R/README.md

In gaow/mmbr: Multivariate Sum of Single Effects Regression

Data

• regression_data.R implements several classes for different input data types and relevant algebraic operations on them.

Model

Implementation of SuSiE model falls roughly in the structure of the SuSiE manuscript. That is, we introduce Bayesian regression model, followed by Single Effect Regression (SER), and finally the SuSiE model. Implementation-wise,

1. *_regression.R implements several classes for Bayesian regression models using different priors.
   • A conventional univariate multiple regression with normal prior.
   • A multivariate regression with fixed multivariate normal prior (prototype not optimized for production).
   • A multivariate regression with fixed multivariate normal mixture prior (the MASH model).
2. single_effect_model.R implements the SER model. It inherits Bayesian regression models.
3. ibss_algorithm.R implements the IBSS algorithm for SuSiE model.

Misc

• mvsusie.R implements interface mvsusie() function.
• utils.R contains some utility functions.

A note on R6 class

R6 is pretty easy to learn just by the length of its documentation.

A couple of quick pointers on R6 class if you don't want to bother reading its documentation:

1. Convention private$<name> refers to private member (ie, variable) or method (ie, function); convention self$<name> refers to public member or method, or, active bindings. Please be careful which to use.
2. If you've got a class A to set B via B=A then modify B, you will also be modifying A! To prevent this from happening you've got to use this: B=A$clone(deep=TRUE).
3. If class A2 is inherited from A1, and suppose a is an instance of A2, then both inherits(a, "A2") and inherits(a, "A1") will be true. If a is instance of A1 then only the latter is true.
4. private cannot have same name as public and active.
5. Using private$ and self$ can have some overhead which can be substential for some applications. I therefore use portable=FALSE and use <<- for class member assignment.

I use the following convention in my code:

1. There is no public member variable. All variables are private.
2. For variables meant to be used outside class I expose them to active.
   • By default active bindings for private variables are read-only, appear as function(); an attempt to assign values to them will raise an error.
   • Some active bindings will allow setting values of variables, and appear as function(v).
   • Private members follows the convention of .<name> and the corresponding exposed variable will be <name> if exposed to active.

Otherwise, the only downside as I found using R6 is that the trace back for errors is not as clear as without using it. But R language in general doesn't have good error trace back (compared to Python which can point to you the exact lines of codes that are problematic), so R6 is not making it that much worse.

gaow/mmbr documentation built on April 24, 2024, 7:12 p.m.