Predictive keyboards are software keyboards that, to conserve screen real estate, display a subset of the full set of alphabetic keys at any one time, predicting the letters to display on the basis of tables of letters' transitional probabilities. Using different types of test texts (names, words, random strings), we evaluated the influence of various manipulations on the efficiency of letter selection with a predictive keyboard. The results of this study indicated that, across tested text types (names, words, random strings), (1) there was little benefit gained from adding a fourth transitional table modeling the likelihood of a letter's use as a function of a space and two preceding letters, (2) there was a potential benefit from increasing the number of displayed letters from six to eight, and (3) for a personal communicator device, an adaptive strategy would probably be less effective than using multiple sets of letter probability tables.