43 / 330
臺大管理論叢
第
27
卷第
2
期
43
Besides, the models are estimated based on double-clustered (year and firm) standard errors
to correct potential cross-sectional and serial correlation of the residuals and/or the
independent variables in the panel regressions.
10
To test H2, which proposes that IT-intensive banks are related to higher earnings
multipliers, first, a proxy for IT intensity, ITD, is defined. Then, following previous studies
that typically include the factor of interest as an interaction variable (Chiang and Mensah,
2004; Chambers, Linsmeier, Shakespeare, and Sougiannis, 2007), the interaction term
between ITD and NI is added to regressions as shown below.
MV
it
= b
0
+ b
1
BV
it
+ b
2
NI
it
+ b
3
ITD
it
+ ε
it
,
Model 2.1
MV
it
= b
0
+ b
1
BV
it
+ b
2
NI
it
+ b
3
NI
it
× ITD
it
+ b
4
ITD
it
+ ε
it
,
Model 2.2
where
ITD
it
= 1 if bank
i
’s IT ratio is above the median among its peers during year t, whereas
IT ratio equals IT-related expenditures divided by total operating expenses. This
variable is the proxy for IT intensity.
According to H2, if b3 in Model 2.1 is significant and positive, it suggests that banks
with higher IT levels (i.e., the higher IT intensity) are related to greater market values. If b3
in Model 2.2 is significant and positive, it can be interpreted that IT intensity enhances the
persistence of earnings. Similarly, Models 2.1 and 2.2 are estimated based on double-
clustered (year and firm) standard errors.
4. Empirical Results
This section reports the results of the descriptive statistics for all variables (Section 4.1),
regression results, and analyses for the value relevance of IT related costs, IT intensity
(Section 4.2), and related robustness checks for our main tests (Section 4.3).
4.1 Descriptive Statistics
The sample consists of annual data from US banks during the 2001-2010 periods, and
10 As suggested by Petersen (2009), despite the common use of panel data sets in finance literature, there is
a wide variation in the methods used to estimate standard errors. He indicates that the most crucial aspect
for the computation of standard errors is the cross-sectional and serial correlation of the residuals and/or
the independent variables in the panel regressions. He further emphasizes the importance of properly
considering the correlations that might exist in the data, and puts forward an approach to address this
issue. Accordingly, to rule out the potential impact from this bias on our empirical results, all regressions
are based on the double-clustered standard errors.