An Open Mind Wants More: Opinion Strength and the Desire for Genetically Modified Food Labeling Policy
Radas, Sonja, Teisl, Mario F., Roe, Brian, The Journal of Consumer Affairs
Two opposing viewpoints exist in the literature; some suggest consumers are unconcerned and do not desire any genetically modified labeling, while others indicate the opposite. The mixed results may be because consumers make finer distinctions than surveys have called for, and have evaluation schemes sensitive to information about the benefits and risks associated with genetically modified foods. We find consumers are quite nuanced in their preferences for genetically modified labeling policy. Unexpectedly, consumers with less-defined views desire mandatory labeling of the most stringent type, while consumers with stronger viewpoints (either pro- or con-genetically modified) are more relaxed in their labeling requirements.
There are two opposing viewpoints regarding consumers' acceptance of genetically modified (GM) foods and their desire for the labeling of these foods. Industry leaders believe consumers accept these foods because the public shows a willingness to consume them. For example, most milk in the United States is produced with the use of bST hormone, even though bST-free milk is available, clearly labeled, and advertised. In fact, except for recent limited gains, initial sales for bST-free milk were so weak it almost disappeared from the market (Webb 2006). In addition, some national surveys indicate that consumer concerns toward GM foods are low and few individuals desire any GM labeling (IFIC 2007). In contrast, as indicated by Noussair, Robin, and Ruffieux (2004), most of the academic literature indicates people are highly concerned about the GM technology (e.g., Huffman et al. 2002; Loureiro and Bugbee 2005), are willing to pay to avoid GM foods (e.g., McCluskey et al. 2003), and would like to see GM foods labeled (e.g., Teisl et al. 2003a).
One problem is that many GM labeling studies (and potentially some current labeling policies) approach the issue as one where the consumer's sole desire for information about GM foods is whether they are, in fact, genetically modified (Teisl and Caswell 2002). This approach may work well for consumers who have lexicographic preferences where the process of GM production must first be resolved before the consumer considers any other quality attributes (Kaye-Blake, Bicknell, and Saunders 2005). However, because the use of biotechnology in food production can have multidimensional effects on product quality (Caswell 2000), consumers who want to know about some or all of the changes in product attributes may find that such a labeling program provides information that is inadequate, irrelevant, or that impedes their decision making (Roe et al. 2001).
Another problem is that many studies often refer to the GM technology in imprecise terms, whereas consumers appear to be capable of making finer distinctions; hence, it is hard to interpret the attitude levels being reported (Fischhoff and Fischhoff 2001). For example, early willingness-to-pay studies commonly assumed that the genetic modification only provided benefits to consumers by lowering prices; only recently have studies (e.g., O'Connor et al. 2005; Onyango and Govindasamy 2005; Hossain and Onyango 2004) looked at situations where individuals may derive nonprice benefits (e.g., improved nutritional characteristics). In turn, it is not surprising that survey respondents would respond negatively to GM content because new technologies are often viewed as having long-term risks. Indeed, when studies include a GM-related benefit, consumers are often willing to buy these foods (e.g., Boccaletti and Moro 2000; Verdurme, Gellynck, and Viaene 2001; Teisl et al. 2003b).
Because consumer acceptance of GM foods is linked to the perceived risks and benefits of these foods (Boccaletti and Moro 2000; Chen and Li 2007; Curtis and Moeltner 2006; Lusk and Coble 2005; Moon and Balasubramanian 2004; Rosati and Saba 2000; Subrahmanyan and Cheng 2000) and because consumers are heterogeneous in how they weigh these risks and benefits (Kaye-Blake, Bicknell, and Saunders 2005), recent authors have focused on segmenting consumers by how they evaluate GM foods (e.g., Baker and Burnham 2001; Ganiere, Chern, and Hahn 2004; Jan, Fu, and Huang 2007; Kontoleon 2003; O'Connor et al. 2005; Rigby and Burto 2005; Roosen, Thiele, and Hansen 2005; Verdurme, Gellynck, and Viaene 2001; Vermeulen 2004). Although Verdurme and Viaene (2003) segment consumers to examine differences in demand for information about GM foods, to our knowledge no one has examined whether consumers' views on GM labeling policy differs across segments.
