Author: Pete Mohanty Date: 10 October 2018
This document discusses possibilities and obstacles to implementing DeclareDesign in Stata in light of two prototypes, both of which are implementations of a simple two-arm design. The first version, hereafter referred to as SO (as in, the Stata-only version), makes use of simulate. The second version, hereafter MW (as in, the Mata-wrapper version), creates a class for design objects and also does core calculations in Mata.
Original R code
SO: detailed description of Stata-only
MW: detailed description of Mata-wrapper
MW comes (or could come) much closer to the corresponding declare R packages in terms of functionality. Importantly, MW allows users to declare aspects of designs in various sequences without simulating or estimating. MW still has serious limits in terms of customizability--in particular, it is not possible to add 'unanticipated' variables to the design class in 'interactive' mode. The more important limitation of MW is from the standpoint of user-friendliness. Though variables can be initialized with one line of code (and the program prompts users with examples), users unfamiliar with MATA or scripting more generally may not have a sense of what's being accomplished.
Of the two, MW has the potential to be faster, both because MATA is a lower level language and because re-implementing common estimators would give one the opportunity to skip calculations which are superfluous to the design. Those speed savings would be hardwon, however, in terms of time spent coding, and because such optimization may make user input more complicated still.
SO, by contrast, is more user friendly. Though MW could be adapted to execute Stata regression commands, SO would do so in a way that's more familar and less error-prone. (That's because Mata executes Stata commands as strings which, like eval(parse()) in R, are more difficult to debug than regular code.) SO (in particular twoarmX.ado, which shows how to incorporate covariates from existing data sets) does go beyond existing Stata functionality for research design. But SO collapses key steps of declaring and diagnosing, limiting the value-added for users with complex or atypical research designs.