Decomposing the effects of time on the social acceptability of biotechnology using age-period-cohort-country models

Abstract

The study of European attitudes toward biotechnologies underlines a situation that is relatively contrasting in Europe. However, as different effects of time can influence the social attitudes (a life-cycle effect, a generational effect, and an exogenous temporal effect potentially affecting the entire population), an appropriate methodology should be used. To this end, age-period-cohort-country models have thus been estimated based on Eurobarometer data from 1991 onward. Applied to different data subsets, these models give similar results underlining the importance of the life-cycle effects as well as the heterogeneity of the link between political affiliation and biotechnologies attitudes across the European countries.

Keywords

biotechnology genetically modified food public understanding of science risk attitudes science attitudes and perceptions

1. Introduction

The analysis of the 2010 Eurobarometer survey on science and technology (Gaskell et al., 2010, 2011) highlights a relatively contrasting situation in Europe concerning the evolution of attitudes toward biotechnologies. It is, therefore, necessary to look at the situation from a long-term perspective in order to understand the originality of our period. As noted by Alakasen and Myhr (2007), “Genetic modifications of crops have primarily been motivated from the production side, in order to increase agricultural output, rather than from a consumer demand and health perspective” (p. 491). Therefore, tackling the social acceptability of genetically modified organisms (GMO), and more generally biotechnology-based food, is of major importance (see also Dannenberg, 2009), especially to design the governance regimes in order to adequately handle uncertain and unknown effects of new technology (Kvakkestad and Vatn, 2011). On this subject, we find mixed results in the literature: does the change come from, as certain authors believe, the emergence of a population that is better educated or more informed about the different technologies (Sturgis et al., 2005) or increased confidence in the different participants in the debate and notably environmental associations (Huffman et al., 2004; Pardo et al., 2009; Priest et al., 2003; Rousselière and Rousselière, 2010)? Perhaps, the change is caused by more citizen implication in the debate, leading to less support of modern genetic applications (Knight and Barnett, 2010), or by a youth that has a greater tendency to favor the adoption of new technologies (Gaskell et al., 2011; Knight and Barnett, 2010).

Time can have different effects and can influence the attitude that Europeans have toward biotechnologies (a life-cycle effect, a generational effect, and an exogenous temporal effect potentially affecting the entire population). An appropriate methodology should be used to give this study a long-term perspective. To our knowledge, the only descriptive study on the subject that does not separate these effects is carried out by Bonny (2008). Another study specific to Australia (Marques et al., 2015) focused on period effect linked to media coverage also has to be mentioned. Building on previous works from Yang (2008) and Yang and Land (2008, 2013), we estimate original cross-classification multilevel age-period-cohort-country models. The use of Eurobarometer data from 1991 onward allows us to develop two original successive analyses, which both shed light on an aspect of the attitude toward biotechnologies: a primary analysis relative to the optimism about biotechnologies and genetic engineering, with the advantage of a greater time period, followed by a second analysis specifically concentrating on the acceptability of GMO. These models give similar results and underline the importance of life cycle while reducing the importance of the generational effect. They highlight major differences in and between countries on the acceptability of different biotechnology applications. These major differences can be explained by the heterogeneity of the link between political affiliation and biotechnology attitudes across the European countries. As discussed in the concluding section, these empirical findings have major implications for public policies.

2. On the effects of time on the social acceptability of biotechnology

Modern biotechnologies (Fiechter, 2000) are a group of techniques involving the manipulation or change to the genetic inheritance of living organisms. They have led to a radical change in the nature of new scientific research, the creation of new chemical entities, and the creation of important products and procedural innovations (Ramani, 2001). They are characterized by the interdisciplinary knowledge base (genetics, molecular biology, biochemistry, microbiology, enzymology, and immunology) and multisectoral applications (health, agrifood, agriculture, chemistry, environment, energy, etc.). The different studies carried out on the social acceptability of biotechnology all underline that the absence of perceived utility is one of the determining factors of the opposition to biotechnologies regardless of the field of application (e.g. Amin et al., 2014; Auer, 2008; Einsiedel and Medlock, 2005; Gaskell et al., 2004, 2011; Klingeman and Hall, 2006; Marris, 2001; Pardo et al., 2009). Pardo et al. (2009) underline, for example, a greater acceptance of genetically modified (GM) plants for obtaining medication or treatments to fight deadly diseases (average approval of 6.8/10), illnesses affecting children (6.4), the effects of aging (4.2), or to obtain cosmetic products (3.4). The development of ornamental horticulture biotechnologies, in other words the use of biotechnologies in an explicitly leisure activity context, is highly controversial (Alston et al., 2006; Auer, 2008; Klingeman and Hall, 2006). Amin et al. (2014) show that the perceived presence of clear benefits of GM salmon enhances the support of the Malaysian public. As part of this utility–risk dilemma under uncertainty and ignorance (Gaskell et al., 2004), which leads the cost–benefit analysis to be less applicable, as developed by Kvakkestad and Vatn (2011), the “ordinary citizen’s” confidence (Joly and Marris, 2003) toward the different parties involved in the debate is seen as a crucial variable (e.g. Barnett et al., 2007; Canavari and Nayga, 2009; Marques et al., 2015; Pardo et al., 2009; Rousselière and Rousselière, 2010).

