Finding Needles in Haystacks: Multiple-Imputation Record Linkage Using Machine Learning
November 2021
Working Paper Number:
CES-21-35
Abstract
This paper considers the problem of record linkage between a household-level survey and an establishment-level frame in the absence of unique identifiers. Linkage between frames in this setting is challenging because the distribution of employment across establishments is highly skewed. To address these difficulties, this paper develops a probabilistic record linkage methodology that combines machine learning (ML) with multiple imputation (MI). This ML-MI methodology is applied to link survey respondents in the Health and Retirement Study to their workplaces in the Census Business Register. The linked data reveal new evidence that non-sampling errors in household survey data are correlated with respondents' workplace characteristics.
Document Tags and Keywords
Keywords
Keywords are automatically generated using KeyBERT, a powerful and innovative
keyword extraction tool that utilizes BERT embeddings to ensure high-quality and contextually relevant
keywords.
By analyzing the content of working papers, KeyBERT identifies terms and phrases that capture the essence of the
text, highlighting the most significant topics and trends. This approach not only enhances searchability but
provides connections that go beyond potentially domain-specific author-defined keywords.
:
estimating,
data,
data census,
microdata,
census research,
respondent,
survey,
employed,
employee,
imputation,
workforce,
record,
household,
census business,
census bureau,
household survey,
sampling,
sample,
census employment,
datasets,
imputation model,
linked census,
linkage
Tags
Tags are automatically generated using a pretrained language model from spaCy, which excels at
several tasks, including entity tagging.
The model is able to label words and phrases by part-of-speech,
including "organizations." By filtering for frequent words and phrases labeled as "organizations", papers are
identified to contain references to specific institutions, datasets, and other organizations.
:
Social Security Administration,
Ordinary Least Squares,
Employer Identification Number,
Department of Economics,
Federal Reserve System,
Survey of Income and Program Participation,
Board of Governors,
Health and Retirement Study,
Longitudinal Employer Household Dynamics,
Alfred P Sloan Foundation,
National Institute on Aging,
Census Bureau Business Register,
Quarterly Workforce Indicators,
Quarterly Census of Employment and Wages,
University of Michigan,
Census Bureau Disclosure Review Board,
Federal Statistical Research Data Center
Similar Working Papers
Similarity between working papers are determined by an unsupervised neural
network model
know as Doc2Vec.
Doc2Vec is a model that represents entire documents as fixed-length vectors, allowing for the
capture of semantic meaning in a way that relates to the context of words within the document. The model learns to
associate a unique vector with each document while simultaneously learning word vectors, enabling tasks such as
document classification, clustering, and similarity detection by preserving the order and structure of words. The
document vectors are compared using cosine similarity/distance to determine the most similar working papers.
Papers identified with 🔥 are in the top 20% of similarity.
The 10 most similar working papers to the working paper 'Finding Needles in Haystacks: Multiple-Imputation Record Linkage Using Machine Learning' are listed below in order of similarity.
-
Working PaperOptimal Probabilistic Record Linkage: Best Practice for Linking Employers in Survey and Administrative Data🔥
March 2019
Working Paper Number:
CES-19-08
This paper illustrates an application of record linkage between a household-level survey and an establishment-level frame in the absence of unique identifiers. Linkage between frames in this setting is challenging because the distribution of employment across firms is highly asymmetric. To address these difficulties, this paper uses a supervised machine learning model to probabilistically link survey respondents in the Health and Retirement Study (HRS) with employers and establishments in the Census Business Register (BR) to create a new data source which we call the CenHRS. Multiple imputation is used to propagate uncertainty from the linkage step into subsequent analyses of the linked data. The linked data reveal new evidence that survey respondents' misreporting and selective nonresponse about employer characteristics are systematically correlated with wages.View Full Paper PDF
-
Working PaperMISCLASSIFICATION IN BINARY CHOICE MODELS
May 2013
Working Paper Number:
CES-13-27
We derive the asymptotic bias from misclassification of the dependent variable in binary choice models. Measurement error is necessarily non-classical in this case, which leads to bias in linear and non-linear models even if only the dependent variable is mismeasured. A Monte Carlo study and an application to food stamp receipt show that the bias formulas are useful to analyze the sensitivity of substantive conclusions, to interpret biased coefficients and imply features of the estimates that are robust to misclassification. Using administrative records linked to survey data as validation data, we examine estimators that are consistent under misclassification. They can improve estimates if their assumptions hold, but can aggravate the problem if the assumptions are invalid. The estimators differ in their robustness to such violations, which can be improved by incorporating additional information. We propose tests for the presence and nature of misclassification that can help to choose an estimator.View Full Paper PDF
-
Working PaperTwo Perspectives on Commuting: A Comparison of Home to Work Flows Across Job-Linked Survey and Administrative Files
