CAIRN-INT.INFO : International Edition

1Does living together benefit women’s health as much as men’s? Are aging men made as vulnerable by losing their (female) life partner as aging women by losing their (male) partner? This article examines those questions not in terms of mortality as is usually done but rather by looking at an eating practice that is strongly recommended for good health and just as strongly dependent on domestic work: eating vegetables.

2Mortality inequalities constitute a simple and virtually unobjectionable way of measuring health inequalities. Ever since demographers and epidemiologists observed an excess mortality rate of single men, they have been investigating relations between conjugal status and mortality (Vallin and Nizard 1977). But comparing men’s and women’s health and longevity across conjugal status is a more complex matter because men at the bottom of the social structure are more likely to be single while the reverse is true for women. In France, with occupational position controlled for, having a life partner seems to be associated with lower mortality for men. The results are not as clear for women: while having a partner also seems beneficial to women’s health (Robert-Bobée and Monteil 2006; Bouhia 2007), it does not seem as beneficial as for men (Cousteaux and Pan Ké Shon 2008).

3In other areas—as suggested by studies of occupational careers—living in a couple seems to benefit men to the detriment of their female partners. Having a female partner improves a man’s career while a woman’s career suffers from living with a man (Wajcman 1996; Gadéa and Marry 2000). Studying a sample of private-sector managers, Sophie Pochic (2005) found that men are often partnered with women who have no career (that is, are unemployed or in a much lower-status occupation than the man’s) while women are more likely to be single or partnered with a man of equivalent occupational status (setting the stage for potential competition within the couple). The contrast between these findings led some sociologists to account for the complex ties between gender, conjugal situation, and social position in terms of intersectionality (Arber 2004). From that perspective, conjugal status is a variable that creates deep disparities within apparently homogeneous categories of men and women, and failing to take it into account can result in public policies that are poorly adapted to large segments of the population. This question is particularly resonant for older persons, many of whom have lost their partner through separation or death—a situation found more often for women than men.

4Through what mechanisms, exactly, does conjugal situation affect health? This is a difficult question, because health depends on complex influences, of which organization of domestic conjugal life is only one. Whether health is measured by the powerful synthetic indicator of life expectancy or more narrow indicators such as risky behaviors (Cousteaux and Pan Ké Shon 2008), it depends in any case on environmental and occupational exposure, access to healthcare and therefore the characteristics of the given healthcare system, and so on—just as men’s and women’s careers depend on company organization, childcare options, and anti-discrimination policies. This makes it hard to determine what role can accurately be ascribed to conjugal situation.

5In this article we proceed indirectly, taking three side steps to apprehend how gender and conjugal situation operate together to produce social inequalities. The first, fairly standard one is to study changes that occur after a union breaks up, the aim being to discover what couple members “have to lose” in losing their partner. The second is to circumscribe the question to the aging population, which enables us to observe a greater number of union dissolutions but also suggests the need to problematize how conjugal situation operates in the context of aging. The third is to look at one of the intermediate links in the causal chain assumed to connect living in a couple and health. Our focus here is on eating—specifically, regular consumption of vegetables, a socially valued daily practice that occurs primarily in the domestic sphere. In the first section, we show that this variable is a good indicator of practices related to social inequalities in health but also, more generally, to what can be identified as a dominant-class (upper- or upper-middle-class) lifestyle. We then draw on longitudinal data to study whether change in a person’s conjugal situation is associated with change in daily vegetable consumption among aging men and women. To do so, we use data from the Gazel epidemiological cohort, [2] collected since 1989 from 20,625 employees of the French gas and electric company EDF-GDF who were born between 1938 and 1953. We examine the tie between conjugal status and vegetable consumption as respondents aged: Does consumption decrease among both men and women if they lose their partner through either divorce or death? Does it decrease regardless of respondent’s own social position and that of his or her partner at the outset of the study? How do these changes affect partners’ vegetable eating as they age?

6In the first section we explain our choice of research topic and briefly present related literature on union dissolution and health and aging. In the second we describe the Gazel cohort and how we used its longitudinal data, particularly how we handled missing data. In the third, we present our findings on how vegetable consumption evolves after union dissolution and by partners’ respective socio-economic positions. We discuss these findings in the fourth section. To facilitate reading, we have reserved discussion of technical statistical questions for the Appendices.

Vegetable consumption, a clear indicator of what is at stake in conjugality

7Our aim was to assess the effect of union dissolution on vegetable consumption among aging men and women. Vegetable consumption is a good way of linking the health and social status issues implicated in eating practices with gender and domestic work. We studied union dissolution to measure and understand the advantage of married over single men and see whether that advantage is found for women too, and to better understand the aging process.

What is at stake in vegetable eating: health, social status, and domestic work

8Examining eating practices is a good way to capture connections between gender, conjugal situation, and social position directly in the home. That is where most daily eating occurs in France: 80% of meals (Escalon et al. 2009, 199), a much higher figure than for other European countries (Warde et al. 2007). INSEE’s 2011 “Budget de famille” survey found that 72% of families’ total food budget went to home meals (Kranklader 2014), a proportion that has declined over time but is particularly high in aging households (Ferrant and Plessz 2015), precisely our study population.

9Still, eating practices are a complex matter. They can be studied in relation to their context (place, schedule, whether people eat alone or with others) as well as their content, which in turn can be analyzed in terms of products purchased, dishes, meals, food groups, or nutrients. Here we concentrate on daily vegetable consumption because this relatively simple indicator is strongly linked to all the issues we are interested in: regular consumption of vegetables is socially desirable today; it is also related to gender and to household organization.

10Eating a lot of vegetables is socially desirable for three reasons. The first is that because vegetables are rich in fiber and micronutrients, they are associated with lower risk of cardiovascular disease and diabetes (World Health Organization 2003); in sum, eating vegetables is good for your health. Health inequality by social position (and sex) has remained stable in France over recent decades (Blanpain 2016; Cambois et al. 2008). Studying vegetable consumption will therefore enable us to better understand how health inequalities are created in the home sphere. According to the WHO report on health, 2.8% of deaths worldwide can be attributed to insufficient consumption of fruits and vegetables, so though our research focus may appear minor or trivial it is actually quite important. However, overemphasizing the health consequences of vegetable consumption risks further individualizing and medicalizing what are in fact public health problems (Lupton 1995; Warde 2015; Aïach and Delanoë 1998).

11The second reason eating vegetables is socially desirable is how important vegetables have become in public health messages on nutrition. “Eat at least 5 fruits and vegetables a day to stay healthy,” instructs one of the main messages of France’s Programme National Nutrition Santé (PNNS), the public health policy program on nutrition launched in 2001. However, according to the health-nutrition barometer survey of France’s Institut National de Prévention et d’Éducation pour la Santé (INPES), “while the proportion of the population aware of the fruits and vegetables benchmark rose from 2.5% in 2002 to 28.1% in 2008, the proportion of French people who had actually eaten the recommended five portions on the day before the survey only rose from 10% to 11.8% over the same period” (Jourdain Menninger et al. 2010, 49). Sociologists of eating practices have shown that reception of public health messages, by adults and children (Régnier and Masullo 2009) or for babies (Gojard 2000), varies by social category: the upper and upper-middle class think of such messages as legitimate and try to comply with them while the underprivileged reject them as incompatible with their budget or living conditions or turn to other sources of advice. In analyzing vegetable consumption, we therefore also capture people’s response to public health recommendations and their ability to comply with them.

12Third, with vegetable eating as with many other aspects of health, scientific recommendations dovetail with upper- and upper-middle-class tastes. Vegetable consumption is highest in those social groups, and was so well before the PNNS program began (Plessz and Gojard 2015; Grignon and Grignon 1981). In fact, vegetable eating was already a dominant-class taste and lifestyle feature before it came to be recommended on the basis of scientific and medical arguments. Eating vegetables regularly is therefore a taste-related vector of social distinction in Pierre Bourdieu’s sense. According to Faustine Régnier and Ana Masullo (2009), diffusion of public health messages encouraging people to eat fruits and vegetables has made consumption of and taste for those foods an even stronger marker of dominant tastes and lifestyle.

13Eating vegetables, then, is linked to social inequalities in health and the social stratification process overall because it is part of the lifestyle and tastes typical of the dominant class. It is also linked to gender and having a life partner. In France, a couple eats more fresh vegetables if the person in charge of the grocery shopping also spends time cooking (Plessz and Gojard 2015). And cooking is one of the most feminized of all domestic activities: women spend an average of seventy minutes a day at it; men, fourteen (Ricroch 2012, 75). Moreover, women with families are generally in charge not just of domestic tasks but also the “care” involved in eating-related health issues (Devault 1991). The “Budget de famille” surveys show that single men eat vegetables less often than single women (Saint Pol 2008). It might be, then, that the presence of a woman in the household is essential to ensuring a diet that complies with health recommendations and dominant tastes—and therefore that having a life partner benefits men more than women in this matter.

