compute_objective_val

Returns the objective value given a design vector as well an an objective function.
This is sometimes duplicated in Java. However, within Java, tricks are played to make
optimization go faster so Java's objective values may not always be the same as the true
objective function (e.g. logs or constants dropped).

Computes experimental designs for a
two-arm experiment with covariates via a number of methods:
(0) complete randomization and randomization with forced-balance,
(1) Greedily optimizing a
balance objective function via pairwise switching. This optimization
provides lower variance for the treatment effect estimator (and higher
power) while preserving a design that is close to complete randomization.
We return all iterations of the designs for use in a permutation test,
(2) The second is via numerical optimization
(via 'gurobi' which must be installed, see <https://www.gurobi.com/documentation/9.1/quickstart_windows/r_ins_the_r_package.html>)
a la Bertsimas and Kallus,
(3) rerandomization,
(4) Karp's method for one covariate,
(5) exhaustive enumeration to find the
optimal solution (only for small sample sizes),
(6) Binary pair matching using the 'nbpMatching' library,
(7) Binary pair matching plus design number (1) to further optimize balance,
(8) Binary pair matching plus design number (3) to further optimize balance,
(9) Hadamard designs,
(10) Simultaneous Multiple Kernels.
In (1-9) we allow for three objective functions:
Mahalanobis distance,
Sum of absolute differences standardized and
Kernel distances via the 'kernlab' library. This package is the result of a stream of research that can be found in
Krieger, A, Azriel, D and Kapelner, A "Nearly Random Designs with Greatly Improved Balance" (2016) <arXiv:1612.02315>,
Krieger, A, Azriel, D and Kapelner, A "Better Experimental Design by Hybridizing Binary Matching with Imbalance
Optimization" (2021) <arXiv:2012.03330>.

Adam Kapelner

GreedyExperimentalDesign

Greedy Experimental Design Construction

David Azriel

Abba Krieger

compute_objective_val function

<dl><dt>X</dt>
<dd>The n x p design matrix</dd>
<dt>indic_T</dt>
<dd>The n-length binary allocation vector</dd>
<dt>objective</dt>
<dd>The objective function to use. Default is <code>abs_sum_diff</code> and the other option is 
<code>mahal_dist</code>.</dd>
<dt>inv_cov_X</dt>
<dd>Optional: the inverse sample variance covariance matrix. Use this
argument if you will be doing many calculations since passing this
in will cache this data.</dd></dl>

Arguments

Author

Computes Objective Value From Allocation Vector — compute_objective_val

<dl>

<dt>X</dt>
<dd>The n x p design matrix</dd>


<dt>indic_T</dt>
<dd>The n-length binary allocation vector</dd>


<dt>objective</dt>
<dd>The objective function to use. Default is <code>abs_sum_diff</code> and the other option is 
<code>mahal_dist</code>.</dd>


<dt>inv_cov_X</dt>
<dd>Optional: the inverse sample variance covariance matrix. Use this
argument if you will be doing many calculations since passing this
in will cache this data.</dd>

</dl>

compute_objective_val: Computes Objective Value From Allocation Vector

Description

Usage

Arguments

Author