Conferences

Student literacy surveys are measuring student achievement in scholastic domains. PISA international survey is conducted every 3 years since 2000. The survey collects 15-year-old student literacy test scores for all OECD and numerous non-OECD countries. PISA is supposed to measure skills needed on the labour market, however, no direct test of such presumptions is possible. The paper relates student test scores to income data trying to assess whether there are any visible similarities. We try to confirm that literacy as measured by PISA is able to predict labour market outcomes. Different sources of income data are employed including United Nations databases with income inequality statistics and EU-SILC individual datasets. The focus, however, is not on comparing country means in literacy and income, but in relating to each other whole distributions and statistics describing their shape. The paper discusses results of such comparative analysis and methodological problems which have to be addressed. The most basic problems are related to differences in variable coding and scaling of background variables which could be potentially used to adjust samples making them directly comparable. More general problems are also discussed, which in this case are mainly associated with discrepancies in complex sampling designs and validity of linking samples of dissimilar populations collected using distinct survey methodologies. Robustness and reliability of different approaches to overcome these problems and doubts are discussed. Finally, the results of comparative analysis are presented. These are based on simple comparisons of basic statistics measuring score and income levels and dispersion, but also on regression and semi-parametric methods of adjusting and comparing two distributions (see Tarozzi, 2006). Main results are in line with intuitions and partially confirm validity of the approach. However, some findings are counter intuitive and their interpretation is uncertain. Examples given make clear that results based on linked datasets are less robust to methodological criticism. In some cases it cannot be assured that they are not artifacts caused by flaws in linking methodology, rather than real phenomena revealed by researchers.