He didn't talk about it in terms of decoherence, but the mathematics of how mixed states arise from pure states by considering subsystems was described in Everett's original paper on Many Worlds (which is not the name that he used---that was Bryce Dewitt).
Suppose that you have a composite system described by a wave function [itex]|\Psi\rangle = \sum_{\alpha, j} C_{\alpha, j} |\alpha\rangle |j\rangle[/itex], where [itex]|\alpha\rangle[/itex] is a complete set of states for the first subsystem, and [itex]|j\rangle[/itex] is a complete set of states for the second subsystem. You can think of [itex]|\alpha \rangle[/itex] as describing the system of interest--maybe an electron--while [itex]|j\rangle[/itex] describes everything else in the universe. Now [itex]K[/itex] be some operator that only affects the first component. That means that its affect on the composite state [itex]|\alpha\rangle|j\rangle[/itex] is this:
[itex]K |\alpha\rangle |j \rangle = \sum_{\alpha'} K_{\alpha' \alpha} |\alpha\rangle |j\rangle[/itex]
Since operators correspond to observables, [itex]K[/itex] represents an observable of the first subsystem alone. Now, let's compute the expectation value of [itex]K[/itex] in the composite state [itex]|\Psi\rangle[/itex]:
[itex]\langle \Psi|K|\Psi \rangle = \sum_{\alpha, \alpha', j, j'} C^*_{\alpha', j'} C_{\alpha, j} \langle \alpha' | \langle j' | K | j \rangle |\alpha\rangle[/itex]
[itex]= \sum_{\alpha, \alpha', j, j'} C^*_{\alpha', j'} C_{\alpha, j} K_{\alpha' \alpha} \delta_{j j'}[/itex]
[itex]= \sum_{\alpha, \alpha', j} C^*_{\alpha', j} C_{\alpha, j} K_{\alpha' \alpha}[/itex]
Now, if we define [itex]\rho_{\alpha \alpha'}[/itex] to be: [itex]\sum_j C^*_{\alpha', j} C_{\alpha, j}[/itex], then we have:
[itex]\langle \Psi|K|\Psi\rangle = \sum_{\alpha, \alpha'} \rho_{\alpha \alpha'} K_{\alpha' \alpha} \equiv Tr(\rho K)[/itex]
So for measurements only involving the first subsystem, the density matrix [itex]\rho[/itex] is all that we need for computing expectation values. It's a mixed state, in general.