OBJECTIVES: We explore the use of state transition models for estimating health state transition probabilities in a Markov model. Using this approach with a longitudinal database, we assess the Markov chain order and determine whether transition probabilities depend on other covariates. METHODS: Markov models rely on the Markovian assumption, i.e. the probability of transitioning from one state to another is independent of a patient's prior history. Transition probabilities can considerably impact a model's outputs. We used current guidelines and a longitudinal adult cohort of asthma patients to characterize their asthma control every six months into three mutually exclusive health states not including death: well-controlled, not well-controlled, and very poorly controlled. We used correlated data multinomial regression methods to fit state transition models that included lag covariates for past control status at six-month time intervals and baseline covariates of interest. We tested the significance of first-order (six-month lag) and second-order (six and twelve-month lags) Markov chains and baseline covariates in predicting six-month transition probabilities. RESULTS: We analyzed 3488 adults (average follow-up 25 months) with severe or difficult-to-treat asthma. First and second-order Markov chains, and baseline severity all significantly predicted present control status (p<0.0001 respectively). The first-order predicted six-month transition probability from well-controlled to well-controlled was 0.626 (95% CI = 0.599, 0.653), but second-order ranged from 0.362 (95% CI = 0.270, 0.467) for those that were very poorly controlled twelve-months ago to 0.738 (95% CI = 0.703, 0.770) for those that were well-controlled twelve-months ago. CONCLUSION: In our example, assuming first-order six-month transition probabilities without regard for a patient's second-order (six and twelve-month past) control status or baseline severity level was a simplification that would likely bias Markov model outputs. All models are simplifications but where data are available, state transition models are tools that can help a Markov model remain simple, but not overly so.