Statistical analysis construction: The mason R package

Written on August 4, 2016

Most analyses follow a similar pattern to how construction/engineering projects are developed: design -> add specifications -> construction -> (optional) add to the design and specs and continue construction -> cleaning, scrubbing, and polishing. I created the mason package to try to emulate this process and make it easier to do analyses in a consistent and ‘tidy’ format.

Basic command flow

The general command flow for using mason is:

1. Start the design of a blueprint for the analysis by specifying which statistical technique to use in your analysis (design()).
2. Add settings/options to the blueprint for the methods of the statistics (add_settings()).
3. Add the variables you want to run the statistics on (add_variables()). These variables include the $y$ variables (outcomes), the $x$ variables (predictors), covariates, and interaction variables.
4. Using the blueprint, construct the ‘mason project’ (stats analysis) so that the results are generated (construct()).
5. Sometimes analyses are too big for one first pass, from blueprint to construction, and needs to add more to the blueprint. Use add_variables() or add_settings() after the construct() to add to the existing results.
6. When you are ready, make the ‘mason project’ cleaned up by scrubbing it down and polishing it up (scrub() and polish_*() commands). The results are now ready for further presentation in a figure or table!

Example usage

Let’s go over an example analysis. We’ll use glm for a simple linear regression. Let’s use the built-in swiss dataset. A quick peek at it shows:

Ok, let’s say we want to several models. We are interested in Fertility and Infant.Mortality as outcomes and Education and Agriculture as potential predictors. We also want to control for Catholic. This setup means we have four potential models to analyze. With mason this is relatively easy. Analyses in mason are essentially separated into a blueprint phase and a construction phase. Since any structure or building always needs a blueprint, let’s get that started.

So far, all we’ve done is created a blueprint of the analysis, but it doesn’t contain much. Let’s add some settings to the blueprint. mason was designed to make use of the %>% pipes from the package magrittr (also found in dplyr), so let’s load up magrittr!

You’ll notice that each time, the only thing that is printed to the console is the dataset. That’s because we haven’t constructed the analysis yet! We are still in the blueprint phase, so nothing new has been added! Since we have two outcomes and two predictors, we have a total of four models to analysis. Normally we would need to run each of the models separately. However, if we simply list the outcomes and the predictors in mason, it will compute a model for each combination! Next, we can construct the analysis using construct().

Cool! This is the unadjusted model, without any covariates. We said we wanted to adjust for Catholic. But let’s say we want to keep the unadjusted analysis too. Since we have ‘finished’ the analysis by cleaning it up, we can still add to the blueprint.

We now have two models in the results. We’re happy with them, so let’s clean it up using the scrub() function.

All scrub() does is remove any extra specs in the attributes and sets the results as the main dataset. We can put it all as a single pipe chain:

There are also additional polish_* type commands that are more or less simply wrappers around commands that you may do on the results dataset, like filtering or renaming. The list of polish commands can be found in ?mason::polish.

Note: This is a slightly modified version of the introduction vignette for mason (vignette('mason') from version 0.2.5). I also created this package to make my analyses easier. So most of the statistics that have been implemented were chosen because that’s what I use. If you would like a particular statistical method included, please fill out an Issue and I will try to implement it!