Longitudinal adherence to a dietary pattern derived by reduced rank regression (RRR) and risk of depressive symptoms in Japanese employees

概要

Background: Depression is a common mental health problem that reduces work productivity, lowers quality of life, and increases mortality. Several studies have reported associations between individual nutrients or foods and depressive symptoms. However, single-nutrient or food analysis in relation to disease is challenging because many nutrients are highly correlated with each other and this analysis may not capture the interactive or synergistic effects among nutrients. To overcome these limitations, dietary pattern analysis has emerged as a complementary approach that better reflects the complexity of diet in daily life and its relation with disease. Two main methodologies facilitate the identification of dietary pattern—a priori defined methods (e.g., diet quality scores) and exploratory methods (e.g., principal component analysis: PCA)—but neither method incorporates existing evidence of the pathway from diet to the specific disease.
　To address this issue, reduced rank regression (RRR) has been proposed. RRR identifies dietary patterns based on disease-specific evidence (e.g., several nutrients or biomarkers) that have been linked to the target disease. However, RRR-derived dietary patterns are not necessary consumed together and could be behaviorally irrelevant. Therefore, the roles of the RRR and PCA methods may be complementary, and provide useful insights when the results are compared. That is, RRR-derived dietary patterns can put PCA findings into perspective by indicating dietary patterns closely related to a target disease and elucidating the possible pathway from diet to the disease by considering disease-specific evidence (e.g., intermediate nutrients or biomarkers), while PCA-derived dietary patterns can put into perspective how behaviorally meaningful RRR-derived dietary patterns are by showing actual intake patterns in the population.
　The primary objective of this current study was to examine the prospective association of a dietary pattern derived by RRR with subsequent development of depressive symptoms in Japanese workers. The secondary goal was to use both RRR and PCA in complementary roles so as to compare dietary structures derived from them, and thereby investigate the cross-sectional and bidirectional prospective association of dietary patterns with depressive symptoms in the same cohort of Japanese workers. I used RRR to derive dietary patterns on the basis of its ability to maximally explain variation in beneficial nutrients against depressive symptoms. A comparison of the food and beverage composition of behavior-based dietary patterns obtained through PCA with those of RRR-derived patterns based on mood-related nutrients will provide perspective on the value of the two kinds of methods (e.i., PCA and RRR), that is, whether prior scientific evidence of mood-related nutrients would increase the strength of the associations of dietary patterns for depressive symptoms compared with solely based on eating behavior. On the basis of previous findings of PCA, I hypothesized that both PCA and RRR techniques would reveal dietary patterns associated with depressive symptoms. However, by incorporating mood-related nutrients as intermediates in the RRR analysis, I hypothesized that the RRR method would uncover dietary patterns more strongly associated with depressive symptoms than PCA-derived patterns.

Methods: The Furukawa Nutrition and Health Study, a nutritional epidemiological survey, was conducted at the time of a periodic health examination among workers of two works of a manufacturing company in Chiba and Kanagawa. Of 2828 health check-up attendees, 2162 agreed to participate in the baseline survey. Of these, 1354 also responded to the follow-up survey.
　Participants with depressive symptoms were defined as those with a score ≥ 16 on the Center for Epidemiologic Studies Depression Scale. Dietary patterns were derived by PCA and RRR at baseline and follow-up survey using a validated, self-administered diet history questionnaire. RRR identifies linear functions of predictors (i.e., the dietary patterns) that explain as much of the variation of response variables (e.g., selected nutrients that is potentially protective against or contributory to the relevant disease). I chose 52 food and beverage items as predictors. We selected the following ten nutrients as potentially protective factors for depressive symptoms (response variables): folate, vitamin C, magnesium, calcium, iron, zinc, vitamin D, vitamin B6, vitamin B12, and n-3 polyunsaturated fatty acids. Biological and epidemiological evidence suggests that these nutrients may be potentially protective factors for depression. PCA works with only one set of predictors (i.e.,52 food and beverage items) and results in the extraction of uncorrelated linear functions of predictors (i.e., the dietary patterns) that explain as much predictor variation as possible.
　The explained percentage of variations in nutrients and food and beverage items attributable to dietary patterns and Pearson correlation coefficients in nutrients, food and beverage items, and dietary patterns scores were calculated. Additionally, the Pearson correlation coefficient between RRR- and PCA-derived dietary pattern scores at baseline was calculated. Multiple logistic regression was used to estimate odds ratios of depressive symptoms according to the baseline dietary pattern score and changes in dietary pattern scores during followup with adjustment for confounders. To examine the reverse causality, I used baseline depressive symptoms as a predictor in a multiple linear regression analysis, with dietary pattern scores at follow-up as outcome.

Results: After excluding participants with severe diseases and missing data for this analysis, I analyzed data among 2006 participants who took part in a baseline surveys for a cross-sectional analysis and among 903 participants without depressive symptoms at baseline who responded to both baseline and follow-up surveys for prospective association. Among RRR- and PCA-derived dietary patterns, first factors (RRR- and PCA-DP1) explained 62.4 % and 48.1 % of variation in mood-related nutrients selected as response variables. Both diets were rich in vegetables, mushrooms, seaweeds, fish, tofu/atsuage, fruits, potatoes, egg and were low in rice. The RRR-DP1 was relatively highly related with fish items, green tea, and natto, but weakly related with mayonnaise/dressing as compared with those of PCA-DP1.
　In a cross-sectional analysis, both RRR-DP1 and PCA-DP1 were associated with decreased prevalence of depressive symptoms. In the prospective analysis, maintaining high or increasing adherence to RRR-DP1 but not PCA-DP1 was associated with a decreased risk of depressive symptoms among Japanese workers. There was no significant association between 3-year changes in PCA-DP1 scores and the risk of depressive symptoms. Additionally, baseline RRR- and PCA-DP1 were not associated with subsequent development of depressive symptoms after 3 years. To explore reverse causality, I examined the prospective association of depressive symptoms at baseline as predictor with the RRR- and PCA-DP1 scores at follow-up survey as outcome, but found no association between them.

