Key words data linkage - claims data - secondary data - cohort study - preterm infants - data
protection
Schlüsselwörter Datenlinkage - Routinedaten - Sekundärdaten - Kohortenstudie - Frühgeborene - Datenschutz
Introduction
Scientific research strongly depends on the nature and the extent of the available
data. By using questionnaires it is possible to collect information specially tailored
to the research question, but on the other hand primary data collection is cost consuming
and may be affected by different kinds of bias. Selection bias, recall bias or social
desirability bias are examples [1 ]
[2 ]
[3 ]. Secondary data constitute an alternative data source. Being data which already
exist, using them bears the potential to be very resource-efficient. Besides the potential
cost efficiency due to the way of data collection, secondary data offer additional
advantages, depending upon the respective provenience. Secondary data of German statutory
health insurances (SHI data), for example, are available with a high degree of timeliness
and completeness. Furthermore, longitudinal analysis is possible [4 ]. However, as secondary data have been collected for other purposes, they might not
include all relevant information. In German SHI data, for example, sociodemographic
and psychosocial information are limited. In addition, as the purpose of SHI data
is billing of services, data may be subject to bias due to upcoding [5 ]. All in all, there are strengths and limitations with all different kinds of data
sources. Thus, linking different data sources on the individual level offers great
opportunities for using the strengths of different data sources synergistically and
overcoming some of their limitations (e. g. [6 ]). The number of research projects that use linkage of different secondary and primary
data to answer research questions increases (examples lidA-Kohortenstudie [7 ], “Gesundes Kinzigtal” [8 ]). The enormous challenge this contains is to realize linkage whilst safeguarding
data privacy.
This paper will provide insight into the methods used in the linkage study EcoCare-PIn
(Early comprehensive Care of Preterm Infants). This cohort study investigated the
short- and long-term consequences of preterm birth with regard to parental stress,
parent-child relationship, family and child quality of life, child development, and
healthcare utilization including costs. For the investigation of this broad spectrum
of clinical, psychosocial and socioeconomic outcomes, data linkage on individual level
is a promising approach. While complying with all relevant data protection requirements,
the study realized an individual-level linkage of
pseudonymized administrative SHI data from a Saxon statutory health insurance on children born between 2007 and 2013 in
Saxony (inpatient and outpatient data),
primary data collected from the parents/caregivers of all eligible very low birth weight (<1500 g)
and low birth weight infants (1500–2500 g) and a matched sample of infants above 2500 g
birth weight,
medical data from kindergarten- and school-entry-examinations of Saxon health authorities (in extension of the EcoCare-PIn study).
SHI data and medical data from kindergarten- and school-entry-examinations of Saxon
health authorities are two examples of already existing data sources that can complement
cost-consuming primary data collections with valuable data sets and open up opportunities
for longitudinal analysis.
Using the study EcoCare-PIn as an example, the aim of this article is to systematically
outline the methods used and the experiences made during the conduct of a linkage
study in the field of pediatric health care research that involved access to and linkage
of three different data sources, among them medical data from health authorities,
which have previously hardly been used for research [9 ].
The article addresses necessary regulatory measures, data access, data validation
and other preconditions for the successful and data protection-compliant data linkage
and discusses challenges and possible solutions. From the experiences made during
EcoCare-PIn, recommendations will be derived that may help researchers in planning
and executing similar studies in (pediatric) healthcare research.
Methods
Study design, data sources and cohorts
The publicly funded cohort study EcoCare-PIn has been registered (Deutsches Netzwerk
Versorgungsforschung: VfD_EcoCare-PIN_13_003463) and described elsewhere [10 ]. The methods and procedures of the study were developed in compliance with the ethical
principles of the Declaration of Helsinki [11 ] and the guidance provided in Good Epidemiologic Practice [12 ] in connection with Good Practice of Secondary Data Analysis [13 ]. Standard Operating Procedures for handling of secondary data guaranteed the fulfilment
of best practice requirements within the Center for Evidence-Based Healthcare (ZEGV)
as data analysis site. The study protocol has been approved by the responsible ethics
committee (EK 67022014) and the data protection officer of the Technische Universität
Dresden as well as the Saxon data protection officer (2-7410-74/1) and the data protection
officer of the SHI.
