I discuss the application of capture-recapture methods to estimating the total number of deaths in armed conflicts, and propose an alternative method based on a trivariate discrete choice model. Data come from the 'Truth and Reconciliation Commission' (TRC) of Peru, around 25000 deaths, classified by three sources of information, geographical strata, and perpetrator: the State and the Shining Path. In these data many killings have been only documented by one source, which makes a projection of killings unfeasible. TRC consultants Ball et al. (2003) tried to overcome this problem by means of a 'residual estimation,' consisting of merging data for different perpetrators. I show theoretically and empirically that this method over-estimates the number of deaths. Using a conditional trivariate Probit I estimate the total number of deaths in around 28000, 60% by the State, 40% by the Shining Path. This number is substantially lower and has a different composition than the around 69000 deaths, 30% by the State, 46% by the Shining Path, and 24% by 'other perpetrators,' calculated by Ball et al.