Discussion: Both RRR- and PCA-DP1 associated with decreased prevalence of depressive symptoms in a cross-sectional analysis. In the prospective analysis, maintaining high or increasing adherence to RRR-DP1 but not PCA-DP1 is associated with a lower risk of depressive symptoms in Japanese employees. This study is the first to compare RRR- and PCA-derived dietary patterns in relation to depressive symptoms, finding that RRR is more appropriate than PCA in identify dietary patterns relevant to depressive symptoms and that a RRR-derived dietary pattern also reflects the eating patterns of Japanese male workers in manufacturing industry.
　In this study, the RRR pattern was closely related to the structure of PCA-pattern. Both diets were rich in vegetables, mushrooms, seaweed, fish, tofu/atsuage, fruits, potatoes, and eggs, and were low in rice. Indeed, baseline RRR- and PCA-DP1 scores were similar (r = 0.87). Considering the PCA-patterns based on eating patterns of the population, those similarities and correlations between PCA- and RRR-derived patterns show that RRR patterns also reflect eating behaviors of the population.
　In my prospective analysis, maintaining high or increasing adherence to RRR-DP1 but not PCA-DP1 was associated with a lower risk of depressive symptoms in Japanese employees. These differential findings may be due to differences in the composition of food and beverage items and explained variation in mood-related nutrients between RRR-and PCA-DP1. Concerning non-concordant food or beverage items between RRR and PCA analysis, RRR-DP1 was relatively high in fish items, green tea, and natto but low in mayonnaise/dressing compared with those of PCA-DP1. Among non-concordant food or beverage items, fish items, green tea, and natto, which possessed relatively high factor loading values for RRR-DP1, were correlated with response nutrients, whereas mayonnaise/dressing, which had relatively high factor loading values for PCA-DP1, were weakly correlated with these nutrients. Indeed, RRR-DP1 were highly correlated with each response nutrient. All Pearson correlation coefficients for RRR-DP1 scores were ≥ 0.69. On the other hand, corresponding values for PCA-DP1 were ≥ 0.41. These differences may account for the difference in associations of RRR- and PCA-derived dietary patterns with depressive symptoms.
　Contrary to expectation, I did not detect a measurable relation between baseline RRR- and PCA-DP1 scores with subsequent development of depressive symptoms after 3 years. The null finding from a prospective analysis was inconsistent with that from a cross-sectional analysis, which found inverse association of RRR- and PCA-DP1 scores with depressive symptoms at baseline survey. This discrepancy between cross-sectional and longitudinal findings raises the possibility of reverse causality. Thus, we examined the prospective association of baseline depressive symptoms as predictor with the RRR- and PCA-DP1 scores at follow-up survey as outcome, but found no association between them. Regarding RRR analysis, the other possibility might be that the protective association of baseline RRR-DP1 with subsequent development of depressive symptoms might be masked by dietary changes during the follow-up period. Alternatively, the observed associations for change in RRR-DP1 scores may not imply causation, but rather simply reflect cross-sectional relations. As to PCA method, I ascribed the lack of association with PCA-DP1 for the risk of depressive symptoms to relatively low attributable variation in response nutrients as compared with that of RRR-DP1. Aside from above mentioned reasons for the lack of a relation between RRR- and PCA-DP1 scores at baseline with subsequent development of depressive symptoms after 3 years, the possible explanation may be that a 3-year follow-up could be long enough to detect a measurable association between them.
　My study has some limitations. First, the large loss to followup may have introduced selection bias. Second, dietary intake assessed at baseline and 3 years might not accurately represent the long-term habitual diet. Third, I cannot rule out the possibility that the observed associations were due to unmeasured and residual confounders. Fourth, because the study was conducted among workers in a Japanese manufacturing company, caution is required in generalizing the findings. Fifth, dietary pattern scores derived by the RRR and PCA methods in this study cannot be reproduced with data from another study population due to the nature of data dependence. Sixth, although identification of a modifiable factor not only for mental health but also for work engagement and productivity is a priority issue in the occupational setting, the present study lacked such data. Seventh, CES-D algorithm may not always match a diagnosis of depression or detect participants who actually have depressive symptoms linked to a major depressive disorder. Eighth, the weakness of the RRR method is that it does not consider the degree of existing evidence levels of each response variable via optimally-weighted each response variable based on their evidence levels. Ninth, RRR requires the availability of response information. The RRR approach cannot be used to evaluate diet–disease associations in the case of missing information regarding intermediate variables. Finally, my exposure variables were defined as changes in dietary pattern scores from baseline to follow-up survey after 3 years and the outcome was defined as the development of depressive symptoms at 3 years. Thus, I was not able to assess exposure status that absolutely preceded the outcome, and this inability limited the drawing of causal inference even though participants were free from depressive symptoms at the beginning of present cohort study.