Using synthetic data to replace linkage derived elements: a case study

Resnick, Dean M.; Cox, Christine S.; Mirel, Lisa B.

doi:10.1007/s10742-021-00241-z

Using synthetic data to replace linkage derived elements: a case study

Published: 23 February 2021

Volume 21, pages 389–406, (2021)
Cite this article

Health Services and Outcomes Research Methodology Aims and scope Submit manuscript

301 Accesses
1 Citation
3 Altmetric
Explore all metrics

Abstract

While record linkage can expand analyses performable from survey microdata, it also incurs greater risk of privacy-encroaching disclosure. One way to mitigate this risk is to replace some of the information added through linkage with synthetic data elements. This paper describes a case study using the National Hospital Care Survey (NHCS), which collects patient records under a pledge of protecting patient privacy from a sample of U.S. hospitals for statistical analysis purposes. The NHCS data were linked to the National Death Index (NDI) to enhance the survey with mortality information. The added information from NDI linkage enables survival analyses related to hospitalization, but as the death information includes dates of death and detailed causes of death, having it joined with the patient records increases the risk of patient re-identification (albeit only for deceased persons). For this reason, an approach was tested to develop synthetic data that uses models from survival analysis to replace vital status and actual dates-of-death with synthetic values and uses classification tree analysis to replace actual causes of death with synthesized causes of death. The degree to which analyses performed on the synthetic data replicate results from analysis on the actual data is measured by comparing survival analysis parameter estimates from both data files. Because synthetic data only have value to the degree that they can be used to produce statistical estimates that are like those based on the actual data, this evaluation is an essential first step in assessing the potential utility of synthetic mortality data.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Practical Lessons from Generating Synthetic Healthcare Data with Bayesian Networks

Generating high-fidelity synthetic time-to-event datasets to improve data transparency and accessibility

Article Open access 23 June 2022

A method for generating synthetic longitudinal health data

Article Open access 23 March 2023

Availability of data and materials

Researchers who wish to obtain access to the linked 2016 NHCS to 2016/2017 NDI file must submit and have an approved research proposal to the NCHS Research Data Center (RDC): https://www.cdc.gov/rdc/index.htm.

Code availability

Code for this analysis is available upon request.

Notes

Responding hospitals were those providing records for at least 50 encounters covering at least 6 months of the year.
Sufficient PII is defined as having two of the following three items: valid date of birth (month, day, and year), name (first, middle, and last), and/or a valid format 9-digit SSN. See (NCHS 2019, p. 5).
Division is coded based on patient home state based on U.S. Census Bureau Schema. See https://www2.census.gov/geo/pdfs/maps-data/maps/reference/us_regdiv.pdf.

References

Breiman, L., et al.: Classification and Regression Trees. CRC Press, Boca Raton (1984)
Google Scholar
Breslow, N.E.: Discussion of the paper by D. R. Cox. J. R. Stat. Soc. B 34, 216–217 (1972)
Google Scholar
Charlson, M., et al.: A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J. Chronic Dis. 40(5), 373–383 (1987)
Article CAS Google Scholar
David, C.R.: Regression models and life tables (with discussion). J. R. Stat. Soc. 34(2), 187–220 (1972)
Google Scholar
Fiscella, K., Fremont, A.M.: Use of geocoding and surname analysis to estimate race and ethnicity. Health Serv. Res. 41(4), 1482–1500 (2006)
PubMed PubMed Central Google Scholar
Landis, J.R., Koch, G.G.: The measurement of observer agreement for categorical data. Biometrics 33, 159–174 (1977)
Article CAS Google Scholar
Lipkovich, I., Ratitch, B., O’Kelly, M.: Sensitivity to censored-at-random assumption in the analysis of time-to-event endpoints. Pharm. Stat. 15(3), 216–229 (2016)
Article Google Scholar
National Center for Health Statistics. National Death Index (NDI). http://www.cdc.gov/nchs/ndi/index.htm
National Center for Health Statistics. Division of Analysis and Epidemiology. The Linkage of the 2016 National Hospital Care Survey to the 2016/2017 National Death Index: Methodology Overview and Analytic Considerations, August 2019. Hyattsville, Maryland. https://www.cdc.gov/nchs/data/datalinkage/NHCS16_NDI16_17_Methodology_Analytic_Consider.pdf
National Center for Health Statistics. National Hospital Care Survey (NHCS). http://www.cdc.gov/nchs/dhcs/index.htm
SAS HPSPLIT documentation. https://documentation.sas.com/?docsetId=stathpug&docsetTarget=stathpug_hpsplit_syntax01.htm&docsetVersion=15.1&locale=en
SAS PHREG documentation. https://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#phreg_toc.htm
Sundararajan, V., et al.: New ICD-10 version of the Charlson comorbidity index predicted in-hospital mortality. J. Clin. Epidemiol. 57(12), 1288–1294 (2004)
Article Google Scholar

Download references

Funding

This work was supported in part with funding from the Department of Health and Human Services’ Office of the Secretary Patient Centered Outcomes Research Trust Fund (OS-PCORTF).

