Correct pitch accent placement can substantially improve the naturalness of text-to-speech synthesis, and has been acknowledged as an important problem in synthesis research. This paper describes an empirical study of the factors that affect pitch accent placement based on the Boston University radio news corpus. Issues examined included accent placement on function words, relative importance of different influencing factors, early accent placement within the word, and variability across speakers. Incorporating the results of these four analyses into questions for decision-tree-based pitch accent prediction results in an algorithm with an accuracy of 95% on syllables marked in all of four read versions of a story and only 7% false assignment of pitch accent to syllables not marked by any of the speakers.