Our objective here is to identify if consumers differ in their risk/benefit evaluation of GM foods and how these differences may translate into different preferences for GM labeling policy.
Examining this issue is important both in terms of policy and research. First, with respect to policy, consumers' attitudes toward GM foods appear to be quite sensitive to information about the potential benefits and risks associated with these foods (Brown and Qin 2005; Huffman et al. 2003c, 2007; Lusk et al. 2004), especially when they are uniformed (Huffman et al. 2007); a condition illustrated across countries (e.g., United States: Hallman and Hebden 2005; James 2004; Shanahan, Scheufele, and Lee 2001; Italy: Boccaletti and Moro 2000; Hungary: Banati and Lakner 2006; China: Li et al. 2002; Belgium: Verdurme, Gellynck, and Viaene 2001; Greece: Arvanitoyannis and Krystallis 2005). Because consumer attitudes toward GM foods and the demand for information about these foods are related to how informed the consumer is (Costa-Font and Mossialos 2005), it is likely that preferences for GM labeling are linked to attitudes toward GM foods. Documented consumer heterogeneity in these attitudes suggests that a similar heterogeneity exists in preferences for GM labeling policy. However, because it is difficult for labeling policies to differ across consumers (i.e., labeling policy is practically restricted to "one size fits all"), differences in labeling preferences could lead to conflicts across consumer groups.
In addition to policy concerns, consumer heterogeneity could provide two reasons why the literature is mixed in terms of preferences for GM foods and GM labeling policy. First, as mentioned, consumers appear to be quite sensitive to information about GM foods and possibly to the framing of survey questions (Cormick 2005). For example, many polls and studies are similar by only asking simple yes/no questions about labeling but differ in how they frame the question (Do you want GM foods labeled?; Do you want mandatory labeling of GM foods?; Do you support U.S. Food and Drug Administration's [FDA]s policy of voluntary labeling of GM foods?). It is unclear from the literature whether respondents who are heterogeneous in their preferences toward GM labeling view these questions as asking different things. Second, mixed views on labeling policy could be explained if consumers who are heterogeneous in their preferences for labeling are also likely to differ in how likely they are to respond to a survey. For example, after reviewing twenty-five studies on GM food preferences, Lusk et al. (2005) indicate that a major factor explaining differences in results is the characteristics of the consumer sample.
For all the above reasons, paraphrasing Rigby and Burton (2005), knowing the "average" preference for labeling is less important than understanding how preferences differ across consumer segments and the relative sizes of these segments.
During the summer 2002, we administered a mail survey to a nationally representative sample of 5,462 U.S. residents and an additional oversample of Maine (710 individuals) residents. The samples were purchased from a frame maintained by InfoUSA. The InfoUSA database contains information about 250 million U.S. residents (more than 95 percent of U.S. households). The address information is continually updated using the U.S. Postal Service's National Change of Address system, allowing InfoUSA's address list to maintain a 93 percent accuracy rate.
The survey was administered with multiple mailings and with an incentive paid for returned completed surveys (incentives were experimentally manipulated and consisted of either a $1 bill, a $2 bill, a $2 value phone calling card, or a $5 phone calling card; for further information about the incentive scheme, see Teisl, Roe, and Vayda 2006). The oversample of Maine residents was added to provide representative results for Maine state policy makers (the Maine Agriculture and Forest Experiment Station provided some of the funding for this research).
In total, 375 Maine residents and 2,012 U.S. (non-Maine) residents responded to the survey for a response rate of 53 and 37 percent, respectively. The overall response rate of 39 percent is marginal, suggesting that individual survey results may not be a valid representation Of the knowledge, practices, and attitudes of the U.S. adult population. However, our purpose here is not to extrapolate our survey results to the aggregate population but to examine differences in labeling preferences across different types of consumers.
Our survey respondents are slightly older, more educated, and have higher incomes compared to the characteristics of the U.S. adult population (Table 1). Although our sample is more likely to be white, the stated differences in race may be reflective of a true underlying difference and/or may reflect differences in the way the race questions are asked across the two surveys. Specifically, our survey only allows respondents to choose one race from a list of five racial categories, while the U.S. Census allows respondents to choose multiple races from a list of fourteen racial categories. These differences in response categories and instructions could lead to differences in reported racial composition.
