Page 111 - 34-1

NTU Management Review Vol. 34 No. 1 Apr. 2024

3.1 Regression Model
The following general model is estimated:

×

× ×
(1)
× ×
,

where i represents firm i and t references time. As explained more fully below, the

dependent variable, Y , is the scaled loss reserve error. The X , are variables used to test
it
it
the hypotheses concerning loss reserve errors. Z is a vector of institutional and firm
it
characteristics variables; and Year indexes time periods. The error term is ε . Estimation
18
t
it
strategy is discussed further in section 3.3.
3.2 Specification of Regression Variables
We first discuss estimation of the dependent variable – i.e., the loss reserve error.
Independent variables used in the regression model are then discussed in the following two

subsections.

3.2.1 Specification of the Dependent Variable
The loss reserve error is defined as the difference between the estimate of total
incurred losses for a company (which includes errors from prior accident years) as of a

given calendar year t (LR ) and the future revised estimate of the same losses in calendar
t
year t+5 (LR ). This definition follows Petroni (1992), and Gaver and Paterson (2004,
t+5
2014), among others. Table 1 provides a detailed explanation on how we calculate the loss

reserve error using an example with actual firm data. The five-year loss reserve error is
scaled by admitted assets, following Beaver et al. (2003), Cheng, Chow, Lin, and Ng (2022),
Gaver and Paterson (2014), and Petroni (1992). The loss reserve error is positive (negative)
if the original estimate of incurred losses is overestimated (underestimated).

18 Previous literature indicates that loss reserving errors are serially correlated (Beaver and McNichols,
1998; Grace and Leverty, 2010, 2012) and heteroscedastic (Grace and Leverty, 2010, 2012). FGLS
rather than full fixed effects is commonly used when these problems exist (Grace and Leverty, 2012).

103