variable is said to be endogenous if it
is correlated with the error term of the data generating process.
Endogeneity caused OLS to be biased and inconsistent, thus making estimates of coefficients and inferences invalid.
31 eliminate the firm-specific effect27. Arellano and Bond (1991) also pointed out that a lagged dependent variable in a traditional fixed effect model is correlated with the
error term, and therefore sufficiently endogenous. Taking first-differences will
eliminate the fixed effects but the first difference of lagged dependent variable, Δ
M_Leverageit-1. , is still correlated with the first difference of the error term, Δυit,
through M_Leverageit-1 andυit, thus the fixed model may not necessarily resolve the
endogeneity problem (Wooldridge, 2002).28 A plausible estimation alternative in a dynamic capital structure context with partial adjustment is the dynamic panel
estimators, by utilizing instrumental variables29 estimation technique, through the Generalized Method of Moments (GMM), bias-corrected least squares dummy
variables (LSDVC)30 or fixed effect instrumental variables (FE IV) methods. The justification for instrumentalizing in a dynamic model is not restricted to the concerns
of endogeneity. Instruments are also necessary to address the fact that regressors may
not be strictly exogenous. In other words, outside events that caused a shock to capital
structure may also impact the regressors, and thus may be indirectly correlated with
the present or past error terms. In an attempt to resolve these issues, Arellano/Bond
27Anderson and Hsiao (1982), suggest a technique of solving the firm- specific effect problem, by transforming
the model into first differences to eliminate firm-specific effects, and then use the second lags of the dependent variable, either differenced or in levels, as an instrument for the differenced one-time lagged dependent variable, to eliminate the correlation with the error term.
28Fixed effect regression takes out the common time-invariant and firm
specific effects out of the error term and through first differencing
eliminate completely the fixed effect component of the model. However, in a dynamic panel estimation, where a lagged dependent is necessary, the first differencing is not sufficient to solve the endogeneity problem
(Roberts/Leary 2005, Arellano/Bond 1991, Greene 2003, chap.13).
29Instrumental variables have to meet two conditions in a fixed effect
model. First, they have to be correlated with the lagged dependent variable or independent variables (i.e. if more proxies are included); second, they have to be uncorrelated with the error term.
32 (1991) estimator suggests using instruments for the lagged dependent variables and
endogenous regressors after first-differencing the model. Arellano/Bond (1991)
estimator is a major leap forward, however, with a highly persistent data series such
as leverage, these estimators are sometimes found to be biased when the co-efficient
on lagged dependent variable is close to unity. Arellano and Bover (1995) and
Blundell and Bond (1998) proposed the “system” GMM estimator. The system
estimator transform the model into first differences, and then estimate the model as a “system” by using either the lagged first differences for level equations or lagged
variables for differenced equations as instruments. The combination of these two sets
of moment conditions is therefore the system (SYS) GMM estimator. The key point to
bear in mind regarding the econometrics of capital structure is that there is no perfect
technique. All the various techniques used in the literature have their relative
strengths and weaknesses. Thus, this chapter starts with all four major dynamic
models used in the literature to independently examine adjustment speed for the three
groups (i.e. DC, MNC, and MNC10). I used the traditional fixed effect instrumental
variables estimator (FEIV), which specifically instrumentalized the lagged dependent
variable with its two own lagged, then the Arellano and Bond (1991) one stage and
second staged differenced GMM estimators, and finally, the LSDVC dynamic
estimator is used to estimate the adjustment speed for MNCs and DCs. According to
Bruno (2005), bias corrected least squares dummy (LSDVC) performs better than
other dynamic models, when estimating unbalanced dynamic panel data, however
LSDVC do not generate system standard errors, and are therefore estimated through a
bootstrapping technique. Generally, traditional statistical inference is predicated on
assumptions about the distribution of the population from which the sample is taken
33 normality assumption. In a small sample, improved standard errors may be obtained
by using bootstrap technique. Many prior studies have used bootstrap standard error
estimates for their parameter estimates, and have suggested that inference based on
bootstrap estimates of standard errors may be more accurate in small samples than
inference based on asymptotic standard error estimates (Bruno 2005). Computer
intensive statistical software programs (i.e. Stata) has commands for obtaining
bootstrapped standard errors by evaluating the distribution of a statistic based on
repeated random resampling of the original dataset. Theoretically, the bootstrapping
works as follows: first, random sample with replacement is repeatedly obtained from
the sample dataset; second, desired statistics corresponding to these bootstrap samples are estimated and third, sample standard deviations‟ are calculated of the sampling
distribution of repeated bootstrap samples . There is a lack of specific consensus in
the literature regarding the number of resampling needed for obtaining bootstrapped
standard errors. I suspect the number of iterations in the literature are most likely
influenced by the amount of computing power, time and sample size. Hence
increasing the number of repeated samples cannot increase the amount of information
in the original data. Consequently, I started with 10 iterations, and by successively
increasing the number of iterations, I found no efficiency gain in higher number of
iterations relative to 10 iterations. In other words, the significance level of the
estimated statistics did not change due to increasing iterations. Using Bootstrapped
standard errors has certain specific benefits including the following:
A) When the theoretical distribution of a statistic under investigation is complicated
B) When the sample size is small for traditional statistical inference.
C) Bootstrapped technique only assumes that the sample is representative of the
34 The fixed effect instrumental variables (FEIV) tend to perform well with unbalanced
panel data without generating too many instruments as in the case of GMM
estimators, and it directly generates system standards errors. Thus, the comparative
integrated dynamic partial adjustment model is estimated with the year-controlled,
fixed effect instrumental variables estimator (FEIV).