Peer effects in education are of interest to parents, policy-makers and researchers alike. This paper uses administrative data on four cohorts of Ontario elementary and middle school students to estimate peer effects in the classroom. I use data on test scores from province-wide assessment of mathematics, reading and writing abilities of all students in public schools which takes place at the end of Grade 3 and then again at the end of Grade 6. I use direct measure of peer ability - test score in Grade 3 - which is immune to the reflection problem because it is observed before the student has changed school. I use variation in the average ability of new peers (those who entered school between Grade 3 and Grade 6) as exogenous source of variation to identify peer effects at the classroom level. The main outcome variable is a change in the test score between grades 3 and 6, thus I estimate a value-added model of  education production function with peer variables. The identification strategy assumes that average peers' ability is not correlated with unobserved variables. To support this claim, I control for the fraction of peers at a school level, for the average test score of new peers at a school level; I also account for time trends and school fixed effects. The main threat to my identification strategy is a non-random assignment of new students across classrooms within school. In order to get insights into how students are allocated into classrooms, I conduct a survey of school principals in Ontario. I also use data from auxiliary sources to show that assignment of students by classrooms as a rule is random and is not based on abilities. I use administrative data from EQAO (Education Quality and Accountability Office) in Ontario. The data cover years from 2006 to 2010 for all students in grades 3 and 6 who participated in mathematics, reading and writing assessments over those years. For three cohorts of students I am able to match data for individual students thus obtaining a short panel. I combine individual level data with data at a school level and further match each school by its postal code to the neighborhood characteristic constructed from Canadian Census. I find that peers' ability matters for own student achievement. For all three subjects, increase in the average peer ability is associated with an improvement in student's own test score in the current period as well as with a significant gain in the test score from Grade 3 to Grade 6. Additionally, I find that boys benefit from the presence of female peers in class, while this is not the case for female students.