1Two European surveys – the Survey of Health, Ageing and Retirement in Europe (SHARE) and the Generations and Gender Survey (GGS) – are widely used in demography, notably to study ageing populations. While they cover different respondent age groups and have different objectives, they share many points in common: both surveys are longitudinal, contain equivalent questions and topics, and include large samples of respondents aged 50-80 in seven European countries. So it should be possible to use them for joint analyses – on condition that they are comparable. In this article, Katherine Keenan, Else FoversKov and Emily grundy test GGS and SHARE for comparability by looking at the consistency of responses across the two surveys. Using as examples two commonly used health indicators, self-rated health and long-standing illness, and their association with level of education and marital status, the authors highlight the importance of weighting, question wording and the placement of questions in the questionnaire to explain the discrepancies between the two surveys.
2Understanding age associated changes in socio-demographic circumstances, health, resources and activity patterns is a key priority in Europe given substantial past and projected future increases in the representation of older people in the population (United Nations, 2013). High quality, representative longitudinal data are required as a basis for developing this understanding. To this end, considerable resources have been devoted to establishing comparable large scale cross-national longitudinal data sets, notably the Survey of Health, Ageing and Retirement in Europe (SHARE) (Börsch-Supan et al., 2013), and the Generations and Gender Survey (GGS) (Vikat et al., 2007), both of which are freely available and widely used. While they both collect data on older people, the surveys have different objectives. The GGS was primarily developed to underpin the study of family and intergenerational processes from young adulthood to old age, while the focus of SHARE is on ageing. Both studies were initiated in the early 2000s and while the quality of the data collected in each has been investigated through comparison with other sources, including national population data and European population surveys such as European Union Statistics on Income and Living Conditions (EU-SILC) (Börsch-Supan et al. 2005; Croezen et al., 2013; Fokkema et al., 2014; Vergauwen et al. 2015), they have not, to date, been systematically compared with each other. We compare data from SHARE and GGS where they cover the same countries and age ranges.
3The aims of this paper are to present sources of differences in the SHARE and GGS surveys and to investigate whether common health measures drawn from the two surveys provide comparable information, either in terms of prevalence or their patterning by socio-demographic characteristics. We start by examining the GGS and SHARE survey methodologies, then compare data on common indicators including age, gender, education, fertility, marriage and health. We go on to assess the comparability of results from multivariate regression modelling of associations between two socio-demographic indicators, education and marital status, with measures of health.
I – Methodology: comparison of SHARE and GGS surveys
Survey design
4Baseline SHARE and GGS surveys have been carried out in ten common countries: Austria, Belgium, Estonia, France, Germany, Hungary, Italy, the Netherlands, Poland and Sweden. We compared measures from seven of these countries for the population aged 50-80 years. We excluded Italy and Austria because their GGS data did not include respondents aged 65 and over. At the time of writing the Swedish GGS data was not available so was initially excluded. We also note that the GGS and SHARE surveys were generally not conducted in the same year, so the time separating any two surveys may vary (by up to eight years). [1] This could limit their comparability.
5Table 1 shows the main characteristics of the sampling procedures and fieldwork for the surveys included in this study: GGS surveys from wave 1, and SHARE surveys from the baseline wave (Estonia, Hungary and Poland joined the SHARE survey at wave 2 or later). All countries except Estonia and the Netherlands (GGS) used sampling strategies based on an external source, with the most common sampling frames being the census or population register. In most countries the sampling frame was designed to provide coverage of the population living in private households, although some countries and surveys also included those living in institutions. [2]
Main characteristics of sampling and fieldwork for GGS Wave 1 and SHARE baseline data (common countries except Austria, Italy and Sweden)(a),(b),(c),(d),(e),(f)
Main characteristics of sampling and fieldwork for GGS Wave 1 and SHARE baseline data (common countries except Austria, Italy and Sweden)(a),(b),(c),(d),(e),(f)
na: Not available.(a) PPS = Probability Proportional to Size; SRS = Simple random sampling; RR = random route; SysR = Systematic sampling with a random start.
(b) Statistics Belgium conducted fieldwork from Feb 08-Apr 09, and TNS Dimarso from Nov 09-May 10.
(c)”Average rate” calculated by Fokkema, Kveder, and Liefbroer (2014).
(d) and their partners/spouses of any age.
(e) Aquitaine, Ile-de-France, Languedoc-Roussillon, Nord-Pas-de-Calais, Pays de la Loire, Rhone-Alpes.