The survey instrument consists of questions used to elicit respondents' perceptions of various food technologies, knowledge of the prevalence of GM foods, perceptions of potential benefits and risks of GM foods, reactions to alternative GM labeling programs, and willingness to pay for or avoid GM foods. The content and wording of questions is based upon an analysis of issues raised in the labeling or consumer perception literature (e.g., Boccaletti and Moro 2000; Hallman and Metcalfe 1994; Hoban 1999; Huffman 2003a; Roe et al. 2001; Rousu et al. 2003; Teisl 2003), state and federal policy needs, and previous focus group research (Teisl et al. 2002). Further, they are based upon conceptualizations of consumer reactions to labeling information as presented in Teisl, Bockstael, and Levy (2001) and Teisl and Roe (1998).
As highlighted in the introduction, a reasonable condition for acquiring insight into consumers' attitudes toward the labeling of GM foods is to first gain an understanding of how consumers evaluate the risks and benefits associated with these foods. We then explore how heterogeneous consumers are in their risk/benefits evaluations and segment them into more homogeneous segments. After segmenting consumers and profiling the segments, we investigate their preferences for alternative GM labeling policies.
To gain an understanding of how consumers view the potential risks and benefits of GM foods, we provided respondents with a list of sixteen potential benefits and sixteen potential risks of GM foods and asked them to rate each one on importance. Ratings were coded on a Likert scale where 1 = not at all important, 3 = somewhat important, and 5 = very important. The potential benefits generally accrue to the consumer (e.g., lower food prices), while others accrue to the food producer (e.g., increased disease resistance in crops). The scope of the risks is broader; some primarily concern consumer health risks (e.g., unknown toxins produced), others concern producer risks (e.g., spread of disease resistance to weeds), while others are focused on environmental (e.g., risks to species diversity), social (e.g., control of agriculture by biotechnology companies), or ethical (e.g., ethical issues with genetic modification) problems.
We use factor analysis on the above benefits and risks to find the set of underlying factors influencing consumers' perceptions. Factor analysis is a data reduction technique used to investigate whether a group of variables have common underlying dimensions and can be considered to measure a common factor. Although the analysis can be used to summarize a larger number of variables into a smaller set of constructs, ultimately the analysis is not a hypothesis testing technique so it does not tell us what those constructs are (Hanley et al. 2005). In turn, the validity of naming the constructs is contingent upon researcher judgment and should be interpreted with some caution (Thompson and Daniel 1996).
For the factor analysis, we used principal components analysis followed by varimax rotation. As is typical, factors with eigenvalues less than one are dropped from further analysis as are variables with factor loadings of less than .6 as these are not considered statistically significant for interpretation purposes. To further verify the reliability of the factor analysis, we compute Cronbach's alpha on the original responses, aiming to have alphas greater than the minimum value of .70 suggested by Nunnally and Bernstein (1994).
The output from the factor analysis provides a reduced set of variables that helps explain heterogeneity in consumer evaluations of the risks and benefits of GM foods (i.e., we use the factor analysis to simplify across variables). We use cluster analysis on the above output to group respondents into homogeneous risk/benefit segments (i.e., we use the clustering analysis to simplify across observations). These two techniques are complementary (Gorman and Primavera 1983) and their combined use is familiar across a variety of disciplines (e.g., Groves, Zavala, and Correa 1987; Panda et al. 2006; Peterson 2002). In fact, this general procedure has recently been used to segment consumers by their attitudes of GM foods (e.g., Arvanitoyannis and Krystallis 2005; Christoph and Roosen 2006; Onyango et al. 2004; Verdurme and Viaene 2003).
