Greater data science at baccalaureate institutions
aa r X i v : . [ s t a t . O T ] O c t Greater data science at baccalaureate institutions
Amelia McNamara, Nicholas J Horton, Benjamin S BaumerOctober 25, 2017Donoho’s paper is a spirited call to action for statisticians, who he points out are losing ground inthe field of data science by refusing to accept that data science is its own domain. (Or, at least, adomain that is becoming distinctly defined.) He calls on writings by John Tukey, Bill Cleveland,and Leo Breiman, among others, to remind us that statisticians have been dealing with data sciencefor years, and encourages acceptance of the direction of the field while also ensuring that statisticsis tightly integrated.As faculty at baccalaureate institutions (where the growth of undergraduate statistics programshas been dramatic [2]), we are keen to ensure statistics has a place in data science and data scienceeducation. In his paper, Donoho is primarily focused on graduate education. At our undergraduateinstitutions, we are considering many of the same questions.We enthusiastically concur with Donoho’s description of a “Greater Data Science” comprised of1. Data Gathering, Preparation, and Exploration2. Data Representation and Transformation3. Computing with Data4. Data Modeling5. Data Visualization and Presentation6. Science about Data Scienceand aim to have our students develop all these key capacities in our courses and major programs.In considering our curriculum development, we have been guided by the 2014 American StatisticalAssociation (ASA)’s
Curriculum Guidelines for Undergraduate Programs in Statistical Science [1]and the 2016
Guidelines for Assessment and Instruction in Statistics Education (GAISE) CollegeReport [13]. Both documents highlight the need for students to work with real problems, messydata, and complex models.Even more recently, a working group (including Baumer) developed the
Curriculum Guidelinesfor Undergraduate Programs in Data Science , which have now been endorsed by the ASA [15].This forward-thinking document addresses one of Donoho’s primary concerns with data scienceeducation—that it may end up being a piecemeal collection of extant courses, with little “long-term direction.” While [15] does provide guidance to institutions working with existing courses, italso outlines a model curriculum with a number of new and reformulated courses.
Data science developments at our institutions
Both the Smith College major in statistical and data sciences and the Amherst College major instatistics have been explicitly structured to introduce, extend, and integrate work in all six of theareas of Greater Data Science. Real problems have been interwoven into our courses at multiple1evels. This has required extensive revision of existing courses along with the creation of a numberof new and courses with complementary learning outcomes.At both Smith College and Amherst College, the introductory course touches on all six GDSelements, with an increased emphasis on visualization and modeling [6, 21]. In subsequent courseslike
Multiple Regression or Intermediate Statistics , students explore, prepare, clean, transform,and visualize data. In the
Communicating with Data , Visual Analytics , and
Multivariate DataAnalysis courses, students learn principles of data visualization and presentation of data. Modelingis reinforced in
Multiple Regression and
Machine Learning . Capstone courses help to integrate priorcourse work with project-based learning while further refining computing and communication skills.Existing Amherst College theory courses such as
Probability and Theoretical Statistics have beenrestructured to integrate computing as an explicit learning outcome (e.g., how to write a function,how to perform simulations, how to undertake empirical problem solving to complement analyticresults, and how to collaborate in groups using GitHub).At Smith College,
Introduction to Data Science , Communicating with Data , Visual Analytics , and
Machine Learning are all new offerings guided by our understanding of data science as its owndiscipline.We would like to draw particular attention to
Introduction to Data Science , a successor to thecourse described in [4] that is offered at both institutions. Donoho makes reference to this course,which teaches data visualization, data wrangling, ethics, SQL, and communication, using a newtextbook [7]. The course is tied together by liberal arts modules , where professor from otherdisciplines outline a question relevant to their discipline, and the students seek to address it usingtheir new-found data skills.As Donoho reminds us, some academic statisticians have long been guilty of eschewing data analysis.But even some programs in data science focus more on tools and skills rather than developing thecapacity to solve real problems. We believe our positions at liberal arts colleges give us a particularability to reach across disciplines, connecting to data in the sciences, social sciences, and thehumanities. The integration of liberal arts modules in