Data from three different sources were used and linked (see [Fig. 1 ]):
Fig. 1 Architecture and variables provided in EcoCare-PIn.
First, the study was based on pseudonymized administrative SHI data that were provided by the German statutory health insurance AOK PLUS for all insured
children within the Federal State of Saxony who were born between January 1st , 2007 and December 31st , 2013 as shown in [Fig. 1 ]. First results of the SHI data analysis have been published recently [14 ].
To enable further analyses, additional primary data were collected on a subgroup of the described administrative insurance data cohort
of children. This subgroup of children was selected as follows: Primarily, all not
deceased children continuously insured from their birth onwards with a Saxon ZIP-code
were defined as being eligible for primary data collection. Out of these, 17,500 children
were to be selected for primary data collection (determination from sample size estimation).
Since the project focused on the long term consequences of low birth weight, all children
with very low and low birthweight were chosen. Control children for the primary data
collection were selected via frequency matching according to birth year, sex and administrative
district. The contents of the primary data collection (postal questionnaire) are shown
in [Fig. 1 ].
On a subgroup of children selected for primary data collection (ages 3 and up), there
is availability of physician-collected medical data from kindergarten- and school-entry-examinations
of Saxon health authorities. As an extension of the EcoCare-PIn study, these standardized medical data from Saxon health authorities were used. All 13 Saxon public health authorities conduct standardized clinical and
developmental tests (fields of vision, hearing, language, fine and gross motor skills)
of 4-and 6-year old children and document the results digitally via using an unitary
software. While the school-entry-examination (at age 6) is compulsory for all children
of the appropriate age, the kindergarten examination (at age 4) introduced in 2003
is voluntary, based on parental consent [15 ]. Therefore, the data source of Saxon health authorities contains files on all 6-year-old
and on a share of 4-year-old children. Due to the update of records, longitudinal
data are available for some children.
Concept of data linkage
A concept of data linkage ([Fig. 2 ]) was developed to realize the individual-level linkage of the three mentioned data
sources in accordance with data protection requirements. The fundamental principle
of the concept of data linkage, which will be outlined below, was to strictly separate
the sites of data collection (I and II) from the data analysis site (ZEGV).
Fig. 2 Concept of data linkage.
Starting point was the site of data collection I (SHI), which submitted pseudonymized
(pseudonym I) administrative outpatient and inpatient data of all children born in
Saxony between January 2007 and December 2013 with health insurance at the AOK PLUS
to the data analysis site (ZEGV).
Data analysis site (ZEGV) then identified individuals eligible for the primary data
collection and transferred pseudonyms (pseudonym I) of all children with very low
birth weight and low birth weight as well as of matched control children (children
with normal birthweight and without recorded birthweight) to the site of data collection
I (SHI). For reasons of better practicability (see below section C technical aspects),
ZEGV additionally transferred a new short pseudonym per individual (pseudonym II).
Data collection site I (SHI) de-pseudonymized these information and sent postal questionnaires
with pseudonym I and II to the caregivers of the selected children insured at the
AOK PLUS. The questionnaires were supplemented by in-depth information about the study
concept and data protection issues (study information) and an informed consent form
for study participation and one (primary data and SHI data) or both (primary data
and SHI data and health authorities data) linkage actions. Additionally, the carer(s)
of all children aged 3 years or older were asked to give informed consent regarding
the transfer and linkage of data from kindergarten- and school-entry-examinations
of Saxon health authorities.
