“Wait, I’m a developer?”: Reflections on developing #Rstats functions
After logging off a two-hour meeting I immediately paused and thought “wait, I’m a developer?”
The meeting was part of the IES-funded Quantifying the Robustness of Causal Inferences project where we discussed the redesign and architecture of the konfound R package. This (kinda) big task came to light when we decided to visually map out the R functions in the package to understand the structure and how we could add in new functions for newly developed statistical methods. And well, it ended up looking a little interesting (figure 1).
We all decided that we wanted to work smarter 💡, not harder so we decided to brainstorm ways into which we could better structure the functions within the `konfound’ package in a way we could easily incorporate functions for newly developed statistical methods. In other words, we wanted a stable back-end structure to add new functions until the end of time 😬
“We built this plane while flying it” - J. Rosenberg, 2022 ✈️
Konfound + THE Question
As of right now, the three main functions in the
konfound package (
mkonfound) mainly support linear models 📈 (some non-linear) for two causal inference sensitivity analysis techniques (omitted variables and replacement of cases). But we have plans for incorporating more techniques that support linear and non-linear models. So the question was “how do we restructure the architecture of the functions to support future development on the back-end and provide the best experience for the user on the front-end?” Womp womp womp. Wait, I’m a developer? 😱
Here are a few (well, let’s be real, 7) things to consider when answering this question:
- The purpose driving the development of your main functions will likely dictate the overarching structure of your package. For example, the purpose of the development of the
konfoundpackage is to provide users with an easy use function to conduct causal inference sensitivity analysis. So it might make sense to set these techniques at the top of the hierarchy structure then build down from there.
- How many arguments is too many arguments? How many arguments should a user specify in order to run the analysis? How many arguments could a user specify to refine the analysis? Which arguments should be default?
- How many functions are too many functions? Should there be a function based on model type (linear or non), technique type, context type (established model, results from already-conducted analysis, meta-analysis) or a combination of these? Here, there can be as few as 3 main functions and as many as❗24 main functions❗ 😳 (calculated by using the rule of product for our future development plans).
- What should be written in error messages? Should there be warning messages? If so, what should be included in these messages?
- Probably the most important consideration, what are the user’s needs? How will they use the functions vs. how we think they will use them? How can we anticipate their needs to minimize user error and increase user experience? ✨(this is the #1 UX principle)✨
- Is there a middle ground between creating a strong back-end structure and a strong front-end structure?
- Will changing the current structure of the functions within the package confuse or drive away current users of the package? If we decide to change the structure in a way that is different from the current structure, how can this be executed in a way that is least burdensome for current users? (the dreaded ⚠️deprecated⚠️ error message) 😰
Final Words: Truly Yours, TBD
As of writing this, I don’t have any immediate answers to these questions. I think it depends on the end goal of your package and functions. I personally have a preference in how we should proceed in our case of the
konfound project, but definitely need some time to throw spaghetti at the wall, so-to-speak.
But, I wanted to share these thoughts and reflections as someone who unknowingly entered the world of #Rstats development.
Keep Tabs on the Project
Stay tuned for updates that will be posted on my blog. You can also follow the konfound project at https://www.konfound-it.org/