Conclusion
Due to the discrepancies in creativity, depth, and multivariate visualization scores between students projects using base R and others employing the tidyverse syntax, we recommend that the tidyverse syntax should be the primary syntax when using R as a supplement for introductory statistics courses. Although the sample is limited by solely analyzing Duke’s STA 101 final projects, the standardization of STA 101 classes provides an optimal baseline for comparison. Another potential source of bias, the project assignment document, can be nullified by the lack of wording differences since the Fall 2013 document (a base R class) supplemented by a similar creativity metric distribution for the Fall 2013 class as other base R section. As such, since there was shown to be higher scores in the three metrics for projects using the tidyverse syntax, the project language is not considered to be a confounding variable.
For our scoring mechanism, initially, we inputted the grades received for the assignment prior to coding the indicator variables. However, after the first two classes were coded, we recognized that the grades might provide a potential source of error, so the grades were subsequently inputted for all future projects after the entire submission was scored. In an attempt to fully expell any bias, we confirmed the variable coding a second time while ignoring the grades received, and tough scoring situations were confirmed via collaboration between Mr. Feder and Dr. Çetinkaya-Rundel.
Since we could not find a significant confounding factor, we can attribute the distinctions in creativity, depth, and multivariate visualization scores to the R syntax. Additionally, since the final projects were randomly scored by course, and then later checked, we are confident that there was no bias in the coding of the indicator variables
We strongly encourage instructors to teach introductory R using the tidyverse syntax, as well as infer for inference tasks, as we believe that the tidyverse’s consistency encourages students to produce more creative and higher quality work while tightly adhering to the GAISE. We decided to create educational materials instead of attempting a form of modeling due to our desire to contribute to the current introductory statistics curriculum.
6.3 Future Work
We encourage others to build upon this analysis in an experimental form, since this study is solely retrospective. Although we have attempted to eliminate all potential sources of bias, a causal analysis can be much more unambiguous in a randomized trial. Perhaps similar to the study performed by (Myint et al., 2019), a randomized experiment conducted through an online educational company, instead focusing on distinct code differences, could be very effective.
In the future, it may be interesting to show tidyverse and base R code snippets accomplishing the same tasks to students with no programming background to determine whether one syntax is easier to initially read and understand. Also, as the infer package grows in popularity, the effectiveness of the infer package in completing inference tasks should be confirmed.