Risco Percebido

download Risco Percebido

of 9

Transcript of Risco Percebido

  • 7/29/2019 Risco Percebido

    1/9

    Appetite 47 (2006) 324332

    Research Report

    Lay peoples perception of food hazards: Comparing aggregated data

    and individual data

    Michael Siegrista,b,, Carmen Kellerb, Henk A.L. Kiersc

    aHumanEnvironment Interaction, ETH, Zurich, ETH Centre CHN J75.1, CH-8092 Zurich, SwitzerlandbDepartment of Psychology, University of Zurich, Zurich, Switzerland

    cHeymans Institute, University of Groningen, Groningen, The Netherlands

    Received 14 October 2005; received in revised form 25 April 2006; accepted 5 May 2006

    Abstract

    The psychometric paradigm has been used to explain the perception of food hazard risks. In past studies, only aggregated data were

    analyzed, and individual differences were neglected. In the present study, both aggregated data and non-aggregated data are analyzed.

    Data stem from a mail survey conducted in Switzerland ( N 448). Analyzing aggregated data, results of past studies were successfully

    replicated. The PCA analysis revealed the two factors unknown risk and dread risk. Results of a three-way component analysis

    (3MPCA) suggest, that two components explain individual differences in the perception of food hazards. The two components were

    labeled unobservable hazards, and familiar hazards. Individual differences in the cognitive representation of hazards were

    correlated with attitudes toward natural foods. Results suggest that people who prefer natural foods differ in perceived risks from people

    who do not prefer natural foods. Results show that methods permitting individual differences are crucial for a better understanding of

    the cognitive representation of food hazards.

    r 2006 Elsevier Ltd. All rights reserved.

    Keywords: Food hazards; Psychometric paradigm; Risk perception; Three-way component analysis; Natural hazards

    Introduction

    The psychometric approach has been utilized to examine

    lay peoples attitudes toward food hazards (Hansen, Holm,

    Frewer, Robinson, & Sandoe, 2003). Psychometric studies

    identify factors that influence the perception of various

    hazards (Fischhoff, Slovic, Lichtenstein, Read, & Combs,

    1978). This research approach has been used to study a

    broad range of hazards including technological risks,

    activities, and food hazards (Slovic, 1987). Studies focusing

    on the perception of food-hazard risks (Fife-Schaw &

    Rowe, 1996, 2000; Sparks & Shepherd, 1994) have

    replicated the results found in the initial studies examining

    a heterogeneous set of hazards. In the psychometric

    paradigm, hazards are located on a cognitive map,

    which depicts the hazards in a space of two or more

    dimensions. The psychometric approach has been criti-

    cized, however, for its neglect of individual differences in

    risk perception (Sjo berg, 1996). Recent research has shown

    that there are substantial individual differences in risk

    perception that are not captured in the traditional

    psychometric paradigm (Siegrist, Keller, & Kiers, 2005).

    In the present study, we examine lay peoples perception of

    various food hazards utilizing a method that allows for

    individual differences.

    The psychometric paradigm

    In studies based on the psychometric paradigm, partici-

    pants evaluate a variety of attribute rating scales for a set

    of hazards (Fischhoff et al., 1978; Sparks & Shepherd,

    1994). Participants assess, for example, how dreadful the

    hazards are, whether the risks are known to science,

    whether people have control over their exposure to the

    hazard. The number of rating scales varies from study to

    study. In most studies utilizing the psychometric paradigm,

    ARTICLE IN PRESS

    www.elsevier.com/locate/appet

    0195-6663/$ - see front matterr 2006 Elsevier Ltd. All rights reserved.

    doi:10.1016/j.appet.2006.05.012

    Corresponding author. Department of Psychology, University of

    Zurich, Switzerland.

    E-mail address: [email protected] (M. Siegrist).

    http://www.elsevier.com/locate/appethttp://localhost/var/www/apps/conversion/tmp/scratch_10/dx.doi.org/10.1016/j.appet.2006.05.012mailto:[email protected]:[email protected]://localhost/var/www/apps/conversion/tmp/scratch_10/dx.doi.org/10.1016/j.appet.2006.05.012http://www.elsevier.com/locate/appet
  • 7/29/2019 Risco Percebido

    2/9

    averages are taken across all participants, and the data

    matrix (hazards rating scales) is submitted to a principal

    component analysis (PCA).

    In a study by Sparks and Shepherd (1994), lay peoples

    perceptions of 25 food hazards were examined. Partici-

    pants assessed these hazards with respect to 23 rating

    scales. The authors found that three factors accounted forthe correlations among the rating scales. The first

    component was labeled severity. The following rating

    scales were highly correlated with the first principal

    component: concern, dread potential, seriousness for

    future generations, and threatening widespread disastrous

    consequences. The second component was labeled un-

    known, and the rating scales known to science and known

    by people exposed were highly correlated with it. The third

    component, called number of people exposed, accounted

    for much less variance than the first two components. Food

    hazards, such as genetic manipulation or pesticide residues,

    were perceived as unknown and high-severe risks. Alcohol

    and excessive calories, for example, were perceived as

    known and low-severe risks. Some of these findings were

    successfully replicated (Fife-Schaw & Rowe, 1996, 2000;

    Kirk, Greenwood, Cade, & Pearman, 2002).

