### Text accompanying preregistration of analysis prior to data collection This preregistration concerns the analysis of data from the MAJIC project, a randomized trial of mental abacus training. Randomization was conducted in fall 2015 at the classroom level. The primary hypothesis of the project is that mental abacus training will improve math outcome measures, consistent with Barner et al. (2016)'s findings. This preregistration is being conducted prior to data collection and data analysis for the SECOND session of data collection, in spring 2016. The goal of the preregistration is to register inclusion criteria, outcome measures, and analyses. We did not preregister prior to data collection in the FIRST session because we were uncertain as to the feasibility of analysis for all computer-based (secondary) outcome measures. The registered preliminary analyses suggest that these measures can be used for mediation analysis as specified below. **Inclusion criteria** + All students for whom we have data and consent from initiation will be included. + Students with missing data will be included for those analyses for which we have data available. + Exclusions for individual tasks will be as in Barner et al. (2016); in general we will only exclude children's data if there is clear evidence or notes that they have not even attempted a task. **Outcome measures** + Primary outcome measures are mathematics: namely, placevalue, in-house arithmetic, and WJ-III arithmetic. + Secondary outcomes are cognitive: spatial working memory, reasoning (matrices), and executive function. **Planned analyses** Analysis group 1: longitudinal mixed models of arithmetic/mathematics of the form: `outcome ~ grade_level + year * intervention + (1 | subid) + (year | class)`. In case of convergence issues, we will prune the year slopes by class, then the class intercepts. The key test of our hypothesis for each of the three measures will be the year * intervention interaction, signaling an increase in performance over the year, due to the intervention. Analysis group 2: longitudinal mixed models of cognitive abilities. Form of the models will be as above, but with the prediction of no interaction (following Barner et al. 2016). Analysis group 3: longitudinal mixed models of mathematics, mediated by spatial working memory. We predict, following Barner et al. (2016), that spatial working memory scores at initiation will predict gains from mental abacus. The model for this will be of the form: `outcome ~ grade_level + year * intervention * swm + (1 | subid) + (year | class)`. Contra Barner et al. (2016), we will code SWM continuously rather than as a discrete variable (though we will use a median split for visualization). These mediations will be carried out for all three math outcome variables. Following Barner et al. (2016) we predict a mediation relationship between SWM and intervention uptake. Analysis group 4: Similar analyses of mediation of math by other variables, namely matrix reasoning and executive function. We informally predict smaller mediation effects than for SWM.