(f) Refers to the sampling characteristic of wave 1 of “Turning points of the Life-Course Program”.
6Reflecting the different objectives of the surveys, a major difference between SHARE and GGS is the target population. In SHARE it was defined as all households with at least one member aged 50 years or older, and within these households, all individuals aged 50 years or older. Respondents’ spouses/ partners were also included, regardless of age (Börsch-Supan and Jürges 2005). By contrast, in the GGS the target population was individuals aged 18-79 [3] years and only one individual from each household was interviewed. Response rates for both the GGS and SHARE surveys were lowest in Belgium (43.8% and 39.2%, respectively). The Netherlands GGS survey also had a rather low response rate of 44.7%. However in most other countries response rates for both surveys were over 60%.
Analysis sample
7We selected men and women aged 50-80 years at the time of survey, based on their reported dates of birth (50-79 in the Hungarian GGS data). Partners outside the SHARE age range (less than 50 years) were excluded because they are not a representative sample. This resulted in samples of between 2,255 and 10,447 individuals for each of the surveys and countries included (Table 2). The proportion of respondents with missing values on the variables considered ranged between 0.1% and 4.1%. The different timings of the GGS and SHARE surveys, especially for Belgium, Estonia, Hungary, and Poland, mean that the surveys include survivors of different birth cohorts (see Table 2), who could be subject to different time trends in some variables. To try to assess whether these different timings affected comparability we made additional comparisons using a subset including only the exact equivalent birth cohorts. However we acknowledge that this means we are comparing cohorts who have survived to different ages, and this selection issue may bias the comparability of health indicators, especially when the period between two surveys is long.
Timing of surveys, sample size and birth cohorts
Timing of surveys, sample size and birth cohorts
Socio-demographic measures
8We compared distributions by gender, age, education, marital status, and number of children. Age was categorized into 5-year groups. Both SHARE and GGS used the International Standard Classification of Education (ISCED) 1997 framework [4] to classify educational level and we distinguish between a low level of education (ISCED 0-2, including no education, primary and lower secondary) versus medium and higher level of education (ISCED 3-6, including upper secondary and non-tertiary post-secondary, and tertiary education).
9We used a de jure measure of marital status which was grouped into four categories: never-married, married, divorced, and widowed. For the main summary and regression analyses this was dichotomized into married versus non-married. In the GGS, marital status was commonly derived from answers to a series of questions on marriage and partnership history, and the resulting variable had the four categories as above. In SHARE we used a single question which asked the respondents if they were 1) married and living together with their spouse, 2) in a registered partnership, 3) married and living separated from their spouse, 4) never-married, 5) divorced or 6) widowed. We grouped together the first three of these categories in SHARE as “married”. In SHARE and GGS, if respondents reported (in other questions) that they were unmarried but living with their partner we coded according to their reported marital status (never married, divorced or widowed).
10As a summary indicator of fertility, we grouped number of children into five categories: 0, 1, 2, 3 and 4+, and for the main summary dichotomized it into childless versus non-childless. Number of children in SHARE was measured by asking “How many children do you have that are still alive? Please count all natural children, fostered, adopted and stepchildren, including those of your husband/your wife/your partner”. The SHARE wave 1 questionnaire also asked specifically whether children were biological or non-biological, but only for the first four children reported. A similar count of number of biological, adopted, foster and stepchildren still alive was obtained in GGS by combining information from the household roster and questions on non-resident children. However, lack of information on non-resident stepchildren in the Polish and Estonian surveys and on deaths of stepchildren in the French survey may result in slight estimation differences for those countries.
Health measures
11We concentrate on two health measures: self-rated health (SRH) and presence of long-standing illness or chronic conditions.
12Differences in question wording mean that SRH was directly comparable across both surveys only in Belgium, France, Germany, and the Netherlands. GGS wave 1 used the European variant of SRH (Robine et al., 2002) which has an ordered scale of 1-5 ranging from “very good” to “very poor”. The SHARE baseline surveys in Belgium, France, Germany, and the Netherlands used both the European variant of SRH and the US variant, which has an ordered scale of 1-5 ranging from “excellent” to “poor”. The SHARE surveys from Estonia, Hungary and Poland collected SRH using only the US version. For the European variant, we dichotomized responses into “good” (those reporting very good or good health) or “poor” (fair, bad, or very bad health). For the US scale, respondents reporting excellent, very good or good health were considered to have good SRH while respondents reporting fair or poor health were considered to have poor SRH. Previous studies have shown that the US and European versions of the SRH question are not exactly comparable (Jürges et al., 2008), and that responses to the US scale usually produce estimates of better health than responses to the European version. Where available, we show distributions of SRH using both European and US variants, which allowed us to see the effect of wording differences.
