Associations between dietary patterns and cardiovascular disease risk in Canadian adults: a comparison of partial least squares, reduced rank regression, and the simplified dietary pattern technique

ABSTRACT

Background

Hybrid methodologies have gained continuing interest as unique data reduction techniques for establishing a direct link between dietary exposures and clinical outcomes.

Objectives

We aimed to compare partial least squares (PLS) and reduced rank regression (RRR) in identifying a dietary pattern associated with a high cardiovascular disease (CVD) risk in Canadian adults, construct PLS- and RRR-based simplified dietary patterns, and assess associations between the 4 dietary pattern scores and CVD risk.

Methods

Data were collected from 24-h dietary recalls of adult respondents in the 2 cycles of the nationally representative Canadian Community Health Survey (CCHS)-Nutrition: CCHS 2004 linked to health administrative databases (n = 12,313) and CCHS 2015 (n = 14,020). Using 39 food groups, PLS and RRR were applied for identification of an energy-dense (ED), high-saturated-fat (HSF), and low-fiber-density (LFD) dietary pattern. Associations of the derived dietary pattern scores with lifestyle characteristics and CVD risk were examined using weighted multivariate regression and weighted multivariable-adjusted Cox proportional hazard models, respectively.

Results

PLS and RRR identified highly similar ED, HSF, and LFD dietary patterns with common high positive loadings for fast food, carbonated drinks, salty snacks, and solid fats, and high negative loadings for fruit, dark green vegetables, red and orange vegetables, other vegetables, whole grains, and legumes and soy (≥|0.17|). Food groups with the highest loadings were summed to form simplified pattern scores. Although the dietary patterns were not significantly associated with CVD risk, they were positively associated with 402-kcal/d higher energy intake (P-trends < 0.05) and higher obesity risk (PLS: OR: 2.09; 95% CI: 1.62, 2.70; RRR: OR: 1.76; 95% CI: 1.44, 2.17) (P-trends < 0.0001) in the fourth quartiles.

Conclusions

PLS and RRR were shown to be equally effective for the derivation of a high-CVD-risk dietary pattern among Canadian adults. Further research is warranted on the role of major dietary components in cardiovascular health.

Introduction

The study of dietary patterns has become an essential instrument for unraveling the complexities of dietary behaviors involved in the development of cardiovascular disease (CVD) (1, 2), the leading cause of global mortality and a major contributor to health care costs (3). Whereas previous approaches to dietary pattern analysis have commonly relied on either hypothesis-oriented (a priori) or data-driven (a posteriori) methods (4, 5), recently proposed hybrid methods combine predefined knowledge with data reduction techniques for the derivation of dietary patterns with a direct link to disease risk (6).

Reduced rank regression (RRR) and partial least squares (PLS) are novel hybrid techniques which aim to create linear combinations of food groups that best explain variation in a set of predefined response variables (e.g., clinical biomarkers, nutrients) that serve as potential mediators in the relation between dietary patterns and disease outcomes (7). Such a methodological approach to exploring diet's multidimensionality allows the incorporation of hypothesized pathways in dietary pattern analysis (8). An interesting approach proposed to circumvent the partial dependence of hybrid methods on study populations is the application of a simplified dietary pattern technique, which is calculated by summing unweighted standardized z scores of food variables (predictors) with the highest loadings for the pattern of interest (9). Although its ability to aid the construction of less data-dependent dietary patterns with improved interpretability and generalizability has previously been confirmed (10, 11), this methodology remains largely under-utilized in nutritional epidemiology.

Findings from several studies have revealed strong associations between dietary patterns derived with hybrid techniques and CVD risk (12–18). Nonetheless, such evidence has so far emerged predominantly from studies using RRR, with no previous attempts at comparing the 2 hybrid methods in the derivation of dietary patterns associated with CVD risk in large-scale nationally representative surveys. To our knowledge, no previous study has simultaneously utilized PLS, RRR, and simplified dietary pattern scores for the exploration of dietary patterns linked to CVD incidence and mortality. This is important because utilizing different techniques for deriving dietary patterns at the national level enables evaluation of real-world eating behaviors upon which dietary pattern recommendations and guidelines are based (19, 20). Therefore, the objectives of the present research were 1) to assess the performance of PLS as compared with RRR in identifying a dietary pattern closely associated with CVD risk [energy-dense, high-saturated-fat, and low-fiber-density (ED, HSF, and LFD)]; 2) to construct simplified dietary patterns based on the PLS- and RRR- derived dietary patterns; and 3) to explore the associations of these 4 dietary pattern scores with CVD incidence and mortality at the national level.

Methods

Study design

Data were obtained from the nationally representative cross-sectional Canadian Community Health Survey (CCHS)-Nutrition, cycles 2004 and 2015 (21, 22). The CCHS provides crucial information on Canadians’ health status, nutrient intakes, and eating habits from all 10 provinces (23). Details about the sampling frame and survey design, as well as differences between the 2 cycles, are available elsewhere (21, 22). Data were collected in accordance with the Statistics Act of Canada and were analyzed at Statistics Canada's Research Data Centre.

The following exclusion criteria were applied to data from CCHS-Nutrition 2015: individuals <18 y of age, individuals with invalid dietary recalls (as defined by Statistics Canada), breastfeeding women, respondents with missing energy intake and physical activity, and those with no food group intake (0 grams) for calculation of energy density. Exclusion criteria for CCHS-Nutrition 2004, on the other hand, included individuals <45 y and >80 y (for survival models), pregnant and breastfeeding women, individuals with invalid dietary recalls (as defined by Statistics Canada), residents of Quebec (for hospitalization records), respondents with missing data on energy intake, individuals with no food group intake (0 grams) for calculation of energy density, and those with heart diseases at baseline. Pregnant women were excluded only in CCHS-Nutrition 2004 because pregnancy status is masked in the 2015 Public Use Microdata File (PUMF) (24). Lack of measured height and weight data was compensated for by the application of a correction factor determined by Statistics Canada (25) on self-reported anthropometric measurements. Missing data on sociodemographic and lifestyle characteristics were imputed in CCHS-Nutrition 2004 (HMisc package in R) but excluded from CCHS-Nutrition 2015 owing to time and resource constraints.

The final sample size for descriptive analyses in CCHS-Nutrition 2004 and 2015 was comprised of 12,313 and 14,020 Canadian adults, respectively. The total number of excluded respondents was 22,794 for CCHS 2004 and 6467 for CCHS 2015. A smaller sample size of 6766 was used in survival analyses owing to the further exclusion of respondents <45 and >80 y of age from CCHS 2004 (Supplemental Figure 1).

All descriptive analyses used the first 24-h dietary recall from CCHS-Nutrition 2015, whereas longitudinal modeling of the impact of dietary patterns on CVD incidence and mortality utilized both days of 24-h dietary recall data available in the CCHS-Nutrition 2004 survey linked to health administrative data (Supplemental Methods).

Exposure and outcome assessment

Dietary intake

Information regarding food and beverage consumption was collected with the use of two 24-h dietary recalls and via the computer-based Automated Multiple-Pass Method (AMPM) (26). Additional information on dietary data collection and procedures can be found elsewhere (21, 22). All food items recorded in the 24-h dietary recalls were processed using the Canadian Nutrient File (CNF; 2015 version) from Health Canada (27).

Outcome ascertainment

CVD incidence and mortality were the primary outcomes, data on which were obtained through linkage of CCHS-Nutrition 2004 with hospital and health-administrative databases (2004–2017). CVD was assessed in accordance with the WHO's definition (28). CVD events, defined as CVD incident hospitalization cases or deaths, represented CVD incidence and mortality.

Energy misreporting

As in previous analyses on CCHS 2004 dietary data (11), we adjusted for misreporting bias by categorizing respondents into under-reporters, plausible reporters, or over-reporters. For the purpose, estimated energy requirements were calculated using the Institute of Medicine (IOM) factorial equations and compared with energy intake (29, 30, 31).

Dietary pattern assessment

Hybrid dietary patterns: PLS and RRR

PLS and RRR regression techniques were applied for derivation of dietary patterns associated with a high CVD risk at the national population level. Briefly, PLS is a flexible multivariable method that identifies dietary patterns that explain the maximum variation in both predictor (explanatory variables) and response variables. RRR, on the other hand, aims to explain as much variation as possible in intermediate disease-specific responses and can derive as many dietary patterns as the number of response variables. In the present study, nonlinear iterative PLS and RRR algorithms were used to derive dietary patterns that explained the maximum variation in 3 CVD-related response variables, namely energy density, percentage of energy from saturated fat (%ESF), and fiber density (FD). These variables were selected in accordance with the WHO and FAO report (32), as well as the relevant literature (14, 33, 34). Predictor variables were 39 standardized (z score) food groups. All dietary variables were scaled (standardized) and centered for the purpose of this research. Predictor variables were food groups which were compiled using Health Canada's Bureau of Nutritional Sciences (BNS) coding system (35) based on culinary usage and nutrient profile to reduce subjectivity (36) (Supplemental Table 1).

