I've become quite a fan of approximate likelihood methods over time, partially because many of them are so creative. A particular favourite is the Whittle likelihood for stationary Gaussian time series, which uses a clever trick to move into Fourier space and greatly reduce the computational cost.