13Long-standing illness (LSI) was measured in the GGS by asking “Do you have any long-standing illness or chronic conditions?” and in SHARE by asking “Do you have any long-term health problems, illness, disability or infirmity?” with response options yes or no. In addition the Estonian GGS specified an illness “lasting 3 months or more”.
Analytical methods and post-stratification weights
14First, to assess representativeness we compared the weighted age and gender distributions in the GGS and SHARE against each other and against national population data as reported in EUROSTAT for the year the survey took place. [5] Next we undertook descriptive analyses to compare the weighted proportions with particular characteristics in SHARE and GGS by country and gender. Comparisons of the non-dichotomized versions of education, marital status, and number of children are presented in appendices A.1 to A.3. We used the post-stratification weights provided with the GGS and SHARE surveys for the comparisons. In the GGS the weighting factors vary between countries, and for some countries such as Hungary no information was available (Fokkema et al., 2014). We provide summary information on GGS weighting factors in Appendix Table A.1. For most countries, the weights aimed to adjust the sample so that it was nationally representative on important aspects such as age, gender, urbanization, region and household size. The GGS weights for Belgium, Estonia and Sweden did not take account of household size, so in those countries we recalculated the weight to include it, using population data on household size distributions from the year of sampling downloaded from EUROSTAT (EUROSTAT, 2015). This adjustment is important because only one person per household was interviewed in the GSS (resulting in higher sampling probabilities for people living alone) while in SHARE all individuals aged 50 and over were sampled. The Poland GGS did not include any weights, so we also weighted the data for household size (but not for other factors such as gender or age). For the SHARE surveys we used the “calibrated cross-sectional weights” which adjust for unequal sampling probabilities, and the known distributions by gender and age in the general population [6] (Mannheim Research Institute for the Economics of Aging, 2013).
15Despite variations in the proportions of respondents with particular characteristics, we would expect the socio-demographic gradient in health to be consistent across the surveys. To assess whether this was the case we fitted multivariable logistic regression models for associations between two sociodemographic indicators (education and marital status) and two health indicators (SRH and long-standing illness). We did not use weights for the multivariable models, but adjusted them for age (continuous measure), marital status, and education. To test whether the multivariable estimates were significantly different in SHARE and the GGS, we combined the datasets, and performed pooled models with an interaction term for data source.
Representativeness
16In previous studies, both surveys have been compared separately with national population estimates to assess their representativeness (Börsch-Supan and Jürges 2005; Fokkema et al., 2014). For the GGS, the use of post-stratification weights reduced (but did not eradicate) deviation from whole population sources for age, gender and region, but did not correct biases by marital status or education (Fokkema et al., 2014). Other studies have found that fertility and marriage rates for older cohorts of the GGS are underestimated when compared with population data even when the sampling weights are applied, although period rates from 1970 broadly approximate those in population data (Kreyenfeld et al., 2013; Vergauwen et al., 2015). In SHARE, use of the post-stratification “calibrated weights” produces estimates that are very similar to the target populations (Börsch-Supan and Jürges, 2005). As a further check we compared weighted distributions by age group and gender for the SHARE and GGS samples with distributions from national population data for the respective survey years (Figure 1). Overall, the SHARE and GGS samples had very similar age structures, with the exception of Hungary which in SHARE had a markedly higher proportion of men and women aged 55-59 years. This is partly due to a 5-6 year difference in the timing of the surveys which meant that the later survey included a larger proportion of people born during the post-war baby boom. The largest disparities between national and survey data were seen for women in the Belgian and German GGS samples, men in the French and Polish GGS samples and the Hungarian SHARE samples. The weights in the Polish GGS did not adjust for age so it was not surprising to see more differences. It was common in the GGS (Estonia, France, and Poland) for weighted survey estimates to under-represent men and women aged 50-54 years. This could be related to the fact that a different age categorization with a larger age bracket (45-64 years) was used to construct the GGS weights whereas in SHARE the weights adjust from age 50. Figure 2 shows the weighted proportion of women in SHARE and GGS compared with national population data. Unsurprisingly the largest disparity was in the Poland GGS where the weights did not adjust for gender. The survey gender distributions were most similar for Germany and Hungary. In no country was the difference larger than 2 percentage points.
