While running panel regressions, it is assumed that someregression coefficients alter across individuals and/or over time in order toaccount for individual differences. There are fixed parameters, even though theregression coefficients are not exactly known. The model is classified as fixedeffect model’ when the coefficients are allowed to change in one or twodimensions. In the current model, the intercept is allowed tochange across individuals (households, firms or countries. However, it isassumed that the parameters of the slope and error variances are constant in thesedimensions.
The random effect model utilizes random quantities unlike the fixedparameters. Random effects are associated with these unsystematic quantities. Inthe model, the intercept and slope parameters are not different while there isa variance in the components of error variances across individuals and/ortimes. If the fixed effect model results in huge losses in degrees of freedom,due to too many parameters in these types of models, then the random effectmodel becomes more appropriate choice. As Judge et al. (1988) and Baltagi(2001) explained in their studies, the random effect model is chosen whenindividuals are randomly selected from a big population. There has been aheated debate on the selection of the fixed effect and random effect model amongeconometricians for so long. The choice of the appropriate model is based onthe assumptions on the interrelationship of the exogenous variables, both cross-sectionaland across time, the error term assumptions, and/or the researcher’s propensityfor increased efficiency and decreased bias in the estimators.
Even though, fixedeffects model is mostly less efficient, the model is known to be moreconsistent and less biased. Random effects model is more restricting than the fixedeffects models. The random effects model, which is a specific case of the fixedeffects model, needs additional assumptions. Fixed effect model deliverscontrol for all variables that do not vary across time. However, the coefficient estimates forvariables that do not vary across time can be calculated by using the randomeffects model. RR It is true that the random effects model has less samplevariability leading to more efficient estimators. However, if the assumptionsare not realized, the model can cause biased estimators. The researchers mostlyprefer fixed effects model due to its unbiased estimators and less restrictive nature.
If there is a necessity for the estimates of coefficients for time invariantvariables, then random effects model is more ideal. It was importantto decide which of these two models can be more appropriate with missing orunbalanced data, such as is the case in the current study. Both the fixed effectsand random effects models are sufficient to work with unbalanced designs of thedata, maintaining degrees of freedom compared to excluding observations inorder to generate a balanced data. In thisstudy, the fixed effects model is expected to be the appropriate method sincewe expect to see the effect of 2008 global crisis on most of the countries inthe sample. Additionally, our sample matches with the population of the study.Finally, our data does not contain any time-invariant regressors.
However, itshould be asserted that in the with the existing literature, the appropriatemodel that fits the sample and the research’s objective must be used. Hausmanand Taylor (1981) test is conducted to decide on the utilization of theappropriate model. In Hausman Test, the correlation between individual effectsand regressors is tested with the null hypothesis that there is no correlationbetween individual effects and regressors.
The Hausman Test results for each ofthe models for two dependent variables are summarized in Table 3.4and Table 3.5respectively.