The Spanish National Health Survey

The dataset consists of self-reported height and weight from the micro-data of the SNHSs (waves of 1987, 1993, 1995, 1997, 2001, 2003 and 2006). This is the only Spanish national and sub-national representative health survey that has been conducted on a regular basis.

Subjects are randomly sampled by means of a complex multi-stage stratified sampling design according to age, sex and place of residence (region, province and municipality size). The adult health survey is taken among the non-institutionalized population aged 16+ and subjects are interviewed face-to-face. The wording of the questions on height and weight has remained constant across all waves. People were asked for their approximate height without shoes on and their approximate weight without clothes on. In the case of pregnant women, they were asked for their weight before pregnancy. For this study, respondents 20 years or older have been selected to prevent both large variations in body weight and height related to the completion of the physical growth process, and younger than 80 years of age in order to avoid sex, age and time/cohort combinations containing fewer than 500 cases. Moreover, previous studies on adult overweight and obesity from self-reported measures have mostly used ages up to 70 or 80 as irregularities are often observed in the information provided by elderly respondents (Kuczmarski et al., Reference Kuczmarski, Kuczmarski and Najar2001). Older people are also under-represented in the non-institutionalized population that is usually targeted for health surveys.

Cross-sectional and quasi cohort approaches

The SNHS is not a panel survey, meaning that people interviewed in one survey wave are not followed up in successive waves. Although this impedes performing a more typical longitudinal cohort study where the cohort acts as a baseline group whose health status is periodically checked, in demography and other social sciences alternative methods are applied to approximate a longitudinal study using cross-sectional data. Two illustrative examples are the calculation of life expectancy or the total fertility rate using synthetic cohorts.

In this study the different SNHSs, which covered almost two decades and included 120,550 persons aged 20–79, were aggregated into one database. Individuals were subsequently designated to so-called quasi birth-cohorts in function of their age and the year they were interviewed. The same dataset was recently used to show that average height – a personal characteristic that should not change over time once adulthood is reached and until the ageing-related shrinkage process sets in – varied insignificantly between surveys for all but the elderly population (Spijker et al., Reference Spijker, Pérez and Cámara2008). Hence, the data were considered suitable for the study of age- and time-specific changes in the average levels of weight and BMI of quasi birth-cohorts.

Notwithstanding, some small sample biases did have to be dealt with. As the elderly and the less-populated regions were oversampled during certain waves, samples were weighted with official population data from the National Statistics Institute (INE). Data were also screened for errors and omissions with respect to the variables age, sex, residence region, height and weight, reducing the sample size to 105,878 persons. In addition, the 1995 and 1997 waves were combined for the cross-sectional analyses due to their low sample sizes (respectively 4790 and 4970 respondents compared with 20,283 in 1987, 15,780 in 1993, 17,083 in 2001, 18,977 in 2003 and 23,995 in 2006).

The sample reliability and its comparability across time and quasi birth-cohorts was further increased by calculating 10-year birth cohorts and aggregating single ages into 10-year age groups, whereby age–cohort combinations with fewer than 500 respondents were also discarded. This provided a final sample size of almost 100,000 (see Table 1).

Table 1.

Number of after-screening sample cases used in the analysis by sex, age and quasi birth-cohort

Source: micro-data from the Spanish National Health Surveys 1987–2003. Own calculations.

On the strength and limitations of cross-sectional self-reported anthropometrics for trend studies

Self-reported anthropometric measures are known to be biased in function of the age, sex and actual anthropometric measures of a respondent. For instance, with regard to height, older adults tend to over-report it (as they recall it from early adulthood), as do those who are short (Gunnell et al., Reference Gunnell, Berney, Holland, Maynard, Blane, Frankel and Smith2000). However, even though an over-reporting of height among the elderly produces an underestimation of the BMI, the bulk of BMI changes between successive age groups can be ascribed to weight changes. This is because shrinkage occurs gradually after age 50 (Birrell et al., Reference Birrell, Pearce, Francis and Parker2005) and mainly at older ages, accounting for a total loss of about 2–5 cm depending on the age reached by the subject (Borkan et al., Reference Borkan, Hults and Glynn1983; de Groot et al., Reference de Groot, Perdigao and Deurenberg1996; Dey et al., Reference Dey, Rothberg, Sundh, Bosaeus and Steen1999). This bias is also less relevant for age-specific cohort differences in BMI. In addition, the average self-reported male height obtained by the SNHSs was also comparable to Spanish military recruitment records (mandatory until the mid-1990s, and accordingly contains height data for male cohorts born before the mid-1970s) and should therefore be taken as the maximum height reached by a given cohort before the ageing-related shrinkage process sets in. This is to say that self-reported height provides a type of age-standardized measure that allows trends in BMI to be interpreted by changes in weight status.

