DESIGN.md

In decisionpatterns/na.tools: Comprehensive Library for Working with Missing (NA) Values in Vectors

DESIGN.md - Design notes for na.actions pacakge

This document tracks opinionated desicion about the na.tools and tidyimpute packages that largely have to do with the design choices made

Goals

The packages shoould handle be the single repository for functions/methods for working with missing values (NA) for all data science workflows.

It should be extensible and be able to handle

• recursive and atomistic values
• types: categorical and continous variables
• classes: specific classes

Naming and Labels

• The explicit value for forcats::fct_explicit_na is (Missing); this is not adopted here because it is needlessly long, '(NA)' is used instead.
• beginning variables with

na.replace

Organization

• na.* works only on atomic vectors;
• na_* functions are for higher-level functions

Style Elements

Behaviors

• Coercion/replacement is type/class-safe; type should not be changes by imputation

• Function names should follow the lower_snake_case naming conventions. This prevents collisions with functions from the stats package. It may make sense for na.* functions operate at a low-level on vectors and similar to the stats package while na_* vectors operate on a higher level.

• Follow tidyverse styles

• Arguments should be lower_snake_case
• Arguments names should often be preceded by a '.' when following tidyverse styles
• NA_explicit_
• cannot be "NA" since it becomes impossible to distinguish it from true NAs when printing to the console.

Simple Imputation

 na.replace( .x, .na=mean )
 na_replace( .tbl, col1 = mean )          Or, 
 na_replace( .tbl, col1 = mean(col1) )

Complex Imputation

Imputation should be preformed when the replacement value is a rhs-formula:

na_replce(tbl, col1 = ~col2, .method=lm )

This has the effect of creating a model for col1 ~ col2

decisionpatterns/na.tools documentation built on May 25, 2019, 4:23 p.m.