To understand the historical evolution of the European situation, it is necessary to underline the four potential contextual effects, found in longitudinal studies, each likely to explain part of the heterogeneity of individual attitudes toward age (H1), generation (H2), period (P3), and country (P4). As the results are mixed and unpredictable a priori, the last two effects have the status of proposition, whereas the first two are hypotheses that can be fully tested. Our analysis will help specifying them.¹

The first potential effect is that of age which can be explained within the framework of the life cycle² and notably when passing from an adolescent to an adult, the accumulation of assets, or an increase in revenue (Fukuda, 2011; Schaninger and Danko, 1993). As age increases, while the utility felt is not any greater, the biotechnology is perceived as being more and more risky and notably over time: for example, as age increases, there is an evolution in people’s lives; they go from being single to a couple and having children. Concerning the use of milk produced using biotechnology techniques (with the injection of recombinant bovine somatotropin hormone), Zepeda et al. (2003) highlight a non-linear relationship between opposition and age, partly due to the birth of children and partly due to specific health problems linked to aging. Dietary practices are generally quite revealing with regard to individuals construction of identity and social relationships (see Recours and Hébel, 2007). Furthermore, the revenue effect contributes to the displacement of individual demand on to products that are not the result of agricultural biotechnologies (Huffman et al., 2004).

We can thus formulate an initial research hypothesis:

H1. Support for biotechnologies will decrease as age increases.

Even if sociologists such as Bourdieu (1990) question the notion of generation, by underlining that internal differences within an age range could be much more important than the collective interests of a generation, it may still be reasonable, however, even for these authors, to assume certain generational effects. The specific socialization of a generation, the fact of having shared similar living conditions or events (compared to other generations), can lead to personal choices or attitudes that are identical and that ultimately have structural results on societies (e.g. Inglehart, 1990). Recent works concerning generational effects on consumer habits (Fukuda, 2011; Mori and Saegusa, 2010; Recours and Hébel, 2007) have highlighted that the generations’ attitudes toward food has to be interpreted with regard to innovations that the person was subjected to at the time they entered active life.³ The succession of generations that are confronted with new technologies on an ever greater scale could lead us to assume that new generations are more likely to accept these new technologies. We can, therefore, formulate the following proposal:

H2. The effect of the cohort would correspond to the succession of generations that are ever more favorable to the use of biotechnologies.

The period effect is based on the idea that all individuals evolve in the same socio-economic and socio-historical environment. Events define period, such as different food crises, periods of shortage, and the generalization of a social behavior style (Recours and Hébel, 2007). Thus, the specific structure of the debate on new technologies and the succession of sanitary crises are aspects likely to produce effects that concern the entire European population over a given period. The social amplification of sanitary crisis or the risks of GMO by the media and consumers (DeRosier et al., 2015; Frewer et al., 2002; Stefani, 2008) can lead to hysteresis. On one hand, the results of Knight and Barnett (2010) and Marris and Rose (2010) highlight that the greater implication of citizens in the debate on modern biotechnologies leads them to be structurally and durably more wary regarding the application of this technology. On the other hand, the “bias assimilation and polarization hypothesis” (Kahan et al., 2011) may lead to a public opinion durably split on this issue. As results are mixed on this topic and unpredictable a priori, this proposition can be formulated in the following manner:

P3. The succession of periods is characterized by major effects on the social acceptability of biotechnologies.

