Beta record linkage methodology for probabilistic bipartite record linkage: the task of merging two duplicate-free
datafiles that lack unique identifiers.
This function runs all the steps of beta record linkage: creates comparisons of the records, runs Gibbs sampler on bipartite matchings,
and derives point estimate of bipartite matching (this determines the final linkage). The parameters of BRL consist of
all the parameters needed to run compareRecords, bipartiteGibbs and linkRecords,
except for intermediate input/output, and in addition to a parameter burn for the burn-in period of the Gibbs sampler.
BRL(
df1,
df2,
flds = NULL,
flds1 = NULL,
flds2 = NULL,
types = NULL,
breaks = c(0, 0.25, 0.5),
nIter = 1000,
burn = round(nIter * 0.1),
a = 1,
b = 1,
aBM = 1,
bBM = 1,
seed = 0,
lFNM = 1,
lFM1 = 1,
lFM2 = 2,
lR = Inf
)two datasets to be linked, of class data.frame, with rows representing records and columns
representing fields. Without loss of generality,
df1 is assumed to have no less records than df2.
a vector indicating the fields to be used in the linkage. Either a character vector, in which case
all entries need to be names of columns of df1 and df2, or a numeric vector
indicating the columns in df1 and df2 to be used in the linkage. If provided as a
numeric vector it is assumed that the columns of df1 and df2 are organized such that
it makes sense to compare the columns
df1[,flds] and df2[,flds] in that order.
vectors indicating the fields of df1 and df2 to be used in the linkage.
Either character vectors, in which case
all entries need to be names of columns of df1 and df2, respectively, or numeric vectors
indicating the columns in df1 and df2 to be used in the linkage. It is assumed that
it makes sense to compare the columns
df1[,flds1] and df2[,flds2] in that order. These arguments are ignored if flds is specified.
If none of flds,flds1,flds2 are specified, the columns with the same names in df1 and df2
are compared, if any.
a vector of characters indicating the comparison type per comparison field. The options
are: "lv" for comparisons based on the Levenshtein edit distance normalized to \([0,1]\), with \(0\)
indicating no disagreement and \(1\) indicating maximum disagreement;
"bi" for binary comparisons (agreement/disagreement); "nu" for numeric comparisons computed as
the absolute difference.
The default is "lv". Fields compared with the "lv" option are first transformed to character
class. Factors with different levels compared using the "bi" option are transformed to factors with the union
of the levels. Fields compared with the "nu" option need to be of class numeric.
break points for the comparisons to obtain levels of disagreement.
It can be a list of length equal to the number of comparison fields, containing one numeric vector with the break
points for each comparison field, where entries corresponding to comparison type "bi" are ignored.
It can also be a named list of length two with elements 'lv' and 'nu'
containing numeric vectors with the break
points for all Levenshtein-based and numeric comparisons, respectively.
Finally, it can be a numeric vector with the break points for all comparison fields of type "lv" and "nu",
which might be meaningful only if all the non-binary comparisons are of a single type, either "lv" or "nu".
For comparisons based on the normalized Levenshtein distance, a vector of length \(L\) of break
points for the interval \([0,1]\) leads to \(L+1\) levels of disagreement. Similarly, for comparisons based on the absolute
difference, the break points are for the interval \([0,\infty)\).
The default is breaks=c(0,.25,.5), which might be meaningful only for comparisons of type "lv".
number of iterations of Gibbs sampler.
number of iterations to discard as part of the burn-in period.
hyper-parameters of the Dirichlet priors for the \(m\) and \(u\) parameters in the model for the comparison data among matches and non-matches, respectively. These can be vectors with as many entries as disagreement levels among all comparison fields. If specified as positive constants, they get recycled to the required length. If not specified, flat priors are taken.
hyper-parameters of beta prior on bipartite matchings. Default is aBM=bBM=1.
seed to be used for pseudo-random number generation. By default it sets seed=0.
individual loss of a false non-match in the loss functions of Sadinle (2017), default lFNM=1.
individual loss of a false match of type 1 in the loss functions of Sadinle (2017), default lFM1=1.
individual loss of a false match of type 2 in the loss functions of Sadinle (2017), default lFM2=2.
individual loss of 'rejecting' to make a decision in the loss functions of Sadinle (2017), default lR=Inf.