    In the studies reviewed, the data cube (hazards rating

    scalesparticipants) was reduced to a matrix (ha-

    zards rating scales). Factor loadings were computed for

    the rating scales, and factor scores were computed for the

    hazards. Because aggregated data were used, we do not

    know whether the model neglects individual differences in

    risk perception. In some studies (Bronfman & Cifuentes,

    2003; Marris, Langford, Saunderson, & ORiordan, 1997;

    Willis, DeKay, Fischhoff, & Morgan, 2005), differentmatrices were examined (e.g., person rating scales,

    hazards rating-scales). Such an approach produces mis-

    leading results, however, when substantial three-way inter-

    actions are present. Only a few studies have reported results

    of non-aggregated data (Siegrist et al., 2005; Vlek & Stallen,

    1981), but these studies did not specifically focus on food

    hazards. The study conducted by Siegrist et al. (2005), used

    a generalization of PCA, three-way principal component

    analysis (3MPCA; see Kiers & Van Mechelen, 2001), which

    is appropriate for the analysis of three-way data. Results of

    that study demonstrated that analyzing the non-aggregated

    data sets of the psychometric paradigm provides additional

    insight into individual differences in risk perception.

    Three-way principal component analysis

    Three-way data are data that can be arranged in a three-

    dimensional array. In the context of our study, the three

    dimensions are the individuals, the hazards and the rating

    scales. The three sets of entities associated with such three-

    way data sets are called the three modes of the array. It

    can be tempting to either analyze three-way data after

    aggregating over one of the three modes or to separately

    analyze all two-way data sets contained in the three-way

    data set. It should be noted, however, that such approaches

    do not yield an explicit description of the three-way

    interaction in the data and therefore can lead to conclusions

    that are incomplete at best. The strength of 3MPCA

    techniques is that they summarize all the information in a

    large, pre-processed three-way data set (i.e., all main effects

    and all interactions together), and they do so efficiently.

    Specifically, three-way methods summarize the entities ofeach mode by means of a few components and describe the

    relations between these components. The present study takes

    participants scores on rating scales with respect to different

    hazards and captures the (most salient) relations between

    individuals, hazards, and rating scales by the relations

    between components summarizing the individuals (indivi-

    dual-mode), hazards (hazard-mode), and rating scales

    (attribute-mode). This is particularly useful in the presence

    of a three-way interaction: Without the use of summarizing

    components, a full description of a three-way interaction

    may require as many terms as there are data points, which is

    not feasible unless the three-way data set is very small. For

    the reconstruction of the original data from the component

    variables related to individuals, hazards and rating scales, a

    small three-way array is used. This array is called the core

    array. The core array indicates the extent to which persons

    who score high on some person component give ratings that

    have high weights on some rating scales component, to

    hazards that have high weights on some hazard component.

    Various techniques for analyzing three-way data sets

    have been proposed; we selected three-mode principal

    component analysis (3MPCA) for the present study

    (Kiers & Van Mechelen, 2001; Kroonenberg & De Leeuw,

    1980; Tucker, 1966). As in PCA, in 3MPCA several choices

    have to be made (for example, how the data are to bepreprocessed, how many components are to be used, and

    what simple structure rotations are to be used). For details,

    we refer the reader to Kiers and Van Mechelen (2001).

    Simple examples explaining the logic of the 3MPCA

    procedure in some more detail can be found elsewhere

    (Siegrist et al., 2005; Van Mechelen & Kiers, 1999).

    Variables affecting risk perception

    External variables can be used in the explanation of

    individual differences in risk perception. Results of the

    study by Siegrist et al. (2005) showed that general trust

    influenced how hazards were perceived. In the domain of

    food, attitudes toward the naturalness of foods seems to be

    important (Rozin et al., 2004). Results of past research

    suggest that perceived naturalness is related to the

    acceptance of novel food (Tenbu lt, de Vries, Dreezens, &

    Martijn, 2005). In addition, self-reported purchase of

    organic foods was related to perceived benefits for human

    health (Magnusson, Arvola, Hursti, Aberg, & Sjoden,

    2003). Based on these results, we hypothesized that

    attitudes toward naturalness shape risk perception. Speci-

    fically, persons who show a preference for natural foods

    perceive food hazards differently compared with persons

    who do not show a preference for natural foods.

    ARTICLE IN PRESS

    M. Siegrist et al. / Appetite 47 (2006) 324332 325

  • 7/29/2019 Risco Percebido

    3/9

    Rationale of the present study

    The aim of present study is to better understand

    individual differences in laypeoples perceptions of food

    hazards. We examined how well results obtained by

    analyzing aggregated data sets represent individual risk

    perception in the domain of food risks. The first goal wasto analyze the aggregated data set and to compare the

    results with those from earlier studies (Fife-Schaw & Rowe,

    1996, 2000; Sparks & Shepherd, 1994). The second

    objective was to utilize a 3MPCA to analyze the individual

    variation that is not covered by the PCA of the aggregated

    data. The third objective was to describe the extent to

    which individual differences in various types of risk

    perception are related to a preference for natural foods.

