Back pain as reported in survey and claims data – comparison of prevalences and concordance by data linkage

Back pain is a burden for those suffering from it as well as for the economy and social insurance system. This explains the great interest in collecting data on back pain in population studies to estimate for instance its prevalence. As part of the lidA study, the agreement between self-reported information on back pain (self-reported prevalence) and health claims data (administrative prevalence) has been examined.

3 identifies a certain group of individuals having few overlaps with one another. Therefore, aiming in identifying the target population for disease-specific health prevention, a combination of both data sources, the prevalence of back pain can be estimated more accurately.

Background
Back pain ranks among the most frequent reasons why people make use of the healthcare system [1]. It is one of the leading causes of work incapacity and years lived with disability (YLDs) [2,3]. It may also lead to early retirement. In addition, back pain creates steep direct costs for treatment and imposes a heavy burden on the economy and the social security system, for instance the pension fund [1,[4][5][6]. Collecting data on back pain in population studies is therefore of great interest in order to estimate prevalence and the need for prevention efforts. However, definitions of back pain show a lack of comparability. Furthermore, a number of different instruments and data pools have been used to collect data on back pain [3,5,[7][8][9][10][11][12][13]. In some instances and as a consequence, prevalences may differ hugely in public studies and are not comparable. To face this dilemma, it would be helpful to have a standard definition to serve as a basis for uniform operationalization in population-based studies [14].
Health and healthcare research is increasingly using health claims data, including data provided by statutory health insurance funds, in addition to individual data collected in surveys and interviews [15,16]. The advantages of both data sources can be combined to overcome some of the limitations of the individual data sources. The individual linkage of survey data with data provided by statutory health insurers offers new opportunities for analyses conducted by current and future health and healthcare researchers. However, the evaluation and interpretation of linked data sources require special methodological standards. In this respect, the validation of the individual data sources used is of 4 paramount importance [17,18]. The best assessment method is a cross validation, i.e., comparing data from different data sources. Generally, evaluating validity requires a gold standard. There is broad consensus that neither survey nor claims data meets this requirement [12,[19][20][21][22][23][24][25]. It has been recommended to use several data sources [20,21,23,26], since the agreement between both data sources depending on the diagnosis examined [24,27]. Self-reported back pain is of limited validity [12], on the other hand there is no need for physicians' contacts if suffering back pain, so claims data also lack on validity.
Due to the missing gold standard, the lack of comparability of definitions of back pain and the limited validity of back pain in different data sources we use several individually linked data sources, to analyze the prevalence of back pain. We have examined -as part of the lidA study -the degree of agreement between subjective information on back pain reported in a computer-assisted personal interview compared to diagnoses documented by doctors (claims data). The recorded diagnoses pertain to claims data provided by statutory health insurers. For cross-validation purposes, both data sources are considered equal after the individual data linkage in order to show the prevalence of back pain more accurately for other content analyses.

Methods
The lidA study is a cohort study on work, age, health, and work participation. It aims at examining the long-term effect of work on the health and work participation of older working employees based on a sample from two age cohorts born in 1959 and 1965, respectively [28]. They present the beginning and the end of Germany's so-called baby boomers who will make up a significant share of the potential older work force in the coming years [29]. The sample was drawn from registry data pertaining to the integrated employment biographies (IEB) of the German Federal Labour Office (Bundesagentur für Arbeit (BA)). It is based on all gainfully employed people covered by social insurance in Germany on December 31, 2009. Public servants and self-employed persons are not part of the study population. The utilization rate is 27.3% [28]. During the computer-assisted interviews (CAPI) of the first two waves in 2011 and 2014, participants were asked for their informed consent to have their interview data linked individual with their claims data. During the second survey in 2014, those participants who had changed statutory health insurance funds were asked again for their consent, as well as those who had not given their consent during the first survey or if other mismatches had occurred [30,31]. A total of 63 percent of all respondents who participated in both survey waves agreed to have their survey and claims data linked [31]. Cooperation agreements with eleven statutory health insurance funds were signed [32]. Finally, data on outpatient and inpatient treatment as well as sick leave and outpatient drug prescription provided by a total of ten statutory health insurance funds [1] were used for the linkage [33][34][35].
A total of 4,244 persons participated in the follow-up [31]. Of these only those participants were included in the validation study for whom linked survey and claims data was available (n=1,031) ( Fig. 1).
During both waves of the primary survey (first wave: 2011, second wave: 2014), study participants were asked exhaustively about their work situation. As regards back pain, we used the questionnaire for the analysis of musculoskeletal symptoms (Nordic questionnaire) [36]. To this end, participants had to answer the following question: "Did you experience any symptoms or pain in the following parts of your body during the past 12 months?" Possible answers included "upper back or thoracic spine" and "lower back or lumbar region". The analysis included all respondents, who affirmed the question of back pain during the past twelve months (self-reported prevalence) in both waves.
The claims data comprised both in-and outpatient data as well as data on sick leave for 6 the period from 2009-2013. Due to unspecific medication and a lack of means to assign it to specific diagnoses, we decided not to use outpatient drug prescription data. Besides, the data would have comprised subscription pain medication only. Data on self-medication using over-the-counter (OTC-)analgetics is not documented in claims data. Back pain is coded using the ICD-10-Code M54 ("back pain") when it is diagnosed either as a principal or secondary diagnosis made by hospitals, as an outpatient diagnosis (classified as 'confirmed' or 'condition after recovery'), and in medical certificates of sick leave. Two different definitions are used to define administrative prevalence in claims data: There seems to be a gap of three years, because claims data were used for the whole years 2009 -2013 but the two questionnaires conducted in 2011 and 2014 respectively.
Since we use all respondents, who reported back pain in both waves, the period of claims data is almost covered. With the definition of back pain in the CAPI data (back pain in both waves) and the definitions of back pain in claims data, especially in Def2, the study looks less at actually back pain than at chronic back pain.

