European Project on Osteoarthritis (EPOSA): methodological challenges in harmonization of existing data from five European population-based cohorts on aging

The data of five existing longitudinal cohort studies are combined using post harmonization procedures. The five cohorts are population based, but vary in recruitment procedures and measurement instruments used. Table 1 provides an overview of the main characteristics of the five cohorts. The design and procedures of all five cohort studies have been approved by the Medical Ethics committee of the respective institutes.

The EPOSA project involves five cohort studies performed in five different countries. While the methodology of each cohort is described in detail elsewhere, here a brief description of their main characteristics is provided.

Germany is represented by the University of Ulm, Institute of Epidemiology and Medical Biometry and AGAPLESION Bethesda Clinic Ulm. The study on Activity and Function in the Elderly in Ulm (ActiFE-Ulm) [8] is embedded in a European funded study on the prevalence of COPD and asthma (Indicators for Monitoring COPD and Asthma - IMCA). A random sample of 7460 persons aged 65 years and over was selected from the population registers in Ulm, Neu-Ulm and Alb-Donau-Kreis. The recruitment phase started in February 2009 and finished in April 2010. In total, 1506 persons agreed to participate in the study. The primary focus is physical activity (as measured by sensor technology) and the consequences of physical activity for cognitive, emotional and social functioning.

Italy is represented by the Department of Medical and Surgical Sciences, University of Padova, and National Research Council (CNR), Aging Branch, Institute of Neuroscience, Padova. The Progetto Veneto Anziani (ProVA) study was originally designed to assess the effects of cardiovascular and osteoarticular diseases on disability [9]. Participants (65+ years, n = 3099) were living in two geographical areas of north-eastern Italy (Camposampiero and Rovigo) near the city of Padova in the Veneto region. Participants were interviewed at home, and invited to the clinic for a general physical examination and additional diagnostic tests. Finally, physicians performed a medical chart review.

The Netherlands is represented by the VU University Medical Center, EMGO Institute for Health and Care Research, Amsterdam. Data were used from the Longitudinal Aging Study Amsterdam (LASA) [10], an ongoing cohort study of predictors and consequences of changes in physical, cognitive, emotional and social functioning in older persons. In 1992/93 the first measurement cycle was completed within a random sample of older persons (55-85 years). The sample was selected from the population registers in 11 municipalities in the Netherlands and stratified for age, sex and level of urbanization. Each measurement cycle consists of a main interview, a medical interview (both face-to-face) and a self-administered questionnaire. For the current study, data were used from 1669 persons aged 65 years and older who participated in the fourth measurement cycle in 2001/2002.

Spain is represented by the Universidad Autónoma de Madrid Centro Universitario de Salud Pública. The study included is Ageing in Peñagrande [11]. Participants (65+) were selected from the Register of Health District Area in the neighbourhood of Peñagrande, which is part of the Fuencarral district in Madrid. Of the 4,244 persons of 65 years and over living in Peñagrande as of July 31, 2007, a random, age and sex stratified sample of 1250 persons was drawn of whom 1110 were eligible (88.8%) and 814 consented to participate. The baseline measurement took place in 2007-2008 and consisted of face-to-face interviews.

The United Kingdom is represented by the University of Southampton, Southampton General Hospital, MRC Lifecourse Epidemiology Unit. The Hertfordshire Cohort Study (HCS) was designed to examine the relationship between growth in infancy and the subsequent risk of common adult diseases, including osteoporosis and osteoarthritis [12]. Potential participants in HCS were born during 1931-1939 in Hertfordshire, and still lived there during the period 1998-2003 (n = 8,650). Persons who consented were visited at home where a structured questionnaire was administered (n = 3,225). Participants then attended a local morning clinic where a variety of investigations were performed (n = 2,997). In 2004/05, a follow-up clinical study was performed in East Hertfordshire (n = 966). In 2007, a postal questionnaire enquiring about clinical events over the follow-up period, was sent to all participants throughout Hertfordshire (n = 2,299). For the current study, data from the baseline measurement was used from persons who visited the clinic and who were 65 years and over (n = 1,879).

