fullmatch: Optimal full matching

In optmatch: Functions for Optimal Matching

Optimal full matching

Description

Given two groups, such as a treatment and a control group, and a method of creating a treatment-by-control discrepancy matrix indicating desirability and permissibility of potential matches (or optionally an already created such discrepancy matrix), create optimal full matches of members of the groups. Optionally, incorporate restrictions on matched sets' ratios of treatment to control units.

Usage

fullmatch(
  x,
  min.controls = 0,
  max.controls = Inf,
  omit.fraction = NULL,
  mean.controls = NULL,
  tol = 0.001,
  data = NULL,
  solver = "",
  ...
)

full(
  x,
  min.controls = 0,
  max.controls = Inf,
  omit.fraction = NULL,
  mean.controls = NULL,
  tol = 0.001,
  data = NULL,
  solver = "",
  ...
)

Arguments

x	Any valid input to match_on. fullmatch will use x and any optional arguments to generate a distance before performing the matching. If x is a numeric vector, there must also be passed a vector z indicating grouping. Both vectors must be named. Alternatively, a precomputed distance may be entered. A matrix of non-negative discrepancies, each indicating the permissibility and desirability of matching the unit corresponding to its row (a 'treatment') to the unit corresponding to its column (a 'control'); or, better, a distance specification as produced by match_on.
min.controls	The minimum ratio of controls to treatments that is to be permitted within a matched set: should be non-negative and finite. If min.controls is not a whole number, the reciprocal of a whole number, or zero, then it is rounded down to the nearest whole number or reciprocal of a whole number. When matching within subclasses (such as those created by exactMatch), min.controls may be a named numeric vector separately specifying the minimum permissible ratio of controls to treatments for each subclass. The names of this vector should include names of all subproblems distance.
max.controls	The maximum ratio of controls to treatments that is to be permitted within a matched set: should be positive and numeric. If max.controls is not a whole number, the reciprocal of a whole number, or Inf, then it is rounded up to the nearest whole number or reciprocal of a whole number. When matching within subclasses (such as those created by exactMatch), max.controls may be a named numeric vector separately specifying the maximum permissible ratio of controls to treatments in each subclass.
omit.fraction	Optionally, specify what fraction of controls or treated subjects are to be rejected. If omit.fraction is a positive fraction less than one, then fullmatch leaves up to that fraction of the control reservoir unmatched. If omit.fraction is a negative number greater than -1, then fullmatch leaves up to |omit.fraction| of the treated group unmatched. Positive values are only accepted if max.controls >= 1; negative values, only if min.controls <= 1. If neither omit.fraction or mean.controls are specified, then only those treated and control subjects without permissible matches among the control and treated subjects, respectively, are omitted. When matching within subclasses (such as those created by exactMatch), omit.fraction specifies the fraction of controls to be rejected in each subproblem, a parameter that can be made to differ by subclass by setting omit.fraction equal to a named numeric vector of fractions. At most one of mean.controls and omit.fraction can be non-NULL.
mean.controls	Optionally, specify the average number of controls per treatment to be matched. Must be no less than than min.controls and no greater than the either max.controls or the ratio of total number of controls versus total number of treated. Some controls will likely not be matched to ensure meeting this value. If neither omit.fraction or mean.controls are specified, then only those treated and control subjects without permissible matches among the control and treated subjects, respectively, are omitted. When matching within subclasses (such as those created by exactMatch), mean.controls specifies the average number of controls per treatment per subproblem, a parameter that can be made to differ by subclass by setting mean.controls equal to a named numeric vector. At most one of mean.controls and omit.fraction can be non-NULL.
tol	Because of internal rounding, fullmatch may solve a slightly different matching problem than the one specified, in which the match generated by fullmatch may not coincide with an optimal solution of the specified problem. tol times the number of subjects to be matched specifies the extent to which fullmatch's output is permitted to differ from an optimal solution to the original problem, as measured by the sum of discrepancies for all treatments and controls placed into the same matched sets.
data	Optional data.frame or vector to use to get order of the final matching factor. If a data.frame, the rownames are used. If a vector, the names are first tried, otherwise the contents is considered to be a character vector of names. Useful to pass if you want to combine a match (using, e.g., cbind) with the data that were used to generate it (for example, in a propensity score matching).
solver	Choose which solver to use. Currently implemented are RELAX-IV and LEMON. Default of "", a blank string, will use RELAX-IV if the rrelaxiv package is installed, otherwise will use LEMON. To explicitly use RELAX-IV, pass string "RELAX-IV". To use LEMON, pass string "LEMON". Optionally, to specify which algorithm LEMON will use, pass the function LEMON with argument for the algorithm name, "CycleCancelling", "CapacityScaling", "CostScaling", and "NetworkSimplex". See this site for details on their differences: https://lemon.cs.elte.hu/pub/doc/latest/a00606.html. CycleCancelling is the default. The CycleCancelling algorithm seems to produce results most closely resembling those of optmatch versions prior to 1.0. We have observed the other LEMON algorithms to produce different results when the mean.controls is unspecified, or specified in such a way as to produce an infeasible matching problem. When using a LEMON algorithm other than CycleCancelling, we recommend setting the fullmatch_try_recovery option to FALSE.
...	Additional arguments, passed to match_on (e.g. within) or to specific methods.