Children whose caregivers agreed with study participation sent the completed questionnaire
and the signed consent form to the data collection site II (trust center Koordinierungszentrum
für Klinische Studien Dresden (KKS)).
Data collection site II (trust center) received the completed questionnaires, deleted
any personal identifiers and sent the pseudonymized datasets to the data analysis
site (ZEGV) for the linkage with health insurance data and statistical analysis.
Furthermore, data collection site II (trust center) realized the personal query of
the medical data of health authorities using the pseudonym for those children with
written informed consent. Initially, the health authorities had been provided with
a list of contemplable children and copies of the respective written informed consent
forms by the trust center. Then the responsible person of the data collection site
II (trust center) asked for the respective datasets in the premises of the health
authorities where they were filtered out and pseudonymized. Only pseudonymized datasets
left the health authorities.
Data collection site II (trust center) electronically provided the pseudonymized datasets
to the data analysis site (ZEGV) for the linkage with health insurance data and questionnaire
data.
Linkage was possible by continuously using pseudonym II, without personal identifiers,
keeping all data protection requirements in mind. As a result of the concept of data
linkage, throughout the study only pseudonymized primary data and/or pseudonymized
secondary data were transmitted to the data analysis site. Solely the trust center
received personal data of the persons involved. The trust center, in turn, had no
access to the SHI data. Once the parties had given their consent and the completed
questionnaires were sent to the trust center, this information was forwarded to the
ZEGV for data analysis only with a pseudonym. Secondary data and personal identifiers
were not in one hand outside the SHI at any time.
Results
Response rates
EcoCare-PIn drew upon pseudonymized SHI data on 139,383 Saxon children born between
2007 and 2013 ([Fig. 3 ]). From this data pool, 17,498 datasets were selected by the data analysis size (ZEGV)
including all very low and low birthweight children as well as a matched sample of
control children. The pseudonyms of these children were transferred to the health
insurance. Due to missing addresses, death or the lack of a contact permission, the
health insurance did not contact the carer(s) of 985 children. The health insurance
approached the carer(s) of the remaining 16,513 children with a postal questionnaire
carrying the pseudonyms, asked them to fill it in and to give informed consent regarding
a linkage of questionnaire and health insurance data. The responses of the primary
data collection (adjusted response rate: n=4,512 (27.3%)) were received by the trust
center. Additionally, the carer(s) of 2,627 children aged 3 years or older (equals
carers of almost 90% of children at this age) gave informed consent regarding the
transfer and linkage of data from kindergarten- and school-entry-examinations of Saxon
health authorities. Finally, for 1,677 of these children data were transferred from
the 13 Saxon health authorities. The difference between the number of children with
informed consent and the number of transferred datasets arised from two causes: either
the children were not yet four years old and therefore had not yet been due for the
kindergarten examination or they were already four years old but no kindergarten examination
had taken place (voluntary examination depending on parental consent; sometimes also
suspended by health authorities in favour of conducting the obligatory school-entry-examinations).
The trust center sent the pseudonymized questionnaire data and the health authorities’
data to the data analysis site.
Fig. 3 Data flow.
Challenges and approaches
A) Regulatory measures in advance to get data access
Comprehensive preliminary considerations and regulatory measures in advance were necessary
to get data access whilst safeguarding data privacy. Several responsible data protection
officers had to be engaged to prove the overall concept of the study including all
study documents. The approval of the study was based on the concept of data linkage
([▶Fig. 2 ]), data protection- and data security concepts of the institutes involved as well
as on study documents for participants of the primary data collection.