14However, meals—the main occasion for eating vegetables—are also an important moment in family life. Qualitative studies have shown how a household’s meals are in fact a reflection of the given family (Kaufman 2005). English-language sociologists (Charles and Kerr 1988; Marshall and Anderson 2002) seeking to characterize the “proper meal” norm stressed the importance of taking into account not only its contents (for example, “one meat and two vegs”) but also its context: a “proper meal” brings together all household members. They also observed that young couples make an effort to adopt this norm, which had seemed less important to them when each member was living alone. Studying table manners in France, Claudine Marenco (1992) noted the importance people attach to the “family meal.” Households that are not “families” are therefore less subject to the meal norm—a point demonstrated by research on students, for example (Grignon et al. 1996). Having a life partner, then, could in and of itself favorably impact vegetable eating. If this is so, then women, too, have more opportunities to eat vegetables when they have a partner, precisely because they are not eating alone. In that case, from the perspective of diet, living together would not be a form of exploitation but a shared benefit.

Studying union dissolution to capture the effects of having a life partner by gender and moment in the aging process

15Studies of union dissolution are found in two distinct bodies of literature that do not overlap much. Demography and epidemiology studies have analyzed union breakups to determine if having a life partner is beneficial to both partners or if, instead, having a female partner is an advantage for men but not vice versa. Sociology of aging, meanwhile, has worked to situate partner loss and its effects within the aging process. Early research in this area tended to study the connection between union dissolution and health quantitatively, generating three points of debate.

16The first was whether or not we are dealing with a causal effect rather than an effect of matrimonial selection among people in relatively good health or who engage in healthy practices. Though this question is particularly difficult to resolve on the basis of retrospective information such as the deaths data used in the earliest studies, it does seem that selection effects play only a minor role (Vallin and Nizard 1977). That finding suggested the relevance of using longitudinal data instead. The second point of debate concerned the type of effect union dissolution has. It can be seen as a crisis, in which case we can expect that the person’s health will get temporarily worse after the breakup and then return to its normal level. The competing hypothesis is that having a partner is a resource that disappears definitively with union dissolution; in that case health can be assumed to decline over a longer period (Williams and Umberson 2004; Umberson et al. 2009). The role of women’s skills and domestic work corresponds to the view of conjugality as a resource, particularly for men. But if conjugal sociability plays a protective role—especially through meal sharing—then we cannot exclude the possibility of a “crisis” effect.

17The third debate was about conjugal status and type of union dissolution (partner’s death or separation). A considerable number of studies have examined whether marriage is more protective than cohabitation, but few take eating into account. Patricia Mona Eng et al. (2005) found that divorce and partner’s death both have a negative impact on men’s vegetable consumption but are non-significant for women’s. All these studies offer comparisons of men and women; some discuss differences in social position. However, this body of work seldom considers how type of partner loss is related to the aging process.

18Sociology of aging has changed profoundly in recent decades, shifting focus from “old age” to aging dynamics. Aging is now thought of as a process and not simply the oldest age period. The aging dynamic can be characterized in two ways: a relatively continuous process associated with a gradual decline in abilities and social activities, or a career marked by life events that in turn indicate successive life stages and signify to the individual and their family that they are growing old (Caradec 2001; Settersten and Angel 2011). In this connection Vincent Caradec (2001) cites the examples of children’s departure from the family home, retiring, losing one’s life partner, as well as health problems and giving up driving. With regard to diet, it has been shown that losing one’s life partner can upset meal preparation and eating, even for people used to cooking (Cardon 2009b; Lhuissier 2012). People react in different ways to their widow(er)hood. Some widows, for example, feel “liberated” from their partner’s eating preferences, which they had to comply with when cooking for the household as a whole. Cross-sectional statistical surveys show that age, household structure, and—for single persons—sex are associated more clearly with vegetable eating than are educational attainment or living standard (Plessz and Gojard 2013, 2015), meaning that vegetable consumption is a good indicator of eating practices in the context of this study. However, those surveys do not enable us to determine >how vegetable consumption develops as individuals grow older or how union dissolution affects that trajectory.

19Existing studies of aging are therefore helpful in that they show how union dissolution is situated in individual life histories, as a milestone in an aging “career” (Arber 2004; Lalive d’Épinay and Cavalli 2007). But they tend to focus on what happens after union dissolution. Moreover, if the hypothesis is correct that women’s daily domestic labor helps explain both their own and their partner’s vegetable consumption, then there should be no difference between partner loss through death and partner loss through divorce. In this study, we first pool the two types of union dissolution together; then test for differences between them.

20Clearly, vegetable consumption encompasses not only issues of health and social distinction but also gender and conjugal sociability. As an indicator, it can be used to contribute to sociology of health inequalities by sex and social position by examining a healthy practice; to contribute to sociology of aging by analyzing a life event characteristic of aging in a large population that has been followed over a long period of time; and to contribute to research on gender and conjugality by questioning the degree to which each partner and the couple itself increase the abilities of the two partners to engage in a socially valued domestic practice.


21To understand how union dissolution affects men’s and women’s vegetable consumption, we modeled how that consumption evolved over the aging process among voluntary respondents from the Gazel cohort living with a partner in 1989 by taking into account changes in their conjugal situation.

The Gazel cohort

22The Gazel cohort is a prospective epidemiological cohort initially designed to study various health factors, with a focus on occupational health (Goldberg et al. 2007). Researchers at the Institut National de la Santé et de la Recherche Médicale (INSERM) invited all permanent employees—40,000 individuals—of France’s then public-sector gas and electricity company EDF-GDF aged 40 to 49 in 1989 to join the cohort. Of that number, 20,625 persons throughout France participated in the survey. As is clear from the recruitment specifications, all Gazel cohort respondents belonged to the same generation. Most were men (the energy distribution sector was highly masculine).

23Most of the variables we used here were taken from follow-up questionnaires sent every year by mail to all Gazel cohort respondents; some were provided by the company’s human resources department. Gazel cohort attrition (respondents definitively lost to follow-up) was very low as the vast majority of employees remained with the company until retirement and either the company or the employee’s retirement fund was apprized of all changes of address. The annual questionnaire non-response rate ranged from 20% to 25%, except in 1989, when all respondents completed it, and in 1990, when the non-response rate was 10%.

24To study union dissolution, we restricted the analysis to individuals aged 40 to 49 in 1989 (this criterion leading us to exclude 2,198 younger women; see note 2) who reported living with a partner at that time (this criterion leading us to exclude 2,036 single persons and 180 non-respondents) and who were still in the cohort in 2014 (this criterion leading us to exclude 1,750 individuals who had died and 439 who had been lost to follow-up). [3]

25Three respondents whose data were censored to keep them anonymous were also excluded, leaving us with a sample of 14,019.

Definitions of our variables

26Questions on eating figure a total of five times; specifically, in the annual questionnaires for 1990, 1998, 2004, 2009, and 2014. However, the time variable we used to model trajectories was respondents’ age. Our age bracket spanned ten birth years, providing us with information on eating at all ages from 41 to 74 even though we only had five answers per respondent. In accordance with earlier findings (Plessz et al. 2015), the relationship between age and vegetable consumption was assumed to be linear, with a break in the slope at age 60.

27We defined living “in a couple”—that is, with a life partner—as having answered either “married” or “living as husband and wife” and in opposition to the other response options (“single,” “separated,” “divorced,” “widow/widower”). Since the respondents included in this study were all living with a partner in 1989, union dissolution was defined as living alone at the time of an eating questionnaire. [4] As explained, we also tested whether there is a difference depending on type of union dissolution, by distinguishing between widow(er)hood and separation, divorce, and singlehood. And we tested the stability of union dissolution effects over time (crisis or resource loss effect) by comparing the effects of living alone in the three years following union dissolution with effects after that period.