Finally, we can ask whether there is a country effect; in other words, an institutional effect specifically concerning the manner in which the biotechnologies’ debate is structured in each European country. Gaskell et al. (2006, 2010) or Bonny (2008) underline that there is a convergence between European countries, on the general attitude toward biotechnologies with the fact that the countries that have, however, recently joined the European Union (EU) would have citizens who are, a priori, more favorable to these technologies. On the contrary, Joly and Marris (2003) highlight the specific structure of the debate within each country. As far as pharmaceutical biotechnologies are concerned, Pardo et al. (2009) underline the existence of major differences between countries, this time due to heterogeneity of perceptions regarding utility of these technologies. Concerning green issues (Clements, 2014), attitudes toward GM food are largely politically driven (Costa-Font et al., 2008). For Durant and Legge (2005), a shift from left to right on the political spectrum increases support for GM foods. Studies have also highlighted a non-linear relationship as there is a resistance not only from the green or the left-wing side but also from the conservative side. Nielsen (1997) and Nielsen et al. (2002) have pointed out these two different groups (modernist green segment and traditionalist blue segment) that based their arguments on very different values and concerns (see also Holmgreen, 2008). For Costa-Font et al. (2008), political affiliation acts as a reference point around which individuals judge new technologies based on little information. But because of the presence of different political strategies and political competition (Green-Pedersen, 2007), this reference point may be country specific (e.g. see Freire, 2008). Using International Social Survey Program (ISSP) data on the environment (2000 and 2010 surveys), Rousselière and Rousselière (2013) show a decline on the political affiliation effect on the opposition to GM crops, with a higher country effect. But the heterogeneous choice model used in this article estimates only additive effects for country and political scales. Contrary to previous works, our empirical strategy will allow us to interpret the interaction of these effects.

We can, therefore, formulate the following proposition:

P4. The country effect is characterized by the existence of high heterogeneity between different European countries. The impact of political affiliation is largely country specific.

3. Econometric strategy: A multilevel age-period-cohort model

Since the initial works carried out by Mason et al. (1973), different econometric strategies have been used to find a way out of the impasse resulting from the perfect colinearity of the three periods that are age, period, and cohort.⁴ The solution proposed by these authors was to set an additional constraint on the parameters to be estimated.⁵ The strategy of Chen et al. (2001) is to choose cohort and age group conditions that are sufficiently large so as to eliminate identification problems. Recours and Hébel (2007) and Hall et al. (2007) use the strategy of Rodgers (1982), who advocated the use of a priori information and parsimony about cohorts or the time periods to help identify the model. The strategy consists of successively testing the validity of different submodels in order to replace one of the variables (generation or period) by variables likely to better explain it (e.g. access to education for generational effect or macroeconomic variables for period effect). Mori and Saegusa (2010) and Fukuda (2006, 2011) use Nakamura’s Bayesian model (developed in Nakamura, 1986), which assume a “gradual change between successive parameters which covers the entire range of all three factors […] in lieu of any single equality constraint arbitrarily chosen” (2010: 50–51).

Taking into account the interest and limits of these various strategies (Chauvel and Schröder, 2015; Luo, 2013), we choose to follow the proposition of the hierarchical age-period-cohort (HAPC; model with random effects and cross-classification) proposed by Yang (2008) and Yang and Land (2008, 2013). In effect, the specific nature of the pseudo panel data that we use creates an additional difficulty (Verbeek, 2008): we do not have longitudinal data on the same individuals but instead a succession of transversal cross-sectional surveys. However, these types of data can be perfectly interpreted as part of a multilevel model due to the hierarchical nature of data created by the sampling procedure: individuals are sampled in a country at a given period and belong to an age group.⁶ The linear nature of the relationship between age, period, and cohort is broken as we consider that the effects are not of the same type: certain effects are specific to the individual (age) and, therefore, concern the finest level, and others are specific to the context in which the individual involved (generation, period, or country) and concern a higher level. The cross-classification model enables us to not emit hypotheses on the hierarchy of these different contextual effects. As was the case with Yang (2008), the variable to be explained is of a dichotomous nature, which justifies the use of a multilevel logit.⁷ The specific nature of our data, however, leads us to propose a more general presentation likely to take the country dimension into consideration (see Bell and Jones, 2014: 341).⁸

We can, therefore, propose the following formula for the model to be estimated

y_{i, j, k, l}^{*} = α + β_{1} \cdot A G E_{i, j, k, l} + β_{2} \cdot A G E_{i, j, k, l}^{2} + \sum_{a = 1}^{A} γ_{a} \cdot Z_{a, i, j, k, l} + f_{j} + v_{k} + w_{l} + u_{i, j, k, l}

