Introduction: High-stake testing plays a crucial role in many educational systems, guiding policies of accountability, resource allocation, and even school choice (OECD, 2013). However, non-representative patterns of attendance can skew how useful these measures are for accomplishing their main objective. Are we really measuring the quality or performance of a school if a non-representative sample of their students are taking the test? In this paper, we study the effect of high-stake testing on student composition of attendance on the day of the test. We implement an event study framework to estimate the effect of testing on attendance using a daily panel from Chile's school system population. We examine the impact on low and high performers between 2nd and 10th grades from the year 2011 to 2018, and estimate the effect of high-stake testing on attendance patterns for different types of students. We are also working on a method for better understanding differences in attendance within schools to help (i) improve current imputation methods, and (ii) identify schools that could potentially be incentivizing non-representative patterns of attendance. Our analysis reveals that the effect of high-stakes testing on attendance is highly heterogeneous across grades, and younger students are more affected than the rest. Most notably, the increased attendance of higher-performing students (measured according to their GPA) seems to be crucial to selective patterns of attendance. In our effort to uncover what lies behind these estimates, three main findings emerge. First, exemptions seem to not play a role in our results, since the event study estimates with and without exempted students are almost identical. Second, testing alone seems not to be a driver of changes in the student test pool for the test. We find no effect of student attendance to a no-stakes test delivered in the same way as the high-stakes test except for the consequences. Finally, we present evidence that low-performing students are less likely to receive information about the test and passing it to their parents. They are also less prepared for the assessment and declare having received incentives to make the school get good results more often, even if individual scores are never published. This paper makes several contributions to the existing literature (Cuesta et al, 2020; Figlio & Loeb, 2011; Cullen & Reback, 2006; Jacob, 2005; Figlio & Getzler, 2002). First, the results starkly contrast with the previous focus on low performers exclusion on test-scores. We show the critical role of students who are not at the bottom of the performance distribution on differential attendance patterns. Second, we document that testing alone does not affect attendance by comparing no-stakes and high-stakes results. Third, our preliminary findings have several implications for how testing agencies should present test-results and prevent absences. Lastly, we are developing a model that uses machine learning techniques to handle big-data, predict attendance, and impute predicted attendance given that some students are not absent because of the test. This model will help us to evaluate how the test is affecting attendance for every school, and to provide useful information to prevent absences. Data Description: We mainly use three datasets. First, we build a registry with all the students and schools that participated in high-stakes and no-stakes tests during 2011-2018. Second, we use the SIGE daily attendance datasets for the same period. Finally, we use a census of student's GPA to classify students according to their academic performance. Table 1 shows the description of our data. Results: Our empirical strategy consists of an event study for each grade tested using the national daily attendance panel described before. As seen in Figure 2, our analysis reveals 4 main results: (1) The effect of high-stake testing is highly heterogenous across grades, and younger students are more affected than the rest; (2) The overall impact of the effect of the test masks substantial heterogeneity across different levels of GPA for high-stake testing; (3) The increased attendance of students in the middle and the top of the GPA distribution is crucial for selective patterns of attendance; and (4) The effect of high-stake testing is also heterogenous across years. We then turn into three exercises to get a better understanding of our findings. We analyze the role of the exemption legislation by running the event study with and without exempted students. Exempted students are generally low-performers and must present a validated certificate crediting permanent special needs (i.e. blindness, autism spectrum disorder or intellectual disability among others). If those students know that their score will not count for the school achievement, they may not have incentives to attend school for the test. Also, we study the possibility that students self-exclude themselves because of the testing situation's disutility and the absence of individual consequences. To handle this, we run an event study for a no-stakes test and compare it to the high-stakes results. We take advantage that the Chilean testing agency delivers a no-stakes standardized test to a sample of schools in several years to improve the test's psychometric properties. Finally, we use unpublished student and parents' questionnaires to give more insights on what might explain our findings. We examine whether low-performing students are less likely to receive differential information or incentives than higher-performing students, among other alternatives. A critical aspect of the impact of testing on attendance is that non-representative attendance does not necessarily imply that schools manipulate the pool of test-takers. Low-performing students might have been absent anyway. In that regard, we are developing a machine learning model to study the change in the attendance distribution of every school. We seek to compare the observed attendance to what it would have had looked if no test had taken place. Then, we want to show the distribution of test-scores in three scenarios: no imputation, imputing all absences, and using the distribution of students who would have come to school if no test had taken place.