    Method

    Participants

    The data for the present study come from a mail survey

    conducted in the German-speaking part of Switzerland. A

    questionnaire and accompanying letter were sent to a

    random sample of addresses from the telephone book. The

    letter and first page of the questionnaire asked that the

    questionnaire be completed by the person in the household

    who was over 18 years of age and who was next in line for

    their birthday. A reminder letter was sent out some time

    later. A second questionnaire was sent to persons who did

    not respond to the letter or reminder. Four hundred and

    forty-eight people completed the questionnaire, with a

    response rate of 37.5%.Fifty-two percent (n 233) of the respondents were

    women, and 48% (n 213) were men. Two participants

    did not report their gender. The mean age was 50.0

    (SD 16.0). Seventeen persons did not report their age.

    Education and income were not measured. Therefore, we

    do not know whether better-educated people were more

    likely to fill out the questionnaire.

    Questionnaire

    Participants were asked to rate 24 possible hazards on

    nine five-point scales. The hazards were similar to those

    used by Sparks and Shepherd (1994), with the addition of

    several new items including diet products, convenience

    foods and vitamin-enriched food. These new items were

    added because both functional foods (van Kleef, van Trijp,

    & Luning, 2005; Verbeke, 2005), and convenience foods

    (Verlegh & Candel, 1999) are increasingly important for

    the food sector. The nine rating scales were adapted from a

    study by Fischhoff and colleagues (Fischhoff et al., 1978):

    (1) Voluntariness of risk (1 voluntary; 5 involuntary).

    (2) Immediacy of effect (1 immediate; 5 delayed).

    (3) Knowledge of risk to those who are exposed

    (1 known precisely; 5 not known).

    (4) Scientific knowledge (1 little; 5 much).

    (5) Control over risk (1 uncontrollable; 5

    controllable).

    (6) Newness (1 new; 5 old).

    (7) Harm for health (1 not at all; 5 very strong).

    (8) Peoples worries (1 not worried; 5 very much

    worried).(9) Probability of health damage (1 not probable;

    5 probable).

    In addition to sociodemographic variables, the ques-

    tionnaire included five items designed to measure pre-

    ference for natural foods (e.g., Whenever possible I buy or

    consume organic food, I feel good when I eat natural

    foods). Respondents were asked to express their agree-

    ment or disagreement with these items using a value

    between 1 (dont agree at all) and 5 (agree absolutely).

    Data analysis

    (a) Data imputation: Data from 34 participants were not

    used because these participants failed to answer more

    than ten questions. For the remaining 414 respondents,

    missing values on an item were replaced by the mean

    value. More sophisticated data imputation procedures

    yielded virtually identical results.

    (b) Response scale recoding: Variables of rating scales 4

    (scientific knowledge), 5 (control over risk), and 6

    (newness) were recoded, so that a high value indicates

    great risk severity.

    (c) Two-way PCA on aggregated data: To replicate the

    results reported in earlier studies (Fife-Schaw & Rowe,2000; Sparks & Shepherd, 1994), a two-way PCA was

    carried out on the data averaged across the individuals.

    (d) Variance component estimation: Before carrying out a

    3MPCA, we checked to see whether the data could be

    analyzed reasonably well by means of a two-way PCA

    on aggregated data. Estimates of the percentages of

    variance for main effects of individuals, hazards, and

    rating scales, as well as for all pairwise interactions,

    were obtained from a fixed effect analysis of variance

    with repeated measures on the 414 24 9 three-way

    data table.

    (e) Three-way component analysis (3MPCA): This type of

    analysis entails three kinds of choices related to: (1) a

    possible preprocessing of the data prior to the actual

    analysis, (2) the choice of the number of the compo-

    nents for each of the three modes, and (3) the choice of

    a simple structure rotation (see Kiers & Van Mechelen,

    2001). The following summarizes the decisions we

    made for the analysis of the present data.

    First, regarding preprocessing of the data: due to the

    incomparability of the labels of the response dimen-

    sions a kind of standardization of the data is required.

    To eliminate the influence of (unknown) neutral points,

    data were centered across individuals (person mode),

    that is, for each hazard, for each rating scale, the

    ARTICLE IN PRESS

    M. Siegrist et al. / Appetite 47 (2006) 324332326

  • 7/29/2019 Risco Percebido

    4/9

    average score across persons is subtracted from the

    scores. An added benefit of this centering is that it takes

    out the very same information that was analyzed by

    the PCA on aggregated data. Thus, the results of the

    3MPCA are fully complementary to the results of the

    PCA. Moreover, given the difference in rating scale

    labels, a correction for artificial differences betweenresponses in scale range is also preferable. Therefore,

    data were scaled within rating scales (attribute mode)

    by dividing each entry by the standard deviation of the

    scores on the corresponding rating scale.

    Second, regarding the choice of the number of

    components (which may differ for individuals, hazards,

    and rating scales), several criteria were used. These

    included percentage of variance accounted for, a

    generalized scree test (Timmerman & Kiers, 2000),

    stability of analyses across two random splits of the

    data, and interpretability of the components. The

    generalized scree test relies on differences in percentage

    of variance accounted for between neighboring solu-

    tions, which in the two-way component analysis case

    correspond to eigenvalues.

