The effect of the Mediterranean Diet on lifespan: a treatment-effect survival analysis of a population-based prospective cohort study in Southern Italy

Abstract

Background

The Mediterranean diet (MedDiet) is associated with good health. We aimed to estimate the effect of levels of adherence to the MedDiet on lifespan by performing treatment effects survival analysis.

Methods

A sample of 5250 subjects aged ≥18 years were randomly selected from the electoral list of Castellana Grotte and Putignano (Apulian Region, Italy). Cohorts were enrolled in 2005–06 and followed-up until December 2018. The adherence to the MedDiet was measured by the relative Mediterranean score (rMED) and categorized as high, medium and low. Time-to-death (all-causes) as estimated by average treatment effect on the treated (ATET), potential outcome mean (POM) and relative efficiency of exposure were the outcomes.

Results

A total of 4896 subjects were included. The median follow-up time was 12.82 (inter quartile range (IQR) 12.22–13.05), 12.91 (IQR 12.21–13.27) and 12.84 (IQR 12.19–13.03) years for high, medium and low rMED subjects respectively. By December 2018, 453 (9.25%) had died. There was a strong effect of medium and low rMED {ATET, −5.10 [95% confidence interval (CI) −9.39, −0.80] and −8.91 (95%CI −13.37, −4.45), respectively}. High rMED has an important effect on mean age at death [POM 90.16 (95% CI 86.06, 94.25)]. The relative effect size for medium and low rMED subjects was a lower lifespan of 5.62% (95% CI 1.01, 10.3) and 9.90% (95% CI 5.30, 5.30), respectively.

Conclusions

We observed an important benefit in additional years of survival from adherence to MedDiet in this southern Italian cohort. Further investigation corroborating our findings in other population groups in other geographic regions will be an important contribution to promoting health and longevity.

Treatment effects survival analysis, Mediterranean diet score, prospective cohort study, lifespan, average treatment effect on the treated, potential outcome mean

Key Messages

• The Mediterranean diet is associated with good health outcomes but the effect on lifespan has not been widely studied

• We investigated the prospective association of adherence to a Mediterranean diet with lifespan in a southern Italian cohort.

• We observed a direct association between adherence to a Mediterranean diet at baseline and lifespan, assessed 12 to 13 years later

• Those with lower adherence to the Mediterranean diet at baseline had a shorter lifespan by five to eight years than those with high adherence, as measured 12–13 years later.

• Greater adherence to the Mediterranean diet can usefully be promoted based on its observed benefits to cardiovascular health and longevity.

Introduction

The Mediterranean diet (MedDiet) is recognized to be a healthy way of eating, first scientifically described by Keys et al.,¹ and associated with positive health outcomes.²

The association between a greater adherence to MedDiet with a reduced risk of all-causes mortality and the incidence of major chronic diseases has been shown in large epidemiologic studies.^2–5 Estimated effects of MedDiet include a reduction of all-causes mortality,⁶ of death/the incidence of cardiovascular (CVD) and cerebrovascular diseases, of death/the incidence of neoplastic diseases, and of the incidence of neurodegenerative diseases,² as well as other clinical outcomes such as stroke and mild cognitive disorders.² Moreover, the beneficial effect of MedDiet for the primary prevention of CVD has been shown in a Spanish multicenter randomized trial;⁷ besides, increasing changes in diet quality index such as the alternate MedDiet score have shown a reduction in overall and CVDs mortality over a period of 12 years.⁸

Much of the research about MedDiet has been performed using survival analysis, as the main interest has been to measure the effect of MedDiet on time-to-event, generally death. In this sense, Cox’s Model has been the most commonly used statistical tool because few assumptions are needed and it is an easy to fit model.⁹

However, two issues arise: the proportional hazard assumption is rarely satisfied in biology and the hazard ratio (HR) is not easy to interpret. Moreover, the assumed multiplicative effect of the exposure on the HRs is only useful if the exposure does not present an effect modification by other variables. Otherwise, the effect measure is not unequivocal.⁹

The defining characteristic of observational data is that the exposure status is not randomized; thus, some common variables affect exposure assignment and exposure-specific outcomes. When all the variables that affect both exposure assignment and outcomes are observable, the outcomes are said to be conditionally independent of the exposure, and treatment effects estimators may be used. Treatment effects offer greater flexibility in estimators and functional forms for the treatment/exposure-assignment models,¹⁰ thus it is possible to estimate a population-level effect. In the survival analysis setting it is feasible to estimate the average difference in time-to-event when everybody, rather than nobody, is exposed. These types of measures are easier to estimate and interpret.¹¹

