|
Abstract:
|
Exact inference for the common odds-ratio, requires the convolution of independent hypergeometric distributions. The computational effort for a direct approach to this problem quickly becomes untenable, which fact has given rise to several approximation methods. Exact inference is available using Meta's (1985) network algorithm, which, for k tables of size N, requires O(k-squared * N-squared) operations. We show that devolving each hypergeometric distribution (one for each table) into the sum of independent Bernoulli random variables, and then reassembling the aggregate across all tables, gives exact results more efficiently than the network algorithm. We also show that this approach is easily implemented in a simple spreadsheet.
|
