linkRecords

0th

Percentile

Bayes Estimates of Bipartite Matchings

Bayes point estimates of bipartite matchings that can be obtained in closed form according to Theorems 1, 2 and 3 of Sadinle (2017).

Usage
linkRecords(Zchain, n1, lFNM = 1, lFM1 = 1, lFM2 = 2, lR = Inf)
Arguments
Zchain

matrix as the output $Z of the function bipartiteGibbs, with n2 rows and nIter columns containing a chain of draws from a posterior distribution on bipartite matchings. Each column indicates the records in datafile 1 to which the records in datafile 2 are matched according to that draw.

n1

number of records in datafile 1.

lFNM

individual loss of a false non-match in the loss functions of Sadinle (2017), default lFNM=1.

lFM1

individual loss of a false match of type 1 in the loss functions of Sadinle (2017), default lFM1=1.

lFM2

individual loss of a false match of type 2 in the loss functions of Sadinle (2017), default lFM2=2.

lR

individual loss of 'rejecting' to make a decision in the loss functions of Sadinle (2017), default lR=Inf.

Details

Not all combinations of losses lFNM, lFM1, lFM2, lR are supported. The losses have to be positive numbers and satisfy one of three conditions:

  1. Conditions of Theorem 1 of Sadinle (2017): (lR == Inf) & (lFNM <= lFM1) & (lFNM + lFM1 <= lFM2)

  2. Conditions of Theorem 2 of Sadinle (2017): ((lFM2 >= lFM1) & (lFM1 >= 2*lR)) | ((lFM1 >= lFNM) & (lFM2 >= lFM1 + lFNM))

  3. Conditions of Theorem 3 of Sadinle (2017): (lFM2 >= lFM1) & (lFM1 >= 2*lR) & (lFNM >= 2*lR)

If one of the last two conditions is satisfied, the point estimate might be partial, meaning that there might be some records in datafile 2 for which the point estimate does not include a linkage decision. For combinations of losses not supported here, the linear sum assignment problem outlined by Sadinle (2017) needs to be solved.

Value

A vector containing the point estimate of the bipartite matching. If lR != Inf the output might be a partial estimate. A number smaller or equal to n1 in entry j indicates the record in datafile 1 to which record j in datafile 2 gets linked, a number n1+j indicates that record j does not get linked to any record in datafile 1, and the value -1 indicates a 'rejection' to link, meaning that the correct linkage decision is not clear.

References

Mauricio Sadinle (2017). Bayesian Estimation of Bipartite Matchings for Record Linkage. Journal of the American Statistical Association 112(518), 600-612. [Published] [arXiv]

Aliases
  • linkRecords
Examples
# NOT RUN {
data(twoFiles)

myCompData <- compareRecords(df1, df2, flds=c("gname", "fname", "age", "occup"), 
                             types=c("lv","lv","bi","bi"))

chain <- bipartiteGibbs(myCompData)

## discard first 100 iterations of Gibbs sampler

## full estimate of bipartite matching (full linkage)
fullZhat <- linkRecords(chain$Z[,-c(1:100)], n1=nrow(df1), lFNM=1, lFM1=1, lFM2=2, lR=Inf)

## partial estimate of bipartite matching (partial linkage), where 
## lR=0.5, lFNM=1, lFM1=1 mean that we consider not making a decision for 
## a record as being half as bad as a false non-match or a false match of type 1
partialZhat <- linkRecords(chain$Z[,-c(1:100)], n1=nrow(df1), lFNM=1, lFM1=1, lFM2=2, lR=.5)

## for which records the decision is not clear according to this set-up of the losses? 
undecided <- which(partialZhat == -1)
df2[undecided,]

## compute frequencies of link options observed in the chain 
linkOptions <- apply(chain$Z[undecided, -c(1:100)], 1, table)
linkOptions <- lapply(linkOptions, sort, decreasing=TRUE)
linkOptionsInds <- lapply(linkOptions, names)
linkOptionsInds <- lapply(linkOptionsInds, as.numeric)
linkOptionsFreqs <- lapply(linkOptions, function(x) as.numeric(x)/sum(x))

## first record without decision
df2[undecided[1],]

## options for this record; row of NAs indicates possibility that record has no match in df1
cbind(df1[linkOptionsInds[[1]],], prob = round(linkOptionsFreqs[[1]],3) )
# }
Documentation reproduced from package BRL, version 0.1.0, License: GPL-3

Community examples

Looks like there are no examples yet.