January 2017
Working Paper Number:
CES-17-34
Commuting flows and workplace employment data have a wide constituency of users including urban and regional planners, social science and transportation researchers, and businesses. The U.S. Census Bureau releases two, national data products that give the magnitude and characteristics of home to work flows. The American Community Survey (ACS) tabulates households' responses on employment, workplace, and commuting behavior. The Longitudinal Employer-Household Dynamics (LEHD) program tabulates administrative records on jobs in the LEHD Origin-Destination Employment Statistics (LODES). Design differences across the datasets lead to divergence in a comparable statistic: county-to-county aggregate commute flows. To understand differences in the public use data, this study compares ACS and LEHD source files, using identifying information and probabilistic matching to join person and job records. In our assessment, we compare commuting statistics for job frames linked on person, employment status, employer, and workplace and we identify person and job characteristics as well as design features of the data frames that explain aggregate differences. We find a lower rate of within-county commuting and farther commutes in LODES. We attribute these greater distances to differences in workplace reporting and to uncertainty of establishment assignments in LEHD for workers at multi-unit employers. Minor contributing factors include differences in residence location and ACS workplace edits. The results of this analysis and the data infrastructure developed will support further work to understand and enhance commuting statistics in both datasets.View Full Paper PDF
-
Working PaperThe LEHD Infrastructure Files and the Creation of the Quarterly Workforce Indicators
January 2006
Working Paper Number:
tp-2006-01
The Longitudinal Employer-Household Dynamics (LEHD) Program at the U.S. Census Bureau, with the support of several national research agencies, has built a set of infrastructure files using administrative data provided by state agencies, enhanced with information from other administrative data sources, demographic and economic (business) surveys and censuses. The LEHD Infrastructure Files provide a detailed and comprehensive picture of workers, employers, and their interaction in the U.S. economy. Beginning in 2003 and building on this infrastructure, the Census Bureau has published the Quarterly Workforce Indicators (QWI), a new collection of data series that offers unprecedented detail on the local dynamics of labor markets. Despite the fine detail, confidentiality is maintained due to the application of state-of-the-art confidentiality protection methods. This article describes how the input files are compiled and combined to create the infrastructure files. We describe the multiple imputation methods used to impute in missing data and the statistical matching techniques used to combine and edit data when a direct identifier match requires improvement. Both of these innovations are crucial to the success of the final product. Finally, we pay special attention to the details of the confidentiality protection system used to protect the identity and micro data values of the underlying entities used to form the published estimates. We provide a brief description of public-use and restricted-access data files with pointers to further documentation for researchers interested in using these data.View Full Paper PDF
-
Working PaperSqueezing More Out of Your Data: Business Record Linkage with Python
November 2018
Working Paper Number:
CES-18-46
Integrating data from different sources has become a fundamental component of modern data analytics. Record linkage methods represent an important class of tools for accomplishing such integration. In the absence of common disambiguated identifiers, researchers often must resort to ''fuzzy" matching, which allows imprecision in the characteristics used to identify common entities across dfferent datasets. While the record linkage literature has identified numerous individually useful fuzzy matching techniques, there is little consensus on a way to integrate those techniques within a single framework. To this end, we introduce the Multiple Algorithm Matching for Better Analytics (MAMBA), an easy-to-use, flexible, scalable, and transparent software platform for business record linkage applications using Census microdata. MAMBA leverages multiple string comparators to assess the similarity of records using a machine learning algorithm to disambiguate matches. This software represents a transparent tool for researchers seeking to link external business data to the Census Business Register files.View Full Paper PDF
-
Working PaperEstimating Measurement Error in SIPP Annual Job Earnings: A Comparison of Census Survey and SSA Administrative Data
September 2002
Working Paper Number:
tp-2002-24
The third chapter investigates measurement error in SIPP annual job earnings data linked to SSA administrative earnings data. The multiple earnings measures provided by the survey and administrative data enable the identification of components of true variation and variation due to measurement error. We find that 18% of the variation in SIPP annual job earnings can be attributed to measurement error. We also find that in both the SIPP and the DER, measurement error is persistent over time. A lower level of auto-correlation in the SIPP measurement error than in the economic error component leads to a lower reliability ratio of .62 for first-differenced earnings.View Full Paper PDF
-
Working PaperNational Experimental Wellbeing Statistics - Version 1
February 2023
Working Paper Number:
CES-23-04
