In an Orthogonal Frequency Division Multiplexing (OFDM) system, the channel is often modeled as a two-dimensional zero-mean Gaussian random process in time and frequency with known second order statistics. With pilot symbols placed across the time-frequency plane, the channel's response at a particular time-frequency point can be derived from its correlation with the nearby pilot locations. However, this a priori channel statistics may not always be available or the channel may not even be stationary to be adequately characterized by a correlation function. In this paper, we introduce a model of time-varying channel based on its delay-Doppler response. We show that a widely adopted time-frequency correlation model, upon which popular channel estimators for OFDM are often based, can be viewed as a special case of such a model with separable statistical profiles imposed on the delay-Doppler plane. Without using any statistical assumption on the channel, we derive a classical estimator based on the two-dimensional delay- Doppler correlator, which can be implemented using Discrete Fourier Transforms (DFT) of the pilot symbol measurements. Simulation results show promising performance even when the pilot symbols' insertion rate is at the channel's maximum delay-Doppler spread.