Pavement distress progression models predict the extent of a distress on pavement sections as a function of age, design characteristics, traffic loads and environmental factors. These models are usually developed using data from in-service facilities to calibrate the parameters of mechanistic deterioration models. The data used for the statistical estimation of such models consist of observations of pavements for which the distress has already appeared. Unfortunately, common statistical methods, when applied to such data sets, produce biased and inconsistent model parameters. This type of bias is known as selectivity bias, and it results from the fact that less durable pavement sections are over-represented in the sample used for model estimation. A joint pavement distress initiation and progression model, consisting of a discrete model of distress initiation and a continuous model of pavement progression is presented. This approach explicitly accounts for the self-selected nature of the sample used in developing the progression model, through the use of appropriate correction terms. Moreover, previous research is extended by accounting for the potential presence of unobserved heterogeneity in the model, which is related to the use of a panel data set for model estimation. This is achieved by using a random effects specification for both the discrete and continuous models. An empirical case study demonstrates the application of this approach for highway pavement cracking models.