28On each of the five eating questionnaires, respondents were asked how often they ate twenty-two types of food. Possible answers were “never or almost never,” “once or twice a week,” “more than twice but fewer than six times a week,” and “every day or nearly.” We defined daily vegetable consumption as the last of these answers. In 1990, the “vegetables” item was formulated thus: “green vegetables (fresh, canned, frozen, etc.).” From 1998 the item read “cooked vegetables as a first course, in soup, or a main course (leeks, cabbage, green beans, etc.).” We added a variable “1990” and checked that the change in item phrasing did not affect the results. [5]

29Other events can occur during aging, and those that affect vegetable consumption can create confounding factors. Drawing on the literature, we controlled for child(ren)’s departure from the family home, retirement, and house moves, particularly from the Région Île-de-France to elsewhere in the country (Cardon 2009a; Wyndels et al. 2011; Plessz et al. 2015). An earlier finding that Gazel cohort respondents lunched out during their work lives (the case for one in two men, one in three women) whereas afterwards they almost systematically lunched at home (Plessz et al. 2015)—a shift that impacted their vegetable consumption—meant that we needed to control for retiring in our analysis. However, there is a strong correlation between child(ren)’s departure and retiring: all of our respondents were working in 1990 and retired in 2014, and in almost all cases their child(ren) had already left home by 2014. We therefore created a cross-dummy variable (working, children living at home (ref.); working, no children living at home; retired, children living at home; retired, no children living at home). Meanwhile, area of residence in questionnaires-on-eating years captures the possible effect of living in a rural area (and having a vegetable garden). Last, we considered a possible health selection effect by re-estimating the model without respondents who reported being in poor health in 1989. To measure respondents’ social positions, we used two coded variables, each with “yes”/“no” response options: educational attainment at least as high as the French baccalauréat secondary school degree, and being a senior manager at age 35. [6]

Modeling: comparing men and women in a repeated observation model

30As specified, our focus here is the relationship between vegetable consumption and union dissolution over the aging process by gender. We drew on Anne-Sophie Cousteaux and Jean-Louis Pan Ké Shon’s 2008 study for our men/women comparisons, estimating a single model for the two sexes while stratifying the variables of interest by sex. We first introduced a sex variable (man/woman) that captures the sizable, persistent gap between men’s and women’s levels of vegetable consumption. To measure the effect of the “now without a partner” conjugal situation, we added two other variables—“man without a partner” and “woman without a partner”—and compared their coefficients. The time-dependent control variables were geographic area, age, whether retired, presence of children in the home, and 1990 questionnaire. [7]

31To model repeated observations such as these, where the number of measures per respondent is quite low compared to the number of respondents (over 14,000), we used Generalized Estimating Equations (GEE) logistic regressions (Agresti 2013, 362; Twisk 2013). [8] The results are to be interpreted as we would those of “naïve” logistic regressions—with three differences. First, standard errors here take into account the correlation between pieces of information reported by the same individual. Second, estimated odds ratios (ORs) are slightly weaker than the ORs that would be obtained with a mixed model for the same data, and they are to be interpreted as average effects for the population in question. Third, GEE are more sensitive to missing data than maximum likelihood estimations (see Appendix A1).

Dealing with missing data

32In Gazel, approximately 20% of annual questionnaires are missing. If we had estimated our models solely on individuals who returned all five annual questionnaires with the section on eating, our sample of 14,019 would have fallen to 7,201 complete individual cases. However, in the data table used to estimate Model 1, only 6% of the cells are empty. [9] If non-response here was random and if its only effect was to shrink sample size, we could ignore it. But we have good reasons to assume that it creates bias; that is, that the distribution of our variables of interest (and typically, vegetable consumption) is not the same for complete and incomplete cases. Indeed, the main investigators for this cohort have shown that not returning the annual questionnaire was associated with sex, age, occupational level, state of health, and alcohol and tobacco consumption in 1989, as well as with being retired (Goldberg et al. 2015). Those factors are also associated with vegetable consumption in the Gazel cohort and other surveys (Lindström et al. 2001; Roos et al. 2001; Plessz and Gojard 2013; Escalon et al. 2009). We therefore had to deal with the missing data before estimating the GEE models: we needed to reduce the effect of that bias on the associations we intended to measure. We chose to impute non-responses by way of Multiple Imputation by Chained Equations or MICE (see Appendix A2 for a description of the procedure and justification of its use in our case; see also White et al. 2011 and Sterne et al. 2009).

33The results of the estimated models on imputed data are presented in Table 3 in the form of odds ratios. We also estimated our models on the 7,201 complete cases (Table 4) in order to compare estimated coefficients for the dataset with the imputed missing data and the set without it, and to test hypotheses not compatible with imputation model specifications. Last, we tested the effect of time elapsed since union dissolution and the effect of dissolution type (separation or partner’s death).


Partnered Gazel cohort respondents: a specific population in 1989

34Of the 20,625 Gazel respondents, 14,019 individuals had a life partner in 1989, were between 40 and 49 that year, and were still alive in 2014. Most were men (11,770) since this employment sector was strongly masculine. As mentioned, we had 7,201 complete cases (that is, cases from which no data is missing): 6,250 men and 951 women.

35Table 1 presents the descriptive statistics for time-independent variables on both complete and incomplete cases. Over 80% of complete-case respondents (92% of women respondents) had initially worked in subordinate positions (men were usually manual workers; women, office workers). These people—especially the men—experienced strong upward socio-occupational mobility through internal EDF-GDF promotions: 15% of women and 47% of men were senior managers upon retiring. The women had lower educational attainment (18% were secondary school graduates, as against 47% of the men). In 2014, 88% of respondents were living with a partner; only 3% had experienced more than two union dissolutions as defined above. Last, the proportions of secondary school graduates and senior managers were lower in cases where at least one variable was missing from the follow-up.

Table 1

Descriptive statistics by sex: time-independent variables (%), complete and incomplete cases

Complete casesIncomplete cases
Birth cohort
Secondary school degree or higher30.017.925.012.0
Occupational level upon entering EDF-GDF
 Senior manager8.
Senior manager at age 3519.25.915.22.5
Senior manager upon retirement46.415.535.58.7
Partner was senior manager in 19898.434.37.832.0

Descriptive statistics by sex: time-independent variables (%), complete and incomplete cases

Source: Gazel cohort; authors’ calculations.

36Table 2 presents time-dependent variables for complete cases. Both men and women ate more vegetables over time, but from 1990 to 2014 there was a difference of nearly 20 percentage points between the two sexes. [10] As explained, we began by restricting the sample to individuals with a life partner in 1989; over the years the percentage of single persons rose slowly, to 6.7% of women and 4% of men in 2014. And some men and women had retired (statutory retirement at EDG-GDF ranged from 55 to 60 by type of career).

Table 2

Descriptive statistics: time-dependent variables (%), complete cases

Eat vegetables every day31.236.252.658.055.1
No longer have a partner1.
*No partner for over 3 years0.024.845.360.783.3
Living with child(ren)
Geographical area
Eat vegetables every day18.621.030.735.636.2
No longer have a partner0.
*No partner for over 3 years0.023.745.260.682.2
Living with child(ren)87.850.920.810.15.1
Geographical area

Descriptive statistics: time-dependent variables (%), complete cases

Note: * Of respondents reporting they no longer have a partner.
Source: Gazel cohort, authors’ calculations. N = 6,250 men and 951 women.

37The Gazel cohort was therefore a quite specific population: all respondents were working in 1989 and all had continuous careers up until retirement—not an ordinary trajectory for women born in the 1940s. Furthermore, they experienced the upward social mobility typical of the baby-boomer generation (Chauvel 2002). The differences between cohort men and women, meanwhile, reflect gender relations on the labor market and in the domestic sphere: women’s educational attainment was lower than men’s; their career advancement was slower; the likelihood of their becoming single between 1990 and 2014 was higher, and their vegetable consumption, more frequent.

Is women’s vegetable consumption less sensitive to union dissolution than men’s?

38In Table 3 we estimated the odds ratios for eating vegetables every day, finding them stronger for women, secondary school graduates, and retired persons. The effect of being one year older is positive; it is weaker after age 60 but nonetheless significant. Occupational position at age 35 is not associated with vegetable consumption when educational attainment is controlled for. Living with children is associated with lower vegetable consumption before and after retirement. [11] Model 1 tests the association between union dissolution of any kind (death or separation in married or un-married couples) and daily consumption of vegetables. That association is negative and significant for both men and women, meaning that union dissolution is followed by a fall in vegetable consumption for both. However, men’s consumption falls more than women’s, and the odds ratio difference is significant (p = 0.017). As this result might have been due to health-related selection bias, we estimated Model 1 without respondents reporting poor health in 1989. The results (not shown) were the same, meaning that the effect for union dissolution is not due to people in poor health at the outset of the study being at greater risk for union dissolution or having a different level of vegetable consumption.

Table 3

Odds ratios for eating vegetables every day: sex and union dissolution among individuals living with a partner in 1989 after imputation of non-responses

Table 3
Model 1 Model 2 1990 questionnaire 1.39*** 1.39*** Geographic Area Île-de-France – – Northeast 1.03 1.03 South 1.24*** 1.25*** Central-eastern 1.32*** 1.32*** West 1.17*** 1.17*** Age By year up to age 60 1.05*** 1.05*** By year beyond age 60 1.01*** 1.01*** Woman 2.37*** 2.07*** Work situation, child(ren) Working, living w/ child(ren) – – Retired, living w/ child(ren) 1.20** 1.20** Retired, not living w/ child(ren) 1.44*** 1.43*** Working, not living w/ child(ren) 1.17*** 1.17*** Secondary degree or higher 1.14*** Senior manager at 35 1.02 Women Men Women Men Secondary degree or higher 1.24* 1.12** Senior manager at 35 0.94 1.02 Conjugal situation Partnered – – No partnered 0.78** 0.62*** Partnered, partner is a manager – – Partnered, partner is not a manager 1.03 0.89* Single, partner was a manager 0.90 0.48*** Single, partner was not a manager 0.76* 0.57*** Number of rows 70,095 70,095 Number of cases 14,019 14,019

Odds ratios for eating vegetables every day: sex and union dissolution among individuals living with a partner in 1989 after imputation of non-responses

 Man-woman difference significant at p < 0.05
 Man-woman difference significant at p < 0.01
Note: * p < 0.05; ** p < 0.01; *** p < 0.001.
See Appendix, Table A2 for Model 1 standard errors.
Source: Gazel cohort, authors’ calculations. Data imputed using multiple imputations (50 imputations).

