INTRODUCTION

Allocation of units to treatments is crucial in designing a study on impact evaluation. The strength of an experiment depends, in addition to formation of the comparison groups (discussed in the previous chapter), on how units are allocated to the two groups. Together they make a design complete in the assessment of the causal effect.

The basic objective in a design is to create a control and a treatment group in such a way that sans the treatment the two would be similar and produce outputs that would be statistically equal to each other. The similarity between the two groups can be achieved in a few different ways. This chapter describes the different types of allocation along with their advantages and disadvantages.

In this context, another approach known as propensity score matching has been gaining popularity. This method, popularly known as Rubin Causal Model, which assists in strengthening a design, has also been discussed. The issue of two distinct situations, depending on whether individuals or clusters are the units of allocation, has been given adequate attention.

ALTERNATIVE TOOLS TO ATTAIN BALANCE

To achieve a balance, one needs to ensure that both treatment and control groups have similar distribution in terms of all the disturbing variables (covariates). There are two distinct approaches to balance the disturbing variables. They are: randomization and matching. The matching can be performed in two ways: a) group matching or stratification and b) pair matching.

8.2.1 Randomization

The objective can be achieved by ensuring that both treatment and control groups are randomly drawn samples from a population under study. That is, either they are drawn with an equal probability selection method EPSEM or, if not, individuals are weighted appropriately so that each provides an unbiased estimate of the population parameter. There is a need to emphasize this, since the issue of unbiasedness of an estimate can be overlooked, particularly when non-response makes the two samples to deviate from an EPSEM. If the pattern of non-response is different in the two samples (treatment and control) it can introduce bias in the estimation of the impact. The difference in the estimates obtained from the two samples will not show the treatment impact, unless the samples are weighted properly to remove the biasing effect created by the non-response.