Local Likelihood for Interval-Censored Recurrent Event Data

1 Summary. Interval-censored recurrent event data arise when event occurence is only determined at periodic assessment times. Nonparametric estimates of rate and mean functions are readily obtained in this settings, but lack the smoothness usually anticipated for the underlying event process. We describe local likelihood methods which generate smooth estimates of rate and mean functions, and associated covariate effects in the context of mul-tiplicative intensity-based models. A modified EM algorithm is described which incorporates kernel smoothing of the rate function under the assumption it can be modeled locally using polynomials. Both Poisson models and mixed Poisson models are considered. Simulations suggest the proposed methods work well, and applications to data from a study of feedwater flow losses in nuclear plants and a study of joint damage in patients with psoriatic arthritis illustrate their use.