    Third, as to the choice of a rotation for the present

    data, we decided to use varimax rotations in order to

    obtain a simple structure primarily for the hazards and

    rating scales; the remaining rotational freedom was

    used for varimax rotation of the core.

    (f) Reliability check: The reliability of the preference for

    natural food scale was checked by calculating Cron-

    bachs alpha. The scale had good internal consistency

    (a :81).

    (g) Correlation with external scales: Correlations acrosspersons were calculated between the 3MPCA person

    component scores and the attitudes toward natural-

    ness.

    Results

    Analysis of the aggregated data

    A two-way PCA of the aggregated data was conducted.

    As Table 1 shows, the first of the two orthogonal

    components of the unrotated factor loadings is highly

    correlated with immediacy of effect, knowledge about risk

    by exposed persons, knowledge to science, and newness.

    This component is labeled unknown risk. The second

    component is associated with the rating scales voluntari-

    ness, harm for health, peoples worries, and probability of

    health damage. This component is labeled dread risk.

    These two components explain 89% of the variance. The

    factor scores of the 24 hazards were computed. Fig. 1 gives

    a two-dimensional plot of these hazards, using factor

    scores as coordinates. Antibiotics in meat, GMO, and food

    irradiation are located high on unknown, and relatively

    high on dread. Salmonella, botulism, and bacterial

    contamination are located high on dread, and relatively

    low on unknown. Alcohol and caffeine are located low

    on unknown, and relatively low on dread.

    Three-mode principal component analysis

    Before carrying out a 3MPCA, an estimation of the

    variance components should be conducted to indicate

    whether the data could also be analyzed reasonably well by

    means of a two-way PCA on aggregated data. For

    example, if all individuals showed roughly the same

    response patterns with respect to all rating scales and

    hazards, one could simply take the averages across

    individuals and analyze the patterns of average responses

    to all hazards with respect to all rating scales. Three-wayanalysis is indicated when the data contain a non-negligible

    three-way interaction; that is, when individuals differ in

    their risk perceptions of different hazards. Table 2 presents

    the results of the variance component estimation. The

    results show that the data averaged across individuals,

    which only represent the rating scales and hazards main

    effects, as well as their interaction, explain only 42.1%

    (11.7%+12.6%+17.8%) of the variance of the original

    data, and 57.9% of the variance is related to individual

    differences. There are two sizable two-way interaction

    terms, as well as the three-way-interaction-plus-error term,

    which cannot be ignored. Although we do not know what

    part of this term is error, it is clear that an important three-

    way interaction is present in the data in addition to two

    important two-way interactions. These results indicate that

    individuals differ in terms of their risk perceptions of

    different hazards. Therefore, in order to examine how

    individuals differ in their risk perceptions, a three-way

    interaction is clearly indicated.

    In selecting the 3MPCA model, several models with

    varying numbers of components were evaluated. For the

    present data, a P 4 (person components), Q 2 (hazard

    components), R 2 (rating scales components) solution

    seems appropriate. It accounts for 19% of the variance not

    explained by the 2-PCA model utilizing aggregated data.

    ARTICLE IN PRESS

    Table 1

    Loadings from two-way principal components analysis over nine rating

    scales averaged across individuals (unrotated solution)

    Rating scale Unknown

    risk

    Dread risk

    Voluntariness (1 voluntary) .51 .80

    Immediacy (1 immediate) .75 .32

    Knowledge of risk to those exposed

    (1 known precisely)

    .96 .16

    Scientific knowledge (1 much) .96 .14

    Control over risk (1 controllable) .62 .72

    Newness (1 old) .93 .05

    Harm for health (1 not at all) .41 .88

    Peoples worries (1 not worried) .00 .96

    Probability of health damage (1 not

    probable)

    .35 .89

    Variance (%) 46.81 42.35

    M. Siegrist et al. / Appetite 47 (2006) 324332 327

  • 7/29/2019 Risco Percebido

    5/9

    Including more components would result in models that

    would be difficult to interpret. That is not to say that such

    models would be meaningless, but only that we have

    chosen to limit the model to the smallest number of

    dimensions that describes the largest differences across

    individuals, hazards and scales. Overall, the selected model

    also appears to be reasonably stable in the split-half

    procedure.

    The component weights for the hazards (hazard mode)

    are presented in Table 3, with weights above .20 set in bold

    face. The 95% confidence intervals were obtained by a

    bootstrap procedure using the optimal target rotations

    method (Kiers, 2004) with 1000 bootstrap samples. These

    indicate the stability of the component weights. Food

    hazards scoring high on the first hazard componentwith

    highest weights for genetically modified plants and animals,

    irradiated food and pesticide residues on foodrefer to

    hazards with unobservable consequences. The second

    component was labeled familiar hazards. The foodhazards with the highest loadings on this component were

    high-sugar diet, excessive calorie intake, high-fat diet,

    alcohol and caffeine. These are food hazards people are

    familiar with.

