This work presents a study of statistical models used for assessment and measurement in the social sciences. The main contributions are: i ) a unified description of how evaluation, assessment, and the theory of measurement evolved within several branches of science; ii ) a review of estimation methods currently employed in nonlinear models; iii ) a general formulation of the maximum likelihood estimation method; and particularly, iv the presentation of the stochastic approximation method for estimation of non linear statistical models in measurement and assessment. Non linear models occurs frequently in the social sciences where it is important to model binary or ordinal response variables. This work deals with item response theory models, logistic regression models and general models with random components. The estimation of these models has been the subject of several recent simulation studies. One cannot say there is a best estimation method. The approximate methods are known to produce biased estimates, numerical integration methods and bayesian methods can present convergence problems in many cases. Stochastic approximation method is a maximum likelihood method that uses the Robbins-Monro algorithm to solve the score equation. As a stochastic approximation method it generates a Markov chain that converges to the desired estimates and can be considered a frequentist MCMC. A simulation study and a comparative estimation study show a good performance, the method producing small bias for the estimates, good precision, and very rare convergence problems.