Only the first dietary pattern derived from PLS and RRR, respectively, was retained because it explained the most variation in the responses (30.8% and 34.0%, respectively) and accounted for the highest interpretability of findings. Several sensitivity analyses were conducted to confirm the nature of these patterns (Supplemental Methods). High scores on both PLS- and RRR-derived dietary patterns were positively correlated with energy density (r = 0.68; r = 0.70) and %ESF (r = 0.33; r = 0.40), and negatively correlated with FD (r = −0.62; r = −0.65) (Table 1). As a result, both of the first dietary patterns derived from PLS and RRR were labeled as “energy-dense, high-saturated-fat, and low-fiber-density” (ED, HSF, and LFD) dietary patterns. The variation explained by energy density, %ESF, and FD in the pattern was 45.3%, 10.0%, and 37.2% in PLS, and 46.3%, 15.5%, and 40.2% in RRR, respectively.

TABLE 1.

Weighted Pearson correlation coefficients between significant predictors (predictor loadings ≥|0.17|), responses, total scores of the PLS- and RRR-derived energy-dense, high-saturated-fat, and low-fiber-density dietary patterns, and their respective simplified dietary pattern scores in adult participants of the Canadian Community Health Survey-Nutrition 2015¹

	Response variables			Total dietary pattern scores
	Energy density	Fiber density	% Energy from saturated fat	PLS	RRR	Simplified PLS2	Simplified RRR2
Predictors
Positive associations
Fast food	0.26	−0.19	0.12	0.40	0.31	0.45	0.45
Carbonated drinks	0.23	−0.22	0.003	0.39	0.29	0.43	0.43
Baked goods	0.18	−0.13	0.09	0.19	0.22	0.07	0.28
Salty snacks	0.23	−0.08	0.014	0.23	0.19	0.30	0.30
Solid fats	0.11	−0.16	0.29	0.20	0.25	0.21	0.24
Inverse associations
Fruit	−0.43	0.38	−0.17	−0.57	−0.59	−0.44	−0.43
Dark green vegetables	−0.22	0.19	−0.04	−0.30	−0.26	−0.36	−0.37
Red/orange vegetables	−0.31	0.23	−0.09	−0.42	−0.37	−0.46	−0.42
Other vegetables	−0.32	0.27	−0.06	−0.42	−0.39	−0.40	−0.41
Whole grains	−0.14	0.31	−0.10	−0.34	−0.33	−0.35	−0.38
Legumes and soy	−0.09	0.23	−0.05	−0.26	−0.24	−0.34	−0.34
Mixed dishes	−0.25	0.09	−0.035	−0.19	−0.23	−0.06	−0.29
Yogurt	−0.18	0.05	0.036	−0.22	−0.15	−0.32	−0.10
Pasta and rice	−0.16	−0.027	−0.18	−0.23	−0.15	−0.38	−0.15
Response variables
Energy density	1.00	−0.48	0.36	0.68	0.70	0.60	0.64
Fiber density		1.00	−0.33	−0.62	−0.65	−0.51	−0.58
% Energy from saturated fat			1.00	0.33	0.40	0.24	0.24
Total dietary pattern scores
PLS	0.68	−0.62	0.33	1.00	0.95	0.90	0.91
RRR	0.70	−0.65	0.40		1.00	0.79	0.85
Simplified PLS	0.60	−0.51	0.24			1.00	0.90
Simplified RRR	0.64	−0.58	0.24				1.00

¹ n = 14,020. All P values < 0.0001 unless otherwise stated based on weighted Pearson correlation analyses. PLS, partial least squares; RRR, reduced rank regression.

²The simplified dietary pattern scores are the sums of standardized intake of food groups with factor loadings ≥|0.17| in PLS and RRR regressions. The simplified PLS score included carbonated drinks, fast food, salty snacks, and solid fats (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, yogurt, legumes and soy, and pasta and rice (all with weights of −1). The simplified RRR pattern score included fast food, solid fats, carbonated drinks, baked goods, and salty snacks (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, legumes and soy, and mixed dishes (all with weights of −1).

³ P = 0.6456.

⁴ P = 0.2525.

⁵ P = 0.0001.

⁶ P = 0.0011.

⁷ P = 0.0287.

We also explored the potential impact on predictor loadings of using survey weights. Our findings revealed a very close similarity between the scores from weighted and unweighted PLS and RRR regressions, and therefore, findings from unweighted analyses are reported in the present research (Supplemental Table 2).

Simplified dietary patterns

To address the arguably limited reproducibility of hybrid dietary pattern methods, we constructed simplified dietary pattern scores as proposed by Schulze et al. (9). Briefly, simplified dietary pattern scores are calculated as the sum of unweighted standardized z scores of food groups with significant factor loadings (≥|0.17|) on a dietary pattern and are based on the concept that leaving out less informative food components improves precision and generalizability. As a result, we constructed 2 simplified dietary pattern scores based on the PLS- and RRR-derived dietary patterns (Figure 1).

FIGURE 1.

Predictor loadings for the energy-dense, high-saturated-fat, and low-fiber-density dietary patterns derived from partial least squares (A) and reduced rank regression (B) analyses (centered and scaled) in adult participants of the Canadian Community Health Survey-Nutrition 2015 (n =14,020).

Statistical analyses

Statistical analyses were carried out using SAS version 9.4 and JMP Genomics version 13.2 (SAS Institute). The bootstrap balanced repeated replication (BRR) method (B = 500) was used for variance estimation to account for the complexity of the sampling framework of CCHS-Nutrition data (37, 38). Survey weights provided by Statistics Canada were used to ensure national representativeness of findings. Statistical significance was defined by a 2-tailed P value < 0.05 and absolute uncorrected P values were reported (Bonferroni-corrected P values = 0.05/n, where n denotes the number of tests).

Dietary pattern scores were considered in both continuous and categorical forms. Correlations between the dietary pattern scores and their subcomponents were evaluated with the use of weighted Pearson correlation analyses (Table 1). All nutritional analyses were performed with consideration of energy intake (density approach), as described previously (39–41). Because dietary pattern scores displayed small ranges, participants from CCHS-Nutrition 2015 were divided into quartiles for all descriptive analyses. In order to assess the performance of the dietary pattern scores, their ability to identify a dietary pattern closely associated with CVD risk (ED, HSF, and LFD) was examined through covariate-adjusted associations with lifestyle and sociodemographic characteristics. This was achieved by using weighted multivariable linear regression and weighted least-square means for continuous (dietary pattern scores, food groups, nutrients, age, and BMI) and categorical (sociodemographic and lifestyle characteristics) variables, respectively (Tables 2–4). The P-trend for continuous variables was defined as the P value associated with the weighted linear regression coefficients. Because obesity is a well-known mediator on the pathway between dietary patterns and CVD (42), the association between dietary pattern scores (quartiles) and obesity risk was evaluated as an additional test for performance using weighted multinomial logistic regression and a generalized logit model (Figure 2). As our previous work (11) has confirmed the importance of differentiation between obesity phenotypes, the association between dietary pattern scores and obesity among those with and without chronic diseases (i.e., high blood pressure, diabetes, heart disease, cancer, and/or osteoporosis) was also studied (Figure 3).

TABLE 2.

