The discussion of the relationship between risk of getting divorced and individual weight change has not advanced much so far. In their initial article Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007) indicate that a higher divorce rate at country level is associated with a lower body mass index (BMI) among middle-aged married Europeans. This assumption was questioned by the re-analysis of the original data applying multilevel models (Schneider & Grimps, Reference Schneider and Grimps2012). We raised, in the main thrust of our comment, objections to the application of traditional regression analysis (OLS) on hierarchical structured data. The analysis of the original data using multilevel modelling led to the suggestion that there is no significant relationship between divorce risk and BMI among middle-aged married Europeans. In response Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2012) conceded that the data are clustered at the country level, but pointed up shortcomings of multilevel models for the considered data – and finally rejected the questioning of their main finding. Hence, it seems we are back to square one. But in effect, two wrongs do not make a right.

Moreover, Lundborg and his colleagues declare, despite the unresolved issues, a previous ‘rather speculative result’ (Lundborg et al., Reference Lundborg, Nystedt and Lindgren2007, p. 540) to statistically backed evidence in their response (Lundborg et al., Reference Lundborg, Nystedt and Lindgren2012) to our comment (Schneider & Grimps, Reference Schneider and Grimps2012). Admittedly, the re-analysis of the considered data via multilevel models may pose problems, as mentioned by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2012). But, apart from this, the initial analysis still suffers from the disaggregation of context variables and underestimation of standard errors. The recently added scatter plots may qualitatively imply a connection between divorce rate and BMI among married Europeans, but they are prone to spurious correlations, and cannot be taken into account for this matter. In fact, the statistical results in the commentary, as well as the results in the response, do not show a significant relation between divorce risk and BMI in married couples. Consequently, the new claims of Lundborg and his colleagues (Reference Lundborg, Nystedt and Lindgren2012) are, once more, unfounded. The crucial factor for the interpretation is not the size of a β-coefficient, but the significance level. Only significant results can be extrapolated to the total population.

In order to take the opportunity to tackle some of the unresolved issues and deepen the analysis, we will clarify the intention of the closing remark from our foregoing commentary as a first step, and then discuss a further country-specific influence factor. Hereafter a more meaningful measurement of the divorce rate is introduced. Then we will turn again to statistics and re-examine the computation of the relationship under debate. Since both traditional regression models (OLS) and hierarchical linear models (HLM) have proved to be problem-afflicted, generalized estimating equation (GEE) regression models are used instead for fitting. The GEE is more robust and allows us to avoid the shortcomings of both OLS and HLM. The objective here is to ensure a more valid and adequate analysis of the baseline data. On the basis of these results we compare the empirical observations with the initial theoretical assumptions, and introduce, in a second step, the share of overweight or obese people of the same sex in a country as a further important, overlooked level-2 variable before we draw our conclusion.

Firstly, the closing remark of Schneider & Grimps (Reference Schneider and Grimps2012) has to be clarified. The remark was not intended as a ‘categorical and concluding argument’ but as a proposal for further research noted. The formulation ‘only if singles have a lower BMI will there be an incentive for the married to slim’ is misleading. This sentence was intended to emphasize the importance of the actual composition of the marriage market. In this context it is especially noteworthy that singles as well as married people belong to the marriage market, as people may keep searching for a new partner independent of their current marital status (Stauder, Reference Stauder2006). Thus, it is necessary to control for the body weight composition of the marriage market, which includes married people as well as singles.

Beyond the importance of the marriage market, there are further reasons to regard the distribution of BMI as a vital influence factor on individual BMI. Since weight change is a long-term process, as likewise noted by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007, p. 533), social norms and learnt behaviour play a role here as well. In particular, food-related health behaviour relies typically on information that is socially learnt, both from mimicking each other's behaviours, and from prevailing social norms. The socio-historical context plays an important role in understanding body weight (Sobal, Reference Sobal and Bjorntorp2001). Even against the backdrop of the standard economic theory, it can be argued that cost and benefit comparisons of investments (or disinvestments) in health and attractiveness are undertaken in the specific community of reference rather than at the individual level, and accordingly depend on local culture (Font et al., Reference Font, Fabbri and Gil2010, p. 1185 ff). An approach to capturing these linkages empirically is to control for the average BMI of persons of the same sex in the respective countries. This cause variable reflects processes of imitation as well as the impact of prevailing social norms.

The OECD Health Database includes information about the distribution of BMI in the necessary countries. In Table 1, the share among the total population of overweight or obese people (BMI≥25) is depicted. In the following models, these values are used to capture the composition of the marriage market and mimicking processes, i.e. social norms in a particular country.

Table 1.

Share of overweight or obese population (BMI≥25)

Source: OECD Health Data (2010).

Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007) introduced in their initial work the proportion of divorced to married people for each country as a proxy for the general divorce risk. This is a rather crude approach to determining the individual risk of returning to the marriage market. Moreover, the same value is used for both men and women by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007), assuming that the risk of divorce is equal for both genders. In fact, however, the risk of divorce differs noticeably between men and women (Hill & Kopp, 2004, p. 299). But since proven data of general divorce rates are available from the United Nations Statistics Division, even separated by age and gender, it is preferred to use these data instead. The UN data can provide more valid measures than a simple approximation using the proportion of divorced to married people. Furthermore, the UN data should reflect the individual divorce risk more accurately, because gender differences are taken into account. Spain has to be excluded from this analysis, because there is only data for 1991 available and we do not want to rely on obsolete data. Table 2 shows the national divorce rates for the 40–59 years age group, subdivided by sex, the year of observation, and the calculated divorce risk by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007) for the purpose of comparison.

Table 2.

National divorce rates for age group 40–59 and divorce risk by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007)

^a United Nations Statistics Division (2007).

^b SHARE wave 1 (release 1).

Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2012) pointed out that the results of multilevel models should be interpreted with particular caution, if there are too few level-2 clusters. In fact, this circumstance was not considered in our comment and could cause the insignificance of the parameter estimator of the national divorce rate and the low amount of the variance at country level. Consequently, we have to step away both from HLM and OLS. As an alternative, we propose to take advantage of GEE to obtain robust standard errors for clustering in estimations (Liang & Zeger, Reference Liang and Zeger1986). The GEE estimates are less efficient than multilevel estimates, but they make weaker assumptions about the structure of the random part of the multilevel model (Hox, 2004). The calculations were performed with SAS 9.2 using PROC GENMOD. The working correlation matrix is set as independent, since an independent working correlation matrix affects only the variance estimates and parameter estimates remain equal to OLS regression (Cole, Reference Cole2001, p. 66). Due to the small number of level-2 units, model-based standard errors will be used for interpretation (Hanley et al., Reference Hanley, Negassa, Edwardes and Forrester2003, p. 371). Again, the initial release of the first wave of the SHARE (the Survey of Health, Aging and Retirement in Europe) serves as baseline data. In order to perform a stricter examination, all models are estimated without Spain and Austria. As mentioned above, for Spain only outdated data are available. Additionally, Austria is excluded, because the country was identified as an outlier by Lundborg and his colleagues (Reference Lundborg, Nystedt and Lindgren2012).

The first model predicts the BMI of married men and women and contains the age-, gender- and country-specific divorce rate. The findings are reported in Table 3. In contrast to the results of our previous article, the divorce rate has a significant effect on the BMI of married men. For married women, a higher divorce rate reduces the BMI in the sample only. Considering this result, our criticism cannot be maintained completely – over here at least. Our former findings prove to be not robust and could be a product of the shortcomings of the multilevel analysis. Notably, married men tend to be slimmer if the country-specific divorce risk is lower. But this is not in line with the theoretical assumption that especially slim women are more rewarded by the marriage market and that incentives to stay slender within marriage may be greater for women (Lundborg et al., Reference Lundborg, Nystedt and Lindgren2012).

Table 3.

Parameter estimates via GEE on BMI (sample without Austria and Spain)

^a ADL: activities of daily living.

^b United Nations Statistics Division (2007).

The second model predicts the BMI of married men and women and additionally contains the share of overweight or obese people in a particular country. The findings are reported in Table 4. Firstly, there is, at least for married men, a negative effect of the national divorce rate on the BMI, as observed by Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007) previously. But the effect is not significant and, therefore, cannot be generalized to the population. By contrast, there is no effect of the national divorce rate on the BMI of women as expected. Secondly, the share of people with a BMI greater than or equal to 25 in a country shows a positive effect on individual BMI. Equally for men and women, the common body weight in a country influences individual BMI positively, as expected by us.

Table 4.

Parameter estimates via GEE on BMI under control of country-specific proportion of overweight people (sample without Austria and Spain)

^a United Nations Statistics Division (2007).

^b OECD Health Data (2010).

In conclusion, under the control of the common body weight the negative effect of divorce rate turns insignificant, i.e. disappears. Instead, the results suggest that common BMI is an important country-specific variable to explain individual BMI. Again, the findings of Lundborg et al. (Reference Lundborg, Nystedt and Lindgren2007) seem to be unfounded and a product of spurious correlation. Theoretically, the obvious correlation between national divorce risk and investment in body shape should be thought over. The individual divorce risk should be measured at the individual level rather than at the country level. For instance, Klein (Reference Klein2011) shows that serious problems in a partnership reduce the BMI compared with relationships without such problems. With such an approach individual divorce risk, namely distance from the marriage market, can be captured more precisely. Additionally, for an understanding of the determinants of body weight a longitudinal perspective should be taken. Only in this way can the results we are striving for be achieved: to find evidence for causal relationships, and not just correlations.