Supplementary Materials for "GraphScape: A Model for Automated Reasoning about Visualization Similarity and Sequencing", submitted to CHI 2017.

1. linear_programming: Edit operation linear program generator
	- lp.js with instruction.

2. formative_study: Formative experiment materials
	- study_design: Screenshots
	- stimuli: Chart images with set_id, chart_id, and whether the chart was fixed at first.
	- analysis: R script with output as comments.
	- Collected data and data table schema information.

3. evaluation_study: Rating evaluation experiment materials
	- study_design: Screenshots
	- stimuli: Sequence images with set_id, chart_id, and whether the sequence was the top sequence in GraphScape.
	- analysis: R script with output as comments + JavaScript code for calculating bootstrapped mean 95% confidence intervals in a manner sensitive to a within-subjects design.
	- Collected data and data table schema information.

4. power_analysis: Power analysis code for formative study and both evaluation analyses
	- power_analysis_binomial.R: Simple power analysis of different proportions of observing one sequence versus another given different N. We find that with 51 people, we see a significant result at least 80% of the time as long when the population proportion is 0.70 or higher
	- power_analysis_mixedE.R: Power analysis that simulates the evaluation study data and runs a mixed effects model to find probability of finding significant effect of graphscape assuming actual observed intercepts and variances. We find that with 55 people (minus incorrect trials ~ 51 people) we have power of 1.
	- power_analysis_friedman_posthoc.R: Power analysis that simulates evaluation study data with different N, and runs a Friedman test followed by posthoc testing with Holm correction to find probability of seeing a significant Friedman test and significant posthoc differences between the graphscape sequence and the other sequences (one, two, three, or all four differences). The simulation samples ratings for non-graphscape sequences assuming the mean and variance we observed on average for a non-graphscape sequence in that task. The simulation assumes the graphscape sequence had ratings that are at least g higher rating than the other sequences, and variance equal to the observed variance for graphscape ratings for that task. We find that 55 people (minus incorrect trials ~ 51 people) results in a power of at least 0.8 for detecting at three differences between graphscape and non-graphscape sequences assuming that the true graphscape rating is g=1 pt higher or more for all tasks, and a power of at least 0.8 for detecting all four differences for 4 out of 6 tasks. 

5. figure4_calculation
 - Full calcuation of each part of sequence costs of sequences in figure 4.