Page 30 - Contributed Paper Session (CPS) - Volume 7

CPS2020 Honeylet T. S.
4. Discussion and Conclusion
Simulation results presented above summarized the effects of various
likely settings in statistical matching with the objective of estimating a count
regression model. Bootstrap-based approaches for estimation of synthetic
data created from various imputation methods are in general at par with or
better than the benchmark regression approach. It might be of interest for
future studies to include a donation class to improve the estimates of the
model when synthetic datasets are created using random hot deck imputation.
Simulations show that MCMC imputation and Poisson regression
imputation produce comparable results when it comes to accuracy of
estimates. Particularly, MCMC imputation yields lowest RBIAS. In terms of
predictive ability of estimated models, Poisson regression imputation and
MCMC imputation still produce comparable results although Poisson
regression imputation has most of the lowest MAE values. Under random hot
deck imputation, RBIAS and MAE values are larger compared to those
produced by the other two matching methods.
Moreover, under Poisson regression imputation and MCMC imputation,
the three estimation procedures yield similar results. Accuracy of estimates
and predictive ability are comparable. In terms of predictive ability, the three
estimation procedures also yield similar results when using Poisson regression
imputation and MCMC imputation although bootstrap within and bootstrap
across produce lower MAE values compared to Poisson regression without
bootstrapping.
The predictive ability of the models estimated using bootstrap within is
good for all sample sizes and ratios of data sources (to total sample) in the
scenario settings. Moreover, regardless of the matching method used,
bootstrap within produces low RBIAS and MAE values. For instance, if random
hot deck is really necessary as the matching method because of the nature of
the data and no donation classes were specified, bootstrap within can yield
lower RBIAS and MAE values compared to Poisson regression without
bootstrapping and bootstrap across method. It is interesting for further study
to incorporate donation classes in random hot deck imputation and evaluate
the performance of bootstrap within and bootstrap across estimation
methods.

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