Statistical Engineering Division
Seminar Series
A Closer Look at Combining a Small Number of Binomial Experiments
Don Malec
The YS Bureau of the Census
Experiments are often expensive and time-consuming with resulting
conclusions being based on relatively small sample sizes. Often,
however, data from related experiments or from pre-experiments are
available and their combined use may help arrive at more confident
conclusions. A way to combine related data, while simultaneously
estimating how much borrowing among data sets should take place, is
through the use of hierarchical models and Bayesian inference. This
approach is illustrated using a series of examples combining two
binomial experiments. The effect of a hierarchical model on
estimates of rates and on the degree to which data is combined is
illustrated. In addition, the phenomenon in which combining data
reduces overall precision is explained. Simpler models based on
finite mixtures are shown to work as well as more computationally
intensive, continuous mixtures. Lastly, an example combining three
concurrent experiments is illustrated.
NIST Contact:
Walter Liggett, x-2851.