Events at CERGE-EI

Master Thesis Chair:
Daniel Münich

Abstract:

People are by nature social beings. Most of us have a complex social network that connects us with other people in numerous aspects of our lives: neighbours, co-workers or peers in schools, and friends. Moreover, it is widely believed that people’s behaviour is to some extent affected by others in their social networks, which is known as peer effects. Therefore, a precise understanding of the behaviour of an individual necessarily includes understanding her interactions with others within her social network.

The first part of this thesis, literature review, summarizes contemporary research on peer effects, shows which aspects of human behaviour may be affected by social interactions, and highlights the importance of peer effects research. In the second part, the estimation of the linearin-means peer effects model, we provide a detailed description of the model, derivations of its alternative formulations, and show the identification conditions. The main contribution of the second part is that we provide a step-by-step analysis of the linear-in-means peer effects model and detailed proofs of theorems in one place.

The third part provides an empirical analysis of peer effects in education in the Czech Republic. In particular, we examine how the test scores of pupils are affected by their classmates. We observe that pupils’ test scores are negatively affected by the test scores of their peers and positively affected by the abilities of their peers. The results are statistically significant; however, they are also excessively high compared with previous research. Therefore, we conduct bootstrap simulation and find that the estimators of standard errors are probably underestimated. Moreover, we conduct a placebo check randomly allocating pupils among classes and show that widely used peer effects estimators are slightly biased in both directions, which could explain high and significant peer effects estimators. Therefore, we conclude that corrected peer effects estimators are likely unsignificant in our setting, which is mainly caused by the small data sample. Finally, we briefly propose possible extensions of the linear-in-means peer effects model, which may give a more realistic description of peer effects in real world.