A controlled trial is used to estimate the effect of an intervention. We present here a decision tree for choosing the most appropriate of five kinds of controlled trial. A time series or quasi-experimental design is used when there is no opportunity for a separate control group or control treatment. In this design, the weakest of the five, a series of measurements taken before the intervention serves as a baseline to estimate change resulting from the intervention. In trials with a separate control group, the usual design is a fully controlled parallel-groups trial, in which subjects are measured before and after their allocated control or experimental treatment. A posts-only design, in which subjects are measured only after their treatment, can be more efficient when poor reliability of the outcome measure over the time frame of the intervention makes large sample sizes unavoidable. Crossover studies, in which all the subjects receive all the treatments, are an option when the effects of the treatments wash out in an acceptable time. In fully controlled crossovers, subjects are measured before and after each treatment, whereas measurements are taken only after each treatment in a simple crossover. Fully controlled crossovers, arguably the best of the five designs, are more efficient if the outcome measure becomes too unreliable over the washout period, and they provide an assessment of the effect of the treatment on each subject. In simple crossovers, individual assessment is possible only by including a repeat of the control treatment.