Previous studies on the effects on students' test scores of their peers' socioeconomic status (SES) reported varying results. A meta-regression analysis including 30 studies on the topic shows that the compositional effect that researchers find is strongly related to how they measure SES and to their model choice. If they measure SES dichotomously (e.g. free lunch eligibility) or include several average SES-variables in one model, they find smaller effects than when using a composite that captures several SES-dimensions. Composition measured at cohort/school level is associated with smaller effects than composition measured at class level. Researchers estimating compositional effects without controlling for prior achievement or not taking into account the potential for omitted variables bias, risk overestimating the effect. Correcting for a large set of not well thought-over covariates may lead to an underestimation of the compositional effect, by artificially explaining away the effect. Little evidence was found that effect sizes differ with sample characteristics such as test type (language vs. math) and country. Estimates for a hypothetical study, making a number of "ideal" choices, suggest that peer SES may be an important determinant of academic achievement.