Results of this study are an important addition to the literature on psychosocial predictors of eating behaviour. By employing a longitudinal design and adjusting for socio-demographic confounders, as well as previous eating behaviour, we found that attitudes, subjective norms, perceived behavioural control, perceived social norms and intention to eat healthier food the next four weeks emerged as significant predictors of one or several eating behaviours eight years later. Of the socio-demographic factors, only education was positively associated with healthy eating for both sexes, while a higher income was associated with a low sugar intake and to be single or divorced was associated with a lower fat intake among men. Overall, the factors examined accounted for 4% to 20 % of the variation in follow-up eating behaviour.
Fruit and vegetable intake
For men, neither subjective norms nor attitude was significantly correlated with fruit and vegetable intake in bivariate analyses. However, in multivariate analyses, these factors appeared to be significantly associated with the intake of fruits and vegetables at follow-up. Because no significant associations existed between psychosocial factors and fruit and vegetable intake among men in bivariate analyses, multivariate associations might be artefacts, but they could also be a suppression phenomenon, see under Internal correlations below. Perceived behavioural control and perceived social norms measured at baseline remained predictive of women's intake of fruits and vegetables. In the 1991 survey, perceived social norms were found to be predictive of healthy eating . In 1999, subjective norms were negatively associated with intake of fruits and vegetables because of the way the question was asked: "Do you believe that your parents, etc. think that you should eat healthier food the next four weeks?" Thus, those already having a healthy diet most likely did not believe their significant others expected them to eat healthier food. Conner and colleagues  investigated the longitudinal relationship between the constructs of TPB assessed at two time points six months apart (time 1 (T1) and time 2 (T2)) and eating behaviour assessed six years after T2. They found that intentions and the interaction between intentions and intention stability (defined as stability in intention between T1 and T2 six years prior to the last follow-up in their study) and between perceived past behaviour and intention stability were predictive of fruit and vegetable intake six years later. In our study, the constructs of TPB were measured only at age 25; hence, we can not address the stability of the psychosocial constructs. In an earlier publication by Conner and colleagues , they concluded that both stability in intentions and perceived behavioural control were important for future dietary behaviour. Our study adds to previous research in that perceived behavioural control, as well as perceived social norms among women measured once several years prior to dietary behaviour assessment, appear to be predictive of dietary behaviour.
Social class and family situation are factors shown to be predictive of dietary habits [15, 29–31]. However, none of the socio-demographic factors in our study emerged as predictors of fruit and vegetable intake when we adjusted for past behaviour. This lack of an association may indicate that the association is mediated through other variables, such as dietary habits and psychosocial factors.
Whole grain intake
Education and previous behaviour remained significant predictors of whole grain intake among men and women in multivariate analyses, while the subjective norms construct was significant only for women. This is consistent with previous findings showing that a higher social class is associated with higher rates of consumption of whole grain foods [14, 32]. As for fruit and vegetable intake, the subjective norms construct was negatively associated with whole grain intake. Patterson and colleagues  did not find that perceived norms, which resembled subjective norms in our study, explained fibre intake three years later.
Intention to eat healthier food measured at baseline was significantly and negatively associated with men's fat intake at follow up, even after adjusting for past behaviour. Perceived behavioural control remained a significant predictor of fat intake among women in multiple regression analyses. This is in contrast to the findings of Conner and colleagues, who found that intentions and the interaction between intentions and intention stability were the only significant predictors of fat intake six years later . While Conner and colleagues' study included both genders, women constituted 83% of their sample. Among men in our study, marital status appeared to be a significant predictor of fat intake; to be single/divorced was associated with a lower fat intake when adjusted for psychosocial factors and past behaviour. This is in contrast to a Finnish study finding that married men and women had diets more in line with the dietary guidelines than not married men and women . We have no explanation for our finding, and it may also be spurious.
In the final model, perceived social norms and household income were predictive of men's sugar intake. In a previous report from the Oslo Youth Study, perceived social norms, represented by partners, were also predictive of healthy eating when adjusted for education . For women, perceived social norms were predictive of sugar intake; however, when we included past behaviour in the model, past behaviour was the only factor that was predictive of women's sugar intake. A cross-sectional study by Grogan and colleagues examined gender differences in attitudes and behaviour using the Theory of Reasoned Action regarding eating sweet snacks . The authors found that perceived social pressure and attitudes toward sweet snacks were associated with women's intentions to eat sweet snacks, while only attitudes were associated with men's intentions. Both men's and women's intentions were associated with reported intake of sweet snacks. However, the perceived social pressure construct in that study resembled the subjective norms construct in our study, and in that respect, our findings were similar. However, in our study, subjective norms lost its significance in predicting sugar intake when we included demographic factors and past intake of food high in sugar in the models.
Past eating behaviour
Scores representing past behaviour were predictive of intake of all dependent dietary measures for both sexes, with the exception of fat. Few studies have investigated the stability of nutrient intakes  and eating habits  during adult years. A study by Mulder and colleagues  on the stability of lifestyle behaviour over four years among adult men 30 – 39 years of age at baseline found the correlation coefficient between the dietary scores (including meal pattern, sweet and salty snacks, fruit and attitudes toward eating fat and fibre) at time one and time two to be 0.57. Our findings indicate that past habits are important predictors of current habits even when adjusting for socio-demographic and psychosocial factors and when taking into account that a different measure to assess food intake was used at baseline of the study. Conner and Armitage  have proposed that past behaviour may predict future behaviour as a moderator of the relationship between the TPB variables, as a source of information and as a mediator of TPB variables. They argue that there are good reasons to incorporate frequency of past behaviour as predictors of current behaviour in the TPB alongside intentions and perceived behavioural control. The results of our study support the view that past behaviour has an independent predictive value of fruit and vegetable intake, whole grain intake and sugar intake when taking the TPB constructs into account. Previous studies have shown that past behaviours were predictive of future behaviours independent of intentions, attitudes, norms and perceived behavioural control (PBC) . Conner and Armitage  reported that after accounting for PBC and intentions, past behaviour, on average, could explain 13% (3% to 28%) of the observed variance in behaviour. In our study, an additional 6.2% on average (ranging from 0.5% to 10.6%) of variance in behaviour was explained by past behaviour beyond what was explained by the TPB variables and demographic factors. Given that the TPB is not often used to investigate dietary behaviours, comparing the explained variance between studies is difficult. The total explained variance in dietary habits in our study is, however, comparable to findings reported by Conner and colleagues . The low explained variances found in our study and in other studies investigating psychosocial and demographic factors' prediction of dietary habits point to other variables having impact on dietary habits. Such factors have not been examined in this study, but physical environment [38, 39], as well as taste, cost and convenience have been proposed as important to dietary behaviours [40, 41].
