Page 29 - Contributed Paper Session (CPS) - Volume 7

CPS2020 Honeylet T. S.
effect on Y, (b) X1 has dominating effect on Z, and (c) X1 and X2 have high
correlation. The only exception is when the ratio of Data Source A: Data Source
B is 90:10. The rest of the scenario cases where X1, X2 have equal effect on Y
have lower RBIAS when bootstrap within and bootstrap across were used.
Under random hot deck imputation, bootstrap across method produces
RBIAS that are much larger than those of the other two estimation methods.
For ratios 10:90 and 90:10, bootstrap within produces lower RBIAS for the
following cases regardless of the common variable used: (1) X1 dominating on
Y and Z, high correlation of X1 and X2 and (2) X1 dominating on Y, X1 and X2
have equal effect on Z, low and high correlation of X1 and X2.
RBIAS are largest when the matching variable used is X2 when X1 has
dominating effect on both Y and Z. Bootstrap across does not perform well
when using random hot deck imputation regardless of the common variable.
With regards MAE values, the three estimation methods produce similar
results regardless of the common variable used. Additionally, MAE values
produced under random hot deck imputation are higher than those of the
other two matching methods.
In terms of predictive ability of estimated models, bootstrap within and
bootstrap across perform better than Poisson regression without
bootstrapping. This is true regardless of the matching methods used.
Particularly, under random hot deck imputation, bootstrap within always has
the lowest MAE values. Moreover, bootstrap within has lowest MAE values
across matching methods for extreme ratios 10:90 and 90:10.

2. When Log Mean of Y and Z are nonlinear functions of X1 and X2
Only MAE will be discussed in this section because the model where the
complete data were generated from is different from the model used to fit the
data.
Regardless of the common variable and matching method used, Poisson
regression, bootstrap within, and bootstrap across estimation methods
produce comparable results. Although results of the estimation methods are
comparable within a matching method, random hot deck imputation has
larger MAE values compared to those of its Poisson regression imputation and
MCMC imputation counterparts.
Furthermore, under Poisson regression imputation and MCMC imputation,
bootstrap within and bootstrap across estimation procedures yield lower MAE
values compared to those produced without bootstrapping. Subsequently
under random hot deck imputation, bootstrap within always produce lowest
MAE values.

18 | I S I W S C 2 0 1 9