39Figure 1 shows the probabilities of eating vegetables every day as predicted by Model 1 for respondents who lost their partner at age 58. Predicted probability falls slightly more for men, who were already less likely than women to eat vegetables every day. The confidence intervals are smaller for men because the number of male respondents is so much higher.

Figure 1

Predicted probabilities of eating vegetables every day by sex, age, and conjugal situation

Figure 1

Predicted probabilities of eating vegetables every day by sex, age, and conjugal situation

Note: Conjugal status was set for “partnered” until age 58 and “no partner” beyond that age. The other variable values were as follows: Île-de-France, not living with children, working, no secondary degree, intermediate or subordinate occupation at age 35.
Source: Authors’ calculations based on Model 1, Table 3 (Gazel cohort data after imputation).

Presence of a woman in the home and social position matter for men’s vegetable consumption

40In Model 2 we tried to understand how vegetable consumption is associated with partners’ respective social positions. We already knew whether a respondent was a senior manager at age 35 and whether his or her partner was a senior manager in 1989 (when the respondent was between the ages of 40 and 49). We then crossed this variable with conjugal situation over time. The reference category was living with a partner who was a senior manager in 1989. We also examined respondents’ educational attainment (their partner’s was not measured).

41Having at least a secondary school degree is associated with frequent vegetable consumption. The effect of educational attainment is positive and significant for both women (OR = 1.24) and men (OR = 1.12). Being a senior manager at 35 has no significant effect when educational attainment and other model variables are controlled for. For respondents still living in a couple, partner’s occupational status affects men’s and women’s vegetable consumption differently, having no effect on women’s (OR near 1). Men whose partner is not or was not a manager eat vegetables significantly less often than men with a manager partner (OR = 0.89). The vegetable consumption level of partnered men therefore depends on their educational attainment and their partner’s social status, while for partnered women only their own educational attainment counts.

42When a man loses his partner, his vegetable consumption drops sharply, regardless of whether the woman was a senior manager in 1989 (respective ORs of 0.48 and 0.57). Union dissolution therefore penalizes men regardless of their former partner’s social characteristics. This is not the case for women: losing a senior manager partner has no effect on their vegetable consumption (OR = 0.90, not significantly different from 1). But when a woman loses a partner who was not a manager, her vegetable consumption falls (OR = 0.76). However, the difference between the two coefficients is not significant (p = 0.319).

43It may be that achieving senior manager status in a generation born in the 1940s (and likely having partnered with someone of the same generation) involves stronger selection effects for women than for men. The women in question may therefore have come from a more privileged social background. If that were so it would be fallacious to compare the effect of having a female manager partner to that of having a male manager partner. So we coded male partners’ and female partners’ career levels differently in estimating our model; i.e., respectively, male partner being a senior manager versus having a different occupational status and female partner being either a senior manager or in an intermediate position versus being in a subordinate position. This led to the same conclusions.

44It therefore seems that men’s vegetable consumption is more strongly associated with how their conjugal situation evolves and the social characteristics of their (female) partner than women’s vegetable consumption. Women seem more “independent” (in that their own characteristics are what govern their vegetable consumption), the exception being women partnered with men of subordinate occupational status.

The circumstances of union dissolution

45The following results investigate a number of hypotheses from existing sociological studies. Testing them gives us a better understanding of the dynamics of eating and diet (here apprehended through vegetable consumption) throughout a person’s conjugal trajectory.

46To begin with, we can evaluate the magnitude of the non-response bias corrected by imputation [12] by comparing Model 1 as estimated on complete cases (Table 4) with Model 1 as estimated on imputed data (Table 3). When Model 1 is estimated on the 7,201 complete cases (Table 4), the odds ratio for eating vegetables every day after union dissolution is still sharply and significantly below 1 but turns out to be comparable for men and women, whereas the imputed data in Table 3 showed a lower odds ratio for men. Clearly, then, imputation changed the results. If we look at available 2014 data, we see that women reporting now being without a partner account for 83% of incomplete female cases while partnered women account for 28%. For men the respective figures are 70% and 26%. Moreover, fewer incomplete (than complete)-case individuals who did answer in 2014 said they consumed vegetables every day. It is therefore plausible that the association between vegetable consumption and conjugal status is biased when measured on complete cases only.

47Second, we tested whether there was a difference between separation and partner’s death. The results, presented in Model 2, Table 4, should be interpreted with caution for women because the estimation was solely on complete cases. [13] The above-noted non-response bias is operative here, and the number of widows and separated women among complete cases is low (respectively forty and twenty-three in 2014).

Table 4

Odds ratios for eating vegetables every day by type of union dissolution: respondents living with a partner in 1989, complete cases

Table 4
Model 1 Model 2 Model 3 Union dissolution Death or divorce Time elapsed since union dissolution 1990 questionnaire 1.48*** 1.48*** Geographic Area Île-de-France – – – Northeast 0.98 0.98 0.97 South 1.23*** 1.18** 1.23*** Central-eastern 1.34*** 1.33*** 1.33*** West 1.16** 1.22*** 1.16** Age Slope to age 60 1.06*** 1.06*** 1.06*** Slope after age 60 1.01*** 1.01*** 1.01*** Woman 2.36*** 2.43*** 2.35*** Work situation, child(ren) Working, living w/ child(ren) – 1.27***- Retired, living w/ child(ren) 1.34*** 1.45*** 1.34*** Retired, not living w/ child(ren) 1.54*** 1.15* 1.54*** Working, not living w/ 1.19*** 1.07 1.19*** child(ren) Secondary ed. or higher 1.11* 1.11* 1.02 Senior manager at 35 1.04 1.04 – Women Men Women Men Women Men Conjugal situation Partnered -- No partner 0.60** 0.61*** – – – – Divorced/separated 0.52* 0.65*** Widow(er) 0.68 0.54*** No partner for 3 years 0.68 0.52*** No partner for over 3 years 0.56* 0,72** Number of rows 36,005 36,005 28,804 N 7,201 7,201 7,201

Odds ratios for eating vegetables every day by type of union dissolution: respondents living with a partner in 1989, complete cases

Key: Difference between cells significant at 5%.
Note: * p < 0.05; ** p < 0.01; *** p < 0.001.
Source: Gazel cohort, authors’ calculations. Complete cases.

48The difference between partner loss through death and partner loss through separation is slight and not significant for men or women. Either this difference is not relevant to vegetable consumption or the number of male respondents in the Gazel cohort, though high, was not high enough to produce a detectable difference.

49Last, we examined whether the effect of union dissolution on vegetable consumption could be temporary; in other words, a “crisis” effect rather than a “resource” effect. Is there a difference in vegetable consumption between the three-year period following union dissolution and after that period? Model 3, Table 4, was estimated for complete cases and only for 1998 to 2014 because conjugal history prior to 1989 was unknown. The coefficients are relatively close for women and the differences not significant. For men, the coefficient associated with having experienced union dissolution in the previous three years is significantly lower than the one associated with having lost a life partner more than three years before answering the survey (p = 0.046). This would seem to suggest a slight crisis effect. Beyond three years, the coefficient remains negative and significant (p = 0.005), meaning there may also be a resource effect. Exploratory analysis of non-partnered men throughout the follow-up seemed to show that many ate vegetables less often than the partnered-in-1989 men studied here. This suggests that after partner loss, now-single men’s behavior pulls closer to the eating practices of single men. But it might also indicate that the crisis of union dissolution is gradually resolved over a longer-than-three-year period.

50Testing the three hypotheses gives us a better understanding of our main result: that men and women—especially men—are likely to eat less vegetables after union dissolution. We have observed that it is important to take non-response bias into account. We have found no difference between partner’s death and separation (divorce). Testing a short-term, “crisis” effect of union dissolution on eating practices yielded ambiguous results. All this suggests that it is indeed in the context of daily activities that conjugality influences eating practices, here apprehended in terms of vegetable consumption.

51* * *

52Who benefits from having a life partner, the man, the woman, or both? When we look at a daily practice—eating vegetables—that involves health and social distinction issues, does conjugality seem a situation that penalizes women and benefits men, or an institution that protects both? Our results lead us to a nuanced conclusion: conjugality benefits both partners, but men more than women. Daily vegetable consumption is viewed here as indicating a lifestyle that conforms to health recommendations and dominant-class tastes. Men’s vegetable consumption is more closely dependent on conjugal situation and partner’s social position than women’s; women are more likely to eat vegetables at all ages.