    The component weights for the rating scales (attribute

    mode) are depicted in Table 4, with weights above .30 set in

    bold face. Similar to the results obtained analyzing

    aggregated data, the first component is associated with

    voluntariness, knowledge of risk to those exposed, scientific

    knowledge, control over risk and newness. Therefore, this

    component was labeled as unknown risk. The second

    component is associated with probability of health

    damage, peoples worries, and harm for health. This

    component was labeled as dread risk. Results indicate

    that the same components can be used in describing both

    the data related to individuals and the aggregated data (the

    data across individuals).

    The chosen 3MPCA model also involved four stable

    person components. This suggests that there are at least

    these four dimensions along which individuals differ

    regarding risk perception of food hazards. Using orthogo-

    nal rotations only (thus leaving the component scores

    uncorrelated) results in a core array that is fairly, but not

    ideally, simple. In the present case, where we have as many

    person components as there are combinations of hazards

    ARTICLE IN PRESS

    Fig. 1. Location of the food hazards within the two-component space.

    Table 2

    Estimated variance components and variance percentages

    Effect SS Percent

    Individuals 6652.51 3.5

    Rating scales 22 163.43 11.7

    Hazards 23 898.85 12.6

    Individuals by rating scales 24 546.46 12.9

    Individuals by hazards 14 285.26 7.5

    Rating scales by hazards 33 771.17 17.8

    I ndi vidu als by r atin g scal es b y haz ards 64 8 64. 94 34.1

    Total 19 0182.62 100

    Note. SS Sum of squares.

    M. Siegrist et al. / Appetite 47 (2006) 324332328

  • 7/29/2019 Risco Percebido

    6/9

    and rating scales components, namely 4, we can find an

    oblique rotation of the person component scores such that

    the core array becomes diagonal. The oblique rotation,

    then, yields person components that are no longer

    uncorrelated, but that have a very simple, one-to-one,

    relation to the hazard and rating scale components, via the

    core. The core resulting from this oblique rotation is given

    in Table 5. The interpretation of each person component

    now is given directly by the combination of hazard and

    rating scale components associated with it, while the

    associated value in the core indicates the importance of

    the component: The highest core entries indicate the largest

    individual differences. It can be seen, however, that

    differences between the core elements are not big. As for

    the interpretation, person component 1 distinguishes

    individuals in terms of their perception of unobservable

    hazards as unknown risk. Individuals scoring high on the

    first person component perceive unobservable hazards as

    highly unknowable, while people with low scores on this

    person component perceive unobservable hazards as quite

    knowable. Individuals scoring high on the second person

    component perceive familiar hazards as highly unknow-

    able, whereas people with low scores on this component

    perceive familiar hazards as quite knowable. The third and

    fourth components are related to the dreadfulness of the

    unobservable hazards. Persons scoring high on the third

    ARTICLE IN PRESS

    Table 3

    Hazard component weights

    Hazard Unobservable hazards 95% Confidence intervals Familiar hazards 95% Confidence intervals

    Pesticide residues in food 0.29 0.21 0.32 0.05 0.08 0.03

    Genetically modified plants 0.38 0.33 0.40 0.11 0.13 0.06

    Genetically modified animals 0.37 0.32 0.39 0.11 0.13 0.06

    Natural toxicants in foods 0.24 0.17 0.27 0.02 0.01 0.09

    BSE (mad cow disease) 0.17 0.08 0.22 0.01 0.04 0.09

    Foods with bacterial contamination 0.14 0.03 0.23 0.05 0.04 0.16

    Food additives 0.25 0.19 0.28 0.10 0.06 0.15

    Salmonella 0.10 0.03 0.20 0.07 0.02 0.19

    Irradiated food 0.34 0.29 0.35 0.05 0.07 0.01

    Food colorings 0.25 0.17 0.30 0.11 0.05 0.18

    Artificial sweeteners 0.13 0.07 0.21 0.20 0.13 0.26

    Caffeine 0.04 0.06 0.01 0.35 0.31 0.37

    Alcohol 0.07 0.10 0.01 0.34 0.28 0.36

    Excessive calorie intake 0.06 0.09 0.01 0.38 0.33 0.40

    High-sugar diet 0.05 0.07 0.00 0.39 0.34 0.40

    High-fat diet 0.05 0.07 0.01 0.37 0.32 0.39

    Antibiotics in meat 0.26 0.18 0.30 0.01 0.04 0.06

    Preservatives 0.25 0.19 0.29 0.11 0.07 0.17

    Botulism 0.10 0.01 0.19 0.08 0.01 0.19

    Nitrates (e.g., in salad) 0.26 0.21 0.28 0.00 0.03 0.05

    High-salt diet 0.05 0.03 0.09 0.28 0.25 0.30

    Diet products 0.06 0.01 0.15 0.24 0.16 0.27

    Convenience food 0.11 0.06 0.19 0.22 0.16 0.26

    Vitamin enriched foods 0.15 0.08 0.23 0.17 0.10 0.23

    Note. Weights larger than .20 set in bold face. Confidence intervals obtained by bootstrap procedure.