For each individual, the effect of the treatment/exposure is obtained by comparing what would happen if the individual received the treatment/exposure vs what would happen if s/he did not. The average treatment effect (ATE), potential outcome mean (POM) and average treatment effect on the treated (ATET), provide measures of the effect, in the time units in which the time-to-event is measured.¹² ATE is the population average of the contrast in outcomes when everyone gets the treatment/exposure and when no one gets it, whereas POM is the potential outcome mean in those subjects not treated/exposed. ATET measures the average treatment/exposure effect in a well-defined, at-risk subpopulation, whereas the ATE/POM or ATET/POM ratios measure the relative importance of the effect.¹²

We used data drawn from two population-based prospective cohort studies conducted in the Apulian Region, Italy. The local version of MedDiet is the most prevalent dietary pattern in this region. In light of previous results, which suggested a lower HR for individuals with a high adherence to MedDiet,¹³ we hypothesized that a higher adherence to MedDiet may grant a longer life. Thus, we aimed to estimate ATET, POM and ATET/POM ratio from observational time-to-event data, using survival treatment effect in individuals with a low and medium as compared with high adherence to MedDiet.

Methods

Details about study populations have been published elsewhere.^14,15 Briefly, two different prospective cohort samples from studies conducted by the Laboratory of Epidemiology and Biostatistics of the National Institute of Gastroenterology, ‘Saverio de Bellis’ Research Hospital (Castellana Grotte, Bari, Italy) were included. The MICOL Study is a systematic 1-in-5 random sample study drawn from the electoral list of Castellana Grotte (≥30 years old) in 1985 and followed up in 1992, 2005–06 and 2013–16. In 2005–06, using the same sampling scheme, this cohort was supplemented with a random sample of subjects (PANEL study) aged 30–50 years, to compensate for the cohort aging. In this paper the baseline for the MICOL cohort was established in 2005–06 to capture all ages and to homogenize follow-up time.

The NUTRIHEP Study is a cohort drawn in 2005–06 from the Putignano General Practitioners (GPs) records (≥18 years old). Using a systematic random 1-in-5 sampling procedure, a sample of the general population ≥8 years of age was drawn from the GP’s list of records. We used the records of GPs, instead of a drawing from the census, because no significant difference was found between the distribution by age and sex of the general population from Putignano and the subjects inscribed in GPs’ records. In Italy, it is stated by law that everybody should have a GP. Therefore, the list of general population in the GP offices corresponds to the census list.

A total of 5152 out of 5250 eligible subjects (98.1% response rate) gave informed written consent to participate. Although presumably there would have been few events in the younger age classes, we combined the two cohorts to reflect a wider range of age groups and exposures, as the MICOL cohort was older than NUTRIHEP.

All procedures performed were in accordance with the ethical standards of the institutional research committee [IRCCS Saverio de Bellis Research and Ethical Committee approval for the MICOL Study (DDG-CE-347/1984; DDG-CE-453/1991; DDG-CE-589/2004; DDG-CE 782/2013)]; and the NUTRIHEP Study in 2005 and 2014 (DDG-CE-502/2005; DDG-CE-792/2014) and with the 1964 Helsinki declaration.

Data collection

Participants were interviewed at baseline in 2004–05 by trained physicians and/or nutritionists to collect information on sociodemographic characteristics, health status, personal history and lifestyle factors including history of tobacco use, food intake, educational level (International Standard Classification of Education),¹⁶ work (International Standard Classification of Occupations, International Labour Office)¹⁷ and marital status.

Weights and heights were measured in underclothing and without shoes. Weights were taken on an electronic balance, SECA^®, and recorded to the nearest 0.1 kg. Height was measured with a wall-mounted stadiometer, SECA^®, and recorded to the nearest 1 cm. Blood pressure (BP) measurement was performed following international guidelines.¹⁸ The average of three BP measurements was calculated.

The European Prospective Investigation into Cancer and Nutrition (EPIC) food frequency questionnaire (FFQ) was used to document the usual food intake of participants at baseline.¹⁹ Nutritionists conducted an in-person structured interview asking participants to report on their frequency of usual intake of 233 foods items over the past year; they reported intakes per day, per week or per year. They were also asked to estimate their portion sizes from photographs; questions were referred to the usual intake in the last year. No further dietary assessments were made during the 12–13 year follow-up of the cohort.