Weighted sociodemographic and lifestyle characteristics by quartile category (Q1 and Q4) of the PLS- and RRR-derived energy-dense, high-saturated-fat, and low-fiber-density dietary pattern scores and their respective simplified dietary pattern scores in adult participants of the Canadian Community Health Survey-Nutrition 2015¹

	PLS		RRR		Simplified PLS2		Simplified RRR2
	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4
Score range	−12.08 to −0.77	0.82–10	−9.98 to −0.66	0.68–9.07	−34.95 to −2.37	2.61–37.71	−37.62 to −2.31	2.42–43.53
Female, %	55.75 ± 2.04	34.35 ± 1.42**	55.94 ± 1.73	35.86 ± 1.57**	53.05 ± 2.36	37.96 ± 1.55*	53.85 ± 1.56	38.37 ± 1.49**
Age, y	49.68 ± 0.42	45.56 ± 0.91*	51.08 ± 0.59	45.77 ± 1.12**	48.62 ± 0.53	45.53 ± 0.69	49.63 ± 0.77	45.0 ± 0.82*
BMI, kg/m2	26.59 ± 0.19	28.14 ± 0.20**	26.66 ± 0.20	28.26 ± 0.21**	26.57 ± 0.19	28.49 ± 0.22**	26.68 ± 0.21	28.09 ± 0.23**
Obesity, %	20.86 ± 0.92	27.38 ± 1.35**	21.58 ± 1.06	27.24 ± 1.26**	20.26 ± 0.98	27.91 ± 1.29**	20.40 ± 0.92	26.34 ± 1.44**
Obese with ≥1 chronic disease, %	13.31 ± 1.25	21.06 ± 1.62*	15.17 ± 1.65	18.74 ± 1.53	12.85 ± 1.45	20.84 ± 1.63*	14.14 ± 1.61	19.43 ± 1.47
Having ≥1 chronic disease, %	50.70 ± 2.14	57.41 ± 2.89*	51.88 ± 2.42	52.59 ± 4.62	49.40 ± 2.62	57.58 ± 3.33	50.99 ± 2.05	54.76 ± 4.55*
Met physical activity guidelines,3 %	47.84 ± 4.08	42.67 ± 2.91*	48.28 ± 4.57	42.31 ± 2.12	47.33 ± 4.66	43.60 ± 2.99	47.62 ± 3.52	43.98 ± 2.86*
Current daily smokers, %	7.20 ± 1.26	24.05 ± 1.78**	7.89 ± 1.77	21.63 ± 2.03**	7.06 ± 0.98	22.53 ± 1.80**	7.35 ± 0.98	21.44 ± 2.0**
Alcohol consumption,4 yes	28.48 ± 3.20	42.33 ± 2.65*	29.34 ± 3.46	39.97 ± 4.04	32.39 ± 3.48	37.24 ± 1.78	31.12 ± 3.44	38.63 ± 2.29
Household education, %
Less than secondary school graduation	4.55 ± 0.41	10.16 ± 0.69**	4.72 ± 0.35	9.76 ± 0.75**	4.26 ± 0.37	10.36 ± 0.72**	4.39 ± 0.36	10.15 ± 0.74**
Postsecondary education	46.34 ± 1.78	26.69 ± 1.42**	45.76 ± 1.59	27.88 ± 1.49**	47.87 ± 1.74	26.12 ± 1.52**	47.29 ± 1.93	26.72 ± 1.56**
Respondent education, %
Less than secondary school graduation	8.28 ± 0.55	16.57 ± 1.57**	8.71 ± 0.76	16.07 ± 1.27**	7.89 ± 0.54	17.55 ± 1.30**	8.03 ± 0.71	16.91 ± 1.75**
Postsecondary education	34.52 ± 2.24	19.32 ± 2.40**	33.67 ± 2.88	20.18 ± 2.07**	35.42 ± 2.18	18.09 ± 1.88**	35.16 ± 3.10	18.88 ± 2.63**
Marital status, %
Married	60.33 ± 7.06	49.92 ± 2.50*	61.31 ± 5.91	49.69 ± 2.68*	61.08 ± 6.34	48.82 ± 2.90*	60.45 ± 6.65	48.54 ± 2.60*
Single	16.38 ± 4.69	23.01 ± 2.51*	15.83 ± 3.94	23.18 ± 2.66*	15.95 ± 4.23	23.80 ± 2.92*	16.32 ± 4.47	24.01 ± 2.71*
Immigrant, %	41.34 ± 5.71	12.94 ± 1.34**	40.14 ± 5.62	14.60 ± 1.84**	38.18 ± 6.41	13.14 ± 1.99**	37.29 ± 6.08	14.69 ± 1.76**
Excellent self-perceived health status, %	22.67 ± 1.04	15.05 ± 0.92**	22.65 ± 1.10	15.72 ± 0.95**	23.68 ± 1.42	14.52 ± 1.05**	23.46 ± 1.35	16.23 ± 1.02**
Highest income, %	16.87 ± 0.91	12.71 ± 0.86*	16.77 ± 0.92	12.64 ± 0.95*	17.14 ± 0.95	12.85 ± 0.91*	16.70 ± 0.93	12.60 ± 0.92*
Breakfast skippers, %	4.51 ± 0.68	16.19 ± 3.04**	5.10 ± 0.73	14.20 ± 2.77**	3.91 ± 0.96	16.39 ± 2.22**	4.44 ± 0.94	15.15 ± 2.09**

¹ n = 14,020. All values are weighted least-square means or percentages ± SEs with bootstrapped variances (500 times). Weighted multivariable linear regression and least-square means were used to explore covariate-adjusted associations of the scores with continuous and categorical variables, respectively. Estimates were adjusted for age and sex. Sex was adjusted for age only and age was adjusted for sex only. Only results for Quartiles 1 and 4 are shown; see Supplemental Table 3 for results across all quartiles. A smaller sample size was used owing to exclusion of “valid skip” and “not stated” responses in the analyses of BMI, kg/m² (n = 9404); household education, % (n = 13,990); respondent education, % (n = 13,934); marital status, % (n = 13,964); immigration status, % (n = 14,009); self-perceived health status, % (n = 14,001); and income, % (n = 14,013). For calculating P-trend, the dietary pattern score was used in its continuous form. P-trend represents the P value associated with the logistic regression coefficient for categorical variables and linear regression coefficient for continuous variables. PLS, partial least squares; RRR, reduced rank regression.

²The simplified dietary pattern scores are the sums of standardized intake of food groups with factor loadings ≥|0.17| in PLS and RRR regression, respectively. The simplified PLS score included carbonated drinks, fast foods, salty snacks, and solid fats (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, yogurt, legumes and soy, and pasta and rice (all with weights of −1). The simplified RRR pattern score included fast food, solid fats, carbonated drinks, baked goods, and salty snacks (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, legumes and soy, and mixed dishes (all with weights of −1).

³Corresponds to 150 min of moderate-/vigorous-intensity physical activity per week in accordance with the current physical activity guidelines for Canadian adults.

⁴Represents any alcohol consumption based on information from the first 24-h dietary recall.

*P-trend < 0.05; **P-trend < 0.0001.

TABLE 4.

Weighted mean daily intakes of important food groups (predictor loadings ≥|0.17|) across the quartile categories (Q1 and Q4) of the PLS- and RRR-derived energy-dense, high-saturated-fat, and low-fiber-density dietary pattern scores and their respective simplified dietary pattern scores in adult participants of the Canadian Community Health Survey-Nutrition 2015¹

	PLS		RRR		Simplified PLS2		Simplified RRR2
	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4
Score range	−12.08 to −0.77	0.82–10	−9.98 to −0.66	0.68–9.07	−34.95 to −2.37	2.61–37.71	−37.62 to −2.31	2.42–43.53
Predictors3
Positive associations
Fast food	56.99 ± 7.35	213.63 ± 8.44	69.86 ± 7.24	175.06 ± 8.86	46.53 ± 9.15	241.03 ± 11.31	47.02 ± 5.98	235.86 ± 9.26
Carbonated drinks	44.40 ± 9.12	337.11 ± 17.47	63.18 ± 8.82	280.41 ± 18.22	40.13 ± 8.99	379.19 ± 17.66	38.39 ± 8.06	351.82 ± 22.40
Baked goods	26.72 ± 3.76	53.46 ± 6.74	26.34 ± 3.61	54.73 ± 7.04	33.75 ± 4.81	39.94 ± 6.324	22.59 ± 3.83	65.18 ± 7.30
Salty snacks	5.47 ± 2.99	17.47 ± 2.20	5.97 ± 2.14	14.60 ± 1.69	4.31 ± 2.17	20.98 ± 2.57	4.74 ± 2.23	19.31 ± 2.22
Solid fats	10.85 ± 1.55	24.69 ± 1.94	11.12 ± 1.19	26.72 ± 2.42	11.09 ± 1.56	25.21 ± 1.68	10.74 ± 1.54	25.79 ± 1.91
Inverse associations
Fruit	291.09 ± 7.67	26.24 ± 8.66	305.80 ± 9.04	19.35 ± 7.15	251.98 ± 9.29	41.86 ± 4.85	259.10 ± 8.52	43.78 ± 5.15
Dark green vegetables	41.49 ± 2.99	6.55 ± 2.55	39.36 ± 3.14	6.84 ± 2.23	46.36 ± 3.55	5.59 ± 2.27	47.56 ± 3.87	4.06 ± 2.14
Red/orange vegetables	104.45 ± 5.46	6.84 ± 2.64	100.86 ± 5.57	7.63 ± 4.21	107.52 ± 5.41	8.09 ± 2.35	105.07 ± 5.54	9.02 ± 2.68
Other vegetables	133.19 ± 5.98	17.60 ± 3.13	128.15 ± 6.36	20.11 ± 3.54	130.01 ± 6.19	20.91 ± 3.66	132.70 ± 5.96	19.38 ± 3.25
Whole grains	60.64 ± 3.85	7.35 ± 2.33	61.37 ± 4.16	7.41 ± 2.21	61.61 ± 4.10	9.49 ± 2.10	65.21 ± 3.75	6.66 ± 2.40
Legumes and soy	28.17 ± 3.06	—5	29.07 ± 3.29	—5	31.43 ± 3.11	1.16 ± 1.21	32.41 ± 2.90	—5
Mixed dishes	90.50 ± 5.38	17.15 ± 6.14	100.49 ± 5.55	14.09 ± 4.84	61.95 ± 4.75	42.21 ± 7.026	107.50 ± 5.53	11.27 ± 5.44
Yogurt	42.37 ± 3.17	0.93 ± 2.52	37.72 ± 3.37	4.88 ± 2.44	54.62 ± 4.0	0.15 ± 3.30	31.08 ± 3.04	10.35 ± 4.80
Pasta and rice	118.82 ± 16.77	26.53 ± 4.92	107.87 ± 17.53	37.43 ± 6.13	135.60 ± 19.02	18.18 ± 5.59	105.68 ± 15.40	41.90 ± 6.81