A vector containing the point estimate of the bipartite matching, as in the output of linkRecords. 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.
Beta record linkage (BRL, Sadinle, 2017) is a methodology for probabilistic bipartite record linkage, that is, the task of merging two duplicate-free datafiles that lack unique identifiers. This is accomplished by using the common partially identifying information of the entities contained in the datafiles. The duplicate-free requirement means that we expect each entity to be represented maximum once in each datafile. This methodology should not be used with datafiles that contain duplicates nor should it be used for deduplicating a single datafile.
The first step of BRL, accomplished by the function compareRecords, consists of constructing comparison vectors for each pair of records from the two datafiles.
The current implementation allows binary comparisons (agree/disagree), numerical comparisons based on the absolute difference,
and Levenshtein-based comparisons.
This can be easily extended to other comparison types, so a resourceful user should be able to construct an object that recreates
the output of compareRecords for other types of comparisons (so long as they get transformed to levels of disagreement), and still be able to run the next step outside
the function BRL.
The second step of BRL, accomplished by the function bipartiteGibbs, consists of running a Gibbs sampler that explores the space of bipartite matchings
representing the plausible ways of linking the datafiles. This sampler is derived from a model for the comparison data and a beta prior
distribution on the space of bipartite matchings. See Sadinle (2017) for details.
The third step of BRL, accomplished by the function linkRecords, consists of deriving a point estimate of the bipartite matching
(which gives us the optimal way of linking the datafiles)
by minimizing the expected value of
a loss function that uses different penalties for different types of linkage errors. The current implementation only supports the
Bayes point estimates of bipartite matchings that can be obtained in closed form according to Theorems 1, 2 and 3 of Sadinle (2017).
The losses have to be positive numbers and satisfy one of three conditions:
Conditions of Theorem 1 of Sadinle (2017):
(lR == Inf) & (lFNM <= lFM1) & (lFNM + lFM1 <= lFM2)
Conditions of Theorem 2 of Sadinle (2017):
((lFM2 >= lFM1) & (lFM1 >= 2*lR)) | ((lFM1 >= lFNM) & (lFM2 >= lFM1 + lFNM))
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.
Mauricio Sadinle (2017). Bayesian Estimation of Bipartite Matchings for Record Linkage. Journal of the American Statistical Association 112(518), 600-612. [Published] [arXiv]
compareRecords for examples on how to work with different types of comparison data,
bipartiteGibbs for Gibbs sampler on bipartite matchings, and linkRecords for examples
on full and partial point estimates of the true bipartite matching that indicates which records to link.
# NOT RUN {
data(twoFiles)
(Zhat <- BRL(df1, df2, flds=c("gname", "fname", "age", "occup"),
types=c("lv","lv","bi","bi")))
n1 <- nrow(df1)
Ztrue <- df2ID
## number of links (estimated matches)
nLinks <- sum(Zhat <= n1)
## number of actual matches according to the ground truth
nMatches <- sum(Ztrue <= n1)
## number of links that are actual matches
nCorrectLinks <- sum(Zhat[Zhat<=n1]==Ztrue[Zhat<=n1])
## compute measures of performance
## precision
nCorrectLinks/nLinks
## recall
nCorrectLinks/nMatches
## the linked record pairs
cbind( df1[Zhat[Zhat<=n1],], df2[Zhat<=n1,] )
## finally, note that we could run BRL step by step as follows
## create comparison data
myCompData <- compareRecords(df1, df2,
flds=c("gname", "fname", "age", "occup"),
types=c("lv","lv","bi","bi"))
## Gibbs sampling from posterior of bipartite matchings
chain <- bipartiteGibbs(myCompData)
## bipartite matching Bayes estimate derived from the loss functions of Sadinle (2017)
Zhat2 <- linkRecords(chain$Z, n1=n1)
identical(Zhat, Zhat2)
# }
Run the code above in your browser using DataCamp Workspace