Data from FFQs were entered by trained dietitians into the National Cancer Institute electronic data base (Milan, Italy) to calculate total intake of food groups, energy and nutrients. Scores for adherence to the MedDiet were calculated using Stata statistical software version 16.0.¹⁹ Details about questions from the FFQ are shown in Supplementary Material, available as Supplementary data at IJE online.

A fasting venous blood sample was drawn, and the serum was separated into two different aliquots. One aliquot was immediately stored at −80°C. The second aliquot was used to test biochemical serum markers by standard laboratory techniques in our central laboratory.

Exposure assessment

Adherence to the MedDiet was measured by the relative Mediterranean²⁰ scoring system which allocates scores (rMED) based on tertiles of reported intake from nine distinct food groups including fruit (excluding fruit juices), vegetables (excluding potatoes), legumes, cereals, fresh fish, olive oil and meat and dairy products, alcohol intake. A score of either 0, 1 or 2 was assigned (with 2 being the high score) for increasing tertiles of each food group, except for meat and dairy products which were scored as 2, 1 or 0 as the high score. Alcohol intake was scored as a 2 for moderate consumers (5–25 and 10–50 g/day for women and men respectively) and 0 otherwise. Scores were summed and rMED ranged from 0 to 18; these were categorized as low (0–6), medium (7–10) and high (11–18) adherence as suggested by the literature.²⁰

The advantages of this scoring system is that it is ‘adjusted for energy’ (based on grams of a food group per 1000 kcal/day), classified on the basis of tertiles of intake of each component (except for alcohol), and includes intake of olive oil of which Italy is a great producer and consumer.²¹

Tracing procedures and outcome assessment

Information about the vital status of participants was obtained from the Municipality of Castellana Grotte and was electronically linked with the database. Inquiries were also made at the Municipalities of current residence of subjects who had migrated.

Statistical analysis

Mean (±SD), median(±IQR) or proportions were used as descriptive measures and tested by ANOVA or Chi-squared test or Fisher exact test as appropriate. As individuals were recruited at random times in their lives and no intervention was assigned, to ensure the best possible adjustment age at death was chosen as the time scale.

ATETs, POMs and the ATET/POM ratio treatment-effects survival analysis estimators were implemented. We used inverse-probability-weighted regression adjustment (IPWRA)^22,23 estimators, which uses estimated weights that control for missing data, for both treatment and outcome models.²⁴

ATET was used because it may be estimated under less restrictive versions of the conditional independence and sufficient overlap assumptions than those required for ATE.²⁴

The outcome (time-to-event) and treatment/exposure assignment (low, medium and high adherence to MedDiet) was modeled as a function of the same covariates, namely, age in years, sex (female vs male), smoking ( never/former vs current), diastolic (DBP) and systolic (SBP) blood pressure, energy daily intake, low-density lipoprotein-cholesterol (LDL-Chol), high-density lipoprotein-cholesterol (HDL-Chol), triglycerides, glycemia and body mass index (BMI), educational level, occupation and employment and marital status. We choose to treat age as categorical variable to reflect the different age structures of the two cohorts. The relative effect size of exposure was estimated as the ATET to POM ratio along with the 95% confidence interval (95%CI).

To test the potential beneficial effect of MedDiet on lifespan we choose the high adherence level as the non-treated/exposed category.

Specification checks for the treatment/exposure assignment model and the overlap condition were performed. Covariate balance was checked by means of raw and standardized differences and variance ratio.^12,25 The overlap assumption about the probability of each treatment/exposure level was graphically inspected and investigated by predicting the propensity score by treatment level. Moreover, kernel density plots were drawn for each covariate (raw and weighted) to probe the balance over treatment/exposure levels.^11,25

Time from enrolment to death, migration or end of the study (31 December 2018), whichever occurred first, was considered as the time-to-event.

Statistical analysis was performed using Stata statistical software version 16.0 (StataCorp, Texas, USA).