¹ n = 14,020. All values (g/d) are weighted least-square means ± SEs with bootstrapped variances (500 times). Weighted multivariable linear regression were used to explore covariate-adjusted associations of the scores with continuous variables. Estimates were adjusted for age, sex, energy intake, and misreporting status. Only results for Quartiles 1 and 4 are shown; see Supplemental Table 5 for results across all quartiles. All P-trends < 0.0001 unless otherwise indicated. PLS, partial least squares; RRR, reduced rank regression.

²The simplified dietary pattern scores are the sums of standardized intake of food groups with factor loadings ≥|0.17| in PLS and RRR regression, respectively. The simplified PLS score included carbonated drinks, fast food, salty snacks, and solid fats (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, yogurt, legumes and soy, and pasta and rice (all with weights of −1). The simplified RRR pattern score included fast food, solid fats, carbonated drinks, baked goods, and salty snacks (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, legumes and soy, and mixed dishes (all with weights of −1).

³Food groups that contributed the most to the PLS- and RRR-derived dietary pattern scores (predictor loading ≥|0.17|).

⁴ P = 0.0011.

⁵Values are not reportable owing to large CVs.

⁶ P = 0.0107.

FIGURE 2.

Weighted multivariate-adjusted ORs (95% CIs) for obesity risk [BMI (in kg/m²) ≥30] across Q categories of the ED, HSF, and LFD dietary pattern scores from PLS (A), RRR (B), simplified PLS (C), and simplified RRR (D) in Canadian adults (≥18 y old) from the Canadian Community Health Survey-Nutrition 2015 (n = 14,020). Estimates were based on the multinomial logistic regression-generalized logit model. All P-trends < 0.0001. ED, energy dense; HSF, high saturated fat; LFD, low fiber density; PLS, partial least squares; Q, quartile; RRR, reduced rank regression.

FIGURE 3.

Weighted multivariate-adjusted ORs (95% CIs) for risk of obesity with and without ≥1 chronic disease across Q categories of the energy-dense, high-saturated-fat, and low-fiber-density dietary pattern scores from PLS (A), RRR (B), simplified PLS (C), and simplified RRR (D) in Canadian adults (≥18 y old) from the Canadian Community Health Survey-Nutrition 2015 (n = 14,020). All models were adjusted for age, sex, misreporting, energy intake, physical activity levels, and smoking status. Estimates were based on the multinomial logistic regression-generalized logit model. P-trends are shown in the following order: 1) obese with chronic disease (n = 1430); 2) obese without chronic disease (n = 813); 3) nonobese with chronic disease (n = 2683). PLS, partial least squares; Q, quartile; RRR, reduced rank regression.

We applied the National Cancer Institute (NCI) method (43) in order to thoroughly identify and address random measurement error (i.e. estimate the usual intake distribution) (Supplemental Methods). Associations between dietary pattern scores and CVD risk were assessed with the use of weighted Cox proportional hazards regression models (with person-days as the time-metric) in the CCHS-Nutrition 2004 survey linked to health administrative data (Table 5). All Cox models used sex interaction terms to test for effect modification. All dietary pattern scores were modeled as continuous variables for better interpretability and HRs and 95% CIs were reported at the midpoint of quintiles, with the 10th percentile serving as a reference. Assumptions of the Cox proportional hazards model were assessed using weighted Schoenfeld residual variables plotted against time-to-event variables and no departure was found from proportionality of hazards over time. This was further confirmed with nonsignificant interaction terms of person-time and dietary factors.

TABLE 5.

Weighted and multivariable-adjusted HRs and bootstrapped 95% CIs of cardiovascular disease events (incidence and mortality) according to midpoints of quintiles of usual intake of the PLS- and RRR-derived energy-dense, high-saturated-fat, and low-fiber-density dietary pattern scores, and their respective simplified dietary pattern scores, in adult participants (45–80 y old) of the Canadian Community Health Survey-Nutrition 2004 linked to the Canadian Vital Statistics Death Database and Discharge Abstract Database (2004–2017)¹

	Mid-value of quintiles of usual intake distribution for dietary pattern score
Dietary pattern scores	10th percentile	30th percentile	50th percentile	70th percentile	90th percentile
PLS
Base model2
Women	1 (Reference)	1.23 (1.07, 1.43)	1.40 (1.11, 1.76)	1.57 (1.15, 2.14)	1.89 (1.22, 2.92)
Men	1 (Reference)	1.11 (0.92, 1.33)	1.17 (0.87, 1.58)	1.24 (0.83, 1.85)	1.36 (0.77, 2.38)
Multivariable-adjusted model3
Women	1 (Reference)	1.12 (0.95, 1.31)	1.20 (0.92, 1.56)	1.28 (0.89, 1.83)	1.42 (0.85, 2.36)
Men	1 (Reference)	1.14 (0.91, 1.42)	1.23 (0.86, 1.76)	1.33 (0.82, 2.16)	1.50 (0.75, 3.00)
RRR
Base model2
Women	1 (Reference)	1.21 (1.04, 1.41)	1.37 (1.07, 1.75)	1.53 (1.10, 2.14)	1.83 (1.14, 2.93)
Men	1 (Reference)	1.07 (0.89, 1.28)	1.12 (0.84, 1.49)	1.16 (0.78, 1.72)	1.23 (0.71, 2.15)
Multivariable-adjusted model3
Women	1 (Reference)	1.11 (0.93, 1.32)	1.19 (0.89, 1.59)	1.27 (0.86, 1.88)	1.40 (0.80, 2.45)
Men	1 (Reference)	1.12 (0.91, 1.37)	1.20 (0.86, 1.68)	1.29 (0.81, 2.04)	1.43 (0.75, 2.73)
Simplified PLS
Base model2
Women	1 (Reference)	1.21 (1.04, 1.41)	1.34 (1.06, 1.70)	1.48 (1.07, 2.03)	1.70 (1.10, 2.62)
Men	1 (Reference)	1.13 (0.98, 1.31)	1.21 (0.97, 1.51)	1.29 (0.96, 1.73)	1.41 (0.94, 2.12)
Multivariable-adjusted model3
Women	1 (Reference)	1.11 (0.94, 1.32)	1.18 (0.91, 1.53)	1.25 (0.89, 1.77)	1.36 (0.85, 2.20)
Men	1 (Reference)	1.12 (0.94, 1.35)	1.20 (0.90, 1.59)	1.27 (0.87, 1.86)	1.40 (0.83, 2.35)
Simplified RRR
Base model2
Women	1 (Reference)	1.18 (1.02, 1.37)	1.28 (1.03, 1.60)	1.41 (1.05, 1.91)	1.63 (1.06, 2.49)
Men	1 (Reference)	1.07 (0.92, 1.24)	1.11 (0.89, 1.39)	1.16 (0.85, 1.57)	1.23 (0.79, 1.90)
Multivariable-adjusted model3
Women	1 (Reference)	1.08 (0.93, 1.25)	1.13 (0.89, 1.43)	1.18 (0.85, 1.64)	1.27 (0.79, 2.05)
Men	1 (Reference)	1.06 (0.90, 1.26)	1.11 (0.84, 1.45)	1.15 (0.78, 1.69)	1.22 (0.71, 2.12)

¹ n = 6766. Weighted HRs (from a Cox proportional hazards model) were calculated using regression calibration and the 95% CIs were calculated by bootstrapping the usual intake estimating models B = 500 times at each step. All Cox models used sex interaction to test for effect modification by sex. PLS, partial least squares; RRR, reduced rank regression.

