In this paper we create different Bayesian networks using expert experience and Markov equivalence condition. Using real-world dataset and holdout sample, we test different Bayesian networks that are not Markov equivalent. All our different configurations give us similar results. Next, we take the simplest configuration and enhance it by adding a hidden node and learning its conditional probability table using the gradient descent algorithm. The results of our experiments indicate that the new Bayesian network with hidden node improves holdout sample performance compared to Bayesian networks without the hidden node. The existence of the hidden node (we call it project characteristics) seems to indicate that software measurement variables only play a partial role in predicting software effort. Actual software effort is expected to vary from one project to another depending on certain unmeasured parameters related specifically to individual projects. Improvement in predictive performance via the use of hidden variable seems to indicate that individual projects should be treated uniquely, and software project management remains a key factor for software effort and cost containment.