Dealing with variability in food production chains: a tool to enhance the sensitivity of epidemiological studies on phytochemicals

Abstract

Background: Many epidemiological studies have tried to associate the intake of certain food products with a reduced risk for certain diseases. Results of these studies are often ambiguous, conflicting, or show very large deviations of trends. Nevertheless, a clear and often reproduced inverse association is observed between total vegetable and fruit consumption and cancer risk. Examples of components that have been indicated to have a potential protective effect in food and vegetables include antioxidants, allium compounds and glucosinolates. Aim: The food production chain can give a considerable variation in the level of bioactive components in the products that are consumed. In this paper the effects of this variability in levels of phytochemicals in food products on the sensitivity of epidemiological studies are assessed. Methods: Information on the effect of variation in different steps of the food production chain of Brassica vegetables on their glucosinolate content is used to estimate the distributions in the levels in the final product that is consumed. Monte Carlo simulations of an epidemiological cohort study with 30,000 people have been used to assess the likelihood of finding significant associations between food product intake and reduced cancer risk. Results: By using the Monte Carlo simulation approach, it was shown that if information on the way of preparation of the products by the consumer was quantified, the statistical power of the study could at least be doubled. The statistical power could be increased by at least a factor of five if all variation of the food production chain could be accounted for. Conclusions: Variability in the level of protective components arising from the complete food production chain can be a major disturbing factor in the identification of associations between food intake and reduced risk for cancer. Monte Carlo simulation of the effect of the food production chain on epidemiological cohort studies has identified possible improvements in the set up of such studies. The actual effectiveness of food compounds already identified as cancer-protective by current imprecise methods is likely to be much greater than estimated at present.

Eur J Nutr 42 : 6772 (2003)

DOI 10.1007/s00394-003-0412-8ORIGINAL CONTRIBUTIONMatthijs DekkerRuud Verkerk Dealing with variability in food production

chains: a tool to enhance the sensitivity of

epidemiological studies on phytochemicals Summary Background Many

epidemiological studies have tried

to associate the intake of certain

food products with a reduced risk

for certain diseases. Results of

these studies are often ambiguous,

conflicting, or show very large deviations of trends. Nevertheless, a

clear and often reproduced inverse

association is observed between total vegetable and fruit consumption and cancer risk. Examples of

components that have been indicated to have a potential protective

effect in food and vegetables include antioxidants, allium com-Received: 15 July 2002

Accepted: 26 January 2003Matthijs Dekker () R. Verkerk

Wageningen University

Product Design and Quality Management

GroupDept. of Agrotechnology and Food SciencesP. O. Box 8129

6700 EV Wageningen, The Netherlands

E-Mail: [email protected] and glucosinolates. Aim

The food production chain can

give a considerable variation in the

level of bioactive components in

the products that are consumed. In

this paper the effects of this variability in levels of phytochemicals

in food products on the sensitivity

of epidemiological studies are assessed. Methods Information on the

effect of variation in different steps

of the food production chain of

Brassica vegetables on their glucosinolate content is used to estimate the distributions in the levels

in the final product that is consumed. Monte Carlo simulations of

an epidemiological cohort study

with 30,000 people have been used

to assess the likelihood of finding

significant associations between

food product intake and reduced

cancer risk. Results By using the

Monte Carlo simulation approach,

it was shown that if information on

the way of preparation of the products by the consumer was quantified, the statistical power of the

study could at least be doubled.

The statistical power could be increased by at least a factor of five if

all variation of the food production

chain could be accounted for. Conclusions Variability in the level of

protective components arising

from the complete food production

chain can be a major disturbing

factor in the identification of associations between food intake and

reduced risk for cancer. Monte

Carlo simulation of the effect of the

food production chain on epidemiological cohort studies has identified possible improvements in the

set up of such studies. The actual

effectiveness of food compounds

already identified as cancer-protective by current imprecise methods

is likely to be much greater than estimated at present. Key words Monte Carlo

simulations epidemiology food

production chain processing

sensitivity glucosinolatesIntroductionMany epidemiological studies have indicated the protective role of vegetable and fruit consumption for various diseases. Epidemiological data considering the effect of vegetable and fruit intake on cancer risk have

been reviewed by several investigators [13].Although there is clearly a positive trend in the protective role of vegetables and fruit for various cancers,

the outcomes of individual epidemiological studies

sometimes conflict. This is clearly illustrated in a review

by Steinmetz and Potter [2], showing statistically significant inverse associations for 15 of the 21 examined