Results

rMED components

The proportion of participants who scored the maximum points for the various rMED components is shown in Table 1. Overall, ∼33% of participants scored maximum points for all components (excluding the score for alcohol intake). Women were likely to score the maximum for vegetables, fruits, legumes, fresh fish, total meat, olive oil and alcohol whereas men scored the maximum for cereals and dairy products. Lower proportions of younger age classes, mainly <35 years old, scored the maximum for all components

Table 1

Scoring criteria of the rMed components and percentange of participants with maximum rMed components scores

Components (g/1000 kcal per day)	Scoring criteria of rMed components			% of participants with maximum component scores
	I Tertile	II Tertile	III Tertile	Age class (years)					Sex
	0 point	1 point	2 points	<35	35–44	45–54	55–64	≥65	Female	Male	Total
				n = 829	n = 1109	n = 945	n = 950	n = 1063	n = 2379	n = 2517	n = 4896
Vegetables	<115	115–209	>209	25.69	30.57	31.85	39.37	38.38	42.50	24.79	33.39
Fruits	<269	269–507	>507	16.28	31.38	36.19	40.21	39.60	38.92	27.89	33.25
Legumes	<15	15–29	>29	27.86	30.48	31.43	38.63	37.63	38.00	28.96	33.35
Cereals	<220	220–340	>340	29.07	34.90	33.97	33.47	34.34	31.06	35.48	33.33
Fresh fish	<19	19–36	>36	31.36	31.83	34.18	36.74	31.89	36.28	30.23	33.17
Total meat	>90	53–90	<53	19.66	22.63	29.31	38.42	47.88	34.93	29.16	31.96
Dairy products	>279	148–279	<148	33.78	30.03	35.34	34.32	28.03	23.83	39.89	32.09
Olive oil	<19	19–31	>31	21.95	17.31	29.84	41.37	52.96	37.16	28.92	32.92
Alcohola	Otherwise	–	5–25 or 10–50	21.11	12.71	11.32	9.47	6.68	20.34	3.97	11.93

a

Female: 5–25 g/day; male: 10–50 g/day.

The cohort

There were 5152 (2851 from MICOL/PANEL and 2301 from NUTRIHEP) subjects eligible for the study but there was complete available information for 4896 (95.0%). A total of 256 FFQ were excluded: 173 because there were >10% of missing answers and 83 because their caloric intake was >5000 kcal daily.

The study base generated a total observation time of 67 714.49 person-years and median follow-up time of 12.82 (12.22, 13.05), 12.91 (12.21, 13.27) and 12.84 (12.19, 13.03) years for high, medium and low rMED subjects respectively.

Baseline characteristics of participants are shown in Table 2. Only 28.3% of the whole cohort scored low rMED. About 55.6% of low rMED scorers were <44 years old whereas 72.8% of high rMED scorers were ≥45 years or older. About 77% of women and 66.7% of men scored high or medium rMED. Less than 1% of subjects were not born in Italy; 81.2% came from Europe (mainly Albania) or Africa (mainly Morocco) and 70.83% of them were medium to high rMED scorers. Subjects with primary or lower secondary education level scored as follows: high (61.2%), medium (63.0%) and low (55.9%) rMED. Married or living together subjects were more likely high (78.0%) rMED scorers than single subjects (13.2%). About 16% of managers and professionals and 59% of craft, agricultural, sales workers or elementary occupations and housewives were high rMED scorers. Most never/formers smokers scored high or medium rMED (84.7%). Low rMED scorers had a better biochemical profile, BMI and lower SBP and DBP.

Table 2

Baseline characteristic of participants by rMed: Castellana Grotte (BA)/Putignano, Italy, 2005–18