53Given these results, the hypothesis that a man’s vegetable consumption depends on his female partner’s domestic work is debatable. For while vegetable consumption does depend on there being a woman in the home, that woman’s social position also matters. This makes sense, because women from privileged social categories are more receptive than others to nutritional recommendations (Régnier and Masullo 2009; Gojard 2000) and more likely to integrate them into what Geneviève Cresson (1995) called their “health domestic work.” For them, feeding vegetables to the household may also be a way of maintaining the family social status, through a lifestyle that conforms to dominant-class tastes (Marenco 1992).

54What is more, the “domestic work” explanation does not account for all of our results. Conjugal sociability probably facilitates vegetable consumption, as mentioned in the introduction and as is further suggested by the possibility of a crisis effect on vegetable eating after union dissolution. Moreover, union dissolution likely reveals women’s post-retirement economic vulnerability (Arber 2004), especially in generations where their occupational careers were modest and often discontinuous (Bonnet et al. 2006). [14] Similarly, some working-class men may facilitate vegetable consumption in their households by growing vegetables, and union dissolution in an aging population may deprive women of that source of vegetables (Weber 1996).

55Our results are based on the Gazel cohort, which, while it ensures the robustness of our analyses, does not cover the entire French population. This is one of the very few longitudinal quantitative data sources on eating practices and diet in France. Respondents here were followed for a considerable length of time (five questionnaires over 25 years) and few were definitively lost to follow-up; the Gazel cohort sample comprises over 20,000 individuals, though women are underrepresented; and the information obtained at the outset of the survey or by means other than the questionnaire made it possible to control for several confounding factors and take into account health-related and non-response selection bias. Multiple imputation is seldom used to reduce non-response bias in sociology. Here, 50.3% of individuals did not answer at least once, but we only had to impute 6% of the data needed to estimate our model. Bias reduction enabled us to bring to light a gender difference in union dissolution effect. Imputing and estimating models on imputed data are time-consuming and statistically challenging procedures, but they force researchers to formulate explicit hypotheses about whether non-response bias is present and if so what form it takes, a problem often handled ad hoc and implicitly. [15]

56It is also important to stress that we have dealt here with a select population within the French population at large, if only in that our sample was restricted to partnered individuals. Furthermore, the Gazel cohort is made up exclusively of persons born in the 1940s who had a permanent job at EDF-GDF in 1989: that generation had better occupational careers and opportunities for promotion than the following ones (Chauvel 2002). And even within it, the status of permanent employee in a public enterprise should not be overlooked, particularly where women are concerned: all these women were working in 1989 and their careers advanced regularly and continuously. Yet despite these factors, which result in a more homogeneous population than the French population at large, we have brought to light differences by sex and occupational position. Those results cannot readily be extrapolated to later generations, however; changes in the labor market and (slower) ones in domestic task division may have since reduced differences between men and women.

57The fact that our survey population was relatively homogeneous by age and occupational position enabled us to discover that respondents ate more vegetables as they grew older. However, it might be that all cohorts during that period did so, including those we did not observe. [16] The period effect may be due, for example, to changes in the food on offer: many vegetable-based products were developed during our period (Plessz and Gojard 2013), though older persons do not seem to have liked them much (Plessz 2013). Or it may be due to the nutritional health messages diffused from 2001 as part of France’s health and nutrition program. Existing studies that attempt to disentangle the effects of age, period, and cohort in eating practices in France (Babayou and Volatier 1997; Hébel and Recours 2007) do not enable us to answer this question. However, it seems unlikely that a period effect of this sort could bias union dissolution effects, as partner loss was distributed across the entire follow-up period and did not concern the entire sample.

58Despite these restrictions, our results broaden the perspectives of sociology of aging and the life course approach in general. Thinking of aging as a process marked by biographical events makes it easier to research steps in what may be thought of as a continuous process. The case we observed here was complex. On the one hand, vegetable consumption increased as the cohort aged: respondents’ behavior thus pulled closer to health norms and socially legitimate eating practices. From this perspective, tastes in food have something in common with cultural tastes, as both are stratified by a combination of social class and age group (Bry et al. 2016). But food preferences also recall that while old age is often associated with reduced activities, recent studies have shown that this translates above all into withdrawal from public, external activities; private domestic activities are affected only later—and increasingly later—in life (Bickel and Lalive d’Épinay 2001). On the other hand, the event under study—union dissolution, a typical event in the aging process—had the opposite effect: it pushed respondents in the direction of poorer health, or less healthy eating practices, working against the aging effect of eating more vegetables. Our study therefore suggests the relevance of broadening and deepening research into tastes and practices throughout the life course while remaining attentive to both their variability and their persistent links to gender and social position.

59Last, there is the question of the benefits of conjugality by sex. This has been studied first and foremost in terms of occupational careers and mortality. Here we have examined it in connection with a daily practice linked to health and social status. In all these matters, conjugal situation proves a factor of heterogeneity and indeed inequality, between the sexes and within each sex. To borrow gender studies vocabulary, this suggests that conjugal situation combined with gender and social class may form a type of intersectionality (Arber 2004) that deserves more attention in research into such diverse topics as eating practices, employment, health, and social exclusion.


Appendix A1. – Estimating with Generalized Estimating Equations (GEE)

60GEE is a modeling strategy used when observations are not statistically independent from each other.

61In our case of longitudinal follow-up, we had five observations for each individual; we therefore had to loosen the hypothesis of independence. With GEE we were able to estimate a matrix of the average correlation between the five observations we had on each individual and take account of those correlations when estimating standard errors of coefficients. We used an unstructured matrix, meaning there were no constraints on the values of the ten correlation coefficients, as illustrated in Table A1 presenting the correlation coefficients estimated when estimating Model 1 (Table 3) on the first imputed file.

62GEE are marginal or population-averaged models, in opposition to conditional models such as mixed (or multilevel or hierarchical) models. For example, a mixed model coefficient is interpreted (conditionally on random effects) as being 60 rather than 59 years old for the given individual, whereas a GEE model coefficient is interpreted as the difference between a 60-year-old man randomly selected from a given sample and a 59-year-old man randomly selected from the same sample. The GEE model coefficient thus describes the average dynamic in the given population rather than the dynamics of its individual cases. The two model types produce the same coefficients in linear regressions, but in logistic regressions GEE yield slightly weaker associations (see Twisk 2013, 138). Table A2 shows the odds ratios and standard errors of Table 3, Model 1.

Table A1

Intra-individual correlation matrix estimated on first imputed file


Intra-individual correlation matrix estimated on first imputed file

Table A2

Standard errors of Table 3, Model 1*

Odds ratiosStandard errors
1990 questionnaire1.390***(0.068)
Geographic Area
 Slope to age 601.054***(0.005)
 Slope after age 601.014***(0.003)
Work situation, child(ren)
 Working, living w/ child(ren)1(.)
 Retired, living w/ child(ren)1.197**(0.066)
 Retired, not living w/ child(ren)1.436***(0.07)
 Working, not living w/ child(ren)@1.166***(0.046)
Secondary ed. or higher1.141***(0.041)
Senior manager at 352.368***(0.045)
Conjugal situation
 No partner, * woman0.783**(0.061)
 No partner, * man0.623***(0.034)

Standard errors of Table 3, Model 1*

Note: Gazel cohort data imputed separately by sex and partner’s occupational status at the start of the survey (m = 50).
* p < 0.05; ** p < 0.01; *** p < 0.001.

Appendix A2. – Multiple imputation of non-responses


63The aim of imputing is to reduce bias due to non-response when estimating model coefficients; specifically, by using variables related to both the probability of a variable being missing and the value it would have had if it had been observed (White et al. 2011; Sterne et al. 2009). An iterative process (equation chain) can be used to impute several variables at a time while identifying the relevant ones (Table A3 below). We did fifty imputations and as many estimations of the model so as to take into account inter-imputation variance. It should be specified that inter-imputation variance was particularly strong in our non-response model because that model could not be very precise.

64Using the state-of-the-art chained equations imputation method, we introduced imputation of all variables into the model, including the dependent variable (White et al. 2011). In accordance with the above-cited literature, we included in the imputation model the auxiliary variables of perceived health and whether respondent was a smoker or drinker as those variables were related to non-response and to the variables of interest. In this way we generated fifty completed files. We then estimated our GEE models on each file and combined the results according to the rules laid out by Donald B. Rubin (1987), which take into account variability of coefficients and of standard error of coefficients from one file to another.

Why is it important to deal with non-responses?

65In both longitudinal and cross-sectional analyses, non-responses can introduce bias. When we estimate a model on complete cases only, we accept to introduce a selection bias due to non-response. With longitudinal data, we may “lose” even more individuals than in a cross-sectional analysis (all it takes is one missing variable from one wave), but we also often have more information on those “incomplete” individuals.

66Missing data generates bias in estimations if the distribution of variables of interest that would have been observed for incomplete cases (or individuals) differs from what is observed for complete cases. Ad hoc solutions like including only complete cases, deleting the variable, or replacing missing values with the average value do not resolve the problem, amounting instead to assuming that non-responses are missing completely at random (MCAR).