This is the U.S. Census Bureau's first release of the National Experimental Wellbeing Statistics (NEWS) project. The NEWS project aims to produce the best possible estimates of income and poverty given all available survey and administrative data. We link survey, decennial census, administrative, and third-party data to address measurement error in income and poverty statistics. We estimate improved (pre-tax money) income and poverty statistics for 2018 by addressing several possible sources of bias documented in prior research. We address biases from 1) unit nonresponse through improved weights, 2) missing income information in both survey and administrative data through improved imputation, and 3) misreporting by combining or replacing survey responses with administrative information. Reducing survey error substantially affects key measures of well-being: We estimate median household income is 6.3 percent higher than in survey estimates, and poverty is 1.1 percentage points lower. These changes are driven by subpopulations for which survey error is particularly relevant. For house holders aged 65 and over, median household income is 27.3 percent higher and poverty is 3.3 percentage points lower than in survey estimates. We do not find a significant impact on median household income for householders under 65 or on child poverty. Finally, we discuss plans for future releases: addressing other potential sources of bias, releasing additional years of statistics, extending the income concepts measured, and including smaller geographies such as state and county.View Full Paper PDF
-
Working PaperThe Sensitivity of Economic Statistics to Coding Errors in Personal Identifiers
October 2002
Working Paper Number:
tp-2002-17
In this paper, we describe the sensitivity of small-cell flow statistics to coding errors in the identity of the underlying entities. Specifically, we present results based on a comparison of the U.S. Census Bureau's Quarterly Workforce Indicators (QWI) before and after correcting for such errors in SSN-based identifiers in the underlying individual wage records. The correction used involves a novel application of existing statistical matching techniques. It is found that even a very conservative correction procedure has a sizable impact on the statistics. The average bias ranges from 0.25 percent up to 15 percent for flow statistics, and up to 5 percent for payroll aggregates.View Full Paper PDF
-
Working PaperSimultaneous Edit-Imputation for Continuous Microdata
December 2015
Working Paper Number:
CES-15-44
Many statistical organizations collect data that are expected to satisfy linear constraints; as examples, component variables should sum to total variables, and ratios of pairs of variables should be bounded by expert-specified constants. When reported data violate constraints, organizations identify and replace values potentially in error in a process known as edit-imputation. To date, most approaches separate the error localization and imputation steps, typically using optimization methods to identify the variables to change followed by hot deck imputation. We present an approach that fully integrates editing and imputation for continuous microdata under linear constraints. Our approach relies on a Bayesian hierarchical model that includes (i) a flexible joint probability model for the underlying true values of the data with support only on the set of values that satisfy all editing constraints, (ii) a model for latent indicators of the variables that are in error, and (iii) a model for the reported responses for variables in error. We illustrate the potential advantages of the Bayesian editing approach over existing approaches using simulation studies. We apply the model to edit faulty data from the 2007 U.S. Census of Manufactures. Supplementary materials for this article are available online.View Full Paper PDF
-
Working PaperConfidentiality Protection in the Census Bureau Quarterly Workforce Indicators
February 2006
Working Paper Number:
tp-2006-02
The QuarterlyWorkforce Indicators are new estimates developed by the Census Bureau's Longitudinal Employer-Household Dynamics Program as a part of its Local Employment Dynamics partnership with 37 state Labor Market Information offices. These data provide detailed quarterly statistics on employment, accessions, layoffs, hires, separations, full-quarter employment (and related flows), job creations, job destructions, and earnings (for flow and stock categories of workers). The data are released for NAICS industries (and 4-digit SICs) at the county, workforce investment board, and metropolitan area levels of geography. The confidential microdata - unemployment insurance wage records, ES-202 establishment employment, and Title 13 demographic and economic information - are protected using a permanent multiplicative noise distortion factor. This factor distorts all input sums, counts, differences and ratios. The released statistics are analytically valid - measures are unbiased and time series properties are preserved. The confidentiality protection is manifested in the release of some statistics that are flagged as "significantly distorted to preserve confidentiality." These statistics differ from the undistorted statistics by a significant proportion. Even for the significantly distorted statistics, the data remain analytically valid for time series properties. The released data can be aggregated; however, published aggregates are less distorted than custom postrelease aggregates. In addition to the multiplicative noise distortion, confidentiality protection is provided by the estimation process for the QWIs, which multiply imputes all missing data (including missing establishment, given UI account, in the UI wage record data) and dynamically re-weights the establishment data to provide state-level comparability with the BLS's Quarterly Census of Employment and Wages.View Full Paper PDF