²Base model: adjusted for the day of the week on which the 24-h dietary recall was collected [weekday/weekend (Friday–Sunday)], sequence of dietary recall (first or second), and baseline age (continuous).

³Multivariable-adjusted model: adjusted for base model covariates in addition to education (less than secondary school graduation; secondary school graduation; some postsecondary; postsecondary graduation), smoking (daily/occasional smoker with 20 ≤ n < 90 cigarettes/d; daily/occasional smoker with <20 cigarettes/d; former daily/occasional smoker and those who smoked a total of ≥100 cigarettes in lifetime; never smoked), misreporting (under-reporter; plausible reporter; and over-reporter), physical activity (daily energy expenditure ≥3; 1.5 ≤ n < 3 daily energy expenditure; and 0 ≤ n < 1.5 daily energy expenditure), marital status (married/common-law partner; widowed/separated/divorced/single; never married), immigrant (yes/no), and alcohol consumption (none; less than once a month / once a month / 2–3 times/mo; once a week / 2–3 times/wk; 4–6 times/wk / every day or being binge drinker: drank >2 times/wk with the frequency of ≥5 drinks being ≥1 times/wk).

Results

Part A: characterizing dietary patterns

Figure 1 presents predictor loadings for the ED, HSF, and LFD dietary patterns derived from PLS and RRR. The PLS-derived ED, HSF, and LFD dietary pattern was characterized by low intakes of fruit (−0.43), other vegetables (−0.30), red/orange vegetables (−0.30), whole grains (−0.24), dark green vegetables (−0.23), yogurt (−0.19), legumes and soy (−0.18), and pasta and rice (−0.17), and high intakes of carbonated drinks (0.29), fast food (0.29), salty snacks (0.19), and solid fats (0.18). The strongest predictor loadings on the RRR-derived ED, HSF, and LFD dietary pattern were for fruit (−0.48), other vegetables (−0.30), red/orange vegetables (−0.29), whole grains (−0.24), dark green vegetables (−0.21), legumes and soy (−0.19), and mixed dishes (−0.18) on one side, and fast food (0.24), solid fats (0.23), carbonated drinks (0.23), baked goods (0.18), and salty snacks (0.17) on the other. In both PLS and RRR, the highest positive correlation between predictors (food groups) and the ED, HSF, and LFD dietary pattern scores was for fast food (r = 0.40 for PLS; r = 0.31 for RRR) (P < 0.0001) and the highest negative correlation was for fruit (r = −0.57 for PLS; r = −0.59 for RRR) (P < 0.0001) (Table 1). Correlations between all 4 dietary pattern scores were strong and positive (r = 0.79–0.95) (P < 0.0001).

Part B: association of dietary pattern scores with lifestyle and nutritional characteristics

Similar sociodemographic and lifestyle characteristics were observed across the quartiles of the PLS- and RRR-derived ED, HSF, and LFD dietary pattern scores and the simplified dietary pattern scores (Table 2, Supplemental Table 3). Moving from the lowest (healthiest) to the highest (least healthy) quartiles of the 4 scores, respondents were less likely to be women, immigrants, or have higher education or excellent health status, and more likely to be obese, frequent smokers, and breakfast skippers (P-trends < 0.05). Of note, the magnitude of significance was slightly attenuated in the simplified dietary pattern scores as compared with the overall PLS- and RRR-derived dietary pattern scores. The highest quartiles of all 4 dietary pattern scores were also mainly characterized by lower beneficial macro- and micronutrient intakes and were positively associated with 402-kcal/d higher energy intake (P-trends < 0.05) (Table 3, Supplemental Table 4). In line with the design of the PLS and RRR algorithms, participants in the fourth quartiles of the dietary pattern scores had significantly higher percentages of energy from saturated fat and energy density, whereas their FD was significantly lower (P-trends < 0.05). Similarly, mean intakes of food groups with positive factor loadings were observed to be significantly higher in the highest quartiles of all dietary pattern scores, whereas the opposite trend was observed for mean intakes of food groups with negative factor loadings (P-trends < 0.0001) (Table 4, Supplemental Table 5).

TABLE 3.

Weighted mean daily intakes of macro- and micronutrients as percentages of energy or per 1000 kcal (nutrient density) by quartile category (Q1 and Q4) of the PLS- and RRR-derived energy-dense, high-saturated-fat, and low-fiber-density dietary pattern scores and their respective simplified dietary pattern scores in adult participants of the Canadian Community Health Survey-Nutrition 2015¹