In EcoCare-PIn, access to SHI data of the health insurance was contract-based and took place on the basis of a detailed
data record description. The health insurance participated as project partner, since
there was strong interest in the assessment of the burden of preterm birth. A positive
internal evaluation of the data protection officer of the health insurance (data owner)
was performed regarding the study on the basis of the legislation valid at the time
of data transfer. The SHI comprehensively examined the necessity of an application
according to §75 of the German Social Code, book X (SGB X). At the time of preparing
data access, this application was deemed not necessary by the SHI. The de-pseudonymization
of the pseudonymized SHI data was regarded impossible outside the SHI and data therefore
considered “de facto anonymous”. According to §3 VI of the Federal Data Protection
Act old version (German: BDSG a.F.), anonymization meant that data can either not
be attributed to an individual or only with a disproportionate amount of time, expense
and labour. However, with the commencement of the General Data Protection Regulation
(GDPR, German: Datenschutzgrundverordnung DSGVO) on 25 May 2018, the latter mitigation
is not provided any more [16 ]. Methodological guidance for the realization of data linkage with regard to the
current data protection regulations is published in Good Practice Data Linkage [17 ].
The regulatory measures to get access to primary data concern a) access to the individuals for primary data collection and b) the collection
and linkage of the data themselves.
Access to individuals for primary data collection was realized via the SHI who postally approached the caregivers of eligible
children with the following documents:
in-depth information about the study concept and data protection issues (study information)
questionnaire for primary data collection
consent form (2x) regarding (1) participation in primary data collection and data
linkage with SHI data and (2) regarding the transfer of data from kindergarten- and
school-entry-examinations of Saxon health authorities and their linkage with SHI and
primary data
nonresponse-form
small incentive (colouring picture for children)
For reasons of data protection, the transfer of information about participation or
non-participation of insured persons to the SHI had to be ruled out. Therefore, the
contribution of the SHI ended with sending the study documents to the caregivers of
eligible children. To collect the returning questionnaires, a trust center was delegated
([Fig. 2 ]).
Primary data collection and linkage was legitimized through obtaining written informed consent from the children’s caregivers
for study participation and one (primary data and SHI data) or both (primary data
and SHI data and health authorities data) linkage actions. In case of participation
in the primary data collection, the initially “de facto anonymized” SHI data had to
be regarded as personal data, whose scientific use was to be covered by informed consent.
Written informed consent was also the legal basis for the use of the health authorities’ data. The head officials of the Saxon health authorities were postally asked to support
the study EcoCare-PIn by providing medical data from kindergarten- and school-entry-examinations
on children whose caregiver(s) had documented their consent. Considerable effort was
made to motivate health authorities (individual invitation, covering letter of the
Saxon State Ministry of Social Affairs, project presentation). Together with IT personnel,
resource-saving procedures for data extraction and transfer were planned and the health
authorities were offered various options of receiving assistance from the trust center
(KKS). Nonetheless, the support of the project meant an additional burden for the
health authorities. Still, all 13 health authorities could be motivated to participate.
Overall, the involvement of a trust center was an absolutely necessary regulatory precondition for the conduct of data collection
and linkage. As an organizationally, spatially and personally independent unit [18 ], it enabled the collection of primary data and health authorities data and their
pseudonymized allocation to the data analysis site.
B) Data protection requirements during the study
To increase the response rate of the primary data collection, a reminder letter was
sent after four weeks. Due to data protection requirements, it was not allowed to
transfer the information to the SHI which children’s caregivers had already answered
and returned the questionnaire. Therefore, the SHI had to repeatedly send the questionnaires
(again supplemented by in-depth information about the study concept and data protection
issues and a consent form) to the whole subcohort of individuals selected for the
primary data collection.
Despite a note in the reminder letter, that the added questionnaire please be ignored
if the initial one had been returned, some participants filled in and sent the questionnaire
a second time. In this case, the more complete one was used for the analysis.
C) Technical aspects
Transfer of health insurance data to the data analysis site (ZEGV)
SHI data were provided via a portal for secure data transfer.
