Increasing the Transparency of Research Papers with Explorable Multiverse Analyses
Total Page：16
File Type：pdf, Size：1020Kb
CHI 2019 Paper CHI 2019, May 4–9, 2019, Glasgow, Scotland, UK Increasing the Transparency of Research Papers with Explorable Multiverse Analyses Pierre Dragicevic Yvonne Jansen Abhraneel Sarma Inria CNRS – Sorbonne Université University of Michigan Orsay, France Paris, France Ann Arbor, MI, USA [email protected] [email protected] [email protected] Matthew Kay Fanny Chevalier University of Michigan University of Toronto Ann Arbor, MI, USA Toronto, Canada [email protected] [email protected] ABSTRACT ACM Reference Format: We present explorable multiverse analysis reports, a new ap- Pierre Dragicevic, Yvonne Jansen, Abhraneel Sarma, Matthew Kay, proach to statistical reporting where readers of research and Fanny Chevalier. 2019. Increasing the Transparency of Research Papers with Explorable Multiverse Analyses. In CHI Conference on papers can explore alternative analysis options by interact- Human Factors in Computing Systems Proceedings (CHI 2019), May 4– ing with the paper itself. This approach draws from two 9, 2019, Glasgow, Scotland UK. ACM, New York, NY, USA, 15 pages. recent ideas: i) multiverse analysis, a philosophy of statistical https://doi.org/10.1145/3290605.3300295 reporting where paper authors report the outcomes of many different statistical analyses in order to show how fragile or robust their findings are; and ii) explorable explanations, 1 INTRODUCTION narratives that can be read as normal explanations but where The recent replication crisis in psychology and other disci- the reader can also become active by dynamically changing plines has dealt a blow to the credibility of human-subject re- some elements of the explanation. Based on five examples search and prompted a movement of methodological reform- and a design space analysis, we show how combining those [70]. Much of this movement calls for more transparency two ideas can complement existing reporting approaches and in the way statistics are reported, so that findings become constitute a step towards more transparent research papers. more trustworthy, more likely to be interpreted correctly, and easier to verify and replicate [29, 69, 72]. Concern for CCS CONCEPTS transparency in statistical reporting has spread to the HCI • Human-centered computing → Human computer in- community, which has published several articles [24, 31, 58] teraction (HCI). and hosted several workshops [56, 57, 96] on the topic. While much of the current discussions around transparent KEYWORDS statistics in HCI focus on how the community can improve Multiverse analysis, explorable explanation, statistics, trans- its practice, it has been suggested that HCI can do more than parent reporting, interactive documents. endorse and promote the transparent statistics movement— it can actively contribute to it by proposing novel user in- terfaces for better doing and better communicating statis- tics [31, 97]. In this article, we consider the research paper as Permission to make digital or hard copies of all or part of this work for a user interface, and seek to understand how we can enrich personal or classroom use is granted without fee provided that copies this user interface to better support and promote transparent are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights statistics reporting. for components of this work owned by others than the author(s) must While there are many ways a statistical report can lack be honored. Abstracting with credit is permitted. To copy otherwise, or transparency, a common and damaging form of opacity is republish, to post on servers or to redistribute to lists, requires prior specific undisclosed flexibility (see Figure 1a), i.e., not reporting the permission and/or a fee. Request permissions from [email protected]. different options that have been tried during the analysis85 [ , CHI 2019, May 4–9, 2019, Glasgow, Scotland UK 98], or the options that would have been chosen had the © 2019 Copyright held by the owner/author(s). Publication rights licensed to ACM. data been different [40]. Undisclosed flexibility is damaging ACM ISBN 978-1-4503-5970-2/19/05...$15.00 because it substantially increases the chances of reporting https://doi.org/10.1145/3290605.3300295 erroneous findings, while being invisible to the reader. Paper 65 Page 1 CHI 2019 Paper CHI 2019, May 4–9, 2019, Glasgow, Scotland, UK Analyzed Reported Analyzed Reported Analyzed Reported A second option consists of summarizing the multiverse in the research paper itself using tables and custom plots. This is the approach put forward by the two papers intro- ducing multiverse analysis [86, 87], which use as examples histograms of p-values [87], grids of p-values [87], and sorted a. traditional analysis b. planned analysis c. multiverse analysis dot-plots of point estimates where color encodes statistical significance [86]. Although such visualizations are very use- Figure 1: Three reporting strategies, from the least trans- parent to the most transparent: a) traditional analysis with ful as overviews, they