We use a two-stage clustering approach; a hierarchical algorithm followed by k-means clustering (Punj and Stewart 1983; Singh 1990). For the hierarchical analysis, we use Ward's method with squared Euclidean distances as prescribed in the marketing research literature (Punj and Stewart 1983; Singh 1990). The resulting hierarchical tree allows us to estimate the number of segments (k); the tree graphically represents increased similarity of observations within a segment by larger linkage distances. In addition, we examine three goodness-of-fit statistics (pseudo-F, pseudo-[t.sup.2], and the cubic clustering criterion) commonly used to determine the number of clusters (SAS 2004). The procedure is as follows: we perform several cluster analyses where the number of segments (k) is set. We then examine the relative size of the three statistics to help identify the number of segments. Larger values of the pseudo-F are considered better, and cubic clustering criteria larger than two or three are preferred, those between zero and two used with caution and those less than zero indicate outliers (SAS 2004). To use the pseudo-[t.sup.2] statistic, one examines how the value of the pseudo-[t.sup.2] statistic changes as you decrease the number of clusters; the appropriate number of segments are identified when the pseudo-[t.sup.2] statistic is markedly smaller than the [t.sup.2] statistic associated with the next smaller number of segments (SAS 2004).
To aid in the interpretation of how the segments differ in terms of their evaluations of the risks and benefits of GM foods, we examine each of the segments in relation to their mean scores on the risk/benefit factor variables. Next, in order to better understand what types of individuals fall into the different risk/benefit segments, we use analysis of variance (ANOVA) and cross-tabulation analysis to examine differences across segments in their socioeconomic profiles, their other (non-GM) food-related concerns, and their food-related behaviors. In addition to respondent characteristics (gender, age, years of education, presence of children under eighteen present in household, respondent's food allergies, household income), we also use two variables to measure respondents' concerns with food production. The first variable is based on a question, designed to measure a respondent's general concern with food production, "How concerned are you about the way foods are produced and processed in the US?" Responses were coded on a Likert scale where 1 denotes "not at all concerned," 3 denotes "somewhat concerned," and 5 denotes "very concerned."
The second variable is based on a factor analysis of seven questions designed to measure concerns related to the use of specific food technologies (use of antibiotics, pesticides, artificial growth hormones, irradiation, artificial colors/flavors, pasteurization, and preservatives). Specifically, in the questionnaire we provided a list of eight technologies (the seven listed above with the addition of "use of genetically modified ingredients"), we then asked respondents to "Review the list and rate how concerned you are with each item." Again, responses were coded on a Likert scale where 1 denotes "not at all concerned," 3 denotes "somewhat concerned," and 5 denotes "very concerned." To investigate whether responses to these latter concerns are correlated, we use factor analysis as a data reduction tool using the same general procedure as presented earlier. If concerns about specific food production technologies are highly correlated, then we can consider them as measuring a smaller set of constructs, which expresses their concerns with food technologies.
We finish our discussion of the segment profiles by examining each segment's food-related behaviors. Here we asked two frequency ("How often do you buy organic foods?," "How often do you read the nutrition labels on the foods you purchase?") and three "activity" ("Do you grow your own vegetables?," "Do you regularly shop at a farmers' market or health food store?," "Do you adhere to a vegetarian diet?") questions. Responses for both frequency questions are coded on a Likert scale where 1 denotes "never" and 5 denotes "always." Responses to the activity questions are binary ("yes," "no").
After describing the consumer segments and their profiles, we use ANOVA and cross-tabulation analysis to explore consumer segments' awareness, experiences, and attitudes toward GM foods and the structure of GM food labeling policies. To measure their awareness and experiences, we asked respondents "Have you ever heard of food being genetically engineered or genetically modified?" (responses are "yes" or "no") and "Have you ever seen a label indicating that a product is 'GMO-free' or 'does not contain genetically modified ingredients'?" (possible responses are "yes," "no," or "don't know"). We measure segments' attitudes toward the following components of GM labeling policy: (1) whether or not GM foods should be labeled, (2) whether the labeling policy should be mandatory or voluntary, (3) whether the GM foods or the non-GM foods should carry the label, (4) which organization should be in charge of overseeing a GM labeling program, and (5) what pieces of information should be included on a GM label.
We had a series of questions to determine respondent preferences for GM labeling policy. First, to determine whether or not a respondent desired a GM labeling policy we asked them the following binary ("yes" or "no") question: "Would you like to see labels on food indicating whether or not the product contains genetically modified ingredients?" Response to the above question gives us some baseline information of whether or not a respondent desires GM labeling; however, we are also interested in how the GM labeling policy should be structured (especially since different structures can lead to different market impacts). In turn, if an individual answered yes to the above question, we then presented him or her with the following information and question:
There are several ways to implement a food labeling program for genetically modified foods.