67If we have reason to suspect bias, we should and can try to reduce it. We only know how to reduce bias caused by missing at random responses (MAR) as opposed to missing not at random ones (MNAR). Data are MAR if the association between non-response probability and the value of the variable that would have been observed can be attributed to one or several observed variables. If, on the other hand, we have reason to believe that data are MNAR (as when relatively high incomes are missing), sensitivity analysis is one solution.

68In analyzing what Gazel cohort respondents ate, we had good reasons to believe that the non-response mechanism operative in a given self-administered questionnaire was not chance; we also had a considerable number of observed variables for all individuals (including responses to earlier and later questionnaires) and existing studies on the basis of which to formulate hypotheses on links between variables (Goldberg et al. 2015).

Doing Multiple Imputation by Chained Equations

69We chose Multiple Imputation by Chained Equations (MICE) to handle our problem of partial non-response. [17] In this procedure, the researcher writes a regression model that predicts each imputable variable. Model specifications have to be compatible with those of the model of interest: incomplete variables and the dependent variable have to be included. Variables associated with both non-response and incomplete variables (including any not included in the model of interest) play a key role here, as they are the ones that will “reduce” non-response bias. These models amount to a statement on assumed ties between all variables, including those involved in the non-response mechanism. The point is that imputed values reflect the fact that the characteristics of incomplete cases are different from those of complete cases.

70The statistical program (in our case, Stata 12) operates as follows:

  1. Replaces all empty cells with arbitrary values;
  2. Estimates the imputation model for variable x1;
  3. Replaces missing x1 values by values that are probable according to the estimated model;
  4. Repeats steps 2 and 3 for all variables to be imputed;
  5. Repeats steps 2 through 4 until the imputed values converge, resulting in a complete data file.

71To do multiple imputation, the entire procedure has to be repeated several times (fifty in our case), generating the same number of complete files, files in which the imputed values that have taken the place of missing values vary slightly each time due to the randomness of step 3. Once the fifty complete data files have been generated we can estimate the model of interest. This procedure (entirely automated in the Stata program) consists first of estimating the model for each file separately, then combining the coefficients thus obtained. The final coefficient reached is the average of the fifty coefficients. The variance is calculated using Rubin’s rules (1987), which combine intra-estimation variance (as in the usual regression model) and inter-estimation variance (the fact that fifty slightly different variances were obtained).

72The chained equations technique offers several advantages. First, we can simultaneously impute several missing variables while taking into account the distribution of each (dichotomous, categorical, continuous) and ties between them. Second, the inter-imputation variance generated by doing multiple imputations rather than just one increases in direct proportion to the imprecision of the imputation model, meaning that the standard errors and confidence intervals obtained when we estimate the model of interest will also reflect the imprecision of our imputations—for example, if we lack observed variables linked to non-responses and incomplete variables. The main drawback is the tedious modeling phase (writing the imputation chained equations) that must be completed before the model of interest can be estimated.

73Since the aim of the imputation model is to reduce bias in the final model, it has to be compatible with the model of interest; that is, all possible associations between that model’s variables also have to be possible in the imputation model. This raises problems for imputing interactions (the procedure assumes one model per variable, whereas interaction requires two variables simultaneously). In our case, interactions were at the center of the analysis so we needed to handle them with care. Interactions with the variable of interest (conjugal status) all brought into play at least one variable for which no data were missing (sex), and to overcome the problem of interaction with partner’s status, we excluded individuals (few in number) for whom partner’s occupational status was unknown. Our solution was to separately impute the four groups of individuals formed by crossing respondent’s sex with partner’s occupational status in 1989. By contrast, we did not impute the interaction between being retired (observed for all respondents) and living with children because it was a control variable (less central to the analysis) and our explorations of complete cases showed that interaction to be negligible. The models for each variable are shown in Table A3 below.

74Last, imputation quality improves if non-missing variables associated with partial nonresponse and with variables of interest are added to the imputation model. In our case those variables were gender, age, conjugal situation at the start of the survey, educational attainment, and occupational situation—information known for all respondents and associated with both vegetable consumption and non-responses. As poor health can affect both non-response and eating practices, we also took into account perceived health and smoking and drinking, all as reported in 1989.

Table A3

Imputation models for model-of-interest variables, Gazel cohort

Table A3
Imputed variable Type Variables included in the imputation model by source Annual Questionnaire Administrative data questionnaire at start of survey (1989) Vegetables_Ta,b Logit Vegs years other Healthc, smoker, Retired_T, ed. than T, Partner T, drinkerd attainment, age_Te, Living cohort, sr manager at w/child(ren)_T age 35 Partner_Td Multinomial Logit Vegs_T, Partner Health, smoker, Retired_T, ed. years other than T, drinker attainment, age_Te, Living cohort, sr manager at w/child(ren)_T age 35 Livew/child_T Multinomial Logit Vegs all years, Live Health, smoker, Retired_T, ed. w/child years other drinker attainment, age_Te, than T, Partner_T cohort, sr manager at age 35 Ed. attainment Ordered logit Vegs all years Health, smoker, Retired_T, ed. drinker attainment, age_Te, cohort, sr manager at age 35 Smoker Logit Vegs all years, Health, smoker, Retired_T, ed. Partner all years, drinker attainment, age_Te, Living w/child(ren) cohort, sr manager at all years age 35

Imputation models for model-of-interest variables, Gazel cohort

Notes: Inclusion criteria: individuals living with a partner in 1989, born between 1939 and 1948, alive in 2014, and partner’s occupational status in 1989 known.
Imputed separately by sex and partner’s occupational status in 1989 (senior manager versus other): 4 imputation groups.
a T: Year: 1990, 1998, 2004, 2009, 2014.
b Vegs: Vegetables: 1 = eats vegetables every day or nearly.
c Health: Perceived health < 6 on a scale of 1 to 9 (9 = “very good health”).
d Drinker: 1 = at least 2 drinks a day for women, 3 for men; 0 if less.
e Age: Quantitative variable, slope break at age 60.
f Partner: 1 = married, living as husband and wife; 0 = single, divorced, separated, widow(er)ed. Partnered individuals only. When conjugal status was known and identical in T-1 and T+1, missing data were completed before imputation. Missing values for 1990 were set at 1 (couple) as in 1989 to preclude prediction problems.

Appendix A3. – Decision to keep individuals for whom we had imputed the dependent variable when estimating the model of interest

75It is important to impute the dependent variable at the same time as all explanatory variables in the model (White et al. 2011). White et al., explain that unless we have good auxiliary variables, individuals for whom the dependent variable has been imputed produce more noise than anything else and so should not be kept when estimating the model of interest. In our case, however, eliminating individuals for whom at least one wave was missing would have led to eliminating almost all “non-respondents,” making the whole procedure useless because the purpose of introducing those individuals was precisely to correct non-response bias.

76Our dependent variable, vegetable consumption, was measured five times for each respondent. These five observations were correlated. Only 685 respondents never answered the vegetable question. Answers from the different waves therefore constituted quite good auxiliary variables for imputing the dependent variable, not to mention perceived health and vegetable consumption at the start of the survey, etc. We therefore chose to keep all individuals in estimating the models of interest.

77As a precaution, we also estimated Model 1 on the following set of different sub-samples, formed, respectively, by

  1. excluding all individuals who answered none of the five annual vegetable questions (685 cases); this gave slightly lower standard errors (to the third decimal) and coefficients identical to Model 1, Table 3 on imputed data;
  2. excluding all individuals who gave either no or only one answer to the vegetable questions (2,035 cases); this gave higher standard errors than for Model 1, Table 3 and very slight changes in coefficients;
  3. excluding all individuals with at least one non-response on the vegetable questions (6,084 cases); this gave results comparable to Model 1, Table 4 on complete cases (a logical outcome because we had excluded nearly all incomplete cases).


