Search

The assumption of cause-effect relationships in ex post facto research is a widely known issue in the field of research methods in social sciences. To address this important limitation, the use of causal inference techniques has become widespread in recent years. Causal inference establishes a set of statistical procedures for drawing causal conclusions in non-experimental research. Despite its wide popularity and diffusion in the social and health sciences, its use in educational research is still marginal. Thus, this paper introduces the main causal inference techniques available to the educational researcher when observational panel data are available. After addressing the key features and potential of propensity score matching, instrumental variables, and regression discontinuity design, we present an example application of each of these techniques. We used the available databases from the PISA 2018 assessments. We included the mathematical competence as the dependent variable in all the three models implemented. Given the different characteristics of each of these techniques, the independent variable used is different in the three models applied: attendance to early childhood education in propensity score matching; student academic expectations in instrumental variables; and size of the community in which the school is located in regression discontinuity design. The article concludes by discussing the potential of this set of techniques, taking into account the needs and methodological procedures most commonly applied in educational research. [ABSTRACT FROM AUTHOR]

The statistical multivariate analysis of Likert response scales, given their widespread use, is a controversial issue in the scientific community, mainly from the specification of the problem of measurement. This work aims to study various conditions of these ordinal scales affect the calculation of the product-moment and tetrachoric-polychoric correlation coefficients. For this purpose, a simulation study was applied in which 90 databases with 10 items each were generated. In the estimation of the databases, the following variables were controlled: number of response categories, symmetrical or asymmetric distributions of data, sample size and level of relationship between items. Thus, 90 matrices (10x10) were obtained which included the difference between the product-moment and tetrachoric-polychoric correlations. The graphical and variance analysis show how the product-moment correlation coefficient significantly underestimates the relationship between variables mainly when the number of response categories of the ordinal scale is small and the relationship between the variables is large. On the other hand, the statistical estimation of both coefficients is very similar when the starting relationship between pairs of variables is small and/or when the number of response options of the variables is greater than 5. The study concludes by making a recommendation to the applied researcher on the most appropriate correlation coefficient depending on the type of data available. Finally, the results are discussed from the previous studies, which reach some similar conclusions. [ABSTRACT FROM AUTHOR]