Data extraction in health authorities
Considerable effort was made to clarify with IT personnel of health authorities how
data extraction is technically possible. As most health authorities had sparsely or
not been involved in research projects, there were no routine procedures implemented
yet. However, as all authorities used the same data software and were managed by the
same IT company, communication could initially be centralized and a unified technical
solution for data extraction could be established. This solution was then communicated
to and applied by IT managers of each single health authority. After data extraction
in the health authorities, the data had to be transferred to the data collection site
II (trust center) in line with data protection requirements. Pseudonymized data were
personally collected by the responsible person of the trust center in the form of
password-protected CD’s.
Electronical reading of questionnaire data
Due to the sample size of EcoCare-PIn, electronic reading of questionnaire data was
opted for. From previous research projects, the reading software was known to have
some difficulty with distinguishing between some letters and numbers, i. e. between
O and 0 or between 1 and 7. To minimize sources for reading mistakes, a second pseudonym
(pseudonym II), comprising of only five numbers, was assigned to every questionnaire
by the data analysis site in addition to pseudonym I, which had been assigned by the
SHI and was a combination of 24 numbers and letters. Only the short pseudonym II was
used for the electronic reading of questionnaires.
Transfer of primary and health authorities data to the data analysis site (ZEGV)
Data collected by the trust center were provided via internal network.
D) Validation of data and data linkage
Considerable effort was spent on data validation in order ensure the high quality
of the final dataset and avoid incorrect data linkage. A number of validation routines
were applied. The validation focused both on plausibility checks within the questionnaire
data (e. g. check whether answers are within valid range of values, comparison of
different variables related to the age of the child) and on consistency checks of
information given in primary and secondary data (i. e. birthweight and age-related
information). In advance, tolerance ranges had been determined together with pediatric
clinicians. For example, differences in birthweight between primary and secondary
data were determined tolerable if ≤ 100 g. Larger differences meant that primary and
SHI data were not linked.
Regarding the comparison of birthweights in primary and secondary data, many initial
“discrepancies” could be eliminated through simple correction, as their cause was
clearly traceable (see [Table 1 ]): the parental confusion of questionnaires between siblings within one family, transposed
digits in hand written birthweight in questionnaires or reading errors of the program
regarding hand written birthweight. Furthermore, some “discrepancies” in birthweight
were caused through the fact that – in the case of healthy newborn multiples – the
SHI assigned the birthweight of one sibling identically to the other sibling(s).
Table 1 Patterns of discrepancies. The table summarizes the different patterns of discrepancies
found while checking the consistency of information given in primary and SHI data.
There might be an overlap within the frequencies.
Types of discrepancies
Examples
Approach
Causes
Prevention
Frequency
1. Birthweight information of two questionnaires with adjacent serial numbers do not
fit to birthweight information in SHI data, the questionnaires seemed to be interchanged
Parents wrote on questionnaires that they confused the questionnaires
Data (pseudonym 2) was assigned correctly and the data were linked (no further differences)
Parents confused the questionnaires for siblings
Name target child in cover letter and study information
n=36
2. Reading error of the program regarding the handwritten birthweight of the primary
data collection
Birthweight in SHI data: 3,430 g Birthweight in the completed questionnaire: 8,430 g
Primary data have been corrected and the data were linked (no further differences)
Reading error with critical numbers (1 and 7, 5 and 8, 3 and 8) or indistinct spelling
Visual inspection during/after the reading process necessary
n=8
3.Transpositions in the hand written birth weight information given in the primary
data collection
Birthweight in SHI data: 2,240 g Birthweight in the completed questionnaire: 2,420 g
Primary data have been corrected and the data were linked(no further differences)
n=20
4. Accumulation of discrepancies of birthweight-information in questionnaire vs. SHI
data for twins/muliples
Birthweight of both children in SHI data: 2,200 g Birthweight in the completed questionnaire:
2,560 g and 2,200 g
Primary data were used and the data were linked
In SHI data in case of healthy twins/multiples the same birthweight (of one of them)
was allocated to all multiples since healthy newborns (and their birthweight) are
coded together with the mother’s delivery.