A mandatory approach would require all food producers to test whether their product contains genetically modified ingredients. Once tested, the program could require either:
* all foods to display whether or not they contain genetically modified ingredients
* only foods containing genetically modified ingredients to display a label
* only foods not containing genetically modified ingredients to display a label
A voluntary approach would allow food producers to voluntarily test whether their product contains genetically modified ingredients. Once tested, the program would allow:
* only foods not containing genetically modified ingredients to display a label
How do you think a testing and labeling, program should be implemented in the United States?
1. Testing is mandatory and all foods must display a label
2. Testing is mandatory and only foods containing genetically modified ingredients display a label
3. Testing is mandatory and only foods not containing genetically modified ingredients display a label
4. Testing is voluntary and only foods not containing genetically modified ingredients display a label
5. Testing and labeling are unnecessary.
The above question allows respondents to tell us how the program should be structured and also allows some respondents to change their mind about labeling (i.e., individuals who at first stated that they desired labeling could now decide it is unnecessary). To aid in the analysis, responses to the above question were coded as five separate binary variables.
Next we consider the choice of certifying agency. For those respondents who still consider GM labeling as necessary, we provided a list of fourteen organizations and asked respondents "Which organization would you prefer to oversee a labeling program for genetically modified foods?" The list of organizations included U.S. Department of Agriculture (USDA), FDA, U.S. Environmental Protection Agency (EPA), Greenpeace, Natural Resources Defense Council, Organic Consumer Association, Identity Preservation Program, Cert ID-Genetic ID Inc., Union of Concerned Scientists, Consumer's Union, National Institute of Health (NIH), American Medical Association (AMA), American Heart Association (AHA), and American Cancer Society (ACS). To simplify the comparison among the three consumer segments, we divided the above fourteen organizations into four broad categories: governmental food agencies (containing USDA, FDA, and EPA), environmental organizations (Greenpeace, Natural Resources Defense Council, Organic Consumer Association), scientific organizations (Identity Preservation Program, Cert ID-Genetic ID Inc., Union of Concerned Scientists, Consumer's Union), and health associations (NIH, AMA, AHA, and ACS), and test for differences in preferences for each broad group across segments. We also test for differences across segments for each governmental organization (small cell sizes preclude these more detailed tests for each of the other organizations in the list).
We finish our examination of labeling policies by analyzing differences in preferences, across segments, for various pieces of information that could be part of a GM label. Here we provided respondents a list of seven items of information (see Table 8 later in this article) that could be included on a GM label and then asked them to rate on a Liken scale how important each piece of information is to them (responses coded as 1 denoting "not important at all," 3 denotes "somewhat important," and 5 denotes "very important").
For all the ANOVA and cross-tabulation analyses, there are a potentially large number of follow-up tests required to determine the full set of differences across segments. A problem with performing such a large number of tests at a specific significance level is that the overall likelihood of inappropriately rejecting a null hypothesis is greater than the specified significance level (called alpha inflation). To reduce the likelihood of committing such a type I error, we used the following procedure. First we test whether a variable is significantly different across all n segments (e.g., we perform an n-way ANOVA to test for significant differences in means across segments or perform an n-way cross tabulation test for significant differences in proportions across segments). The significance for these "n-way" tests is set at the 10 percent level. If we find a significant difference across the n segments, we then test individual pairings of segments for significant differences. The significance for these latter pair-wise tests ([[alpha].sup.p]) is set by using the formula: 1 - [(1 - [[alpha].sup.p]).sup.n] =. 10. This procedure maintains the overall probability of committing a type I error to 10 percent (Hand and Taylor 1987).
The factor analysis on the benefit variables indicates that two factors (Table 2) explain respondent reactions. (1) Kaiser's overall measure of sampling adequacy is high (.95) indicating the factor model is appropriate; values greater than .80 are considered sufficiently high for analysis (SAS 2004). We call factor 1 own benefits (OB) because the variables loading highly on this factor mostly accrue directly to consumers. We call factor 2 producer benefits (PB) because these generally impact the producer. Note that at the time of the survey, most approved GM foods were "first generation" (primarily providing producer-valued attributes--Schneider and Schneider 2006); second-generation foods (those providing consumer-valued attributes) are only now beginning to be developed. Factor OB explains 50 percent of variance, while factor PB explains 9 percent of variance.
