4. The Empirical Analysis
4.3 Econometric Issues
Fixed Effect versus Random Effect: The individual effect can be controlled using fixed
effect or random effect approaches. A fixed effect model assumes that the individual effect is
correlated with the independent variables in the model, while a random effect model assumes
that there is no correlation between the individual effect and the independent variables. Hausman
(1978) proposes a test to check a more efficient model against a less efficient, but consistent,
model. Under the null hypothesis, both fixed effect and random effect estimates are consistent,
but a random effect estimate is more efficient; whereas, under the alternative hypothesis, the
fixed effect estimate is consistent, but the random effect estimate is not consistent. In this paper
the Hausman test shows that fixed effect model is appropriate for affiliated, non-affiliated and all
companies together.
Endogenous Variable: With the presence of moral hazard, the reinsurance coverage may
make insurer take less precaution controlling risks, which leads to higher incurred losses. Hence
the right-hand side explanatory variable of LRi,tis endogenous. It is correlated with the error term of the equation (12). To correct the endogeneity, the first stage estimation equation (13) is
used: t i t i D t i t i t i D t i REINS REINS LR DPW LR, =α20 +β21 ,−1+β22 ,−2 +β23 ,−1 +β24ln( ), +ε, (13) REINSi,t−1=One lag of reinsurance purchase for the primary insurer iin yeart−1;
t i,
From Wooldridge (2002), the OLS estimators will be biased if the endogenous variables
are included in the estimated model. To conduct an endogeneity test, a set of suitable instrument
variables (IV hereafter) is needed for this potential endogenous variable. The regression-based
approach introduced by Wooldridge (2002) is applied. An appropriate IV needs to be correlated
to the endogenous variable and uncorrelated with the error term in the model. Intuitively, the
direct loss is positively related to the direct written premium by the primary insurers, and the
direct written premium can serve as an IV per se. Therefore, log of direct written premium is
used as one IV for loss incurred. In addition, one and two lags of reinsurance purchase are
included in the model as one instrumental variable. This inclusion can be used to test its effect on
the concurrent losses incurred which may arise due to moral hazard with the reinsurance
coverage.
The reduced form of direct incurred loss is estimated by using all the independent
variables in the estimation (12) and four IVs as the independent variables. After the residual of
this estimate is obtained, the dependent variable, the reinsurance purchase in the equation (12), is
regressed on all the independent variables and the obtained residual, as well. The insignificant
robust t-statistic of estimated coefficient for the error term indicates that the direct incurred loss
is not an endogenous variable in the estimation and the corresponding results are unbiased. The
corresponding p-value is 0.00 which implies that the variable of concurrent loss incurred is
endogenous in the estimation equations, and the OLS estimators are biased. We need to apply
IVs to fix the endogeneity issue.
In addition, equation (13) partly reflects the potential of moral hazard on the part of the
primary insurer. With the presence of moral hazard, the higher level of reinsurance purchase in
primary insurer if other firm characteristics are controlled. Hence, the estimated coefficient of
lag of reinsurance purchase is expected to be positive if moral hazard does exist in the
reinsurance market.
Heteroskedasticity: If the error terms do not have constant variance with each observation,
the heteroskedasticity problem arises. In this case, the OLS estimators are unbiased and
consistent but inefficient because the assumption of the constant variance for error terms is
violated. In the presence of hetoroskedasticity, the variance of the coefficients obtained from
OLS tends to be underestimated, so the OLS standard error is not valid for constructing
confidence intervals and t statistics. To solve this problem, Weighted Least Square (WLS)
estimators or robust standard errors are usually adopted to improve efficiency.
In the estimation, the White test is employed to detect the possible heteroskedasticity
problem. The White test statistics is 2261.24 and corresponding p-value is 0.00. This result
rejects the null hypothesis that the residuals in the model are homoskedasticity. Therefore, the
heteroskedasticity issue occurs when estimating the model, and the estimators are unbiased and
consistent but inefficient. In addition, the normal standard errors are invalid to construct the
confidence intervals and the t-statistics. Therefore, the robust standard errors are used instead to
improve the estimator efficiency in the presence of heteroskedasticity.
Individual Effect versus Pooled OLS: The error term ui,t in equation (12) can be
decomposed asui,t =ai +νi,t, where aiis called individual effect, νi,tis idiosyncratic error and
t i
u, is composite error. The individual effect is usually unobservable. If the unobserved individual effect is correlated with other independent variables in the model, the pooled OLS estimators are
biased and inconsistent. If the individual effect is a random variable and is uncorrelated with
As a result, the presence of the individual effect to choose the appropriate estimation method
needs to be tested. Breusch and Pagan (1979) derive the Lagrange Multiplier (LM) test to detect
the presence of individual effect. Based on the residuals from the equation (12), the LM test
statistics is 5217.6, which reject the null hypothesis of the absence of individual effect. In the
presence of individual effect, the pooled OLS estimation is not appropriate for our model.
Overidentifying Test: To test the model identification, Anderson Canon and Cragg-
Donald tests are undertaken by using STATA code of “xtivreg2”. The small p-value of these
tests shows the model proposed is identified.