Page 22 - Contributed Paper Session (CPS) - Volume 6
P. 22
CPS1468 Takeshi Kurosawa et al.
̂ ̂
calculated (̂, |).
pp
̂ ̂
Table 1 shows the values of (̂, |) fo r each model. suggested
pp
pp
that the models involving only one or both of the two binary predictors C2 and
S2 have low predictive power. This result is not surprising given one would
expect a model fitted based on continuous as opposed to categorical
predictors would offer better predictive performance when the response itself
is not categorical. Also, the inclusion of both W and Wt simultaneously seems
to not be favorable compared to including either one only, which is perhaps a
reflection of the fact that W and Wt are highly correlated (Takahashi and
̂ ̂
Kurosawa, 2016). Our proposed measure (̂, |) exhibited similar trends
pp
to AIC, although noting that AIC also takes account the penalty for the number
of parameters into it.
References
1. Agresti, A. (2002). Categorical Data Analysis, Second Edition. John Wiley
and Sons.
2. Eshima, N. (2004). Canonical exponential models for analysis of
association between two sets of variables. Statistics & Probability Letters,
66:135–144.
3. Eshima, N. and Tabata, M. (2007). Entropy correlation coefficient for
measuring predictive power of generalized linear models. Statistics &
Probability Letters, 77:588–593.
11 | I S I W S C 2 0 1 9