Factor analysis of risks similarly yields two factors; again, Kaiser's overall measure of sampling adequacy is high (.90). We call factor 1 health/ environmental risk (HER) as it relates to risks directly impacting the consumer through perceived deteriorations in food safety or are related to potential negative environmental impacts. We call factor 2 producer risk (PR) as it describes risk born by the producer. HER explains 63 percent of variance, while factor PR explains 11 percent.
An important component of many of the variables loading highly on HER seems to be the uncertainty of long-term impacts. The high level of concern surrounding unknown long-term impacts is a consistent theme explaining consumers' negative reactions to new food technologies. This hypothesis is consistent with Slovic (1987) who indicates a major factor impacting a consumer's evaluation of a new technology is the degree to which risks are "unknown"; that is, risks that are not observable or evident, have effects that are delayed, or are not definitively known to science (Marks 2001). For example, concerns about long-term health impacts seem to explain consumers' initial opposition to pasteurization (Huffman 2003a) and seems to be a factor in consumer acceptance of GM foods (Hoban 1997). In general, consumers trust food scientists' abilities to determine the short-term safety of new food technologies but understand the limitations scientists face in determining long-term impacts (Levy and Derby 2000). Interestingly, the level of technical knowledge about a new food technology does not seem to impact consumers' concerns; it is the lack of experience with the technology (Levy 2001).
Computation yields Cronbach's alphas of .91, .83, .92, and .92 for OB, PB, HER, and PR, respectively, indicating our analyses have a high degree of reliability. Further, all item-to-total correlations equal .66, .52, .74, and .74 for OB, PB, HER, and PR, respectively (indicating a high degree of internal consistency). Note that all factor loadings are positive (Table 2) indicating that each of the four factor scores are positively correlated to the variables originally used in their construction. In turn, although each of the four factor scores are normalized to mean zero, the direction of each score is positively correlated to the direction of the original variables. Hence, higher (lower) factor scores indicate a higher (lower) level of importance for that risk/benefit factor.
[FIGURE 1 OMITTED]
To determine the number of segments, we begin by inspecting the hierarchical tree from the cluster analysis of the risk/benefit factors (Figure 1); three relatively large linkage distances suggest three relatively homogeneous segments. The pseudo-F statistic indicates that three or four segments are appropriate (Table 3), while the pseudo-[t.sup.2] indicates either three or six possible segments. Finally, the cubic clustering criteria indicate either a two- or three-segment solution. Hence, we conclude there are three segments of respondents and consequently perform k-means segmentation, where k = 3. This number of segments is similar to those found in the literature (two segments: Mora et al. 2007; three segments: Baker and Burnham 2001; Jan, Fu, and Huang 2007; Onyango et al. 2004; Vermeulen 2004; four segments: Christoph and Roosen 2006; Ganiere, Chern, and Hahn 2006; O'Connor et al. 2005, 2006; Verdurme and Viaene 2003).
By examining the segments in relation to their mean scores on the risk/ benefit factor scores (Figure 2), we describe the segments as:
* Risk avoiders: this segment is intermediate in size (n = 677) and not interested in the potential benefits of GM foods but are very worried about health and environmental risks potentially associated with GM food.
* Risk dismissers are the smallest segment (n = 482) and rate their OBs and PBs higher than health and environmental risks; in fact, they are the least worried about these potential risks. It seems these respondents believe the technology can bring benefits at a low personal risk.
* Balanced but interested is the largest group (n = 896) and finds both benefits and risks important; unlike the other segments, these respondents are not strongly committed to any of the above points of view.
Characterization of the above segments is similar to other studies. For example, Verdurme and Viaene (2003) find four segments: "halfhearted," "green opponents," "balancers," and "enthusiasts." Their enthusiasts are similar to our risk dismissers in that they both discount health and environmental risks and have some interest in GM-related benefits. Their balancers are similar to out balanced but interested segment in that they both place importance on the benefits and risks of GM foods. Finally, the halfhearted and green opponent segments are similar to our risk avoiders in that they are both concerned about health and environmental risks and discount any GM-related benefits.
[FIGURE 2 OMITTED]
The factor analysis (Table 4), investigating whether respondent concerns for specific food technologies