case-control studies between one or more vegetable/

fruit categories and colon cancer. Four of the studies did

not show significant inverse associations for any vegetable/fruit category and in two studies statistical sig- EJN 41268 European Journal of Nutrition, Vol. 42, Number 1 (2003) Steinkopff Verlag 2003nificance was not reported. Many sources of uncertainty

and variation can be responsible for these discrepancies.Accurate assessment of dietary intake is of importance to investigate associations between vegetable/

fruit consumption and cancer incidence. Several methods are available to assess intake, but they all have their

specific measurement errors. Dietary assessment methods can be roughly divided into methods assessing current diet, consisting of records and 24h recalls and

methods assessing habitual diet, consisting of dietary

history and food frequency questionnaires. The intake

data derived by these methods combined with analytic

data from food composition tables or from chemical

analysis, result in individual nutrient (and non-nutrient) intakes. In addition to these methods, some biological indicators (biomarkers) of dietary exposure

have been developed.Food composition tables are important tools for epidemiological studies on nutrients and non-nutrient

phytochemicals in relation with diseases. However, accurate knowledge of the nutrient and non-nutrient intake of individuals and groups of people requires information on the contents of prepared foods, in other

words prior to consumption. Unfortunately, dietary calculations are frequently made on the basis of foods as

brought into the kitchen. Also, for many components of

potential interest for health protection (like flavonoids

and other polyphenols, lignans, carotenoids, glucosinolates, peptides, lactic acid bacteria) no or only limited

food composition data are available [46].It is demonstrated that many steps in the food production chain, like cultivation, storage, processing and

preparation of the vegetables (Figs. 1 and 2), can have a

large impact on levels and thus intake of phytochemicals[7]. This variation has been investigated for the group of

glucosinolates from Brassica vegetables but is likely to

occur for most classes of health-promoting phytochemicals. The potential health benefits of glucosinolates

have been reviewed extensively in [8]. Epidemiological

studies on Brassica vegetables are given in [3, 9, 10].Epidemiology often does not take into account the effects of food processing and preparation on glucosinolate content, and thus of the consequences for potential

health-protection. It can be hypothesized that the high

variable levels of glucosinolate intake explain the inconsistent epidemiological findings. It is impossible to systematically gather information on all possible variations

in all steps of the production chain. Therefore, a predictive modeling approach was developed to estimate the

effects of variation in conditions and processes on the

level of glucosinolates and breakdown products [7, 11,12].Fig. 1 Schematic representation of the food production chain of Brassica vegetablesFig. 2 Distribution in the content of total glucosinolates of different Brassica

products (data from [11])Simulation strategy The steps in the simulation experiments of the epidemiological cohort studies are schematically represented in

Fig. 3. The hypothesis of the study is that whether an individual will develop cancer depends on his/her absolute risk of cancer which depends in part on the intakeM. Dekker et al. 69

Dealing with variability in food production chainswithandwhere x, and are continuous parameters. The distributions were truncated (0<x max), the parameters

used for the variation at the different steps of the production chain are given in Table 1. The values for cultivation give a distribution in level in the raw material (in

mole/100 g fresh weight).The resulting variation in the glucosinolate content

in the products that are consumed were calculated by

multiplying each randomly picked level of raw material

with the randomly picked industrial and consumer processing effect. The component intake was calculated by

multiplying the sampled product intake with the calculated (as described above) glucosinolate content in the

product.To calculate the relative risk of each consumer, a relation was assumed between the intake of a health-promoting component in the diet (glucosinolates in this

study) and the reduced risk for the development of cancer. This dose-response relation was described by the

equation:(2)Fig. 3 Schematic representation of the applied strategy for the simulation of epidemiological cohort studies on the protective effect of foods on cancer incidenceof protective compounds from foods. The simulation

strategy as depicted in Fig. 3 will give information as to

whether or not the cohort study will give a significant

association between product intake and the risk on developing cancer.Because of the availability of information of the distribution of the levels of glucosinolates and of some important effects of the food production chain, these compounds have been taken as the protective compounds.Three consumer groups were simulated with low

(0200),medium (200400) and high (4001000 g/week)

Brassica vegetable intakes. In total 30,000 consumers

were simulated with a distribution in intakes as shown

in Fig. 4.Distributions of the glucosinolate content in raw materials and of the effect of industrial processing, storage

and consumer processing were estimated from published experimental data for Brassica vegetables [7, 11,12].The distributions were described by log-normal distributions:(1)in which RR is the relative risk for cancer development,

