fit_inspre_from_h5X: Fits inverse sparse regression model.

In brielin/inspre: INverse SParse REgression

Fits inverse sparse regression model.

Description

See also inspre::inspre() for more details.

Usage

fit_inspre_from_h5X(
  X,
  X_control,
  X_ids,
  X_vars,
  targets,
  weighted = TRUE,
  max_med_ratio = NULL,
  filter = TRUE,
  rho = 100,
  lambda = NULL,
  lambda_min_ratio = 0.01,
  nlambda = 20,
  alpha = 0,
  gamma = NULL,
  its = 100,
  delta_target = 1e-04,
  verbose = 2,
  cv_folds = 0,
  mu = 10,
  tau = 1.5,
  solve_its = 10,
  ncores = 1,
  min_nz = 0.01,
  warm_start = FALSE,
  constraint = "UV",
  DAG = FALSE
)

Arguments

X	H5D. Example: ‘X=hfile[[’X']]' where hfile is an hdf5r object containing a data matrix stored under label 'X'.
X_control	Matrix. features by control observations. Usually pulled from X but stored in memory so you don't have to pull the same control observations off of disk for every instrument.
X_ids	Sequence of strings with length equal to number of columns in X. Entries correspond to the instrument applied to that entry.
X_vars	Sequence of strings with length equal to the number of rows in X. Name of the feature measured in that row.
targets	List. Entries correspond to variables in X ('X_vars'), names correspond to the instrument targeting that variable ('X_ids'). These targets will be used to calculate causal effect sizes.
max_med_ratio	Float or NULL. Passed through to 'make_weights', NULL for no weights.
filter	Bool. True to filter the produced TCE matrix with 'fitler_tce'.
rho	Float. Initial learning rate for ADMM.
lambda	Float, sequence of floats of NULL. L1 regularization strength on inverse of X. If NULL, a logarithmicly spaced set of values between the maximum absolute off diagonal element of X and lambda_min_ratio times this value will be used.
lambda_min_ratio	Float, ratio of maximum lambda to minimum lambda.
nlambda	Integer. Number of lambda values to try.
alpha	Float between 0 and 1 or NULL. If > 0, the model will be fit once with gamma = 0 to find L0, then all subsequent fits will use gamma = alpha * L0 / D. Set to NULL to provide gamma directly.
gamma	Float or sequence of nlambda floats or NULL. Determinant regularization strength to use (for each lambda value). It is recommended to set alpha rather than setting this directly.
its	Integer. Maximum number of iterations.
delta_target	Float. Target change in solution.
verbose	0, 1 or 2. 2 to print convergence progress for each lambda, 1 to print convergence result for each lambda, 0 for no output.
cv_folds	Integer. Number of cross-validation folds to perform.
mu	rho modification parameter for ADMM. Rho will be increased/decreased when the dual constrant and primal constraint are off by a factor of > mu.
tau	rho modification parameter for ADMM. When called for, rho will be increased/decreased by the factor tau.
solve_its	Integer, number of iterations of bicgstab/lasso to run for each U and V update.
ncores	Integer, number of cores to use.
warm_start	Logical. Whether to use previous lambda value result as starting point for next fit.
constraint	One of "UV" or "VU". Constraint to use.
DAG	Bool. True to resitrict solutions to approximate DAGs.
vars_to_use	Sequence of strings. Entries in X_vars to keep. Default NULL to keep all.
obs_to_use	Sequence of bools. Indicator of columns of X to use in calculations. Useful for cross validation.
train_prop	Float between 0 and 1. Proportion of data to use for training in cross-validation. NOT USED.

brielin/inspre documentation built on Dec. 3, 2023, 4:55 a.m.