Snapshot samples

Abstract

We consider a coverage model where an initial event that occurs at some point in time triggers an activity of random duration that leads to some subsequent event. A snapshot sample is constructed at a fixed point in chronological time either by sampling only subjects where the initial event has occurred but the subsequent event has yet to occur ( active subjects), or by sampling only subjects where both the initial and subsequent events have occurred ( inactive subjects). The biases inherent in snapshot sampling can be neatly characterized by the properties of two random variables: the history H (defined as the time the initial event occurs as measured into the past from the chronological point of sampling), and the active time A (defined as the length of time between the initial and subsequent events). Though snapshot samples are biased, recognizing the biases enables correct inferences to be drawn from snapshot-sampled data. Considering only the case where H and A are independent random variables, this paper presents the probability models associated with snapshot sampling, demonstrates the problems that can occur, offers procedures for overcoming these problems, and applies the methods to interesting data sets.