  • [1]
    The analyses presented here were done during an eighteen-month research stay in Unit UMS 011 Cohortes Épidémiologiques en Population (INSERM/Université Versailles-Saint-Quentin-en-Yvelines), under the supervision of Marie Zins. Our thanks to her, and to Marcel Goldberg, Sébastien Czernichow, and Séverine Gojard for their precious advice throughout our work on this article, and to the entire UMS 011 team and the EDF-GDF works council, which produced the Gazel cohort data.
  • [2]
    The annual questionnaires can be consulted at (visited March 30, 2016). The cohort included women aged 35 to 39 in 1989 (for a total of 20,625 respondents); we excluded that group to ensure comparison of people of similar ages.
  • [3]
    Respondents “lost to follow-up” were individuals who left EDF-GDF before retiring or requested not to be sent any more questionnaires. Excluding those cases and respondents who died during the follow-up period probably led to overestimating the overall level of vegetable eating in the cohort as early mortality is statistically associated with poorer health habits. However, we were interested here in how vegetable consumption evolves as people age, so it was important to keep the sample identical sample through follow-up, with all individuals present from the first to the last wave.
  • [4]
    We do not know whether respondents changed partners between two questionnaires; nor do we know partner’s sex. According to INSEE, 0.6% of partnered individuals in France live with a person of the same sex (Buisson and Lapinte 2013). This percentage applied to our sample equals 84 of 14,019 respondents. We therefore treated all couples as heterosexual.
  • [5]
    See Plessz et al. 2015 and Model 3, Table 4, estimated without the 1990 questionnaire.
  • [6]
    The company’s human resource department classifies occupations as exécution (routine manual or non-manual work; here “subordinate”), maîtrise (technicians, lower management; here “intermediate”), or cadre (executive; here “senior manager”).
  • [7]
    We checked that there was no cause to stratify these control variables by sex.
  • [8]
    Hierarchical or multilevel models with fixed or random effects (where a time-independent parameter is estimated for each individual to capture their specificities) are not as well adapted here because we had relatively little information for estimating those parameters (five 0 or 1 observations for each individual). In direct contrast, we had a great number of observations (several thousand) for estimating each of the ten parameters of the correlation matrix (between the five observation years) needed for the GEE model. First-difference models (difference between “before” and “after” union dissolution) would have enabled us to identify the effect of dissolution but not to situate it within the aging process as we sought to do here.
  • [9]
    This table comprises 10 variables x 14,019 individuals x 5 rows per individual, generating 700,950 cells, 42,263 of which are empty due to non-responses.
  • [10]
    The percentage of daily vegetable eaters is lower among incomplete cases who answered the 2014 questionnaire (than among complete cases): 49.5% of women and 29.2% of men. In that year, complete cases included 241 now non-partnered men and 63 now non-partnered women. For all respondents (all complete and incomplete cases available in 2014), the corresponding numbers were 837 men and 391 women.
  • [11]
    It may be that nutrition recommendations for children are counterbalanced by the additional complexities of living with children (lack of time, having to take into account the tastes of all household members). See Devault 1991 and Plessz et al. 2016.
  • [12]
    We may not have corrected all bias and we cannot check because by definition we do not know non-respondents’ answers. In cases where there is a suspicion that non-responses are “missing not at random” (or where no observed variables are available to correct bias), sensibility tests can be done.
  • [13]
    We can only test these hypotheses in the imputed data model if we relax them in the imputation model, which we did not do (White et al. 2011).
  • [14]
    Another explanatory hypothesis is that women like vegetables more than men, but then women’s vegetable consumption should increase after union dissolution, which is not the case.
  • [15]
    We cannot exclude the possibility that even though we took into account the main confounding factors, including respondents’ health, our results are due in part to a non-observed heterogeneity between respondents who experienced union dissolution and the others. This explains why our discussion focuses not so much on the causal effect of union dissolution as on men-women comparisons and how the dissolution event may be related to the aging process.
  • [16]
    In an earlier article using very similar methodology (Plessz et al. 2015), year of birth (in two categories) was controlled for. It did have an effect, but the age effect was of the same magnitude as found here.
  • [17]
    Total non-responses are often reweighted (using inverse probability weighting, for example).

Is living together beneficial to both men and women or to men rather than women? How does a major life event such as loss of one’s partner impact the aging process? We investigate these questions by examining a mainly domestic practice—vegetable consumption—that is also a clear indicator of a lifestyle that complies with current health recommendations and conforms to dominant-class tastes. We used data from INSERM’s epidemiological Gazel cohort, comprising 20,625 respondents followed from 1989. Our results: after union dissolution, men eat fewer vegetables than women and their consumption level is more sensitive to their (female) partner’s social position than women’s is to their (male) partner’s; women’s vegetable consumption does not decline when the dissolved union was with a man of relatively low socio-occupational status. Our conclusion: in the aging population studied, conjugality benefits both partners but men more than women. The article makes a methodological contribution on the question of non-response in cohort data and a theoretical contribution through our discussion of a possible intersection between gender, class, and conjugal status in the domestic sphere.

  • aging
  • conjugal situation
  • eating practices
  • gazel cohort
  • gender
  • multiple imputation