Introduction to Data Science can be used asa model for similar courses.Another learning outcome in all of our courses is to produce students who learn how to learn. Aswith many disciplines, data science is evolving quickly. The tools we teach our students today maynot be relevant in five years. In fact, several of the R packages referenced by Donoho ( reshape2 and plyr ) have now been supplanted by others ( tidyr and dplyr ) [26, 27]. As instructors, wedo our best to stay on top of the current computational trends to provide our students with themost contemporary methods, which requires us to continually modify our curriculum. However,the focus is on generalized problem-solving that can be applied using different tools in differentsettings.Ethical precepts are an important part of any data science program. Donoho alludes to this withhis detailed coverage of the University of California–Berkeley Master’s program, which includes acourse now titled “Behind the Data: Humans and Values” (formerly “Legal, Policy, and EthicalConsiderations for Data Scientists”) [23]. At Amherst ethics is now included as a learning outcomein the
Intermediate Statistics course with subsequent extension and reinforcement in elective andcapstone courses. Ethics is also a component of the
Introduction to Data Science courses. Studentsconsider questions like those posed in [8]: what are the ethical implications of data science products?Who has access to data science, and who does not? What are our ethical obligations to our clients,2urselves, and our subjects? These higher-level questions make up a key part of the capstonecourses.At all levels, our courses emphasize best practices of statistical computing and reproducible re-search. These efforts build upon scholarly work that goes back at least to Don Knuth’s literateprogramming [17] and Donoho’s previous work on reproducibility [12]. Baumer and McNamara areformer faculty fellows on Project TIER: Teaching Integrity in Empirical Research [3], which aimsto spread good computing and data practices to the social sciences. We are now seeing evidenceof adoptions at our institutions, and others, where faculty members in economics, psychology, andenvironmental science and policy integrate reproducible research into their coursework, furtherstrengthening our pool of data-capable students.
Data science scholarship
Beyond our interest in the pedagogy of data science, we are also researchers. However, this is an areathat is also undergoing development. Since it is an emerging field, institutions must determine howto judge new types of scholarly production. Like many problems of data science, this is somethingthat applied statisticians have been wrestling with for decades. However, not all data science workis precisely applied statistics (thus, the new degree programs and scholarship).Much like Donoho’s notion of Science about Data Science, Jeff Leek has been proposing the idea ofData Science as a Science [18]. While Donoho’s examples focus on meta-analysis, Leek’s conceptionincludes hands-on research. Calling on examples like Cleveland’s study of graphical perception [14],Leek advocates for data scientists experimenting to learn how software syntax impacts learning,and how practitioners are actually working (like [22]).As a case study of scholarly production in data science, consider Hadley Wickham’s many contri-butions. Wickham’s work often centers on a profoundly useful R package. However, each pieceof software fits into a higher-level framework of intellectually-weighty ideas. The ideas behind ggplot2 were articulated in a book on implementing Wilkinson’s Grammar of Graphics [25, 28]. Inaddition to tidyr , Wickham wrote a article in the Journal of Statistical Software on the conceptof tidy data, which transcends the language it is implemented in [26]. Although these works arehighly-cited, they do not fit cleanly into the traditional fields of statistics (having nothing to dowith modeling, estimation, or inference) nor computer science (software engineering?). We submitthat these are early, influential works of scholarship in data science.Another set of exemplary papers can be found in a recently-published collection of articles—curatedby Jenny Bryan and Hadley Wickham—entitled
Practical Data Science for Stats (to which theauthors all contributed) [11]. These articles discuss meta-data science topics like how to packagereproducible analytical work [19], how to organize data in a spreadsheet [9], how to share data forcollaboration [16], and how to implement a version control system [10]. Our contributions discussedsurviving as an isolated data scientist [5], and wrangling categorical data [20].The collection also contains an article on evaluating scholarly work in data science, focusing partic-ularly on data science faculty in traditional statistics and biostatistics departments [24]. Can theseexemplary scholarly contributions in data science be neatly categorized into statistics or computerscience research? If not, this further strengthens the notion that data science exists as a field ofresearch unto itself. 3 ituating greater data science
This brings us to our final question. If Donoho’s vision of ‘Greater Data Science’ takes hold, onewonders whether the current academic departmental alignments will (or should) continue. Of theauthors, one is situated within a Department of Mathematics and Statistics (Horton), while theother two are appointed in a Program in Statistical and Data Sciences. Which approach is mostfruitful?Clearly, there are many other academic areas that use data and data science methods. As we’vediscussed, our colleagues across the disciplines are embracing it. However, if data science is its owndiscipline, it cannot be solely situated within data-generating departments. Its unique teachingand scholarship indicate it may need to become a separate entity.
