Maximum marginal likelihood estimation of multidimensional item response theory (IRT) models has been hampered by the calculation of the multidimensional integral over the ability distribution. However, the researcher often has a specific hypothesis about the conditional (in)dependence relations among the latent variables. Exploiting these relations may result in more efficient estimation algorithms. A well-known example is the bi-factor model, in which each item measures a general dimension and one of K other dimensions, for which Gibbons and Hedeker (1992) showed that full information maximum likelihood estimation only requires the integration over two-dimensional integrals. In this paper, it is shown how the approach of Gibbons and Hedeker (1992) can be placed into a graphical model framework. The advantage of the graphical model framework is that efficient estimation schemes can be derived in a fully automatic way by applying algorithms to the graphical representation of a statistical model. This renders the approach fairly generally applicable, and tedious derivations by hand are no longer involved. The generality of the approach is demonstrated by applying it to a multidimensional IRT model with a second order dimension. It turns out that full information maximum likelihood estimation for such a model also requires the evaluation of two-dimensional integrals only. Two appendices are included: (1) Identification Restrictions for the Bi-factor and the Testlet Model; and (2) Computations for the Bi-Factor Model. (Contains 6 figures.)