Author information

Authors and Affiliations

NORC at the University of Chicago, Bethesda, MD, USA
Dean M. Resnick & Christine S. Cox
Centers for Disease Control and Prevention, National Center for Health Statistics, Hyattsville, MD, USA
Lisa B. Mirel

Authors

Dean M. Resnick
View author publications
You can also search for this author in PubMed Google Scholar
Christine S. Cox
View author publications
You can also search for this author in PubMed Google Scholar
Lisa B. Mirel
View author publications
You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Dean M. Resnick.

Ethics declarations

Conflict of interest

The authors have no conflicts of interest to report.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix

NHCS 2016 Survival Estimates

Comparison of Actual to Synthetic Data

Comparison for All Deaths

#	Variable	Estimate	Estimate (Synth.)	StdErr	StdErr (Synth.)	ProbChiSq	ProbchiSq (Synth.)
Age group (rounded to nearest 5-year, reference category: 70 years old)
1	0	− 3.37	− 3.34	0.036	0.035	< 0.0001	< 0.0001
2	5	− 4.12	− 4.07	0.059	0.057	< 0.0001	< 0.0001
3	10	− 3.97	− 3.95	0.063	0.062	< 0.0001	< 0.0001
4	15	− 3.47	− 3.41	0.048	0.047	< 0.0001	< 0.0001
5	20	− 2.75	− 2.78	0.032	0.032	< 0.0001	< 0.0001
6	25	− 2.34	− 2.33	0.025	0.025	< 0.0001	< 0.0001
7	30	− 2.13	− 2.15	0.023	0.023	< 0.0001	< 0.0001
8	35	− 1.79	− 1.80	0.021	0.021	< 0.0001	< 0.0001
9	40	− 1.48	− 1.50	0.019	0.019	< 0.0001	< 0.0001
10	45	− 1.19	− 1.16	0.017	0.016	< 0.0001	< 0.0001
11	50	− 0.90	− 0.87	0.014	0.014	< 0.0001	< 0.0001
12	55	− 0.62	− 0.62	0.012	0.012	< 0.0001	< 0.0001
13	60	− 0.38	− 0.39	0.011	0.011	< 0.0001	< 0.0001
14	65	− 0.21	− 0.22	0.011	0.011	< 0.0001	< 0.0001
15	75	0.21	0.21	0.010	0.010	< 0.0001	< 0.0001
16	80	0.48	0.47	0.010	0.010	< 0.0001	< 0.0001
17	85	0.79	0.78	0.010	0.010	< 0.0001	< 0.0001
18	90	1.12	1.11	0.011	0.011	< 0.0001	< 0.0001
19	95	1.52	1.50	0.012	0.013	< 0.0001	< 0.0001
20	Age missing	0.26	0.15	0.041	0.043	< 0.0001	0.0004
Conditions not present (reference category: condition not present)
21	Myocardial Infarction	0.03	0.03	0.016	0.016	0.0804	0.0309
22	Diabetes without complications	− 0.31	− 0.33	0.023	0.023	< 0.0001	< 0.0001
23	Diabetes with complications	− 0.03	− 0.06	0.036	0.037	0.3462	0.1008
24	Paraplegia and Hemiplegia	0.33	0.31	0.021	0.022	< 0.0001	< 0.0001
25	Renal Disease	0.06	0.06	0.014	0.015	< 0.0001	< 0.0001
26	Cancer	0.56	0.54	0.015	0.016	< 0.0001	< 0.0001
27	Moderate or Severe Liver Disease	0.59	0.58	0.076	0.076	< 0.0001	< 0.0001
28	Metastatic Carcinoma	1.30	1.30	0.029	0.031	< 0.0001	< 0.0001
29	AIDS/HIV	− 0.08	− 0.12	0.045	0.047	0.0598	0.0081
30	Congestive Heart Failure	0.39	0.38	0.009	0.010	< 0.0001	< 0.0001
31	Peripheral Vascular Disease	− 0.15	− 0.15	0.016	0.016	< 0.0001	< 0.0001
32	Cerebrovascular Disease	− 0.05	− 0.06	0.012	0.012	0.0001	< 0.0001
33	Dementia	0.50	0.48	0.010	0.011	< 0.0001	< 0.0001
34	Chronic Pulmonary Disease	0.03	0.03	0.009	0.009	0.0002	0.0013
35	Connective Tissue Disease-Rheumatic Disease	− 0.05	− 0.03	0.020	0.020	0.0203	0.1004
36	Peptic Ulcer Disease	− 0.06	− 0.03	0.025	0.025	0.0189	0.2764
37	Mild Liver Disease	0.82	0.82	0.016	0.017	< 0.0001	< 0.0001
Charlson Index Summary (reference category: 1)
38	0	− 0.64	− 0.65	0.010	0.010	< 0.0001	< 0.0001
39	2	0.38	0.37	0.011	0.011	< 0.0001	< 0.0001
40	3	0.59	0.58	0.016	0.017	< 0.0001	< 0.0001
41	4	0.78	0.77	0.022	0.024	< 0.0001	< 0.0001
42	5	0.85	0.84	0.030	0.032	< 0.0001	< 0.0001
43	6	0.62	0.59	0.040	0.043	< 0.0001	< 0.0001
Imputed Race/Ethnicity (reference category: White)
44	Asian	− 0.21	− 0.20	0.017	0.017	< 0.0001	< 0.0001
45	Black	− 0.04	− 0.04	0.007	0.007	< 0.0001	< 0.0001
46	Hispanic	− 0.24	− 0.23	0.009	0.009	< 0.0001	< 0.0001
47	Amer. Ind	0.06	0.14	0.070	0.070	0.3613	0.0505
48	Sex: Female (reference category: Male)	− 0.27	− 0.26	0.005	0.005	< 0.0001	< 0.0001
Division (reference category: Mid-Atlantic)
49	Northeast	− 0.05	− 0.03	0.018	0.018	0.0128	0.1139
50	East North Central	0.12	0.12	0.009	0.009	< 0.0001	< 0.0001
51	West North Central	0.10	0.10	0.013	0.013	< 0.0001	< 0.0001
52	South Atlantic	0.21	0.22	0.009	0.009	< 0.0001	< 0.0001
53	East South Central	0.17	0.19	0.011	0.011	< 0.0001	< 0.0001
54	West South Central	0.89	0.90	0.012	0.012	< 0.0001	< 0.0001
55	Mountain	0.18	0.19	0.013	0.013	< 0.0001	< 0.0001
56	Pacific	0.40	0.41	0.010	0.010	< 0.0001	< 0.0001
Hospital bed size (reference category: > 500)
57	0–25	− 0.37	− 0.38	0.038	0.038	< 0.0001	< 0.0001
58	26–100	− 0.19	− 0.19	0.017	0.017	< 0.0001	< 0.0001
59	101–500	− 0.08	− 0.09	0.006	0.006	< 0.0001	< 0.0001
Hospital ownership (reference category: non-profit)
60	For profit	− 0.55	− 0.54	0.019	0.019	< 0.0001	< 0.0001
61	Government	− 0.03	− 0.03	0.009	0.009	0.0005	0.0043
Hospital type (reference category: general acute)
62	Children’s	− 0.14	− 0.16	0.046	0.045	0.0027	0.0004
63	Psychiatric	0.37	0.35	0.024	0.024	< 0.0001	< 0.0001
64	Long-term acute care, rehab., etc	0.17	0.20	0.026	0.027	< 0.0001	< 0.0001
Hospital urban rural classification (reference category: large central metropolitan)
65	Large Fringe Metropolitan	− 0.10	− 0.09	0.008	0.008	< 0.0001	< 0.0001
66	Medium Metropolitan	− 0.05	− 0.05	0.007	0.007	< 0.0001	< 0.0001
67	Small Metropolitan	0.25	0.22	0.011	0.011	< 0.0001	< 0.0001
68	Micropolitan	0.19	0.20	0.013	0.013	< 0.0001	< 0.0001
69	Non-core	0.10	0.12	0.045	0.045	0.0218	0.0078