force the paper author to collapse the undisclosed flexibility; b) planned analysis; c) multiverse results of each analysis into a single point estimate, p-value analysis. Each branching represents a choice between differ- or statistical significance verdict. In doing so, a lot ofthe ent analysis options [71]. richness and nuance of well-crafted statistical reports is lost. A third approach, which we propose here, consists of One response to the issue of undisclosed flexibility has making the results section of the research paper interac- been to encourage researchers to commit to a single sta- tive. More specifically, we introduce and explore the concept tistical analysis that has been planned [17, 31] and ideally of explorable multiverse analysis report (emar). An emar is registered [24] ahead of time (Figure 1b). Although plan- a multiverse analysis report that is presented in the form ning eliminates the problem of undisclosed flexibility, some of an explorable explanation in Bret Victor’s sense [94]. Ex- statisticians and methodologists are starting to argue that plorable explanations are explanations that “enable and en- more transparency can be achieved by letting researchers try courage truly active reading” and allow “the reader to play many analyses and report all of them in their paper [86, 87] with the author’s assumptions and analyses, and see the (Figure 1c). This is partly motivated by evidence that dif- consequences” [94]. At the same time, they can be read like ferent researchers who analyze the same data will make normal prose: “the reader is not forced to interact in order to different choices and will thus get slightly—and sometimes learn” [94]. Consistent with this design philosophy, an emar widely—different results [84]. looks like a regular results section, but readers are given In a multiverse analysis, researchers identify sets of de- the possibility to change some of the analysis options and fensible analysis choices (e.g., different ways of excluding immediately see the results within the research paper itself. outliers, different data transformations), implement them all, We posit that emars can be a useful and compelling com- and then report the outcomes of all analyses resulting from plement to existing multiverse reporting approaches. Our all possible choice combinations. This approach increases goal with the present paper is to provide a better under- transparency because readers can appreciate the “fragility standing of the design space of possible approaches. We will or robustness of a claimed effect”86 [ ] by checking whether explore this design space through five examples of short the findings are dependent on arbitrary analysis choices. interactive papers1 we have written to demonstrate the con- Multiverse analysis promises an unprecedented level of cept of explorable multiverse analysis report. We will discuss transparency for research papers, but the idea is in its infancy. the trade-offs between different emar reporting strategies, Writing papers with a multiverse analysis are difficult and as well as the benefits and challenges raised by emars com- there is currently very little guidance. An important part of pared to alternative reporting strategies such as the use of the difficulty lies in reporting the outcomes of potentially supplemental material or static multiverse analysis reports. hundreds or thousands of analyses in a single research paper (past examples contain between 120 and 1728 analyses [86, 2 RELATED WORK 87]), causing challenges for authors, reviewers, and readers. We regard the research paper as an interactive medium and There are currently two ways to approach this problem. focus on how this medium can support the communication of A first option consists of sharing multiverse analyses as multiverse analyses. We review prior literature on interactive supplementary material, letting readers look under the hood documents and interactive statistical reports, and discuss the of a default analysis and try alternative analysis options state of academic publication practices. in a different environment. This approach has long been promoted by the reproducible research movement, and has Interactive Documents been the subject of a vast body of work [21, 41, 73]. Although We use “document” to refer to any information artifact that is supplementary material is crucial for reproducibility and constructed around a textual narrative. Since the invention reuse, casual readers are very unlikely to engage with it. Thus of hypertext [16], the HCI community has never stopped proponents of multiverse analysis argue for acknowledging, reporting and discussing the multiverse in the research paper itself [86, 87]. 1available at https://explorablemultiverse.github.io/. Paper 65 Page 2 CHI 2019 Paper CHI 2019, May 4–9, 2019, Glasgow, Scotland, UK to explore how interactivity can be used to enhance doc- snippets of code to re-compute figures of a research article in uments, e.g., for supporting annotation tasks [81, 101] or a side panel [61]. To date, Distill [1], mentioned previously, non-linear navigation within document content.