	rMeda
Variable	High	Medium	Low	P-value
n	1001 (20.4%)	2509 (51.2%)	1386 (28.3%)
Age at enrollment (years)	54.59 (15.54)	52.81 (15.46)	45.49 (15.51)	<0.001
SBP (mmHg)	127.86 (17.13)	123.95 (18.19)	120.85 (16.51)	<0.001
DBP (mmHg)	79.35 (9.31)	76.17 (9.69)	76.11 (9.40)	<0.001
BMI (kg/m2)	27.24 (4.98)	27.92 (5.24)	26.93 (5.01)	<0.001
Triglycerides (mmol/L)	1.36 (0.91)	1.39 (1.00)	1.37 (1.03)	0.78
Total cholesterol (mmol/L)	5.18 (1.04)	5.12 (1.00)	5.00 (0.99)	<0.001
HDL-Chol (mmol/L)	1.36 (0.36)	1.34 (0.36)	1.31 (0.34)	<0.001
LDL-Chol (mmol/L)	3.20 (0.89)	3.16 (0.86)	3.07 (0.84)	<0.001
Glucose (mmol/L)	5.89 (1.52)	5.89 (1.40)	5.82 (1.34)	0.26
ALTb (µkat/L)	0.26 (0.24)	0.28 (0.23)	0.29 (0.19)	0.022
Age at death (years)c	68.79 (56.33, 78.42)	66.00 (52.68, 77.03)	55.69 (47.15, 67.59)
Follow-up (years)c	12.82 (12.22, 13.05)	12.91 (12.21, 13.27)	12.84 (12.19, 13.03)
Gender
Female	635 (26.7%)	1196 (50.3%)	548 (23.0%)
Male	366 (14.5%)	1313 (52.2%)	838 (33.3%)	<0.001
Age class (years)
<35	125 (15.1%)	334 (40.3%)	370 (44.6%)
35–44	147 (13.3%)	561 (50.6%)	401 (36.2)
45–54	203 (21.5%)	462 (48.9%)	280 (29.6%)
55–64	247 (26.0%)	538 (56.6%)	165 (17.4%)
≥65	279 (26.2%)	614 (57.8%)	170 (16.0%)	<0.001
Country of birth
Italy	996 (20.5%)	2479 (51.1%)	1373 (28.4)
Europe	5 (15.1%)	19 (57.6%)	9 (27.3%)
Africa	0	5 (83.3%)	1 (16.7%)
America	0	3 (60.0%)	2 (40.0%)
Asia	0	3 (75.0%)	1 (25.0%)	0.973
Educationd
Primary education	369 (24.2%)	861 (56.5%)	295 (19.3%)
Lower-secondary Education	244 (16.9%)	720 (49.8%)	481 (33.3%)
Upper-secondary education	275 (19.5%)	663 (47.0%)	471 (33.4%)
Tertiary education	65 (19.4%)	168 (50.2%)	102 (30.4%)
Illiterate	39(27.3%)	78 (54.5%)	26 (18.2%)
Without information	9 (25.0%)	17 (47.2%)	10 (27.8%)	<0.001
Marital status
Single	132 (16.0%)	399 (48.3%)	295 (35.7%)
Married or living together	781 (21.9%)	1865 (52.2%)	925 (25.9%)
Separated or divorced	23 (17.4%)	67 (50.8%)	42 (31.8%)
Widower	59 (21.3%)	166 (59.9%)	52 (18.8%)	<0.001
Job
Managers and professionals	54 (15.7%)	169 (49.1%)	121 (35.2%)
Craft, agricultural and sales workers	186 (13.7%)	676 (49.8%)	496 (36.5%)
Elementary occupations	223 (19.0%)	609 (51.7%)	345 (29.3%)
Housewives	178 (27.7%)	327 (50.9%)	138 (21.5%)
Pensioners	297 (27.5%)	620 (57.4%)	163 (15.1%)
Jobless	50 (19.8%)	89 (35.2%)	114 (45.1%)
Without information	0 (0.0%)	3 (100.0%)	0 (0.0%)	<0.001
Smoker
Never/former	877 (21.7%)	2096 (52.0%)	1059 (26.3%)
Current	117 (13.9%)	401 (47.7%)	322 (38.3%)	<0.001
Status
Alive/censored	924 (20.8%)	2243 (50.5%)	1276 (28.7%)
Dead	77 (17.0%)	266 (58.7%)	110 (24.3%)	0.004

a

Cells reporting subject characteristics contain mean (±SD) or n (%), unless otherwise indicated.

b

ALT, alanine aminotransferase.

c

Cells report median (IQR).

d

Educational level as classified by the International Standard Classification of Education (ISCED-97): Primary education includes level 0 (pre-primary) and 1 (primary education); Lower-secondary education includes level 2; Upper-secondary education includes level 3 (upper secondary education) and level 4 (post-secondary and non-tertiary education); Tertiary education includes level 5 (first stage of tertiary education) and level 6 (second stage of tertiary education).

A description of each cohort as well as a comparison among subjects with complete and incomplete data are shown in Supplementary Tables S1–S3, available as Supplementary data at IJE online. Missing data refer to those subjects who were excluded due to >10% missing data.

Results from treatment effects survival analysis are shown in Table 3.

