The aim of the present study, based on data from 20 countries, is to identify the pattern of variables (at country, school and student levels), which are typical of students performing below the low international benchmark compared to students performing at the advanced performance benchmark, in the Progress in International Reading Literacy Study (PIRLS) 2006. The dependent variable of the analysis is a dichotomous variable, the values of which represent the two different performance groups of students. The independent variables are two sets of Organisation for Economic Co-operation and Development educational indicators, variables from PIRLS 2006 Reading Curriculum Questionnaire data, variables and indices based on data obtained from questionnaires for teachers, schools, parents and students. The analysis is based on classification and regression trees, which is a full hierarchical non-parametric method suited to detecting and interpreting complex reciprocal influences between a large number of independent variables. The results show that pupils' performance can be predicted by the relationships between country-level variables (the changes in teachers' salaries at the top of the salary scale), school context variables (e.g. percentages of students from economically affluent home) and home educational resources and variables at student level (e.g. students' reading self-concept).