Weighted age distributions of men and women aged 50-80 years in SHARE and GGS, compared with national population data for the respective survey years
Weighted age distributions of men and women aged 50-80 years in SHARE and GGS, compared with national population data for the respective survey years
Proportion of women in the population aged 50-80 years according to SHARE and GGS, compared with national population data for the survey years
Proportion of women in the population aged 50-80 years according to SHARE and GGS, compared with national population data for the survey years
II – Results
Distributions by socio-demographic characteristics and health
17Tables 3 and 4 show the weighted (but unadjusted) distributions of sociodemographic and health indicators in SHARE and GGS, for women and men respectively. The differences between SHARE and GGS in the proportion married and childless were relatively small (results were similar using the non-dichotomized measures of marital status and number of children). In Belgium, the GGS sample had proportionally more childless respondents than the SHARE sample. Overall the largest differences between the surveys in the distributions were seen for education and long-standing illness, with similar differentials for men and women.
Weighted distributions by socio-demographic variables in GGS and SHARE, women aged 50-80 years(a)
Weighted distributions by socio-demographic variables in GGS and SHARE, women aged 50-80 years(a)
na: Not available.(a) Questions were worded differently in GGS and SHARE.
Weighted distributions by socio-demographic variables in GGS and SHARE, men aged 50-80 years(a)
Weighted distributions by socio-demographic variables in GGS and SHARE, men aged 50-80 years(a)
na: Not available.(a) Questions were worded differently in GGS and SHARE.
18In Estonia and Hungary, higher proportions of people reporting a low educational level were recorded in the GGS samples, whereas in the Netherlands and Poland, higher proportions were recorded in SHARE. In the Polish GGS, 30% of women had low education, compared with 51% in SHARE. The proportion of Polish women with higher education was also twice as high in the GGS as in SHARE (12% versus 5%) (see Appendix Figure A.1). These differences might reflect inadequate weighting in the Polish GGS resulting in different age structures, which suggests the importance of adjusting for basic demographic variables such as age and sex when comparing cross-country estimates. When the analysis was repeated using equivalent birth cohorts instead of age groups to assess if any differences were introduced by the different timing of the surveys, the differences in education completely disappeared for Estonia (where the SHARE and GGS surveys were conducted 5-6 years apart). In Hungary and Poland (where surveys were conducted 6-7 and 4-5 years apart, respectively), the differences were attenuated but remained substantial.
19When using the European variant of SRH in both surveys, there were few substantial differences. The US variant of SRH produced a lower prevalence of poor health than when using the European variant, which is consistent with previous studies (Jürges et al., 2008). The proportions with poor SRH, however it was measured, were highest in the Eastern countries, Estonia especially.
20As expected from the different question wording, the proportions reporting LSI were consistently different between the surveys (higher in SHARE than in GGS). All countries except the Netherlands and France had differences in excess of 10 percentage points, and in Germany and Estonia it approached 30 percentage points. The proportional differences between the surveys were not correlated with the overall prevalence of reported poor health in each country.
21Figures 3 and 4 show results from several multivariable logistic regression models fitted to explore the associations between socio-demographic factors and poor health in the different countries and surveys. Figure 3 shows (unweighted) associations between low education (ISCED score 1-2) and two binary outcomes – reporting poor SRH, and reporting having an LSI, adjusted for age and marital status. All odds ratios in Figure 3 are above one, meaning that low education is associated with higher odds of reporting poor SRH or LSI. For some countries (Hungary and Poland particularly), the GGS sample was larger than the SHARE sample, and standard errors accordingly lower, which may affect differences in significance levels. Models fitted to data from each survey pooled across countries showed that odds of poor SRH and LSI associated with low education were slightly larger in the GGS than SHARE although the confidence intervals overlap. For SRH, the same pattern of effect was seen for men and women, but there was some gender variation in results from the models for LSI. Some of the country-specific coefficients show substantial differences, most notably in Hungarian men and women where the SHARE analysis showed higher odds of poor health by low education than the GGS results. This cannot be explained by large differences in the prevalence of poor health in general. In very few countries were the odds from both surveys similar (possibly France, the Netherlands and Germany were most consistent). As described in the methods section, we combined the SHARE and GGS data, and tested for significant differences in the associations by survey. In most cases the differences were non-significant, but notable exceptions were SRH and LSI in Hungary, and SRH in Poland (see Appendix Table A.2).