As for the weight component of BMI, overweight and obese persons are more likely to under-report their weight, particularly young adult females (Rowland, Reference Rowland1990; Nawaz et al., Reference Nawaz, Chan, Abdulrahman, Larson and Katz2001; Ezzati et al., Reference Ezzati, Martin, Skjold, Vander Hoorn and Murray2006). Self-reported measures may therefore not be valid for epidemiological or clinical studies where a high accuracy at the individual level is required. However, given the high correlation between both types of data – even among the elderly who are less susceptible to the cultural pressures that lead to under-report weight (Lawlor et al., Reference Lawlor, Bedrod, Taylor and Ebrahim2002) – self-reported measures may still be used to establish long-term trends within large populations. Likewise, earlier tests on the reliability of self-reported weight from the SNHSs showed that, at least at the aggregated level, age- and sex-specific patterns match standard biological growth and maturation cycles (Cámara & Spijker, Reference Cámara and Spijker2008). As the human physical maturation process concludes around age 25, any increase in weight observed for the aggregated age interval 20–29 years may be partly attributable to it. After that, weight is more dependent on the individual's control irrespective of existing differences in metabolism that influence the exposure to overweight and obesity and therefore any weight change should be mostly read in environmental terms. Additionally, for women both pregnancy prevention (e.g. contraceptive pills contain oestrogens that may cause metabolic changes) and motherhood (women permanently keep some of the weight gained during pregnancy) are also factors that contribute to weight gain (Bogin, Reference Bogin1991).

Regarding Spanish literature on reported height and weight biases, a regional health survey held in Catalonia (north-eastern Spain) in 2006 included both self-reported and actual anthropometric measures, thus allowing potential biases of self-reported height and weight of this Spanish subpopulation aged 15–65 to be estimated (Gil & Mora, Reference Gil and Mora2009). It was found that, as expected, the oldest age group (56–65) tended to over-report its height the most (1.42 cm). The largest bias for weight was found among those aged 25–35. In total, authors estimate a mean underestimation of 1.42 kg in self-reported weight and a 0.64 cm gap between self-reported and measured height (i.e. individuals over-reported their statures). In terms of BMI this corresponds to a 3.5% and 2.1% bias for women and men aged 15–65, respectively. Similar values were obtained when BMI categories were analysed. For instance, if corrections based on this regional sample were applied to the last SNHS in 2006, obesity prevalence would rise from 15% to about 18%.

Finally, beyond the nature and specific problems of self-reported data, an additional limitation of using serial cross-sectional data for the analysis of cohort trends is the selective effect of mortality. This is especially the case for underweight or obese individuals. For instance, while the proportion of obese persons of a particular birth cohort can be calculated, the proportion of that cohort who already died from obesity is unknown. This bias would be most relevant for the eldest cohorts and will be subject of specific comments in the discussion section.

After the required data from the seven SNHSs were screened, the BMI was calculated by dividing the respondent's weight by the square of their height (kg/m²). Body mass index has become a widely used indicator to assess the prevalence of overweight and obesity once adulthood has been reached (Rolland-Cachera, Reference Rolland-Cachera, Ulijaszek, Johnston and Preece1998). Indeed, the weight and height measurements are precise (highly repeatable), accurate (close to the true value) and valid (representing what they are thought to present), correlating highly with fatness (% fat, fat mass) but less well with height (Norgan, Reference Norgan1994). One important criticism, however, of only using BMI in epidemiological studies is the lack of potentially important information on body composition (e.g. % fat, skinfold thickness) and shape (relative skeletal dimensions of the body). For this reason Norgan (Reference Norgan1994), who examined the relationships between BMI and body composition and shape in different population groups, concluded that while differences in the relationships of body composition to BMI are minor over the range of BMI 20–25 kg/m², to interpret BMI in terms of body composition more specifically it is necessary to take into account sex, age and ethnicity (due to its relation to body shape). Although the aforementioned limitations are present in the SNHS data (i.e. on body composition and shape) the current study sample was mainly composed of Spanish nationals who are very homogenous in terms of ethnicity (Spanish nationals comprised 97% or more of the interviewees of the 1987 to 2003 surveys and 89% of the 2006 survey) and men and women and age groups were analysed separately. It is therefore unlikely that differences in body composition and shape will affect the results of BMI trends in a substantial manner.

Both continuous and the customary categorical BMI values are analysed, i.e. underweight (<18.5 kg/m²), normal weight (18.5–24.9 kg/m²), overweight (25.0–29.9 kg/m²) and obesity (>30.0 kg/m²).