    Table 4

    Rating scale component weights

    Dimension Unknown

    Risk

    95% Confidence intervals Dread risk 95% Confidence intervals

    Voluntariness (1 voluntary) 0.42 0.30 0.48 0.15 0.08 0.23

    Immediacy (1 immediate) 0.08 0.06 0.20 0.23 0.35 0.08

    Knowledge of risk to those exposed

    (1 known precisely)

    0.44 0.35 0.49 0.15 0.23 0.07

    Scientific knowledge (1 much) 0.50 0.31 0.63 0.01 0.14 0.13

    Control over risk (1 controllable) 0.47 0.34 0.57 0.05 0.05 0.14

    Newness (1 old) 0.39 0.27 0.46 0.02 0.13 0.08

    Harm for health (1 not at all) 0.01 0.05 0.05 0.47 0.37 0.51

    Peoples worries (1 not probable) 0.07 0.01 0.13 0.54 0.47 0.58

    Probability of health damage (1 not

    probable)

    0.04 0.08 0.00 0.62 0.56 0.65

    Note. Weights above .30 set in bold face. Confidence intervals obtained by bootstrap procedure.

    M. Siegrist et al. / Appetite 47 (2006) 324332 329

  • 7/29/2019 Risco Percebido

    7/9

    component perceive unobservable hazards as more dread-

    ful than persons scoring low on this component. Finally,

    persons scoring high on the fourth component perceive

    familiar hazards as more dreadful, whereas persons with

    low values on this component perceive unfamiliar hazards

    as more dreadful. It can be seen that the two person

    components associated with unknown risk are mildly

    correlated (r :36), just as with the two components

    related to dread risk (r :

    48), while these components

    mutually are hardly correlated (rp:11).

    The person component scores were correlated with the

    attitude preference for natural foods. A large correlation

    was observed for preference for natural foods and person

    component 3 (r :45, N 399, 95%CI: [.37, .52]).

    Individuals, who show a preference for natural foods tend

    to perceive unobservable hazards as more dreadful than

    individuals who do not show a preference for natural

    foods. The food attitude scale was also rather strongly

    correlated with person component 4 (r :25, N 399,

    95%CI: [.16, .34]). This means that the more preference for

    natural food is expressed, the more that familiar hazardsare perceived as dreadful risks. Finally, a small but

    statistically significant correlation between preference for

    natural foods and person component 1 was observed

    (r :14, N 399, 95%CI: [.04, .23]). This result suggests

    that there is a small tendency for people with a preference

    for natural foods to perceive unobservable hazards as more

    unknown risks than people not having a preference for

    natural foods. The correlation between person component

    2 and preference for natural foods was very small and not

    significant (r :05, N 399, 95%CI: [.05, .15]).

    Discussion

    The psychometric paradigm has been used to address the

    question of why people perceive various hazards differ-

    ently. In this paradigm, participants are asked to evaluate a

    list of hazards with respect to a number of rating scales.

    The big advantage of this approach is that perceptions of

    different hazards can be compared with each other. Several

    studies have examined the perceived risks associated with

    potential food hazards (Fife-Schaw & Rowe, 1996, 2000;

    Kirk et al., 2002; Sparks & Shepherd, 1994). One weakness

    of these past studies was the neglect of individual

    differences.

    The present study employed an adaptation of the ratings

    scales used by Fischhoff et al. (1978). The set of hazards

    that was used is comparable to the hazards included in the

    study by Sparks and Shepherd (1994). Initially, aggregated

    data were analyzed as in most psychometric studies. The

    PCA analysis of the rating scales revealed two factors that

    were labeled as unknown risk and as dread risk. These

    components are similar to those elicited by other research-

    ers utilizing the psychometric paradigm to examine food

    hazards (Fife-Schaw & Rowe, 2000; Kirk et al., 2002;

    Sparks & Shepherd, 1994). In the present study, the

    location of the food hazards within the two-component

    space is very similar to the study by Sparks and Shepherd

    (1994). This similarity is remarkable, given that the two

    studies differ in a number of aspects. Data collection

    occurred more than 10 years apart, and the two studies

    used different rating scales. Results of the present study

    suggest, therefore, that on an aggregated level the patterns

    produced by the psychometric paradigm are very stable.

    Results of the 3MPCA indicate that at least two

    components can be used sensibly for explaining individualdifferences in the perception of hazards. The two compo-

    nents were labeled unobservable hazards, and familiar

    hazards. Food hazards such as genetically modified plants

    or animals, irradiated food, and pesticide residues on food

    loaded highly on the component unobservable hazards,

    and hazards like alcohol, excessive calorie intake, and high-

    sugar diet loaded low on this component. Hazards like

    high-sugar diet, and caffeine had high loadings on the

    component familiar hazards. Low loadings on this

    component were observed for hazards such as GMO and

    irradiated food. The rating scales can be described by the

    two factors unknown risk, and dread risk. Results

    suggest, therefore, that the same components that explain

    aggregated data can also be used effectively to explain

    individual differences in risk perception.