Unweighted associations between low education and reported poor self-rated health/long-standing illness, adjusted for age and marital status (ORs and 95% CIs)(a),(b)
Unweighted associations between low education and reported poor self-rated health/long-standing illness, adjusted for age and marital status (ORs and 95% CIs)(a),(b)
(a) SRH question: European variant in GGS, US version in SHARE.(b) GGS and SHARE surveys were conducted more than 5 years apart.
Notes: OR: Odds ratio. Reference group: medium or high education. CI: Confidence interval.
Significance levels: * p<0.05 ** p<0.01 *** p<0.001.
Unweighted associations between unmarried status and reported poor self-rated health/long-standing illness by country, gender and survey; adjusted for age and education (ORs and 95% CIs)(a),(b)
Unweighted associations between unmarried status and reported poor self-rated health/long-standing illness by country, gender and survey; adjusted for age and education (ORs and 95% CIs)(a),(b)
(a) SRH question: European variant in GGS, US version in SHARE.(b) GGS and SHARE surveys were conducted more than 5 years apart.
Notes: OR: Odds ratio. Reference group: medium or high education. CI: Confidence interval.
Significance levels: * p<0.05 ** p<0.01 *** p<0.001.
22Figure 4 shows associations between being unmarried (never married, divorced or widowed) and the two health outcomes, adjusted for education and age. Results from analyses of samples pooled across countries showed that being unmarried was associated with poor health in both men and women. However, country-specific analyses showed that this association only reached conventional levels of statistical significance in Belgium, Germany and, for women only, the Netherlands. There was more similarity between the surveys in the country-specific odds ratios than appeared the case on the analyses of differentials by low education, particularly for Belgium and Germany and, to a lesser extent, France and Hungary. Estimates for Poland appeared inconsistent, being higher in the GGS than SHARE. We found evidence that the estimates were different between SHARE and GGS for Polish women and the outcome of LSI.
III – Discussion and conclusion
23This article makes a unique contribution by comparing the sociodemographic and health data of the population aged 50-80 years from the common SHARE and GGS surveys, and complements previous studies which have compared SHARE or GGS data to national population data or other surveys such as EU-SILC (Croezen et al., 2013; Fokkema et al., 2014; Kreyenfeld et al., 2013; Sauer et al., 2012; Vergauwen et al., 2015). In addition, it extends previous work by comparing the socio-demographic gradient in health variables in the two surveys.
24We assumed that estimates of basic demographic variables from both surveys would be similar after applying post-stratification weights but this was not always the case. Comparisons of weighted estimates of age from the two surveys with national data (Figure 1) show that the post-stratification weights provided, particularly in the GGS, were not always successful at adjusting to the target population. The factors used in the weights were inconsistent between the GGS surveys, and between the GGS and SHARE, so differences seen could partly be a consequence of this. The surveys are relatively comparable in terms of distributions by age, gender, marital status, proportions childless, and poor SRH (provided the same variant of SRH was used). Differences in education in Estonia, Hungary and Poland could be related to the different timing of surveys. In these countries the GGS and SHARE surveys were conducted 4-7 years apart, and changes in compulsory schooling occurring during the Soviet period could have resulted in changed distributions of education for older cohorts (Pennar et al., 1971; Róbert, 1991; Szebenyi, 1992). However, it is worth noting that even when comparing the equivalent birth cohorts some differences remained for Hungary and Poland. There were also some differences for the Netherlands despite the two surveys being conducted during a similar time frame. These differences may suggest inconsistencies in matching responses to ISCED codes.
25The differences according to survey in the prevalence of LSI highlight how sensitive health reporting is to question wording and question order (Dubuc et al., 2004; Freedman et al., 2004; Jette, 1994; Picavet and van den Bos, 1996). In SHARE the question on LSI mentioned “disability and infirmity”, implying permanence which usually leads to lower reporting of health problems (Picavet and van den Bos, 1996). However, survey content and the ordering of questions are also known to be important (Bowling and Windsor, 2008; Freedman et al., 2004), with some studies suggesting that question order may have a stronger effect on older people’s health assessments compared with younger people’s (Crossley and Kennedy, 2002). As commonly recommended, the question on SRH in both surveys was asked at the start of the health module, but this may produce biases in different cultural contexts (Lee and Grant, 2009). The GGS collected a rather limited range of information on health whereas in SHARE it was a major part of the survey. Respondents may therefore have been more focussed on considering their health than in the GGS. Our finding that SHARE respondents report better SRH is consistent with a previous study comparing SHARE with other surveys, namely the Health Interview Surveys (HIS), the European Social Survey (ESS), and the EU Statistics on Income and Living Conditions (EU-SILC) survey (Croezen et al., 2013).