		PLS		RRR		Simplified PLS2		Simplified RRR2
	Model	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4	Quartile 1	Quartile 4
Score range	—	−12.08 to −0.77	0.82–10	−9.98 to −0.66	0.68–9.07	−34.95 to −2.37	2.61–37.71	−37.62 to −2.31	2.42–43.53
Energy intake, kcal/d	a	1814.92 ± 24.03	2276.25 ± 35.11**	1760.36 ± 19.91	2319.14 ± 25.59**	1917.56 ± 23.91	2107.47 ± 39.46*	1825.04 ± 25.45	2222.51 ± 35.44**
	b	2229.55 ± 17.40	2425.79 ± 18.53**	2204.01 ± 17.72	2439.13 ± 19.81**	2254.07 ± 20.82	2387.30 ± 17.75**	2230.93 ± 21.38	2407.75 ± 17.52**
Carbohydrate, % of energy	a	50.92 ± 0.81	45.08 ± 0.55**	52.34 ± 0.76	43.08 ± 0.61**	50.25 ± 0.75	46.02 ± 0.52**	49.99 ± 0.70	46.33 ± 0.85**
	b	50.25 ± 0.78	44.97 ± 0.52**	51.86 ± 0.73	43.11 ± 0.60**	49.70 ± 0.74	45.58 ± 0.47**	49.27 ± 0.67	46.10 ± 0.80**
Fiber density, g/1000 kcal	a	13.89 ± 0.19	6.19 ± 0.11**	14.25 ± 0.16	6.10 ± 0.10**	12.99 ± 0.22	6.77 ± 0.17**	13.67 ± 0.21	6.65 ± 0.12**
	b	13.71 ± 0.18	6.23 ± 0.09**	14.15 ± 0.17	6.19 ± 0.09**	12.78 ± 0.22	6.62 ± 0.14**	13.49 ± 0.20	6.65 ± 0.10**
Added sugar,3 % of energy	a	6.12 ± 0.17	12.90 ± 0.43**	6.46 ± 0.24	11.96 ± 0.45**	6.73 ± 0.18	12.81 ± 0.35**	6.62 ± 0.26	12.77 ± 0.55**
	b	5.90 ± 0.19	12.75 ± 0.42**	6.26 ± 0.24	11.82 ± 0.44**	6.65 ± 0.20	12.73 ± 0.34**	6.41 ± 0.29	12.62 ± 0.53**
Total fat, % of energy	a	29.28 ± 0.50	35.48 ± 0.44**	28.35 ± 0.46	37.10 ± 0.77**	29.83 ± 0.50	35.02 ± 0.36**	29.81 ± 0.37	35.28 ± 0.65**
	b	29.91 ± 0.56	35.54 ± 0.49**	28.83 ± 0.50	37.03 ± 0.84**	30.34 ± 0.61	35.42 ± 0.40**	30.41 ± 0.43	35.43 ± 0.70**
Saturated fat, % of energy	a	8.59 ± 0.15	12.22 ± 0.17**	8.41 ± 0.15	12.77 ± 0.29**	9.19 ± 0.20	11.76 ± 0.15**	9.06 ± 0.13	11.87 ± 0.27**
	b	8.78 ± 0.16	12.20 ± 0.17**	8.54 ± 0.17	12.70 ± 0.29**	9.36 ± 0.23	11.89 ± 0.15**	9.26 ± 0.15	11.89 ± 0.26**
MUFA, % of energy	a	11.07 ± 0.28	13.20 ± 0.19**	10.55 ± 0.26	13.85 ± 0.27**	10.99 ± 0.19	13.10 ± 0.17**	11.06 ± 0.19	13.21 ± 0.29**
	b	11.38 ± 0.33	13.28 ± 0.23**	10.80 ± 0.30	13.87 ± 0.32**	11.25 ± 0.25	13.32 ± 0.20**	11.35 ± 0.24	13.32 ± 0.33**
PUFA, % of energy	a	6.81 ± 0.12	7.02 ± 0.10	6.66 ± 0.12	7.33 ± 0.21*	6.79 ± 0.12	7.13 ± 0.10*	6.80 ± 0.10	7.15 ± 0.10*
	b	6.88 ± 0.14	7.0 ± 0.12	6.72 ± 0.13	7.30 ± 0.24*	6.82 ± 0.15	7.16 ± 0.11*	6.86 ± 0.11	7.15 ± 0.12*
Linoleic acid, % of energy	a	5.69 ± 0.15	6.06 ± 0.09*	5.55 ± 0.13	6.35 ± 0.21**	5.65 ± 0.13	6.19 ± 0.10**	5.67 ± 0.11	6.19 ± 0.09**
	b	5.76 ± 0.17	6.05 ± 0.11*	5.61 ± 0.14	6.32 ± 0.23**	5.70 ± 0.15	6.21 ± 0.11**	5.73 ± 0.13	6.18 ± 0.11**
Linolenic acid, % of energy	a	0.79 ± 0.02	0.74 ± 0.01*	0.79 ± 0.02	0.74 ± 0.01	0.80 ± 0.02	0.75 ± 0.01*	0.81 ± 0.02	0.76 ± 0.01*
	b	0.79 ± 0.02	0.74 ± 0.01*	0.79 ± 0.02	0.74 ± 0.01	0.80 ± 0.02	0.75 ± 0.01*	0.80 ± 0.02	0.75 ± 0.01*
Protein, % of energy	a	17.75 ± 0.18	15.55 ± 0.27**	17.14 ± 0.18	16.31 ± 0.20*	17.59 ± 0.20	15.54 ± 0.16**	17.65 ± 0.19	15.53 ± 0.25**
	b	17.52 ± 0.19	15.60 ± 0.35**	16.82 ± 0.19	16.37 ± 0.21	17.44 ± 0.19	15.43 ± 0.19**	17.44 ± 0.19	15.54 ± 0.31**
Alcohol, % of energy	a	2.05 ± 0.34	3.88 ± 0.28**	2.17 ± 0.34	3.52 ± 0.30*	2.33 ± 0.37	3.42 ± 0.25**	2.55 ± 0.43	2.86 ± 0.19
	b	2.32 ± 0.23	3.89 ± 0.43**	2.48 ± 0.25	3.50 ± 0.47	2.52 ± 0.22	3.57 ± 0.26**	2.87 ± 0.30	2.94 ± 0.30
Cholesterol density, mg/1000 kcal	a	135.04 ± 3.91	149.35 ± 4.19*	128.55 ± 4.67	161.92 ± 3.76**	140.69 ± 4.66	141.10 ± 3.95	143.24 ± 5.84	142.35 ± 4.86
	b	134.0 ± 4.94	150.74 ± 5.47*	125.68 ± 6.06	163.12 ± 4.70**	141.02 ± 5.73	141.58 ± 5.36	142.85 ± 6.95	143.57 ± 6.21
Calcium density, mg/1000 kcal	a	452.73 ± 11.53	412.76 ± 6.48*	446.58 ± 10.60	421.67 ± 7.13*	462.42 ± 16.81	405.93 ± 6.65*	456.20 ± 11.09	406.85 ± 6.91**
	b	445.16 ± 10.79	412.80 ± 6.84*	437.73 ± 9.79	422.61 ± 7.47	458.02 ± 16.22	402.67 ± 7.08*	449.92 ± 10.32	406.31 ± 7.31**
Vitamin A density in retinol activity	a	501.15 ± 23.32	271.35 ± 8.13**	499.01 ± 21.59	283.62 ± 6.53**	480.58 ± 24.62	266.34 ± 9.34**	507.84 ± 23.53	261.90 ± 9.31**
equivalent, µg/1000 kcal	b	493.75 ± 26.85	275.56 ± 15.92**	491.35 ± 25.08	289.61 ± 12.58**	475.71 ± 30.01	263.18 ± 15.48**	502.72 ± 27.85	265.44 ± 16.89**
Vitamin D density, µg/1000 kcal	a	2.68 ± 0.20	2.27 ± 0.07*	2.61 ± 0.10	2.44 ± 0.09	2.69 ± 0.08	2.16 ± 0.06*	2.67 ± 0.14	2.24 ± 0.06*
	b	2.66 ± 0.17	2.30 ± 0.07	2.57 ± 0.09	2.47 ± 0.08	2.68 ± 0.08	2.15 ± 0.07*	2.66 ± 0.12	2.27 ± 0.08*
Vitamin C density, mg/1000 kcal	a	88.30 ± 3.32	34.55 ± 1.97**	90.22 ± 3.38	33.30 ± 2.13**	83.17 ± 3.55	35.94 ± 1.42**	86.34 ± 3.64	35.54 ± 1.71**
	b	87.10 ± 3.73	35.73 ± 2.69**	89.53 ± 3.70	35.06 ± 2.85**	82.47 ± 4.07	35.58 ± 1.52**	85.44 ± 4.09	36.48 ± 2.33**
Sodium density, g/1000 kcal	a	1456.23 ± 22.29	1497.56 ± 15.60	1477.59 ± 25.50	1470.44 ± 16.27	1421.81 ± 23.22	1533.15 ± 28.62*	1502.68 ± 22.35	1481.89 ± 17.87
	b	1427.87 ± 22.79	1496.65 ± 16.33*	1447.91 ± 25.11	1473.58 ± 17.30	1403.98 ± 23.72	1519.35 ± 28.32**	1478.07 ± 23.15	1478.46 ± 18.46
Thiamin density, mg/1000 kcal	a	0.91 ± 0.02	0.80 ± 0.01**	0.92 ± 0.02	0.77 ± 0.01**	0.90 ± 0.02	0.82 ± 0.01**	0.91 ± 0.02	0.80 ± 0.01**
	b	0.89 ± 0.02	0.79 ± 0.01**	0.91 ± 0.02	0.78 ± 0.01**	0.89 ± 0.02	0.81 ± 0.01**	0.89 ± 0.02	0.79 ± 0.01**
Riboflavin density, mg/1000 kcal	a	1.08 ± 0.02	1.04 ± 0.02	1.07 ± 0.02	1.03 ± 0.01	1.09 ± 0.03	1.0 ± 0.02	1.09 ± 0.02	0.99 ± 0.01*
	b	1.05 ± 0.03	1.04 ± 0.02	1.04 ± 0.03	1.04 ± 0.01	1.08 ± 0.03	0.99 ± 0.02*	1.07 ± 0.02	0.99 ± 0.02*
Niacin density in niacin equivalents,	a	22.37 ± 0.30	19.18 ± 0.21**	21.90 ± 0.30	19.83 ± 0.34**	21.81 ± 0.42	19.31 ± 0.33**	22.05 ± 0.31	19.23 ± 0.24**
mg/1000 kcal	b	22.12 ± 0.30	19.32 ± 0.23**	21.55 ± 0.30	19.99 ± 0.27**	21.65 ± 0.40	19.21 ± 0.29**	21.82 ± 0.32	19.30 ± 0.23**
Vitamin B-6 density, mg/1000 kcal	a	1.14 ± 0.02	0.73 ± 0.02**	1.15 ± 0.02	0.73 ± 0.02**	1.07 ± 0.03	0.75 ± 0.02**	1.11 ± 0.02	0.73 ± 0.01**
	b	1.13 ± 0.02	0.74 ± 0.02**	1.14 ± 0.02	0.75 ± 0.02**	1.06 ± 0.03	0.74 ± 0.02**	1.10 ± 0.02	0.74 ± 0.02**
Vitamin B-12 density, µg/1000 kcal	a	2.19 ± 0.08	2.14 ± 0.09	2.08 ± 0.08	2.21 ± 0.05*	2.13 ± 0.07	2.10 ± 0.07	2.20 ± 0.10	2.09 ± 0.10
	b	2.13 ± 0.10	2.12 ± 0.08	2.01 ± 0.09	2.19 ± 0.07*	2.09 ± 0.09	2.06 ± 0.08	2.16 ± 0.09	2.07 ± 0.09
Naturally occurring folate density,4	a	165.57 ± 3.04	86.05 ± 1.49**	165.73 ± 2.71	87.80 ± 1.93**	162.06 ± 2.77	87.22 ± 1.45**	166.41 ± 3.07	85.19 ± 1.67**
µg/1000 kcal	b	163.12 ± 2.86	87.04 ± 2.16**	163.63 ± 2.69	89.45 ± 2.64**	160.15 ± 2.74	85.89 ± 1.65**	164.47 ± 2.88	85.92 ± 2.38**
Folacin density from food sources,5	a	216.93 ± 2.64	149.24 ± 1.86**	217.21 ± 2.67	147.19 ± 1.82**	215.31 ± 2.73	152.42 ± 2.58**	216.36 ± 2.67	153.32 ± 1.91**
µg/1000 kcal	b	214.31 ± 2.76	150.39 ± 2.04**	215.10 ± 2.77	149.07 ± 2.08**	213.40 ± 2.83	151.12 ± 2.44**	213.88 ± 2.75	153.94 ± 2.18**
Phosphorus density, mg/1000 kcal	a	749.88 ± 6.43	636.89 ± 8.08**	740.33 ± 8.81	660.61 ± 9.44**	745.40 ± 7.76	633.58 ± 5.54**	746.03 ± 6.84	638.31 ± 8.43**
	b	738.17 ± 6.79	636.24 ± 9.89**	726.03 ± 8.85	660.91 ± 11.40**	736.50 ± 8.75	626.64 ± 6.01**	735.02 ± 7.42	636.19 ± 9.96**
Magnesium density, mg/1000 kcal	a	211.38 ± 2.11	132.75 ± 1.36**	210.63 ± 2.23	135.66 ± 1.78**	202.35 ± 1.84	135.01 ± 1.81**	208.56 ± 2.0	134.61 ± 1.40**
	b	207.74 ± 2.32	133.55 ± 1.68**	207.01 ± 2.48	137.12 ± 1.63**	199.31 ± 1.93	132.74 ± 1.37**	205.18 ± 2.24	134.79 ± 1.94**
Iron density, mg/1000 kcal	a	7.49 ± 0.10	5.98 ± 0.08**	7.52 ± 0.08	5.86 ± 0.07**	7.33 ± 0.07	6.24 ± 0.07**	7.53 ± 0.09	6.17 ± 0.08**
	b	7.39 ± 0.10	5.98 ± 0.08**	7.43 ± 0.09	5.88 ± 0.07**	7.25 ± 0.08	6.17 ± 0.08**	7.43 ± 0.09	6.16 ± 0.10**
Zinc density, mg/1000 kcal	a	6.10 ± 0.17	5.19 ± 0.14**	5.90 ± 0.12	5.32 ± 0.10**	6.0 ± 0.10	5.26 ± 0.11**	6.18 ± 0.14	5.25 ± 0.21**
	b	6.01 ± 0.18	5.17 ± 0.16**	5.79 ± 0.14	5.31 ± 0.12**	5.93 ± 0.12	5.20 ± 0.13**	6.09 ± 0.15	5.22 ± 0.23**
Potassium density, mg/1000 kcal	a	1873.03 ± 17.47	1216.80 ± 12.79**	1894.59 ± 22.11	1198.02 ± 12.48**	1791.0 ± 17.38	1243.41 ± 13.05**	1845.63 ± 15.73	1211.93 ± 13.55**
	b	1843.97 ± 29.05	1230.50 ± 23.46**	1869.58 ± 35.44	1219.74 ± 17.39**	1771.31 ± 27.03	1230.05 ± 19.01**	1821.49 ± 24.71	1220.5 ± 25.59**
Caffeine density, mg/1000 kcal	a	84.72 ± 3.38	114.65 ± 7.85*	89.05 ± 4.97	98.96 ± 4.97	85.32 ± 3.56	104.86 ± 7.73*	87.41 ± 3.05	99.65 ± 4.12
	b	73.32 ± 3.58	115.32 ± 7.35**	77.28 ± 4.59	101.36 ± 4.62*	79.68 ± 3.37	100.72 ± 7.09*	78.10 ± 3.12	99.04 ± 3.84*
Moisture density,6 g/1000 kcal	a	1778.27 ± 34.27	1220.71 ± 30.70**	1767.18 ± 30.30	1171.66 ± 30.26**	1608.73 ± 26.58	1301.25 ± 33.60**	1701.07 ± 24.92	1242.23 ± 20.23**
	b	1666.57 ± 34.93	1248.92 ± 22.82**	1648.59 ± 29.80	1219.56 ± 22.60**	1547.12 ± 26.03	1258.05 ± 24.73**	1600.59 ± 22.33	1249.63 ± 24.02**
Glycemic index density,7 per 1000 kcal	a	29.98 ± 0.94	28.33 ± 0.92	31.31 ± 0.86	26.60 ± 0.85*	27.64 ± 0.89	31.33 ± 0.85**	29.66 ± 0.87	29.20 ± 1.19
	b	26.58 ± 0.80	29.39 ± 0.60**	27.50 ± 0.73	28.25 ± 0.52*	25.95 ± 0.75	30.19 ± 0.43**	26.57 ± 0.78	29.55 ± 0.80**
Energy density,8 per 1000 kcal	a	0.75 ± 0.01	1.21 ± 0.05**	0.75 ± 0.01	1.19 ± 0.02**	0.73 ± 0.01	1.29 ± 0.04**	0.76 ± 0.01	1.19 ± 0.03**
	b	0.63 ± 0.01	1.23 ± 0.04**	0.61 ± 0.02	1.22 ± 0.02**	0.67 ± 0.02	1.25 ± 0.03**	0.66 ± 0.01	1.19 ± 0.03**