Table 3

Effect of adherence to MedDiet (rMED score) on lifespan: Castellana Grotte (BA)/Putignano, Italy, 2005–18

rMED score	Coefficient	SE	P-value	95% CI
ATET
Medium vs high	−5.10	2.19	0.020	−9.39, −0.80
Low vs high	−8.91	2.27	0.000	−13.37, −4.45
POM
High	90.16	2.09	0.000	86.06, 94.25
ATET Medium/POM	−0.056	0.023	0.014	−0.103, −0.011
ATET low/POM	−0.099	0.023	0.000	−0.144, −0.053

The ATET for medium vs high rMED was −5.10 (95%CI −9.39, −0.80) and for low vs high rMED was −8.91 (95%CI −13.37, −4.45), whereas POM for high rMED was 90.16 (95% CI 86.06, 94.25). The relative effect size for medium rMED subjects was estimated as 5.62% (95% CI 1.01, 10.3) less in the time to death whereas for low rMED this effect was 9.90 % (95% CI 5.30, 14.4).

In short, in this at-risk subpopulation of treated/exposed subjects (medium and low rMED) the average time to death was estimated to be 6.21 and 8.28 years earlier respectively as compared with subjects with high rMED. The average time to death when everyone in the population scored high rMED is 90.16 years.

Results of exploring the balance of covariates are shown in Table 4.

Table 4

Treatment effects survival analysis diagnostic check for the treatment model covariates. Raw and weighted standardized differences and variance ratios for average treatment effect on the treated by rMED: Castellana Grotte/Putignano (BA), Italy, 2005–18

	Medium rMED				Low rMED
	Standardized differences		Variance Ratio		Standardized differences		Variance ratio
	Raw	Weighted	Raw	Weighted	Raw	Weighted	Raw	Weighted
Age (categorical, years)	−0.058	−0.036	0.954	0.939	−0.305	0.062	0.657	1.330
Gender	0.318	0.015	1.072	0.999	0.513	−0.024	1.016	1.001
Smoker	0.120	−0.009	1.284	0.984	0.297	−0.021	1.692	0.961
SBP, mmHg	−0.218	−0.024	1.132	1.090	−0.412	0.033	0.914	1.146
DBP, mmHg	−0.328	−0.003	1.088	1.052	−0.329	−0.018	1.002	1.064
Total cholesterol (mmol/L)	−0.049	0.034	0.915	0.976	−0.164	0.033	0.905	0.962
Triglycerides (mmol/L)	0.026	−0.030	1.209	1.107	0.001	−0.005	1.331	1.173
Glucose (mmol/L)	0.009	−0.047	0.873	0.829	−0.079	0.003	0.655	1.061
BMI, kg/m2	0.140	−0.086	1.120	0.929	−0.077	−0.012	0.991	1.071
Educational levela	0.003	0.032	0.936	1.008	0.187	0.013	0.747	0.938
Jobb	−0.226	0.004	1.023	0.989	−0.404	0.003	1.108	1.007
Marital statusc	−0.013	−0.022	1.162	1.045	−0.197	0.079	1.063	1.362
Energy daily intake	0.465	−0.098	1.368	0.869	0.748	−0.145	1.573	0.823

a

Educational level as classified by the International Standard Classification of Education (ISCED-97): Primary education includes level 0 (pre-primary) and 1 (primary education); Lower-secondary education includes level 2; Upper-secondary education includes level 3 (upper secondary education) and level 4 (post-secondary and non-tertiary education); Tertiary education includes level 5 (first stage of tertiary education) and level 6 (second stage of tertiary education).

b

Job as classified by International Standard Classification of Occupations, International Labour Office includes managers and professionals; craft, agricultural and sales workers; elementary occupations, housewives, pensioners, jobless and without information.

c

Marital status includes single, married or living together, separated or divorced and widower.

Weighted standardized differences are much closer to 0 than the raw values and the weighted variance ratios are closer to 1 than the raw variance ratios, indicating that weights balanced covariates. The density of the propensity scores for low, medium and high rMED subjects shows a clear distribution offering the same support as weighted standardized differences and variance ratios, indicating that there is no violation of the overlap condition (Figure 1).The predicted propensity score shows that the maximum for each treatment level is <1 (Table 5).