Dietary differences between men and women have previously been demonstrated, with women generally reporting healthier eating habits than those reported by men . Our results at baseline agree with this finding, as women had higher scores on fruit and vegetable intake and lower scores on fat and sugar intake compared to scores among men. However, at follow-up we observed no statistically significant differences between men and women's dietary intakes. In a representative and random sample of Norwegians 16 to 79 years of age in 1993 that applied the same method as the 1999 follow up of the Oslo Youth Study, researchers reported that women had higher intakes of vegetables and of fruit and berries compared to men, while men had higher intakes of cereals compared to women . Also, fat and sugar intake differed between men and women in the previous study, while in our study the differences were not statistically significant. Even though most findings from our study showed the same patterns as in previous studies, differences between men and women did not reach statistical significance. This may be due to differing age groups and places of residence, as these factors also influence dietary intakes assessed with this method .
Gender differences in the TPB constructs regarding healthier eating have not been reported previously, but a study applying the Theory of Reasoned Action found women were under more social pressure not to eat sweet snacks than were men . Barker and colleagues demonstrated that fat-phobic and fibre-philic attitudes were more prevalent among women, and that fat-phobic attitudes were inversely related to fat intake among women, but not among men . A Norwegian study showed that, compared to men, women were more prone to consider foods that were in accordance with dietary recommendations as healthy, and less prone to consider fat- and protein-rich foods as healthy . However, the gender difference disappeared when including "trust in experts" in the model, indicating that women's higher trust in experts might be one reason they ranked fish, fruits, vegetables and potatoes as healthy foods. This is similar to Grogan and colleagues' findings that women felt more pressure from health experts than did men to avoid eating sweet snacks . Our results support this finding by indicating that men and women might use different psychosocial bases to carry out certain behaviours, but as we can see no certain patterns regarding gender, we can not conclude about how TPB constructs predict men's and women's eating behaviour differently.
The relatively high attrition rate in this study is comparable to attrition rates found in previous longitudinal studies with similar follow-up periods [47, 48]. There were, however, no baseline differences between participants and drop-outs, and we do not think that the attrition seen in this study is a threat to the validity of the observed prospective relationships.
The psychosocial factors in this study were constructed around healthier eating defined as "food low in fat, sugar and salt." Being aware that fruit, vegetables and whole meal bread was not included in the definition of a healthy diet in the 1991 survey is important. In 1991, the dietary focus in Norwegian society, as well as in the Oslo Youth Study, was on reducing the intake of fat and salt. The dietary focus in 1991 influenced the way the TPB questions were phrased and this might have influenced the observed associations between psychosocial factors and dietary habits. Baranowski and colleagues  claim that the influence on dietary habits varies by foods, and that the predictive value of TPB appears to be higher in predicting intake of a single food item or narrow categories of foods. Also the variability in measurement may play a role in the prediction of, for instance, fat eating patterns , which can be measured by means of total fat intake in grams per day, as per cent of total energy from fat or as foods high in fat. All this will contribute to differing prediction of different behaviours by psychosocial factors.
The time period for which intentions and attitudes toward healthier eating was applied at baseline in this study was "the next four weeks." It is therefore remarkable that the constructs of TPB contributed to the prediction of eating habits assessed eight years later. Results might point to the stability of the underlying psychosocial constructs. These findings are in agreement with previous studies investigating the stability of TPB constructs over time and as predictors of food choice over time [7, 8].
The methodology of assessing diet in 1991 and 1999 differed. The dietary method used in 1999 made it possible to compute total intake of energy, macro- and micronutrients, while the questionnaire used in 1991 only enabled assessment of intake frequencies. Prediction of single nutrients in 1999 by means of previous intake was difficult as we did not have measures of the same nutrient eight years earlier. However, the single food items measured in 1991 are good sources for the particular 1999 nutrients . Despite the differing methods used to assess dietary intakes, past behaviour was predictive of current behaviour for all items except fat intake, indicating a high degree of stability in dietary habits. Dietary scores composed of intake of specific food items and nutrients are shown to be valid for evaluating diet quality among adults [50, 51].
In multivariate analyses of prediction of fruit and vegetable for men, the high correlations between subjective norms and attitudes might explain why variables not significant in bivariate analyses became significantly associated with the diet under investigation in multivariate analyses. Subjective norms might act as suppressors on attitudes, and vice versa, in bivariate analyses between each of these constructs and the intake of fruit and vegetables (negative confounding). However, in the multivariate analyses, these constructs will be mutually adjusted and the association between each of them and fruit and vegetable intake will become significant. However, the observed associations between attitude, subjective norms and fruit and vegetable intake among men in the multivariate analyses might also be an artefact. The other modest correlations between independent variables in this study are not regarded as a threat to the validity of the results.