Sampling Based Clerical Review Methods in Probabilistic Matching

1
Sampling Based Clerical Review Methods in
Probabilistic Matching
Sampling Based Clerical Review Methods in
Probabilistic Matching
2
Clerical Review
File A
Automatically assign as a non-link
File B
Send for manual clerical review



Automatically assign as a link
3
Output of Record Pair Comparisons

• Set of matched records with an associated
  comparison weight
• Lots of these, high weights, low weights and
  in-between

4
Frequency
Comparison Weight
5
... but

• Clerical review can be time consuming
• Thousands or tens of thousands of clerical review
  pairs
• High level of repetitive VDU based tasks can lead
  to health and safety issues

6
Frequency
Comparison Weight
7
Acceptance Sampling

• Allows quantification of uncertainty in sampling
• Methods
• AS 1199 Sampling Procedures and Tables for
  Inspection by Attributes
• DIY calculations

proc power onesamplemeans mean
5 10 ntotal 150 stddev
30 50 power . plot xn min100
max200 run
8
(No Transcript)
9
Producer's Risk (?), risk of having to review a
batch with a large number of number of non-matches
AQL Match rate for automatic rejection
10
RQL Match rate that is unacceptable to
automatically accept as non-links
11
P(send for manual review)?
100
80
60
40
20
0
0
20
40
60
80
100
Acual Quality Level
12
(No Transcript)
13
Setting a Single Cut-off

• Sometimes there are not enough fields to do
  meaningful clerical review
• Particularly, when we are not using names and
  addresses
• In these cases we want to meaningfully set a
  single cut off

14
Estimated Cumulative Matches Linked
Estimated Cumulative Non - Matches Linked
Comparison Weight
15
Case Study

• Migrants Settlement Database to Census 2006
  linkage
• 131,000 records identified for clerical review
• Sampling scheme was 50 batches from which 65
  records pairs was selected
• Only 39 batches were actually inspected for a
  total of 2,535

16
Final Remarks

• Sampling based clerical review can very
  significantly reduce the amount of clerical
  review
• Can be used to rigorously set up cut-offs
• Provide information on linkage quality
• Can introduce missed and false links, but the
  extent of these can be estimated

17
Thank you for your attention
Any Questions?