Stock assessments based on fitting age-structured population dynamics models using the integrated approach usually require data on the length-/age-structure of fishery removals and age-length data to estimate key population parameters such as growth rates, recruitment, natural mortality rates and selectivity. The errors in the estimates of these population parameters directly impact the ability to estimate biomass and hence recommended catch limits based on harvest control rules. The performance of a stock assessment method therefore depends in part on the quality of these “composition” data. Simulation is used to evaluate the effect of a range of effective sample sizes for the length-composition and conditional age-at-length data on the errors in estimating spawning biomass, depletion, recruitment, selected population model parameters, and key model outputs, including catch limit recommendations. Multiple operating models, including those that result in the estimation method being mis-specified, are considered, and the simulations account for the effects of life history, catch history and current stock status. As expected, the errors in estimates of management-related quantities are less for greater effective sample sizes, but effective sample size is seldom the major determinant of the magnitude of estimation errors. The effect of sample size on relative errors is neither linear nor exponential, and spatially-unbalanced sampling is more detrimental than infrequent collection of composition data but with larger sample sizes. The quantitative results are likely case-specific, as they depend on several aspects of the operating model; the framework of this paper can be applied to specific fisheries, in combination with an algorithm that links effective to actual sample sizes.