(1)

with

y_{i, j, k, l} = 1 if y_{i, j, k, l}^{*} > 0 and y_{i, j, k, l} = 0 if y_{i, j, k, l}^{*} \leq 0

where $f_{j}$ is a random effect observed in the country j with $f_{j} ~ N (0, τ_{j}^{2})$ , $v_{k}$ is a random effect observed in the cohort k with $v_{k} ~ N (0, τ_{k}^{2})$ , w_l is a random effect observed in the period l with $w_{l} ~ N (0, τ_{l}^{2})$ , $u_{i, j, k, l}$ is a random effect observed on the individual i in the cohort k living in the country j in a period l with $u_{i, j, k, l} ~ N (0, σ^{2})$ , and $Z_{a, i, j, k, l}$ is one of the a control variables observed on the individual i in the cohort k living in the country j in a period l.

We can also estimate the same model with a random slope for a variable D, which we can suspect to vary across country

y_{i, j, k, l}^{*} = α + β_{1} \cdot A G E_{i, j, k, l} + β_{2} \cdot A G E_{i, j, k, l}^{2} + (β_{3} + ς_{j}) \cdot D_{i, j, k, l} + \sum_{a = 1}^{A} γ_{a} \cdot Z_{a, i, j, k, l} + f_{j} + v_{k} + w_{l} + u_{i, j, k, l}

(2)

where $ς_{j}$ is a random effect observed in the country j with $ς_{j} ~ N (0, d_{j}^{2})$ .

Yang (2008) and Yang and Land (2008, 2013) have shown the interest of these models even in the case where the number of groups for the different levels is relatively low (from 5; in other words, a low number of periods or cohorts), this being notably due to the very different sizes of groups (known as the “unbalanced data design effect”).⁹ Note although there is an ongoing discussion on the interests and limits of these models (Bell and Jones, 2014; Chauvel and Schröder, 2015; Luo, 2013; Reither et al., 2015),¹⁰ we retained the HAPC methodology as it clearly fits with our literature review.¹¹ We acknowledge the fact that any constraint put on APC model is—ultimately—based on theoretical assumptions, and therefore, our own strategy is theory driven.

We have estimated these models via Markov Chain Monte Carlo (MCMC) Bayesian modeling (Browne, 2009). This type of estimator has been shown as being particularly good for models containing dichotomous or category-related response variables (Ng et al., 2006), cross-classification models (Browne et al., 2001), as well as for cases where the number of categories in the upper echelon is low.¹² A total of 100,000 iterations were necessary in order to obtain, as per the diagnosis results of Raftery and Lewis (1992), Markov chains that were sufficiently long to carry over sufficiently accurate parameters. Finally, estimations can be sensitive to the distribution function of prior parameters. As Cameron and Trivedi (2005) note, estimations based on large samples are not affected by the choice of different priors because of the “sample information” domination (p. 425; Bernstein–Von Mises Theorem). As our various subsets contain between 57,659 and 140,119 observations, we can be reasonably confident in our estimations, and we simply use the diffuse (or “flat”) prior parameters proposed by Browne (2009: 4–5).

The cohort, period, or country effect can been assessed through the calculation of the residual intraclass (or conditional) correlation coefficient (ICC). This coefficient gives the variance percentage in the acceptance taken into account by the inclusion of a level. The ICC observed at level k is, therefore, written as follows

I C C_{k} = \frac{τ_{k}^{2}}{σ^{2} + \sum_{j = 1}^{K} τ_{j}^{2}}

(3)

where $τ_{j}^{2}$ the variance between groups observed at a level j with $j = 1, 2, 3$ ,¹³ $τ_{k}^{2}$ the variance between the groups observed at a level k, and $σ^{2}$ the variance within the group observed on the individual level (within the case of a multilevel logit $σ^{2} = π^{2} / 3$ ).

We have

I C C_{k, j} = I C C_{k} + I C C_{j}

(4)

where $I C C_{k, j}$ corresponds to the percentage of variance taken into account by the inclusion of levels k and j.

An alternative measurement can be used, especially if we are looking to combine these effects. This is the median odds ratio (MOR; Larsen et al., 2000). In our case, this ratio measures the impact of switching from one group to another of the same level on the probability of being favorable to biotechnologies on a given level k when we compare two individuals drawn randomly from the population

M O R_{k} = O R {(k)}_{m e d i a n} = e x p (\sqrt{2 τ_{k}^{2}} \cdot ϕ^{- 1} (\frac{3}{4}))

(5)

where $τ_{k}^{2}$ the variance between the groups on level k and $ϕ^{- 1}$ the inverse of the distribution function of the normal.