¹ n= 14,020. All values are weighted least-square means or percentages ± SEs with bootstrapped variances (500 times). Weighted multivariable linear regression and least-square means were used to explore covariate-adjusted associations of the scores with continuous and categorical variables, respectively. Estimates were adjusted for age and sex (model a), with the addition of misreporting status (model b). Only results for Quartiles 1 and 4 are shown; see Supplemental Table 4 for results across all quartiles. For calculating P-trend, the dietary pattern score was used in its continuous form. P-trend represents the P value associated with the linear regression coefficient for continuous variables. PLS, partial least squares; RRR, reduced rank regression.

²The simplified dietary pattern scores are the sums of standardized intake of food groups with factor loadings ≥|0.17| in PLS and RRR regression, respectively. The simplified PLS score included carbonated drinks, fast food, salty snacks, and solid fats (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, yogurt, legumes and soy, and pasta and rice (all with weights of −1). The simplified RRR pattern score included fast food, solid fats, carbonated drinks, baked goods, and salty snacks (all with weights of 1), and fruit, other vegetables, red/orange vegetables, whole grains, dark green vegetables, legumes and soy, and mixed dishes (all with weights of −1).

³Because added sugars are not included in the Canadian Nutrient File (CNF), the method proposed by Brisbois et al. (44) was used to derive estimates of added sugars.

⁴Naturally occurring folate included various forms of folate that are naturally present in food.

⁵Sum of quantities of naturally occurring folate in addition to folic acid without consideration of their differing bioavailabilities.

⁶Based on water content in food sources.

⁷Estimated by assigning the mean values reported in the International Glycemic Index Table (45, 46) to each of the Bureau of Nutritional Sciences (BNS) food categories (35), as described previously (47, 48).

⁸The dietary energy density was calculated by dividing the total energy from foods (kcal) by the total weight of foods (g) (with the exclusion of beverages) (49, 50).

*P-trend < 0.05; **P-trend < 0.0001.

Association of dietary pattern scores with obesity

The multivariate-adjusted odds of being obese were largely and significantly increased in the fourth (unhealthiest) compared with the first (healthiest) quartiles of the ED, HSF, and LFD dietary pattern scores derived from PLS (OR: 2.09; 95% CI: 1.62, 2.70) and RRR (OR: 1.76; 95% CI: 1.44, 2.17) (P-trends < 0.0001) (Figure 2A, B). The risk of obesity mutually adjusted for all potential confounders (Model 4) was similar across the quartile categories of the PLS-derived and RRR-derived simplified dietary pattern scores (P-trends < 0.0001) (Figure 2C, D). Jointly stratifying respondents by chronic disease status in addition to obesity risk (Figure 3) revealed increased odds of being unhealthy obese in the fourth quartiles compared with the first quartiles of the PLS-based dietary pattern scores (PLS: OR: 2.15; 95% CI: 1.55, 2.99; simplified PLS: OR: 2.22; 95% CI: 1.42, 3.46) (P-trends < 0.05), whereas associations in the RRR-based dietary pattern scores were attenuated and not statistically significant. The probability of being obese and without a chronic disease (healthy obese) was significantly increased in the highest quartiles of all 4 dietary pattern scores (P-trends < 0.05), whereas no significant associations were observed for nonobese with chronic disease.

