Stochastic compartmental models are usually based on continuous-time Markov processes. Markovian models assume that the compartments have exponential retention times, which is known not to hold in certain applications such as calcium clearance from bone. A number of semi-Markov models with restricted families of retention times have been proposed recently. This paper presents a general, tractable procedure for determining from suitable data the estimated retention time distributions for assumed non-Markovian phenomenological models. The procedure is based on using phase-type distributions. These distributions can describe the long tails observed in calcium clearance data, and they frequently lead to complex eigenvalues that are often overlooked in practice. The procedure is illustrated on several proposed models for describing a particular data set.