Rights and permissions

Reprints and permissions

About this article

Cite this article

Resnick, D.M., Cox, C.S. & Mirel, L.B. Using synthetic data to replace linkage derived elements: a case study. Health Serv Outcomes Res Method 21, 389–406 (2021). https://doi.org/10.1007/s10742-021-00241-z

Download citation

Received: 30 April 2020
Revised: 09 November 2020
Accepted: 18 January 2021
Published: 23 February 2021
Issue Date: September 2021
DOI: https://doi.org/10.1007/s10742-021-00241-z

Keywords

Access this article

Log in via an institution

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Using synthetic data to replace linkage derived elements: a case study

Abstract

Access this article

Similar content being viewed by others

Practical Lessons from Generating Synthetic Healthcare Data with Bayesian Networks

Generating high-fidelity synthetic time-to-event datasets to improve data transparency and accessibility

A method for generating synthetic longitudinal health data

Availability of data and materials

Code availability

Notes

References

Funding

Author information

Authors and Affiliations

Corresponding author

Ethics declarations

Conflict of interest

Additional information

Publisher's Note

Appendix

Rights and permissions

About this article

Cite this article

Keywords

Navigation

Using synthetic data to replace linkage derived elements: a case study

Abstract

Access this article

Similar content being viewed by others

Practical Lessons from Generating Synthetic Healthcare Data with Bayesian Networks

Generating high-fidelity synthetic time-to-event datasets to improve data transparency and accessibility

A method for generating synthetic longitudinal health data

Availability of data and materials

Code availability

Notes

References

Funding

Author information

Authors and Affiliations

Corresponding author

Ethics declarations

Conflict of interest

Additional information

Publisher's Note

Appendix

Appendix

Rights and permissions

About this article

Cite this article

Share this article

Keywords

Search

Navigation