Statistical analyses
First, self-reported and administrative prevalence of back pain was determined descriptively. The agreement of survey and claims data was determined using Cohen's Kappa. A Cohen's Kappa < 0.40 indicates a low level of agreement, a Kappa between 0.41 -0.60 means a moderate level of agreement and a Kappa between 0.61 -1 means a high level of agreement [38]. The overall prevalence of back pain was calculated as the sum of self-reported and administrative prevalence. In order to examine possible differences between the two cohorts as well as between women and men, Cohen's Kappa was then 7 analyzed carefully according to cohort affiliation and gender. Sensitivity and specificity as well as the positive predictive value and the negative predictive value were not calculated because of the missing gold standard and deviations in both directions.
The analyses were performed using IBM SPSS 24 ©.
[1] The data oft he following health insurance funds are part oft the cumulative data set:

Discussion
Due to its design -an individual data linkage of survey and claims data bases in informed consent -the lidA study offered an opportunity to compare data from both data sources in the scope of a cross validation to analyze the prevalence of back pain. To this day, we are not aware of any comparable analysis of back pain based on data from German cohort studies [32]. Since there is no applicable gold standard, both data sources were considered equal. Self-reported prevalence hardly differed from administrative prevalence Tisnado et al. [12] also compared self-reported data with medical records regarding back pain and clinical services. At a Kappa of 0.3 and 0.1, respectively, they also arrived at only low levels of agreement. This low value may be explained by the fact that two different groups were logged in the two data sources, back pain in whole versus back pain diagnosed by a physician. On the one hand, respondents subjectively report back pain without having received a diagnosis from a physician. On the other hand, respondents might only receive a M54 diagnosis if doctors see an indication to treat them. Since self-reported information may be incorrect [12], for instance by exhibiting a recall bias, we consider the use of self-reported information due to its limited validity as a limitation for healthcare research. The question regarding pain within the past 12 months may also have been cause for misunderstanding, which may have led to an over-coverage of back pain in the CAPI data. It is also conceivable that respondents mention pain sooner and more frequently if asked about the diseases diagnosed by a doctor. Some people often experience pain but do not automatically go to see a doctor. According to Green et al. [39], only 25% of individuals having health problems will go to the doctor. In this light, a question about a disease diagnosed by a physician would be more specific in order to specify chronic back pain or pain compromising a person's health. Schmidt et al. [9] also assume that the phrasing of a question about back pain within the past twelve months may have contributed to the high prevalence. Then again, the Nordic questionnaire is an internationally recognized instrument that has been tested on a wide spectrum of occupational groups and validated using clinical data [36,40]. Another factor to keep in mind is the temporal dimension of the question with regard to the period of the diagnosis.
During each of the two waves, respondents were asked about pain within the past twelve months, which means that two years (2011 and 2014) of experiencing back pain were measured. The claims data considers the M54 diagnosis in the period from 2009 to 2013, which means back pain is recorded for a period of five years.
Admittedly, these reasons may also have led to an under-coverage in the claims data because having complete data depends, among other things, on a doctor's accuracy and thoroughness in terms of documentation [12]. What is more, claims data only comprises claim-relevant diagnoses. If a person experiences back pain but does not mention it or does not get treatment, it will not appear in the data. Consequently, the data only comprises individuals who actually went to the doctor. That means there is a utilization 11 bias. It is also possible that respondents experienced other reportable diseases as more relevant and subjectively more impairing and stressful.

Strengths and limitations of the lidA study
Analyses of the representativeness of the lidA study indicate a largely unbiased sample [28,31,41]. Despite the fact that survey data were linked with claims data provided by only ten out of currently 109 statutory health insurers [42], March et al. [32] and Stallmann et al. [43] found selectivity effects to be negligible.

Conclusions
In conclusion, the low level of agreement between both data sources suggests that a data linkage may present a benefit since each data source identifies a certain group of individuals having few overlaps with one another. There are limitations in each of the used data sources which can be minimized by data linkage [44]. We are aware that the period in both data sources does not completely overlap and different things in both data sources were measured. Nevertheless, within a combination of both data sources, the prevalence of back pain can be estimated more accurately. The data linkage captures both persons with back pain who may not have gotten treatment but should have been treated (selfreported) and persons whose back pain was diagnosed by a physician. For the lidA study we interviewed middle-aged employees covered by social insurance. Regarding back pain the cohort itself is a limitation of the study as the prevalence during employment increases with age [3,14]. Especially with regard to health problems such as back pain where self-reported and claims data may differ, it is worth linking several data sources in order to arrive at the true prevalence. Which definition to use for further analyses of back pain in claims data depends on the question to be asked. For analyses of chronic back pain, we recommend to use a more conservative approach such as Def2 while Def1 is more suitable for studies to estimate the need for prevention measures.

Consent for publication
Not applicable.

Availability of data and materials
Not applicable. For reasons of data protection and deletion deadlines the data is no longer available after the end of the project.

Competing Interests
None declared.

Funding
The lidA study was supported by the German Federal Ministry of Education and Research

Authors' contributions
All authors contributed to the design and concept of the study. AF drafted main parts of the manuscript and performed the statistical analysis. SM was responsible for the claims data and data linkage of survey and claims data. All authors contributed to the interpretation of the results. All authors revised the manuscript critically. All authors read and approved the final manuscript.