The goal in this paper is to explain the academic performance for PhD students as measured by the the number and type of publications and conference papers co-authored during the last three years. In the literature, performance in creative jobs such as those involving knowledge creation is explained from different types of variables. Some authors stress the role of social background variables; some other authors use mainly attitudinal variables such as job satisfaction or job motivation, and some others focus on network variables such as centrality, density or closeness. Very rarely have all types of variables been used together. The data were collected with a web-administered questionnaire from the whole population of PhD students at the University of Girona (Spain). The egocentered networks of both the PhD student and his/her supervisor were merged into what is known as a nosduocentered network (Coromina et al. 2005). The population size did not make it possible to use formal measurement error models for the attitudinal variables, which were measured using Summated Rating Scales (SRS). Appropriate reliability measures were computed from exploratory factor analysis models and correlations were corrected for attenuation. Since the complete population was available, formal statistical tests were not interpretable and the explanator power of the variables was assessed by means of standardized regression coefficients, partial correlations and adjusted R2. We started by specifying three regression models, one for each group of variables. The best background variable model included supervisor performance, seniority at the department, field of study, having children and age. The best attitudinal variable model included different motivations to start a PhD and job satisfaction. The best network variable model is the maximum density for nosduocentered networks. Then, we fitted three regression models combining the predictors of all possible pairs of groups and a regression model with the predictors of all three groups. The comparison of the adjusted R2 statistics made it possible to see which groups of predictors added explanatory power with respect to the other groups. The final model to explain PhD student performance is composed by background and attitudinal SRS variables. The lack of predictive power of network variables cannot be understood as the network being completely irrelevant. In fact, the most relevant predictor in the model is the supervisors’ academic performance, and the supervisor should be a very important source of social capital for the PhD student. The fact that the field of study is significant still reveals very different traditions of publishing. Age, having children and the amount of years working in the department also have a positive effect on performance. The last variables with predictive power for performance are motivations to start a PhD such as greater work autonomy and career advantages.