Validity assessment of binary models for improving targeting of Conditional Cash Transfer programs through posterior simulation and multilevel modeling.

Conditional Cash Transfer (CCT) programs are a recent and popular approach in developing countries to improve health and education prospects of poor families through conditional financial incentives.

However, due to sub-optimal transfers and administrative difficulties in identifying needy families, many CCT programs suffer from incomplete coverage of needy families and/or low financial efficiencies.

De Janvry and Sadoulet (2003) have shown that their targeting method (the DJS method), which uses binary regression models to identify needy families and optimal transfer amounts, can effectively overcome these problems.

This dissertation tackles two issues encountered in implementing the DJS method: the need for an appropriate methodology to compare binary models, and the need to develop binary models with high external validity.

For targeting of CCT programs, binary models need to be analyzed comprehensively beyond the overall goodness-of-fit.

For example, policy makers may need to know the rates of false positive and false negative predictions of various models and choose the most appropriate model for given policy priorities.

This dissertation intensively reviews the binary model classifiers and proposes either expected or simulated Confusion Matrix depending on the context of analysis.

The advantage of this methodology over traditional classifiers is that it can easily incorporate policy preferences into model evaluation through inclusive diagnosis of model performance.

In the DJS method, student data in previous periods is used to develop models, which determine the eligibility and the transfer amounts of the new/current cohort. Because regression models in general predict poorly with new/external data, models with high external validity should be employed.

This dissertation applies posterior-simulation to simulated datasets to evaluate the internal and external validity of various binary models, including the models used by the DJS method.

The study concludes that the multilevel model, which accounts for the correct group structure in conjunction with the relevant group level predictor, has the highest external validity.

In addition, the multilevel models exhibit higher external validity with small/minority groups, who are often the targets of CCT programs.

Furthermore, the findings were consistent with the cases where pilot CCT programs are expanded to wider populations.