    Four person components were used to describe the

    individual differences. These four components are at odds

    with the results of the two-way PCA, which assumes that

    there are no individual differences. The results of the two-

    way PCA indicate that people generally associate unobser-

    vable hazards as unknown and dreadful risks, and familiar

    hazards as known and not dreadful risks. Results of the

    3MPCA indicate, instead, that for each combination

    of a hazards component and a rating scales component, a

    stable component can be found that describes individual

    ARTICLE IN PRESS

    Table 5

    Core array

    Unknown risk Dread risk

    Person component Unobservable hazards Familiar hazards Unobservable hazards Familiar hazards

    Component 1 60.0 0 0 0

    Component 2 0 69.5 0 0Component 3 0 0 67.6 0

    Component 4 0 0 0 61.8

    M. Siegrist et al. / Appetite 47 (2006) 324332330

  • 7/29/2019 Risco Percebido

    8/9

    differences in the extent to which the associated type of

    hazards are perceived as unknown risks or dreadful risks.

    The person component scores were correlated with the

    external variable preference for natural foods. Results

    suggest that individuals who show a preference for natural

    foods perceive unobservable food hazards as more dreadful

    than individuals who do not show a preference for naturalfoods. Past research suggests that people who are

    concerned with the naturalness of food are more likely to

    buy organic food (Lockie, Lyons, Lawrence, & Grice,

    2004). Results of the present research suggest that attitudes

    toward natural food also shape risk perception of food

    hazards. In other words, people for whom naturalness of

    food is important perceive food risks differently from

    people for whom naturalness of food is not important.

    The present study examined the perception of food

    hazards not included in other studies. Newer risks, such as

    convenience food or functional foods (i.e., vitamin-

    enriched food), are viewed as rather unknown risks with

    low dreadfulness. Perceived risk seems not to be a problem

    for these new foods.

    Products or food technologies with tangible benefits for

    the consumer are viewed as less dreadful (e.g., artificial

    sweetener, convenience food) than products or food

    technologies without obvious consumer benefits (e.g.,

    GM foods, food irradiation). It might be tempting to

    assume that consumers accept new food technologies when

    products with desirable benefits are created. It should be

    noted, however, that novel foods with clear benefits might

    not be appealing to all consumers. Results of the present

    research suggest that attitude preference for natural foods

    shaped perception of food hazards. Therefore, attitudesmay also influence acceptance of novel foods with clear

    benefits for consumers.

    Past research suggests that psychometric studies based

    on aggregated data do no fully explain individual risk

    perception (Siegrist et al., 2005; Willis et al., 2005).

    Analysis of variance results show that individual differ-

    ences are responsible for a substantial part of the variance.

    About 58% of the variance is associated with individual

    differences or with measurement errors. Results of the

    present study are in line with results of earlier studies,

    therefore, in suggesting that laypeoples risk perception

    cannot be explained by a model based on aggregated data

    (Siegrist et al., 2005). Results of the analysis of variance

    strongly suggest that techniques for analyzing three-way

    data sets should be utilized in order to get a complete

    description of the data.

    The present study focusing on food hazards yielded very

    similar results to the study by Siegrist et al. (2005) in which

    risk perception of a very heterogeneous set of hazards was

    examined. In the Siegrist et al. (2005) study the model

    explained 13% of the variance, whereas in the present

    study the proposed model explains 19% of the variance. In

    the present study, participants responded to more specific

    hazards, and this resulted in a larger part of the variance

    being explained by the model. To some, an explained

    variance of 19% may still seem low. However, the results

    suggest that the present 3MPCA solution gives a usefully

    interpretable, parsimonious, and stable description of

    individual differences in risk perception. It should also be

    emphasized that there were 216 hazard and rating scale

    combinations on which individuals could differ. That only

    four components account for as much as 19% of thevariance is therefore quite a surprising result.

    Some limitations of the present study should be

    addressed. The response rate was 37.5%, even though

    households were contacted three times. It could be, that

    better-educated people were more likely to fill out the

    questionnaire. In view of the fact that participants were

    being asked to respond to more than 250 items, the

    response rate seems acceptable. Furthermore, in most

    studies utilizing the psychometric method for studying risk

    perception, convenience samples were examined. The

    sample utilized in the present study is clearly less biased

    than a pure convenience sample.

    The psychometric method requires that the attributes

    that might be important for risk perception be explicitly

    stated. For measuring food hazards, we adapted rating

    scales that were used to measure other hazards in earlier

    studies (Fischhoff et al., 1978). It is possible, however, that

    the attributes we included were not all relevant to the

    perception of food hazards. Perceived naturalness is an

    important factor in the domain of foods (Rozin, 2005), for

    example, that was not included as an attribute in the

    present study. Future studies should include additional

    attributes that might be important for lay peoples risk

    perception.

    Results of the present study strongly support the notionthat three-way-component methods should be used for a

    better understanding of lay peoples risk perception. In

    future studies, the food hazards should be described in

    greater detail. This might result in a higher proportion of

    variance explained by the model. In the present research,

    preference for natural food was used as an external

    variable. Future studies should test other factors that

    might explain individual differences in risk perception.

    References

    Bronfman, N. C., & Cifuentes, L. A. (2003). Risk perception in adeveloping country: The case of Chile. Risk Analysis, 23, 12711285.