RM is the minimum risk, EF is the effectiveness of the

component and I is the intake of the component. To assess the statistical power of epidemiological studies, different effectiveness factors (EF) for protecting against

cancer (EF = 120) and a fixed minimum risk (MR

= 0.2) were taken (Fig. 5).For each simulated consumer the relative risk was

transformed in an absolute risk value by multiplying the

RR by 0.01 (this value is estimated from the reportedFig. 4 Distribution of the 30,000 individuals in the three Brassica intake groupsTable 1 Distribution curves and parameters used for the three steps in the production chain taken into account in the Monte Carlo simulations ( mean; standard deviation)Step in chain Type of distribution max.Cultivation Truncated log normal 100 100 1000Industrial processing Truncated log normal 0.7 0.6 3Domestic cooking Truncated log normal 0.5 0.4 170 European Journal of Nutrition, Vol. 42, Number 1 (2003) Steinkopff Verlag 2003Fig. 5 Relative Risk (RR) for developing cancer as a function of the glucosinolate

intake, as calculated with Eq. (2) for five different values of the Effectiveness Factor

(EF) as shown within the graphcases of colon/rectal cancer in a recent cohort study[10]). In order to simulate whether an individual will develop cancer this calculated absolute risk was compared

with a fate of life factor that was randomly picked from

a uniform distribution between 0 and 1. If this fate of

life factor was lower than the absolute risk value the

consumer was marked as a cancer patient in the study.All Monte Carlo simulations were performed using

Pallisade @RISK software as an add-in for Microsoft

Excel.Simulation results and discussionFrom the three product intake groups (0200, 200400,

4001000 gram Brassica vegetables per week), the intake

of glucosinolates was calculated by taking into account

all the variation that could be due to three steps in the

production chain (cultivar, industrial processing and

domestic cooking). The resulting compound intake distributions are shown in Fig. 6. A considerable overlap in

compound intake levels can be seen between the groups,

owing to the large possible variation caused by the production chain. In reality some of the variation caused by

cultivar differences will level out because a range of different cultivars will be consumed over the years that

these studies take. The type of products and the way of

domestic cooking, however, can be expected to be not so

variable for each individual and therefore the variation

caused by these steps will not level out. Because of the

fact that only three sources for variation resulting from

the food production chain were taken into account,

while the complete chain will consist of at least 6 steps

(Fig. 1) with each multiple sources of variation [7], the

real variation might in fact be underestimated by the approach taken here.By using the compound distributions as shown in

Fig. 6 the risk of cancer of each individual was calculated. By the strategy described above this absolute risk

could be converted to the number of cancer patients in

the three product intake groups in the cohort.This assessment was done for three scenarios:A. The ideal situation, in which the production chain

causes no variation (median values of glucosinolate

content and processing effects were used, as given in

Table 1).B. The real situation, in which the three steps that are

considered in the production chain give the variation

as can be expected from experimental observations

(the whole distributions as given in Table 1 were

taken into account).C. An intermediate situation, in which the first two

steps of the production chain give the expected variation, but the domestic cooking effect is assumed to

be known for each individual, and therefore no additional variation is introduced by this step.The resulting relative risks for the three product intake groups is given in Fig. 7 for a value of the effectiveness factor (EF) of five. As can be seen in Fig. 7A, the

health protecting effect of Brassica vegetables could be

very significantly assessed if it were the case that the

food production chain gave products with a constant

level of glucosinolates. For the real situation, however,

in which the production chain gives an enormous variation in the levels of glucosinolates in products, no

health protective effect could be seen from the cohort

study (Fig. 7B). So even though the glucosinolates present in Brassica vegetables have quite a potent health

protective effect (see Fig. 5 for EF = 5), this protective effect cannot be seen in a large cohort study where only

the product intake is assessed as an input parameter. If

the study design were to include information on the

cooking behavior of the individuals the unknown variation in the levels of glucosinolates in the consumed

products could be reduced. In this case the cohort study

would conclude a significant health protective effect

(p < 0.05) of the consumption of Brassica vegetables on

the development of cancer (Fig. 7C).To establish the improvement in the statistical power

of cohort studies, the simulations were performed withFig. 6 Calculated glucosinolate intake of the three Brassica intake groups

(: 0200, : 200400, : 4001000 g Brassica/week)M. Dekker et al. 71

Dealing with variability in food production chainsvalues of EF between 1 and 20.The significance of the resulting health protecting effect is shown in Fig. 8, in

which the p-value between the relative risks of the high

and low consumption groups is plotted versus the EF

value. From this figure it can be concluded that the statistical power can be at least doubled by incorporating

information of cooking habits in the design of the study.