When various samples are used to study one particular hypothesis, the problem of heterogeneity between samples emerges. To cope with these differences, two approaches are possible: centralized analyses, which include harmonization of the datasets to create one large dataset in which the analyses are done, or coordinated independent analyses, in which the analyses are done in each dataset and the results per dataset are pooled afterwards [13]. In the current study, preference was given to centralized analyses, because this approach enables analyses of which covariates explain differences between the countries. Post harmonization guidelines were developed to overcome the differences in measurement instruments between the cohorts. The harmonization guidelines used in the Comparison of Longitudinal European Studies on Aging (CLESA) project served as an example for the current study [14].

Figure 1 shows a schematic presentation of the harmonization procedures. Data from all cohorts were merged based on the common denominator available for each variable if at least three cohorts had the information relative to that specific measure available. If not all countries had the same detail of information, detailed information was lost in the harmonized variable. Alternatively, the same variable could be created for less than all five but more than two countries, if this led to preservation of detailed information. The cohort without detailed information was coded as having missing values on this more detailed variable. Cut-off values for categorization of variables were based on cut-off values from the literature. If no cut-off values were available, cut-off values were based on biological criteria or the distributions in the cohorts. Figure 1 also lists all harmonized variables of which the harmonization was successful (i.e. no or very little loss of information). Although information on radiographic OA and site-specific clinical OA was available in two cohorts only, it was included in the harmonized database, as these are main variables in the assessment of OA.

The involved partners sent all relevant data with variable and value labels in English to the coordinating center (VU University Medical Center, Amsterdam, the Netherlands). The coordinating center then developed the harmonized dataset in cooperation with the partners. A code book (harmonization guidelines) was made explaining how the variables were harmonized (see additional file 1: Harmonization guidelines).

Datasets were merged, but sample sizes and age and sex distributions differed across cohorts due to differences in recruitment procedures and inclusion criteria. Sample weights were calculated to adjust for differences in recruitment procedures and inclusion criteria. The weights were calculated per sex and per five-year age category, using the following formula: W = Nexp/Nobs, with the Nobs being the number of persons in a specific age/sex category in the cohort, and Nexp being the number of persons in a specific age/sex category in the population [14]. The expected number of persons were derived from the European Standard Population [15]. Nexp for the European population were based on the numbers for 2001 (average baseline year) and included 27 countries. The weights for each of the countries were based on the numbers for that country given their specific baseline year.

Table 1 shows the main characteristics of the five cohorts that were included in the EPOSA study. The differences in recruitment criteria become evident in the differences in baseline characteristics of the five cohorts. Table 2 shows the characteristics of the specific measurement cycle of each cohort that is used in the EPOSA study. The ActiFE, ProVA, LASA and Peñagrande studies were similar regarding age distribution, however, in HCS, the mean age was approximately 10 years lower. In ActiFE and HCS the percentage of men was higher than in the other studies. Note that the years of baseline in EPOSA differ across the cohorts. For each cohort, the most recent measurement cycle that contained comprehensive information on OA was chosen as the baseline measurement in EPOSA. For example, 2001/02 was chosen as the baseline year in LASA, because this was the most recent measurement cycle in which both self-reported and general practitioners information on the presence of OA was available.

In this paper it is not possible to describe the harmonization process for all variables included in the EPOSA database. Please refer to the harmonization guidelines in additional file 1 for a detailed description of the algorithms used per variable. Here, harmonization of OA, ADL limitations and social participation is described in detail to illustrate the challenges and solutions in the harmonization process.

OA case definitions used in the literature vary by site of joint involvement (knee, hip, hand, generalised, unspecified) and by method (self-report, clinical diagnosis/classification, plain radiography or other imaging, combinations of these). Each of these different approaches to OA case definitions was evident in the five cohorts, highlighting the lack of accepted standard case definitions in current European studies. As OA is believed to be a collection of disorders with shared features rather than a single disease entity, several definitions of OA may be the best strategy. Thus, it was decided not to develop one harmonized variable for OA, but rather develop several OA variables (site-specific/non-specific self reported OA, site-specific/non-specific clinical OA and radiographic OA) which can be used according to specific research questions. Table 3 shows which questions or measurements were used to construct the self-reported, clinical and radiographic definitions for knee, hip, hand and non-specific OA in each of the cohorts. Note that not all definitions were available for all cohorts. However, because OA is our main variable of interest in this study, we made an exception regarding the harmonization procedure: all types of OA definitions were constructed, even when data from only two countries were present.

