Recent work has shown that rodent decision-making relies on distinct latent or hidden states that switch on a timescale of tens to hundreds of trials [Ashwood 2021; Bolkan, Stone 2021; Weilnhammer 2021]. However, the neural basis for these states, and the contributions of different brain regions to state-dependent decision-making strategies, remains an important open problem. Here we address this challenge by performing a model-based analysis of decision-making data acquired during multi-region laser-scanning inactivation of dorsal cortex. We analyzed optogenetic inactivation data from mice performing a visual-evidence-accumulation task while navigating in a virtual environment (a publicly available dataset from Pinto et al. 2019). In this dataset, 29 different individual cortical regions were bilaterally inactivated during a randomly interleaved subset of trials. We fit choice data from these experiments using a Hidden Markov Model (HMM) with Bernoulli Generalized Linear Model (GLM) observations. The resulting GLM-HMM framework describes choice behavior with state-specific GLMs that quantify how the animal combines different features (e.g., sensory evidence, bias, choice history) to make decisions in each state, and a matrix of transition probabilities governing the switches between states. To incorporate the effects of neural perturbations on behavior, we grouped the 29 inactivation sites into three clusters. We then extended the model by adding GLM weights for each cluster, allowing inactivation of each cluster to have distinct, state-specific effects on choice. Our preliminary analyses revealed that a multi-state GLM-HMM substantially outperformed a basic GLM at predicting behavior, and suggested that the effects of bilateral inactivations in dorsal cortex were highly state-dependent, with the sign and strength of the effect on choice varying substantially across states.