Figure 1

Average treatment effect on the treated. Propensity scores by rMed. Castellana Grotte/Putignano (BA), Italy, 2005–18

Table 5

Average treatment effect on the treated. Propensity score by levels of rMED: Castellana Grotte/Putignano (BA), Italy, 2005–18

rMED	Observations	Mean	SD	Minimum	Maximum
High	1001	0.279	0.130	0.022	0.689
Medium	2509	0.199	0.112	0.000	0.746
Low	1386	0.168	0.101	0.001	0.579

Kernel density plots for each of the covariates in the treatment/exposure assignment model show that the weighted covariate distributions do not vary over treatment levels (Supplementary Figure S1–S13, available as Supplementary data at IJE online).

Results from treatment effect sensibility analysis both from treatment/exposure assignment and outcome models are shown in Supplementary Tables S4 and S5, available as Supplementary data at IJE online, respectively. It is worth noting the importance of adding BMI to the treatment assignment and energy daily intake to the outcome models.

Discussion

In health sciences studies, outcomes are frequently time-to-event, and treatment effects survival analysis provides a measure of the treatment/exposure effect in the same time units in which the outcome is measured. This study shows that, in this population, subjects with medium or low adherence to MedDiet have a decline of 5.62 % (95% CI 1.04,10.2) or 9.78 % (95% CI 5.19,14.35,), respectively, in time-to-death as compared with high adherence to MedDiet. Specification checks for the treatment assignment model and the overlap condition showed that the assumptions behind the model hold.

The health benefits of the MedDiet, including prevention of prevalent chronic diseases have been recognized.^1,7,26–28 Research about beneficial effects of MedDiet has included several different approaches² and is continuing to explore its interactions with and effect modification by other variables such as environmental conditions, physical activity, seasonality and conviviality, which have recently been incorporated into the MedDiet frame.^2,29

The strongest evidence about the relationship between MedDiet and health outcomes pertains to mortality and CVD; cancer incidence, neurodegenerative diseases and diabetes and other outcomes require stronger evidence.³⁰ Evidence derives from meta-analyses and umbrella reviews of both observational and randomized clinical trials^13,30,31 conducted in different settings³² and populations.^33,34 Overall, there is evidence of an inverse relationship between adherence to MedDiet and negative health outcomes, measured by a variety of indices.³⁵ In this study, low rMed scorers were younger, had a better biochemical and anthropometric profile, yet 64% of deaths in this low scoring group were due to cancers or CVDs (data not shown).

Previous work about MedDiet effect has used classical survival analysis, namely Cox’s model⁹ or similar statistical techniques that estimate the effects in a multiplicative scale (relative risk). The assumptions behind these methods often fail to find the biological plausibility to support them. Moreover, the relative risk estimator is difficult to understand and to transfer to common language and clinical practice. We performed our survival analysis using treatment effects, which presents at least two advantages, namely parameters are estimated in their natural scale and the differences constitute the natural measure used in public health, and are easier to communicate and more easily understood by the public.

Few studies have tried to estimate the effect of MedDiet on lifespan.^36–39 Among those studies that have investigated survival, the duration of follow-up was similar in all studies including ours, but they differ from ours in exposure assessment and statistical methodology. Mayet al.³⁷ and Struijk et al.³⁸ performed a two-part model (logistic and linear models) to estimate the impact of several modifiable health behaviors or diet only on a combination of diseases and mortality using disability-adjusted life years. They found that persons adhering to four healthy lifestyle characteristics lived a minimum of 2 years longer in good health,³⁷ and in persons with high adherence to MedDiet only 51 days of healthy life were lost.³⁸ The rate advancement period (RAP) approach⁴⁰ used by Fresán et al.³⁶ and van den Brandt³⁹ showed that the best RAP (3.1 years gained) was associated with the Mediterranean dietary pattern³⁶ or with RAP of 15.1 years in women and 8.4 years in men who adhered to a four item healthy lifestyle score including adherence to MedDiet.³⁹ However, it is not possible to compare these results with ours, although quantitatively similar, as RAP is a model-based estimate⁴⁰ whereas treatment effects use the potential outcome framework.⁴¹

The potential outcome or counterfactual framework may be traced back to Hume's work in the 18th century, but it was only in the 20th century that intensive theoretical development occurred. These models formalize notions of cause and effect found in much of philosophy and epidemiology.^41,42

This is a treatment effect survival analysis applied to real-life data to estimate the effect of MedDiet on lifespan. ATET is easy to estimate and interpret. Estimates do not need specific knowledge about the values of covariates and, in the clinical setting, are able to formulate a prediction about a particular patient. Moreover, these types of measures are useful not only for designing but also for assessing the effects of population-based health policies.