Association between dietary pattern scores and CVD events

The same ED, HSF, and LFD dietary patterns (as in CCHS 2015) were derived using nationally representative CCHS 2004 data, despite the intervening 11 years (Supplemental Table 6). Further comparison of pattern scores of participants in the 2 surveys confirmed the similarities (Supplemental Table 7). Over a total follow-up of 80,471 person-years, 748 CVD events occurred. In the base model, among females, there was a consistent significant association of higher scores on PLS- and RRR-derived dietary patterns and their respective simplified dietary pattern scores with CVD events (HR: 1.89; 95% CI: 1.22, 2.92 for comparison of the 90th with the 10th percentile of usual intake of the PLS-derived pattern) (Table 5). The associations were weaker after adding potential confounders (multivariable-adjusted models) and not statistically significant. Effect modification by sex was significant in both base and multivariable-adjusted models across all dietary pattern scores (P-interaction < 0.0001 in all).

Discussion

In this nationally representative study in Canadian adults, we observed almost identical results using PLS and RRR for derivation of dietary patterns associated with CVD risk. The 2 simplified dietary pattern scores showed associations slightly weaker than those of the PLS and RRR scores, but of very similar magnitude and direction (for both obesity and CVD events). All 4 dietary pattern scores examined were positively, although nonsignificantly, associated with CVD incidence and mortality in CCHS 2004 (n = 6766). However, the same dietary patterns were significantly associated with a 75%–109% increase in obesity risk in CCHS 2015 (n = 14,020).

Only a small number of previous studies have assessed the ability of PLS and RRR to derive dietary patterns with potential implications for cardiovascular health (13, 17, 18, 51, 52). Generally, a consensus that PLS-derived dietary patterns, as compared with RRR-derived patterns, account for a similar amount of variation in responses but explain a much higher percentage of variation in predictors has led to the understanding that PLS offers more flexibility than RRR. On the other side, evidence on which of the 2 methods could be superior in explaining associations between dietary patterns and CVD outcomes has been largely inconsistent, with some studies favoring PLS (13, 18), whereas others favored RRR (17, 52). In the present research, we revealed that PLS and RRR performed highly comparably in identifying a high-CVD-risk dietary pattern at the population level, signifying that the 2 hybrid methods could potentially have an interchangeable application in large-scale analyses.

In the two major ED, HSF, and LFD dietary patterns derived, fast food, carbonated drinks, salty snacks, and solid fats were identified as the most important contributors to CVD risk (loading criterion), whereas fruit, all vegetables, whole grains, and legumes and soy were shown to exert the most prominent protective effects in Canadian adults. In particular, the predictor loading for fruit was double those for legumes and soy, yogurt, mixed dishes, and pasta and rice, which were also significant. This indicates that a 1-SD change in fruit intake, when the consumption of all other foods is held constant, has double the effect on total dietary pattern scores as compared with a similar change in other relevant food groups. These findings emphasize the need for an increased consumption of nutrient-dense plant-based foods at the expense of highly processed ready meals high in saturated fat, sodium, and sugar, and are in line with previous research on dietary patterns in Canadians using CCHS-Nutrition (11,31, 41, 53–55), as well as reviews and general practice guidelines for CVD prevention (56–59).

It is of note that despite the positive and consistent association we observed between the highest levels of compliance to an ED, HSF, and LFD dietary pattern and obesity risk, the associations were not statistically significant in relation to CVD risk. This may be attributed to the small number of CVD events observed in this study and the more distal nature of CVD risk on the diet–disease pathway, as compared with obesity, which may require examination/inclusion of disease mediators (e.g., measured blood pressure, lipid profiles) as response variables for improving discrimination. Nevertheless, the association between the ED, HSF, and LFD dietary pattern and CVD risk was in the expected direction. In comparison, previous research on the use of PLS or RRR for deriving dietary patterns with a potential role in CVD outcomes has predominantly involved RRR (but not PLS) and remained inconclusive, with some studies reporting a significant association (12, 14–16) and others mixed results (60, 61) or no such evidence (33, 51, 62).

Analyses comparing PLS with RRR are sparse and highly heterogeneous, involving specific study populations, different response variables (nutrients, clinical biomarkers, or both), and varying CVD endpoints (13, 17, 18, 51, 52). Thus far, 1 large-scale study, the UK Biobank prospective cohort study (14), has investigated CVD events (total and fatal) among middle-aged adults. Using RRR and response variables similar to ours (ED, %ESF, FD, and free sugars), a pattern almost identical to our ED, HSF, and LFD dietary pattern was derived and found to be positively and significantly associated with both fatal and total CVD. Other cohort studies (12, 15, 16) have also reported positive significant associations between CVD risk and unhealthy dietary patterns (i.e., Western patterns) using RRR and clinical biomarkers as response variables (instead of dietary components). To the best of our knowledge, the exact combination of response variables of the present research has previously been used only in the Swedish Obese Subjects study by Johns et al. (33), which applied RRR regression for the investigation of CVD incidence in adults with severe obesity. Similarly to our study, an ED, HSF, and LFD dietary pattern was derived with nonsignificant association in relation to CVD incidence. However, differences in sample size and participants’ characteristics between the 2 analyses make comparisons of the results somewhat limited.

Construction of simplified dietary pattern scores based on the PLS- and RRR-derived dietary patterns was a unique aspect of this study and applied in the context of CVD risk analysis for the first time in a general population. As hypothesized, we observed the PLS- and RRR-based simplified dietary pattern scores to be adequate approximations of the original PLS- and RRR-derived dietary patterns and less population-dependent because they included only the most informative food groups on the pathway to CVD development. These findings are in line with previous studies that utilized a simplified dietary pattern approach in relation to different health outcomes in individual PLS (11, 34) and RRR (10) regression analyses.

This is the first and the largest nationally representative study comparing the performance of PLS and RRR for deriving dietary patterns with a potential role in CVD incidence and mortality. The main methodological challenges we addressed, which are often ignored, included defining regression calibration methods for handling random measurement error (43) and controlling for systematic error through adjustment for misreporting status (30). A notable strength in our approach was the construction of simplified dietary pattern scores from both hybrid methods, which enables generalization of results while still preserving the data-driven advantages of hybrid dietary patterns (9). Using measured anthropometry, objective health measures (CVD events retrieved from the universal health care system data), inclusion of a complete a priori list of confounding variables in regression analyses, and sensitivity analyses were other strengths of this research.

Our findings should be considered in light of some limitations. To start with, the Discharge Abstract Database (hospitalization record) excluded Quebec (the second most populated province in Canada), leading to potentially attenuated statistical power. In addition, owing to computational limitations and complexity, regression calibration analyses were conducted unweighted (after confirming the similarity of weighted and unweighted dietary patterns). Limited statistical power, on the other hand, did not allow for any dietary pattern subgroup analyses, nor separate evaluation of CVD incidence and mortality. We also acknowledge the risk of making type I error due to the multiple hypothesis tests performed. Finally, the absence of repeated dietary intake data in CCHS-Nutrition cycles, lack of access to measured clinical biomarkers (e.g., blood pressure, lipid profiles), residual confounding due to omission of covariates such as medication use, family history of chronic diseases (CVD, BMI, and diabetes), as well as differences in data collection between the 2004 and 2015 cycles limited causal inference.

In conclusion, using a nationally representative nutrition survey linked to health administrative databases, our study confirmed the somewhat equal applicability of PLS and RRR for deriving dietary patterns with potential implications for reducing chronic disease risk at the population level. The simplified dietary pattern scores closely approximated the originally derived dietary patterns, highlighting their potential for evaluation in subsequent nutrition studies on CVD prevention with much-improved interpretability.

Notes

Supported by Canadian Institutes of Health Research Grant 378193 (to MJ), the Canada Research Chair Program, and Banting Foundation Discovery Award (to MJ). The funders had no role in the design, implementation, analysis, and interpretation of the data.

Supplemental Figure 1, Supplemental Methods, and Supplemental Tables 1–7 are available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/ajcn/.

Abbreviations used: CCHS, Canadian Community Health Survey; CVD, cardiovascular disease; ED, energy dense; ESF, energy from saturated fat; FD, fiber density; HSF, high saturated fat; LFD, low fiber density; PLS, partial least squares; RRR, reduced rank regression.

Data Availability

Data described in the article, code book, and analytic code will be made available upon request pending application and approval.