Details

If passing an already created discrepancy matrix, finite entries indicate permissible matches, with smaller discrepancies indicating more desirable matches. The matrix must have row and column names.

If it is desirable to create the discrepancies matrix beforehand (for example, if planning on running several different matching schemes), consider using match_on to generate the distances. This generic function has several useful methods for handling propensity score models, computing Mahalanobis distances (and other arbitrary distances), and using user supplied functions. These distances can also be combined with those generated by exactMatch and caliper to create very nuanced matching specifications.

The value of tol can have a substantial effect on computation time; with smaller values, computation takes longer. Not every tolerance can be met, and how small a tolerance is too small varies with the machine and with the details of the problem. If fullmatch can't guarantee that the tolerance is as small as the given value of argument tol, then matching proceeds but a warning is issued.

By default, fullmatch will attempt, if the given constraints are infeasible, to find a feasible problem using the same constraints. This will almost surely involve using a more restrictive omit.fraction or mean.controls. (This will never automatically omit treatment units.) Note that this does not guarantee that the returned match has the least possible number of omitted subjects, it only gives a match that is feasible within the given constraints. It may often be possible to loosen the omit.fraction or mean.controls constraint and still find a feasible match. The auto recovery is controlled by options("fullmatch_try_recovery").

In full matching problems permitting many-one matches (min.controls less than 1), the number of controls contributing to matches can exceed what was requested by setting a value of mean.controls or omit.fraction. I.e., in this setting mean.controls sets the minimum ratio of number of controls to number of treatments placed into matched sets.

If the program detects that (what it thinks is) a large problem, a warning is issued. Unless you have an older computer, there's a good chance that you can handle larger problems (at the cost of increased computation time). To check the large problem threshold, use getMaxProblemSize; to re-set it, use setMaxProblemSize.

Value

A optmatch object (factor) indicating matched groups.

References

Hansen, B.B. and Klopfer, S.O. (2006), ‘ Optimal full matching and related designs via network flows’, Journal of Computational and Graphical Statistics, 15, 609–627.

Hansen, B.B. (2004), ‘Full Matching in an Observational Study of Coaching for the SAT’, Journal of the American Statistical Association, 99, 609–618.

Rosenbaum, P. (1991), ‘A Characterization of Optimal Designs for Observational Studies’, Journal of the Royal Statistical Society, Series B, 53, 597–610.

Examples

data(nuclearplants)
### Full matching on a Mahalanobis distance.
( fm1 <- fullmatch(pr ~ t1 + t2, data = nuclearplants) )
summary(fm1)

### Full matching with restrictions.
( fm2 <- fullmatch(pr ~ t1 + t2, min.controls = .5, max.controls = 4, data = nuclearplants) )
summary(fm2)

### Full matching to half of available controls.
( fm3 <- fullmatch(pr ~ t1 + t2, omit.fraction = .5, data = nuclearplants) )
summary(fm3)

### Full matching attempts recovery when the initial restrictions are infeasible.
### Limiting max.controls = 1 allows use of only 10 of 22 controls.
( fm4 <- fullmatch(pr ~ t1 + t2, max.controls = 1, data=nuclearplants) )
summary(fm4)
### To recover restrictions
optmatch_restrictions(fm4)

### Full matching within a propensity score caliper.
ppty <- glm(pr ~ . - (pr + cost), family = binomial(), data = nuclearplants)
### Note that units without counterparts within the caliper are automatically dropped.
### For more complicated models, create a distance matrix and pass it to fullmatch.
mhd <- match_on(pr ~ t1 + t2, data = nuclearplants) + caliper(match_on(ppty), width = 1)
( fm5 <- fullmatch(mhd, data = nuclearplants) )
summary(fm5)

### Propensity balance assessment. Requires RItools package.
if (require(RItools)) summary(fm5,ppty)

### The order of the names in the match factor is the same
### as the nuclearplants data.frame since we used the data argument
### when calling fullmatch. The order would be unspecified otherwise.
cbind(nuclearplants, matches = fm5)

### Match in subgroups only. There are a few ways to specify this.
m1 <- fullmatch(pr ~ t1 + t2, data=nuclearplants,
                within=exactMatch(pr ~ pt, data=nuclearplants))
m2 <- fullmatch(pr ~ t1 + t2 + strata(pt), data=nuclearplants)
### Matching on propensity scores within matching in subgroups only:
m3 <- fullmatch(glm(pr ~ t1 + t2, data=nuclearplants, family=binomial),
                data=nuclearplants,
                within=exactMatch(pr ~ pt, data=nuclearplants))
m4 <- fullmatch(glm(pr ~ t1 + t2 + pt, data=nuclearplants,
                    family=binomial),
                data=nuclearplants,
                within=exactMatch(pr ~ pt, data=nuclearplants))
m5 <- fullmatch(glm(pr ~ t1 + t2 + strata(pt), data=nuclearplants,
                    family=binomial), data=nuclearplants)
# Including `strata(foo)` inside a glm uses `foo` in the model as
# well, so here m4 and m5 are equivalent. m3 differs in that it does
# not include `pt` in the glm.

optmatch documentation built on Sept. 19, 2024, 9:06 a.m.