  • Agresti, Alan. 2013. Categorical Data Analysis. 3rd ed. New York: Wiley.
  • Aïach, Pierre, and Daniel Delanoë. 1998. L’Ère de la médicalisation: Ecce homo sanitas. Paris: Anthropos.
  • OnlineArber, Sara. 2004. “Gender, Marital Status, and Ageing: Linking Material, Health, and Social Resources.” Journal of Aging Studies 18 (1): 91-108.
  • Babayou, Patrick, and Jean-Luc Volatier. 1997. “Les effets d’âge et de génération dans la consommation alimentaire.” Cahier de recherche du CRÉDOC 105: 1-63.
  • OnlineBickel, Jean-François, and Christian Lalive d’Épinay. 2001. “Les styles de vie des personnes âgées et leur évolution récente: une étude de cohortes.” In La retraite: une révolution silencieuse, edited by Monique Legrand, 243-80. Toulouse: Érès (Pratiques du champ social).
  • Blanpain, Nathalie. 2016. “Les hommes cadres vivent toujours 6 ans de plus que les hommes ouvriers.” INSEE Première 1584.
  • OnlineBonnet, Carole, Sophie Buffeteau, and Pascal Godefroy. 2006. “Disparités de retraite entre hommes et femmes: quelles évolutions au fil des générations?” Économie et statistique 398-399: 131-48.
  • Bouhia, Rachid. 2007. “Les personnes en couple vivent plus longtemps.” INSEE Première 1155.
  • Bry, Xavier, Nicolas Robette, and O. Roueff. 2016. “A Dialogue of the Deaf in the Statistical Theater? Addressing Structural Effects within a Geometric Data Analysis Frame- work.” Quality & Quantity 50 (3): 1009-20.
  • Buisson, Guillemette, and Aude Lapinte. 2013. “Le couple dans tous ses états.” INSEE Première 1435.
  • OnlineCambois, Emmanuelle, Caroline Laborde and Jean-Marie Robine. 2008. “La double peine’ des ouvriers: plus d’années d’incapacité au sein d’une vie plus courte.” Population & Sociétés 441. Published simultaneously in English as “A Double Disadvantage for Manual Workers: More Years of Disability and a Shorter Life Expectancy,” Population & Societies 441.
  • Caradec, Vincent. 2001. Sociologie de la vieillesse et du vieillissement. Paris: Armand Colin.
  • OnlineCardon, Philippe. 2009a. “Les effets de la mobilité résidentielle des retraités sur leur alimentation.” Recherches familiales 1: 105-15.
  • OnlineCardon, Philippe. 2009b. “‘Manger’ en vieillissant pose-t-il problème? Veuvage et transformations de l’alimentation des personnes âgées.” Lien social et politiques 62, special issue, “Vieillir pose-t-il vraiment problem?”: 85-95.
  • Charles, Nickie, and Marion Kerr. 1988. Women, Food, and Families. Manchester: Manchester University Press.
  • Chauvel, Louis. 2002. Le destin des générations: structure sociale et cohortes en France au xxe siècle. Paris: Presses Universitaires de France.
  • OnlineCousteaux, Anne-Sophie, and Jean-Louis Pan Ké Shon. 2008. “Le mal-être a-t-il un genre? Suicide, risque suicidaire, dépression et dépendance alcoolique.” Revue française de sociologie 49 (1): 53-92. Published in English as “Is Ill-being Gendered? Suicide, Risk for Suicide, Depression and Alcohol Dependence,” Revue française de sociologie 51 (5) 2010: 3-40.
  • Cresson, Geneviève. 1995. Le travail domestique de santé: analyse sociologique. Paris: L’Harmattan.
  • DeVault, Marjorie L. 1991. Feeding the Family: The Social Organization of Caring as Gendered Work. Chicago: The University of Chicago Press.
  • OnlineEng, Patricia Mona, Ichiro Kawachi, Garrett Fitzmaurice, and Eric Rimm. 2005. “Effects of Marital Transitions on Changes in Dietary and other Health Behaviours in US Male Health Professionals.” Journal of Epidemiology and Community Health 59 (1): 56-62.
  • Escalon, Hélène, Claire Bossard, and François Beck, eds. 2009. Baromètre santé nutrition 2008. Saint-Denis: INPES.
  • Ferrant, Coline, and Marie Plessz. 2015. “Structure des budgets alimentaires dans l’enquête Budget de famille 2011.” Aliss Working Paper 2015-02;
  • OnlineGadéa, Charles, and Catherine Marry. 2000. “Les pères qui gagnent.” Travail, genre et societés 1: 109-35.
  • OnlineGojard, Séverine. 2000. “L’alimentation dans la prime enfance. Diffusion et réception des normes de puériculture.” Revue française de sociologie 41 (3): 475-512.
  • OnlineGoldberg, Marcel, Annette Leclerc, Sébastien Bonenfant, Jean-François Chastang, Annie Schmaus, Nadine Kaniewski, and Marie Zins. 2007. “Cohort Profile: The GAZEL Cohort Study.” International Journal of Epidemiology 36 (1): 32-39.
  • OnlineGoldberg, Marcel, Annette Leclerc, and Marie Zins. 2015. “Cohort Profile Update: The GAZELCohort Study.” International Journal of Epidemiology 44 (1): 77-77g.
  • Grignon, Claude. 1981. “Alimentation et stratification sociale.” Cahiers de nutrition et de diététique 16 (4): 207-17.
  • Grignon, Claude, Louis Gruel and Bernard Bensoussan. 1996. Les conditions de vie des étudiants. Paris: La Documentation française.
  • OnlineHébel, Pascale, and Fanette Recours. 2007. “Effets d’âge et de générations: transformation du modèle alimentaire.” Cahiers de nutrition et de diététique 42 (6): 297-303.
  • Institute for Digital Research in Education (IDRE). “Multiple Imputation in Stata” UCLA: lode. [accessed May 22, 2017].
  • Jourdain Menninger, Danièle, Gilles Lecoq, Jérôme Guedj, Pierre Boutet, Jean-Baptiste Danel, and Mathieu Gérard. 2010. Évaluation du programme national nutrition santé PNNS 2 2006-2010. Paris: Inspection générale des affaires sociales.
  • Kaufmann, Jean-Claude. 2005. Casseroles, amour et crises: ce que cuisiner veut dire. Paris: Armand Colin.
  • OnlineKehily, Mary Jane, Lydia Martens, Kate Burningham, Susan Venn, Ian Christie, Tim Jackson, and Birgitta Gatersleben. 2014. “New Motherhood: A Moment of Change in Everyday Shopping Practices?” Young Consumers 15 (3): 211-26.
  • Kranklader, Élodie. 2014. “Où fait-on ses courses? Les achats en ligne progressent, excepté pour l’alimentation.” INSEE Première 1526.
  • OnlineLalive d’Épinay, Christian, and Stefano Cavalli. 2007. “Changements et tournants dans la seconde moitié de la vie.” Gérontologie et société 121 (2): 45-60.
  • OnlineLhuissier, Anne. 2012. “Weight-Loss Practices of Working Class Women in France.” Food, Culture and Society: An International Journal of Multidisciplinary Research 15 (4): 643-64.
  • OnlineLindström, M., B.S. Hanson, E. Wirfält, and P.-O. Öostergren. 2001. “Socioeconomic Differences in the Consumption of Vegetables, Fruit and Fruit Juices.” European Journal of Public Health 11 (1): 51-59.
  • Lupton, Deborah. 1995. The Imperative of Health: Public Health and the Regulated Body. London: Sage.
  • Marenco, Claudine. 1992. Manières de table, modèles de moeurs. Cachan: Les Éditions de l’École Normale Supérieure de Cachan.
  • OnlineMarshall, David W., and Annie S. Anderson. 2002. “Proper Meals in Transition: Young Married Couples on the Nature of Eating Together.” Appetite 39 (3): 193-206.
  • OnlinePlessz, Marie. 2013. “Les légumes transformés: diversité des produits, diversité des usages sociaux.” Revue d’études en agriculture et environnement 1: 13-37.
  • OnlinePlessz, Marie, Sophie Dubuisson-Quellier, Séverine Gojard, and Sandrine Barrey. 2016. “How Consumption Prescriptions Affect Food Practices: Assessing the Roles of Household Resources and Life-Course Events.” Journal of Consumer Culture 16 (1): 101-23.
  • OnlinePlessz, Marie, and Séverine Gojard. 2013. “Do Processed Vegetables Reduce Socio-Economic Differences in Vegetable Purchases? A Study in France.” European Journal of Public Health 23 (5): 747-52.
  • OnlinePlessz, Marie, and Séverine Gojard. 2015. “Fresh is Best? Social Position, Cooking, and Vegetable Consumption in France.” Sociology 49 (1): 172-90.
  • OnlinePlessz, Marie, Alice Guéguen, Marcel Goldberg, Sébastien Czernichow, and Marie Zins. 2015. “Ageing, Retirement and Changes in Vegetable Consumption in France: Findings from the Prospective GAZEL Cohort.” British Journal of Nutrition 114 (6): 979-87.
  • OnlinePochic, S. 2005. “Faire carrière: l’apport d’une approche en termes de genre.”Formation emploi 91: 75-93.
  • OnlineRégnier, Faustine, and Ana Masullo. 2009. “Obésité, goûts et consummation: Intégration des normes d’alimentation et appartenance sociale.” Revue française de sociologie 50 (4): 747-73.
  • Ricroch, Layla. 2012. “En 25 ans, moins de tâches domestiques pour les femmes, l’écart de situation avec les hommes se réduit.” In Regards sur la parité, édition 2012, edited by INSEE. Paris: INSEE.
  • OnlineRobert-Bobbée, Isabelle, and Christian Monteil. 2006. “Différentiels sociaux et familiaux de mortalité aux âges actifs: quelles différences entre les femmes et les hommes?” Économie et statistique 398-399: 11-31.
  • OnlineRoos, Gun, Lars Johannsen, Anu Kasmel, Juraté Klumbiené, and Ritva Prättälä. 2001. “Disparities in Vegetable and Fruit Consumption: European Cases from the North to the South.” Public Health Nutrition 4 (1): 35-43.
  • Rubin, David B. 1987. Multiple Imputation for Nonresponse in Surveys. New York: Wiley.
  • Saint Pol, Thibaut de. 2008. “La consommation alimentaire des hommes et femmes vivant seuls.” INSEE Première 1194.
  • OnlineSterne, Jonathan A.C., Ian R. White, John B. Carlin, Michael Spratt, Patrick Royston, Michael G. Kenwood, Angela M. Wood, and James R. Carpenter. 2009. “Multiple Imputation for Missing Data in Epidemiological and Clinical Research: Potential and Pitfalls.” British Medical Journal 338: b2393
  • Settersten, R.A., and Jacqueline L. Angel, eds. 2011. Handbook of Sociology of Aging. New York: Springer (Handbooks of Sociology and Social Research).
  • Twisk, Jos W.R. 2013. Applied Longitudinal Data Analysis for Epidemiology: A Practical Guide. 2nd ed. Cambridge: Cambridge University Press.
  • OnlineUmberson, Deborah, Hui Liu, and Daniel A. Powers. 2009. “Marital Status, Marital Transitions, and Body Weight. Journal of Health and Social Behavior 50 (3): 327-43.
  • OnlineVallin, Jacques, and Alfred Nizard. 1977. “La mortalité par état matrimonial. Mariage sélection ou mariage protection.” Population 32 (1): 95-125.
  • OnlineWajcman, Judy. 1996. “The Domestic Basis for the Managerial Career.” The Sociological Review 44 (4): 609-29.
  • Warde, Alan. 2015. “On the Sociology of Eating.” Revue d’études en Agriculture et environnement 96 (1): 7-15.
  • OnlineWarde, Alan, Shu-Li Cheng, Wendy Olsen, and Dale Southerton. 2007. “Changes in the Practice of Eating.” Acta Sociologica 50 (4): 363-85.
  • OnlineWeber, Florence. 1996, “Réduire ses dépenses, ne pas compter son temps. Comment mesurer l’économie domestique?” Genèses 25 (1): 5-28.
  • OnlineWhite, Ian R., Patrick Royston, and Angela M. Wood. 2011. “Multiple Imputation Using Chained Equations: Issues and Guidance for Practice.” Statistics in Medicine 30 (4): 377-99.
  • OnlineWilliams, Kristi, and Deborah Umberson. 2004. “Marital Status, Marital Transitions, and Health: A Gendered Life Course Perspective.” Journal of Health and Social Behavior 45 (1): 81-98.
  • World Health Organization. 2003. Fruit and Vegetable Promotion Initiative/A Meeting Report, 25-27/08/2003. Geneva: World Health Organization.
  • OnlineWyndels, K., J. Dallongeville, C. Simon, V. Bongard, A. Wagner, J.-B. Ruidavets, D. Arveiller, J. Ferrières, P. Amouyel, and L. Dauchet. 2011. “Regional Factors Interact with Educational and Income Tax Levels to Influence Food Intake in France.” European Journal of Clinical Nutrition 65 (9): 1067-75.
Marie Plessz
Centre Maurice Halbwachs (CMH)
INRAE-CNRS-ENS-EHESS-PSL Research University
48, boulevard Jourdan
75014 Paris
Alice Guéguen
UMS 011 Cohortes epidémiologique en population
Hôpital Paul Brousse – Bâtiment 15/16
16, avenue Paul Baillant-Couturier
94807 Villejuif
Translated by
Amy Jacobs-Colas
This is the latest publication of the author on cairn.
Uploaded on on 24/06/2020
Distribution électronique pour Presses de Sciences Po © Presses de Sciences Po. Tous droits réservés pour tous pays. Il est interdit, sauf accord préalable et écrit de l’éditeur, de reproduire (notamment par photocopie) partiellement ou totalement le présent article, de le stocker dans une banque de données ou de le communiquer au public sous quelque forme et de quelque manière que ce soit.
Loading... Please wait