Added becomes_ta() to determine the sample size when a unit enter the
take-all stratum.
Internal changes to the way classes are instantiated. No user-visible changes.
Updated maintainer email.
Added a vignette.
A single sample size is now recycled for each stratum when drawing samples or calculating inclusion probabilities.
Added a cutoff argument that puts units with sizes above the cutoff into a take-all stratum.
sps objects now inherit from numeric instead of integer.
Replacement methods for length and levels no longer mangle sps objects.
prop_allocation() gets a new argument for breaking ties. The default now breaks ties according to the references; the old behavior can be had by setting ties = "first".
Argument names for prop_allocation(), expected_coverage(), and sps_repweights() have changed to be either more descriptive, or consistent with the names for other functions.
Fixed a bug when calculating inclusion probabilities that could result in ties not breaking according to position (as documented) when alpha > 0.
Simplified the codebase.
Added a new argument alpha for calculating inclusion probabilities. It can be used to place units with inclusion probabilities close to 1 into the take-all stratum. This was implicitly 0 in previous versions, but the current default is 1e-4.
All functions can now accept size vectors with zeros.
sps() is now faster, especially for large populations.
Added a function factory to generate other order sampling methods.
This version has a number of non-backwards compatible changes to address undesirable behavior with some functions.
Largest-remainder rounding could result in prop_allocation() giving visibly non-proportional allocations, so it has been removed. The default is now the Jefferson/D'Hondt method.
The method argument in prop_allocation() has been replaced with the divisor argument to supply a divisor function for rounding.
The initial argument in prop_allocation() is recycled to ensure the initial allocation is feasible when supplying a single value. This change is not strictly backwards compatible as values that could not be ordinarily recycled no longer give an error.
Supplying a vector of permanent random numbers to sps() or ps() that are generated with a given seed now gives the same result when setting that seed prior to calling sps() or ps(). This means that setting the seed to a given value can give a different sample compared to older versions, although permanent random numbers should be used for reproducible samples.
The argument name for specifying strata is now strata instead of s in all functions. Partial matching means this change won't break existing code.
prop_allocation() gains two new arguments:
initial sets the initial allocation for each stratum. This saves from having to manually keep track of an initial allocation and adjust the sample size passed to prop_allocation().
method selects the apportionment method used to round a proportional allocation to integer values. In particular, highest-averages methods can be used in place of the largest-remainder method.
The defaults for these arguments do not change the behavior of the function from previous versions.
A few convenience functions have been added:
ps() for drawing ordinary Poisson samples with the same interface as sps().
inclusion_prob() for generating inclusion probabilities for a frame.
expected_coverage() for calculating the expected number of strata when sampling from a frame.
Some of the internals have been updated to improve performance with large frames.
Most functions do a little more argument checking. In particular, it is no longer possible to pass length 0 arguments to sps(), prop_allocation(), or sps_repweights().
Added the option to draw samples with permanent random numbers.
Added methods for Math, Ops, [<-, and [[<- that strip attributes from sps objects so as not to treat the result as a sample.
sps() now returns an integer vector (when possible), rather than a double.
printing the matrix of replicate weights now shows row names (if any).