Scientific Computing

• I propose to use R/RStudio enhanced with Rcpp/Eigen for scientific computing in applied microeconomics. There are three reasons for this choice.
• First, tidyverse, especially dplyr, seems to be currently the most efficient and elegant way for cleaning and transforming data.
  • Python's pandas is not at this level of elegance.
• Second, RStudio is a well-designed IDE.
  • I strongly recommend to develop a project as an R package, in which functions are built and loaded as if we use methods in installed packages. RStudio project is nicely integrated with this package development system.
  • Python + Jupyter, Python + vscode or Python + PyCharm may be an alternative. Unless there is a strong need for machine learning modules only available in Python, R/RStudio still dominates Python for applied microeconomics projects.
• Third, Rcpp offers an easy and elegant approach to integrate C/C++ functions.
  • This will be only relevant for those who conduct a heavy computational task.
  • Rcpp automatically generates a wrapper function to call a C/C++ function from R. With Rcpp, we are never bothered by the way to call C/C++ functions and the way to pass objects between R and C/C++ applications.
  • As I explain in the subseq5uent section, if we follow some best practices, functions written in C/C++ can be used almost in the same way as R functions.
  • Moreover, R objects can be directly passed to C/C++ as Eigen objects. Thus, if functions to be translated are written with vectors and matrices, they can be literally translated into C/C++ functions using Eigen methods.
  • I tried several libraries in Python of similar functionalities. But I did not find any of them worked as easy as Rcpp did.
  • Julia may be an alternative to speed up scientific computation. I am exploring this option. However, at this point, I do not choose Julia, because it does not have rich enough methods for data cleaning and transformation or statistical analysis.

Version Control

• I propose three version control configurations depending on the nature of the project.

1. git/GitHub only > Configuration 1
   • This is for a light project in which the codes can run in a few minutes from the data cleaning to the end of the analysis.
   1. Use git to version-control codes and use GitHub to host its remote repository.
   2. Collaborators share the codes through the GitHub remote repository.
   3. Raw data is hard copied among collaborators.
   4. The output of the codes are not shared across collaborators, but are generated locally by running the shared codes.
2. git/GitHub + AWS s3 > Configuration 2
   • This is a scalable configuration that works for a project of any size.
   1. Use git to version-control codes and use GitHub to host its remote repository.
   2. Collaborators share the codes through the GitHub remote repository.
   3. Directly read and write data in AWS s3 bucket.
   4. Collaborators share the data through the AWS s3 bucket.