In ProVA, persons were classified by physicians as having definite, possible or no OA in the knees, hips or hands. In the other countries, however, dichotomous definitions for clinical OA were used (yes/no). To test whether the possible-category in the ProVA data should be interpreted as either definite/yes or no, two statistical procedures were used: (1) optimal scaling technique, and (2) regression models with known OA-correlates. The optimal scaling technique used is called Princals, which is based on non-linear factor analysis [16]. This technique quantifies the distance between the scaling points (i.e. definite, possible, no). Regression analyses were done to test the associations with use of pain/anti-inflammatory drugs, physical performance, ADL limitations, and grip strength (left and right). The optimal scaling analysis and regression analyses showed smaller distances between 'definite' and 'possible' than between 'possible' and 'no' for clinical knee, hip and hand OA (data not shown). Therefore, 'possible' was interpreted as 'definite clinical OA' in the ProVA cohort.

Table 4 presents the prevalence rates for knee, hip, hand and non-specific OA per definition used. The lowest site-specific prevalence rate was found for radiographic knee OA in the UK, the highest prevalence rate was found for clinical knee OA in ProVA. In LASA and Peñagrande, prevalence rates for self-reported OA seem to be higher than for clinical OA, however, in Italy, a reverse pattern was found. The agreement between self-reported and clinical OA varied between 77 and 82%. In HCS, prevalence rates were higher for clinical than radiographic knee OA, but lower for clinical than radiographic hand OA. The agreement between clinical and radiographic OA was 64% for the hand and 79% for the knee.

In all five cohorts, a variety of Activity of Daily Living (ADL) variables were assessed in structured interviews (Table 5). First it was investigated which items were available in all cohorts. The item "do you have difficulty with walking up and down from a staircase?" was available in all cohorts and therefore included in the harmonized ADL score. In the LASA study, the item "do you have difficulty with taking a shower or bath?" was not available, but a proxy was used from the quality of life index score of the EQ-5D: "how much difficulty do you have with washing or dressing yourself?" Subsequently, this item was available in all cohorts and included in the harmonized ADL score. Second, it was investigated which items were available in most cohorts, which resulted in the selection of seven items in total (see Table 5). Of these seven items, five items were used and per item 0 to 2 points were assigned according to the level of difficulty with that activity (0: no difficulty, 1: with difficulty, 2: unable to do alone). Thus, an ADL score ranging from 0-10 was made in all cohorts, but the actual items used within this score partly differ across cohorts. However, we tried to match the items as closely as possible according to the underlying physical demands. For example, the ActiFE and LASA studies did not have the item "Do you have difficulty with using your fingers to catch or handle small objects?". Instead, the item "Do you have difficulty cutting your toenails?" was used, because this activity also requires fine motor skills of the hands.

All five cohorts had some information on social participation. In the ActiFE study the Lubben Social Network Scale was used, which assesses social isolation in older adults by measuring perceived social support received by family and friends. Two examples of items of this questionnaire are "How many relatives do you see or hear from at least once a month?" and "How many relatives do you feel at ease with that you can talk about private matters?". If social participation is regarded as an objective measurement of social activity, it is not possible to include data from ActiFE, as the questionnaire measures self perceived social support rather than engagement in social activities. In ProVA, Peñagrande, LASA and HCS, the frequency of participation in social activities was measured. In HCS the social activity score was derived from the social health questionnaire which includes the participation and frequency of 13 activities, such as positions of office or visiting friends. In Peñagrande, only two questions were asked regarding the frequency of caring for sick persons and caring for children. In LASA, respondents were asked whether they were involved in social organizations such as a political party or sport club and how often they participated in activities or meetings of these organizations. Although the frequency of social activities was assessed in four cohorts, the response categories varied: e.g. in ProVA the number of weekly hours devoted to social activities was asked, while in HCS the response categories were "weekly", "monthly" or "less often". Thus, the number and types of activities and the response categories differ greatly across the cohorts. It was therefore concluded that the social participation variables from the different cohorts were too heterogeneous to be harmonized.