26The observed differences in the educational or marital status gradient on health questions could in many cases be explained by sampling variability. The main exception was Hungary, and to some extent Poland, where large, significant differences remain which might reflect differences between the surveys in distributions by education. There were also more differences between the surveys for multivariable associations with education compared to marital status, where the underlying distributions of the independent variable were more unequal. Overall this points to the importance of examining the distribution of underlying variables in detail when interpreting multivariable results.
27The comparison has highlighted important differences between the surveys in their objectives and target populations which should guide researchers when choosing which survey to use. For studying the population over 50, the weighted estimates in SHARE usually better approximate age and sex distributions in the target population, however this could be resolved for the GGS surveys by developing more effective weights. Measurement consistency is crucial for cross national comparisons, and in some respects SHARE is more straightforward because the same instrument was used in all countries, rather than being harmonized post-hoc as in the GGS surveys. The SHARE survey asked a larger variety of health questions whereas the GGS has greater breadth in other areas such as attitudes. Our comparisons of health variables show that (contingent on similar wording) distributions and multivariable associations are relatively similar, but that researchers using the surveys jointly should pay attention to differences in question wording and representativeness when analysing the data and interpreting results.
Acknowledgements
This research was funded by a European Research Council Advanced Grant (FAMHEALTH) awarded to Professor Emily Grundy, under the European Union’s Seventh Framework Programme (FP7/2007-2013)/ ERC, reference number 324055. GGS data were obtained from the Generations and Gender Programme Data Archive and were created by the organizations and individuals listed under each respective country name at http://www.ggp-i.org/online-data-analysis.html.This paper uses data from SHARE Wave 4 release 1.1.1, as of March 28th 2013 (DOI: 10.6103/SHARE.w4.111) and SHARE Waves 1 and 2 release 2.6.0, as of November 29th 2013 (DOI: 10.6103/SHARE.w1.260 and 10.6103/SHARE.w2.260). The SHARE data collection has been primarily funded by the European Commission through the 5th Framework Programme (project QLK6-CT-2001-00360 in the thematic programme Quality of Life), through the 6th Framework Programme (projects SHARE-I3, RII-CT-2006-062193, COMPARE, CIT5- CT-2005-028857, and SHARELIFE, CIT4-CT-2006-028812) and through the 7th Framework Programme (SHARE-PREP, N° 211909, SHARE-LEAP, N° 227822 and SHARE M4, N° 261982). Additional funding from the U.S. National Institute on Aging (U01 AG09740-13S2, P01 AG005842, P01 AG08291, P30 AG12815, R21 AG025169, Y1-AG-4553-01, IAG BSR06-11 and OGHA 04-064) and the German Ministry of Education and Research as well as from various national sources is gratefully acknowledged (see www.share-project.org for a full list of funding institutions).
Results of testing for significant differences in odds ratios estimated from SHARE and GGS for the associations shown in Figures 3 and 4
Results of testing for significant differences in odds ratios estimated from SHARE and GGS for the associations shown in Figures 3 and 4
ns: Non-significant (p ≥ 0.05).
Marital status of men and women aged 50-80 years in SHARE and GGS
Marital status of men and women aged 50-80 years in SHARE and GGS
Notes
-
[*]
London School of Economics.
Correspondence: Katherine Keenan, London School of Economics, Department of Social Policy, Houghton Street, London, WC2A 2AE, e-mail: k.keenan1@lse.ac.uk -
[1]
The Swedish SHARE wave 1 was conducted in 2004; the Swedish GGS wave 1 was conducted in 2012-2013.
-
[2]
The Estonian GGS, and the German, Hungarian, Netherlands SHARE stated that institutionalized people were included in the sampling frame. However, the proportion of such individuals aged 50-80 years in our analysis was negligible: in the Estonia GGS there were 7 (<0.1%), and in SHARE they were not identifiable.
-
[3]
In the GGS countries which sampled to age 79, the data includes some people aged 80, who had a birthday between sampling and fieldwork. The exceptions to this were GGS Estonia which deliberately sampled 80 year-olds, and GGS Hungary which sampled only to age 78.
-
[4]
In GGS, many questionnaires used their own country-specific classifications and these were post-coded into ISCED.
-
[5]
Population data from EUROSTAT reports age distributions on 1 January, whereas in the survey data we calculated age at the time of survey using dates of birth.
-
[6]
We chose the SHARE “calibrated weights” in preference to the “design weights” which adjust only for unequal sampling probabilities.
