This document proposes STEPHY, a method for improving LDA topic models by allowing topic distributions to vary over time. STEPHY conducts multistage variational Bayesian inference over LDA models with increasing levels of time dependence, starting with a time-independent model and progressing to models where topic distributions change over timestamps. This multistage approach allows efficient inference by initializing parameters from the previous model. The method is evaluated on several datasets, showing improved perplexity over standard LDA.