This can be seen in Fig. 8 by looking at the EF value for

which the study design gives a statistical significant difference in the relative risks. If the design were to allow

for assessment of all the possible variation of the food

production chain, the statistical power could be increased by at least a factor of five.In the design of epidemiological studies, the information that can be gathered from the individual on their

cooking habits will be limited by practical constraints

like the number of questions in a questionnaire or diary

and the accuracy of individuals in describing their

habits. However, in combination with a predictive modeling approach as described by Dekker et al. [7] and

Verkerk et al. [11], the quantification of cooking effects

on health parameters can be made in a computerized

way, using cooking time and the ratio of water to vegetable as inputs.This simulation study clearly shows that, by improving the design of epidemiological studies, health protective compounds in foods can be identified with more

statistical power. In addition, a simulation approach as

presented here can also be used to establish the minimum size of the cohort to be used in an epidemiological

study. Another important conclusion is that the underlying protective effect of certain compounds in foods

will be much larger than is shown in the present epidemiological studies that identified a significant protection by the intake of certain food product categories.Fig. 7 The relative risk for cancer for the three Brassica intake groups as observed

in the simulated cohort studies with three different scenarios: A, B and C. See text

for explanation of the scenariosFig. 8 Significance of the observed protective effect of Brassica vegetables on cancer incidence as a function of the Effectiveness Factor of glucosinolates for the three

study scenarios as explained in the text72 European Journal of Nutrition, Vol. 42, Number 1 (2003) Steinkopff Verlag 2003References1. Steinmetz KA, Potter JD (1991) Vegetables, fruit and cancer. I. Epidemiology.

Cancer Causes and Control 2:3253572. Steinmetz KA, Potter JD (1996) Vegetables, fruit and cancer prevention: a review. J Am Diet Assoc 96:102710393. Verhoeven DTH, Goldbohm RA, van

Poppel G, Verhagen H, van den Brandt

PA (1996) Epidemiological studies on

Brassica vegetables and cancer risk.

Cancer Epidemiology, Biomarkers &

Prevention 5:7337484. Hertog MGL, Hollman PCH, Katan MB

(1992) Content of potentially anticarcinogenic flavonoids of 28 vegetables

and 9 fruits commonly consumed in

the Netherlands. J Agric Food Chem 40:

237923835. Hertog MGL, Hollman PCH, van de

Putte B (1993) Content of potentially

anticarcinogenic flavonoids of tea infusions, wines and fruit juices. J Agric

Food Chem 41:124212466. Hollman PCH, Arts ICW (2000)

Flavonols, flavones and flavanols-nature, occurrence and dietary burden.

J Sci Food Agric 80:108110937. Dekker M, Verkerk R, Jongen WMF

(2000) Predictive modelling of health

aspects in the food production chain: a

case study on glucosinolates in cabbage. Trends in Food Science and Technology 11:1741818. Mithen, RF, Dekker, M, Verkerk, R,

Rabot S, Johnson, IT (2000) The nutritional significance, biosynthesis and

bioavailability of glucosinolates in human foods. J Sci Fd Agr 80:9679849. Michaud DS, Spiegelman D, Clinton SK,

Rimm EB, Willett WC, Giovannucci EL

(1999) Fruit and vegetable intake and

incidence of bladder cancer in a male

prospective cohort. J Natl Cancer Inst

(Bethesda) 91:60561310. Voorrips LE, Goldbohm RA, van Poppel

G, Sturmans F, Hermus RJJ, van den

Brandt PA (2000) Vegetable and fruit

consumption and risks of colon an rectal cancer in a prospective cohort study.

The Netherlands cohort study on diet

and cancer. American Journal of Epidemiology 152 (11):1081109211. Verkerk R (2002) Evaluation of glucosinolate levels throughout the production

chain of Brassica vegetables; towards a

novel predictive modelling approach.

PhD Thesis, Wageningen University,

The Netherlands12. Verkerk R, Dekker M, Jongen WMF

(2001), Post-harvest increase of indolyl

glucosinolates in response to chopping

and storage of Brassica vegetables. J Sci

Fd Agric 81:953958