    Fife-Schaw, C., & Rowe, G. (1996). Public perceptions of everyday food

    hazards: A psychometric study. Risk Analysis, 16, 487500.

    Fife-Schaw, C., & Rowe, G. (2000). Extending the application of the

    psychometric approach for assessing public perceptions of food risks:

    Some methodological considerations. Journal of Risk Research, 3,

    167179.

    Fischhoff, B., Slovic, P., Lichtenstein, S., Read, S., & Combs, B. (1978).

    How safe is safe enough? A psychometric study of attitudes towards

    technological risks and benefits. Policy Sciences, 9, 127152.

    Hansen, J., Holm, L., Frewer, L., Robinson, P., & Sandoe, P. (2003).

    Beyond the knowledge deficit: Recent research into lay and expert

    attitudes to food risks. Appetite, 41(2), 111121.

    Kiers, H. A. L. (2004). Bootstrap confidence intervals for three-way

    methods. Journal of Chemometrics, 18, 2236.

    ARTICLE IN PRESS

    M. Siegrist et al. / Appetite 47 (2006) 324332 331

  • 7/29/2019 Risco Percebido

    9/9

    Kiers, H. A. L., & Van Mechelen, I. (2001). Three-way component

    analysis: Principles and illustrative application. Psychological Methods,

    6, 84110.

    Kirk, S. F. L., Greenwood, D., Cade, J. E., & Pearman, A. D. (2002).

    Public perception of a range of potential food risks in the United

    Kingdom. Appetite, 38, 189197.

    Kroonenberg, P. M., & De Leeuw, J. (1980). Principal component analysis

    of three-mode data by means of alternating least squares algorithms.Psychometrika, 45, 6997.

    Lockie, S., Lyons, K., Lawrence, G., & Grice, J. (2004). Choosing

    organics: A path analysis of factors underlying the selection of organic

    food among Australian consumers. Appetite, 43, 135146.

    Magnusson, M. K., Arvola, A., Hursti, UKK., Aberg, L., & Sjoden, P. O.

    (2003). Choice of organic foods is related to perceived consequences

    for human health and to environmentally friendly behaviour. Appetite,

    40(2), 109117.

    Marris, C., Langford, I., Saunderson, T., & ORiordan, T. (1997).

    Exploring the Psychometric Paradigm: Comparisons between

    aggregate and individual analysis. Risk Analysis, 17, 303312.

    Rozin, P. (2005). The meaning of natural. Psychological Science, 16,

    652658.

    Rozin, P., Spranca, M., Krieger, Z., Neuhaus, R., Surillo, D., Swerdlin,

    A., et al. (2004). Preference for natural: Instrumental and ideational/moral motivations, and the contrast between foods and medicines.

    Appetite, 43, 147154.

    Siegrist, M., Keller, C., & Kiers, H. A. L. (2005). A new look at the

    psychometric paradigm of perception of hazards. Risk Analysis, 25,

    209220.

    Sjo berg, L. (1996). A discussion of the limitations of the psychometric and

    cultural theory approaches to risk perception. Radiation Protection

    Dosimetry, 68, 219225.

    Slovic, P. (1987). Perception of risk. Science, 236, 280285.

    Sparks, P., & Shepherd, R. (1994). Public perceptions of the potential

    hazards associated with food production and food consumption: An

    empirical study. Risk Analysis, 14, 799806.

    Tenbu lt, P., de Vries, N. K., Dreezens, E., & Martijn, C. (2005). Perceived

    naturalness and acceptance of genetically modified food. Appetite, 45,

    4750.

    Timmerman, M. E., & Kiers, H. A. L. (2000). Three mode principalcomponent analysis: Indicating the numbers of components and

    sensitivity to local optima. British Journal of Mathematical and

    Statistical Psychology, 53, 116.

    Tucker, L. R. (1966). Some mathematical notes on three-mode factor

    analysis. Psychometrika, 31, 279311.

    van Kleef, E., van Trijp, H. C. M., & Luning, P. (2005). Functional foods:

    Health claim-food product compatibility and the impact of health

    claim framing on consumer evaluation. Appetite, 44, 299308.

    Van Mechelen, I., & Kiers, H. A. L. (1999). Individual differences in

    anxiety responses to stressful situations: A three-mode component

    analysis model. European Journal of Personality, 13, 409428.

    Verbeke, W. (2005). Consumer acceptance of functional foods: Socio-

    demographic, cognitive and attitudinal determinants. Food Quality and

    Preference, 16(1), 4557.

    Verlegh, P. W. J., & Candel, M. J. J. M. (1999). The consumption ofconvenience foods: Reference groups and eating situations. Food

    Quality and Preference, 10(6), 457464.

    Vlek, C., & Stallen, P.-J. (1981). Judging risks and benefits in the small and

    in the large. Organizational Behavior and Human Decision Processes,

    28, 235271.

    Willis, H. H., DeKay, M. L., Fischhoff, B., & Morgan, M. G. (2005).

    Aggregate, disaggregate, and hybrid analyses of ecological risk

    perceptions. Risk Analysis, 25, 405428.

    ARTICLE IN PRESS

    M. Siegrist et al. / Appetite 47 (2006) 324332332