dbglm: Fast generalized linear model in a database

In dbglm: Generalised Linear Models by Subsampling and One-Step Polishing

Description

Fast generalized linear model in a database

Usage

dbglm(formula, family = binomial(), tbl, sd = FALSE,
weights = .NotYetImplemented(), subset = .NotYetImplemented(), ...)

Arguments

...	This argument is required for S3 method extension.
formula	A model formula. It can have interactions but cannot have any transformations except factor
family	Model family
tbl	An object inheriting from tbl. Will typically be a database-backed lazy tbl from the dbplyr package.
sd	Experimental: compute the standard deviation of the score as well as the mean in the update and use it to improve the information matrix estimate
weights	We don't support weights
subset	If you want to analyze a subset, use filter() on the data

Details

For a dataset of size N the subsample is of size N^(5/9). Unless N is large the approximation won't be very good. Also, with small N it's quite likely that, eg, some factor levels will be missing in the subsample.

Value

A list with elements

tildebeta	coefficients from subsample
hatbeta	final estimate
tildeV	variance matrix from subsample
hatV	final estimate

References

http://notstatschat.tumblr.com/post/171570186286/faster-generalised-linear-models-in-largeish-data

dbglm documentation built on June 23, 2021, 9:07 a.m.