This study compares the estimated grammar school effect in different regression models, and explains why previous evidence of the effectiveness of grammar school is mixed. Like most studies of school effectiveness evaluation, previous research on grammar school effect usually applies regression to control for confounding between-school factors and determines whether attending grammar schools is associated with an academic benefit. While this value-added approach is very feasible and widely adopted, there is usually substantial variation in the evidence produced when statistical choices differ. Applying the national data with more than 149,000 pupils, the study presents the sensitivity of grammar school effectiveness when baseline variables, outcome variables, and regression techniques differ. The findings will help researchers understand the mixed evidence of the effectiveness of grammar schools so far, and the unsettled debate on grammar schools nationally. While inconsistent estimations due to passive design is a common trait in school effectiveness evaluation studies, grammar schools' effectiveness might be even more sensitive to statistical choices due to their special student intakes. Interpreting the variation between different statistical models, the findings caution researchers in dealing with the evidence in grammar school effectiveness and call for stronger designs as well as more carefully selected explanatory variables.