This approach can replace multiplicative methods in some areas of biomedical research. One area of application could be research in public health, as the estimates obtained are on an additive scale, which seems to be the natural scale for public health research.⁴³ In addition, outcomes from survival analysis in years of lifespan lost or gained are relatively simple for patients and the public to understand.

This study has several strengths, particularly the cohort design, and the large random population sample from a geographic area where the MedDiet is the cultural norm to varying degrees. Exposure assessment was performed using an internationally recognized FFQ and previously published methods for scoring for adherence to a MedDiet.¹⁹ In a 1997 validity study conducted by the EPIC study group in Italy, they compared results of the FFQ with 24-h recall diet assessment and a biomarker for protein intake.⁴⁴ Close agreement between the methods was found to be good only for alcohol, but the FFQ overestimated intakes of vegetables, meats, fish and protein and underestimated pasta and rice, cheeses, legumes, cakes, added fats and soft drinks. Subsequent modifications were made to the questionnaire by the EPIC study team, which would be expected to decrease these measurement errors. In our study, the accuracy of the point estimates of food and nutrient intakes was of less interest. Rather, we sought to classify respondents into tertiles of intake, drawing on the strength of the FFQ method to rank-order subjects and thus reduce misclassification error. We acknowledge that study participants may also have changed their diets over the 12–13 year follow-up, which may have influenced the outcome. Indeed, a recent study showed lower adherence to the MedDiet among older age groups in a southern Italian population, although most of the difference in adherence was attributed to lower olive oil intake with aging. We believe our dietary assessments were sufficiently accurate for the purposes of our analyses.

Our study did not include a measure of physical activity, a potential serious limitation given previous research findings linking physical activity with all-cause and cause-specific mortality. Thus, we may have over or underestimated the effect of diet due to confounding or effect modification of physical activity.^45,46 On the other hand, ATET may be still estimated if some unmeasured variables increase (or decrease) the likelihood of assignment to the treatment/exposure, increase (or decrease) the time-to-event in the treatment/exposure group, and has no effect when not in the treatment/exposure group.²⁴

In our study, neither socio-economic status, as measured either by level of education or by occupation nor country of birth did not modify the effects of adherence to MedDiet on survival.

We have included all the available variables related to the treatment/exposure and the outcome (years to death), and diagnostic tools used to verify both conditional independence and the sufficient overlap assumption confirmed a good fit of the model. Whereas the sufficient overlap assumption was not satisfied for ATE (Figure 1), ATET may be estimated under less restrictive assumptions. This weaker version only requires that each individual in the treated/exposed subpopulation have a positive probability of not getting treated (being exposed). In particular, ATET but not ATE can be estimated when some individuals have zero chances of getting the treatment (being exposed).^23,47

In conclusion, high adherence to a MedDiet appears to have an important beneficial effect on lifespan in a southern Italian population. This adds longevity to evidence of health benefits of such a diet and will be useful in promoting the diet to patient and population groups alike. Although our results showed a positive effect on survival, unmeasured confounding/effect modification may have influenced the survival advantage we observed from adherence to the MedDiet. As well, our results may not be generalizable to samples in other geographic areas where a MedDiet is atypical, and among populations of different characteristics. Further studies using our analytical methods would be informative in confirming the survival benefits of the MedDiet across populations and geographic regions.

Supplementary data

Supplementary data are available at IJE online.

Funding

MICOL III: This research was supported by a public grant from the Ministry of Health, Italy (Progetto Finalizzato del Ministero della Salute, ICS 160.2/RF 2003, 2004/2006).MICOL IV: This research was supported by a public grant from the Ministry of Health, Italy (Ricerca Corrente DDG 045 del 24.01.2017). NUTRIHEP: This research was supported by a public grant from the Ministry of Health, Italy (Progetto Finalizzato del Ministero della Salute- Progetto no. 37-2004).

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.

Author contributions

A.C., G.M. and A.M contributed equally to the conception and design of the research and drafted the manuscript; M.G.C., L.R.A., L.V.A.S., A.B., I.F., P.S., C.Bu., A.M.C., M.N. and V.M.B.G. contributed equally to the acquisition, editing and cleaning of the data; C.Bo contributed to the formal analysis and A.R.O. contributed to writing, reviewing, and editing the manuscript; A.R.O. was responsible for supervision and project administration.

Conflict of interest

None declared.

Author notes

Joint first authors