n=29
5. Discrepancies of birthweight/ age-information in questionnaire vs. SHI data, that
cannot be clarified
Age of 4 years in SHI data vs. 7 years in the completed questionnaire Birthweight
of 2,450 g in SHI data and 2,600 g in the completed questionnaire
Age: Evaluation was implausible, data were not linked Birthweight: consultation of
clinical experts → differences of maximal 100 g were defined to be acceptable, children
with larger differences were excluded Of 4,074 children both birthweight-information
exist (SHI and and primary data) no difference in both data sources: 82.8% differences
of maximal 100 g: 11.5% differences of more than 100 g: 5.7%
Birthweight: DRG-Upcoding [15 ] Wrong child, parents confused the questionnaires for siblings Wrong child, i. e.
care givers of children selected as control child from the NBW group coincidentally
also had a preterm child within the family and therefore filled in the questionnaire
for the preterm child
Name target child in cover letter and study information Slightly modified study information
for the control group
n=58
After having identified and corrected all of these “false alarms”, there remained
58 questionnaires which were obviously not filled in for the correct child and therefore
were not used for data linkage.
To validate the linkage of the different data sources several variables (e.g. age)
were used that are concordantly recorded in all data bodies to be linked.
Discussion
The present article outlined the methodological approach used in the linkage study
EcoCare-PIn. The experiences of linkage on an individual level within the study shows
the feasibility of integrating data from the health authorities and may serve as a
blueprint for subsequent studies which want to link data from various data sources
individually. Linkage is possible and worthwhile to create a solid database for the
further development of perinatal healthcare in Germany. Derived from the experiences
of EcoCare-PIn, the following points summarize the main insights that may be of interest
for researchers who intend to conduct similar studies:
The complexity of the data-flow concept and of the validation process increases with
increasing number of data sources that are aimed to be linked. The providers of secondary
data and their data protection officers therefore need to be involved from an early
stage.
There was considerable effort to motivate all 13 health authorities to participate,
since the project meant an additional workload. Therefore, supporting them with the
implementation of a data extraction routine was highly appreciated. Special structural
institutions (trust center) are indispensable to link the data sources on individual
level in accordance with data protection requirements.
The complexity of the study also led to high requirements in the preparation of study
documents to describe the procedure in a comprehensible manner for the participants.
A written informed consent was the legal basis for usage and linkage of the secondary
and primary data. The study highlights the value of data from health authorities,
which have so far received little attention from research [8 ].
When selecting participants for the survey from SHI data it should be taken into account
that - due to missing addresses, death or the lack of a contact permission - the health
insurance may not be able to contact the whole target population. In the case of EcoCare-PIn,
the SHI did not contact the carer(s) of 985 children. This is a considerable loss
of a substancial group of children chosen for primary data collection that is necessary
to be taken into account in sample size estimation and also in considerations regarding
the generalizability.
The resending of all study documents at the reminder, which was important for data
protection reasons, caused confusion and displeasure among some participants and moved
some participants to repeatedly fill in the questionnaire. This resulted in an increased
effort in the documentation of the return and in the validation process.
The existence of several variables that are concordantly recorded in all data bodies
to be linked proved useful for validation purposes. However, in case of discrepancies,
decisions have to be made regarding the degree of data non-conformance that is still
acceptable or is considered to indicate incorrect data linkage. Thus, data validation
is time consuming and has to be scheduled in advance. For the definition of tolerance
ranges, interdisciplinary exchange with clinicians and data providers is indispensable.
Recommendations for practice
schedule enough time in advance to clarify data access and establish/involve trust
center
select a sufficiently high number of SHI patients based on sample size estimations,
taking into account that probably not all insurants can be contacted by the SHI
if possible: collect several variables via primary data collection although they are
also available in secondary data (for validation purposes)
schedule enough time for data and data linkage validation
include clinicians to determine reasonable tolerance ranges for data validation