References [1] American Statistical Association (2014). . .[2] American Statistical Association (2015). A peek into thelargest, fastest-growing undergraduate statistics departments. http://magazine.amstat.org/blog/2015/02/01/undergraduatedepts_feb2015 .[3] Ball, R. and Medeiros, N. (2012). Teaching integrity in empirical research: A protocol fordocumenting data management and analysis. The Journal of Economic Education , 43(2):182–189.[4] Baumer, B. (2015). A data science course for undergraduates: Thinking with data.
TheAmerican Statistician
Technology Innovations inStatistics Education , 8(1).[7] Baumer, B., Kaplan, D. T., and Horton, N. J. (2017).
Modern data science with R . Chapman& Hall/CRC.[8] boyd, d. and Crawford, K. (2012). Critical questions for big data.
Information, Communication& Society , 15(5):662–679.[9] Broman, K. W. and Woo, K. H. (2017). Data organization in spreadsheets. ,5:e3183v1.[10] Bryan, J. (2017). Excuse me, do you have a moment to talk about version control? , 5:e3159v2.[11] Bryan, J. and Wickham, H., editors (2017).
Practical Data Science for Stats . PeerJ.412] Buckheit, J. B. and Donoho, D. L. (1995). Wavelab and re-producible research. Technical Report 474, Stanford University. http://statistics.stanford.edu/~ckirby/techreports/NSF/EFS%20NSF%20474.pdf .[13] Carver, R. et al. (2016).
Guidelines for Assessment and Instruction in Statistics Education:College Report 2016 . American Statistical Association.[14] Cleveland, W. S., McGill, R., et al. (1985). Graphical perception and graphical methods foranalyzing scientific data.[15] De Veaux, R. D. et al. (2017). Curriculum guidelines for undergraduate programs in datascience.
Annual Review of Statistics and Its Application , 4(1):1–16.[16] Ellis, S. E. and Leek, J. T. (2017). How to share data for collaboration. ,5:e3139v5.[17] Knuth, D. (1992). Literate programming.
CSLI Lecture Notes, Stanford University , 27.[18] Leek, J. (2016). Data science as a science. In
Joint Statistical Meetings .[19] Marwick, B., Boettiger, C., and Mullen, L. (2017). Packaging data analytical work reproduciblyusing R (and friends). , 5:e3192v1.[20] McNamara, A. and Horton, N. J. (2017). Wrangling categorical data in R. ,5:e3163v1.[21] Pruim, R., Kaplan, D. T., and Horton, N. J. (2017). The mosaic Package: Helping Studentsto ‘Think with Data’ Using R.
The R Journal , 9(1):77–102.[22] Silberzahn, R. et al. (2017). Many analysts, one dataset: Making transparent how variationsin analytical choices affect results.[23] UC Berkeley School of Information (2017). Master of information and data science: Curricu-lum. https://datascience.berkeley.edu/academics/curriculum/ .[24] Waller, L. A. (2017). Documenting and evaluating data science contributions in academicpromotion in departments of statistics and biostatistics. , 5:e3204v1.[25] Wickham, H. (2009). ggplot2: Elegant Graphics for Data Analysis . Springer Verlag: NewYork, NY.[26] Wickham, H. (2014). Tidy data.
The Journal of Statistical Software , 59(10). http://vita.had.co.nz/papers/tidy-data.html .[27] Wickham, H. and Francois, R. (2016). dplyr: a grammar of data manipulation . R packageversion 0.5.0.9000.[28] Wilkinson, L. (2005).