We also have

M O R_{k, j} = O R {(k, j)}_{m e d i a n} = e x p (\sqrt{2 (τ_{k}^{2} + τ_{j}^{2})} \cdot ϕ^{- 1} (\frac{3}{4}))

(6)

where $M O R_{k, j}$ corresponds to the MOR for two randomly drawn individuals belonging to the same group on level k and the same group on level j.

As the total residual variance is no longer constant as in random-intercept models, “general” MOR or ICC cannot be calculated in a random-slope model. For these models (see Table 1), we only report the ICC and MOR for the reference situation.¹⁴

Table 1.

Descriptive statistics.

	All periods					Mean by period
	N	M	SE	Minimum	Maximum	1991	1993	1996	1999	2002	2005	2010
biotechpositive	140,119	0.512	0.500	0	1	0.493	0.460	0.448	0.397	0.445	0.636	0.548
encourageGMO	69,674	0.312	0.463	0	1			0.430	0.327	0.320	0.252	0.210
Age	140,119	45.156	18.082	15	99	42.758	43.092	43.403	44.017	44.940	46.566	47.348
Left	140,119	0.241	0.428	0	1	0.241	0.224	0.253	0.243	0.254	0.230	0.246
Center	140,119	0.311	0.463	0	1	0.286	0.298	0.293	0.278	0.328	0.345	0.312
Politicalposition_refusal	140,119	0.209	0.407	0	1	0.176	0.229	0.168	0.218	0.225	0.210	0.222
Right	140,119	0.238	0.426	0	1	0.296	0.249	0.286	0.261	0.194	0.214	0.220
Female	140,119	0.529	0.499	0	1	0.514	0.507	0.516	0.526	0.530	0.552	0.532
Male	140,119	0.471	0.499	0	1	0.486	0.493	0.484	0.474	0.470	0.448	0.468
Educ15	69,674	0.250	0.433	0	1			0.278	0.270	0.254	0.238	0.205
Educ16_19	69,674	0.388	0.487	0	1			0.367	0.388	0.380	0.407	0.402
Educ20+	69,674	0.260	0.439	0	1			0.243	0.234	0.263	0.263	0.302
Student	69,674	0.102	0.302	0	1			0.112	0.109	0.103	0.092	0.091
Independent	69,674	0.080	0.272	0	1			0.089	0.084	0.077	0.074	0.074
Manager	69,674	0.100	0.301	0	1			0.098	0.092	0.094	0.108	0.109
Employed	69,674	0.107	0.309	0	1			0.102	0.106	0.103	0.115	0.108
Worker	69,674	0.211	0.408	0	1			0.219	0.222	0.220	0.202	0.191
Housewife/man	69,674	0.107	0.309	0	1			0.120	0.114	0.125	0.097	0.083
Jobseeker	69,674	0.062	0.241	0	1			0.064	0.061	0.054	0.044	0.079
Pensioner	69,674	0.233	0.423	0	1			0.200	0.216	0.223	0.268	0.266

The model selection is based on the deviance information criterion (DIC; Spiegelhalter et al., 2002). DIC is the generalization of the Akaike information criterion (AIC; Akaike, 1974) for multilevel models. DIC is the addition of two terms that measure the “fit” and the “complexity” of a particular model. Thus, we have

D I C = \bar{D} + p_{D}

(7)

where $\bar{D}$ is the average deviance value and $p_{D}$ the number of parameters of the model.¹⁵

Its reduction underlines an increased performance of estimations. The DIC can be seen as an approximation of the Bayes factor, involving a particular prior on the parameter space, defined by Burnham and Anderson (2004) as “savvy.” However, as we discussed above, when the sample size is sufficiently important, different priors lead to the same results. Therefore, Jeffery’s rule of thumb can be used (Burnham and Anderson, 2004; Kass and Raftery, 1995; Spiegelhalter et al., 2002). A difference of 10 between two DICs might definitely rule out the model with the highest DIC, as it involved that the model with lowest DIC has approximately a posterior odds of 150:1 to be the true model (see Kass and Raftery, 1995: 777).

4. Presentation of data

Special Eurobarometer surveys have been the subject of in-depth theme-based studies carried out for the European commission or other European institutions and integrated into standard Eurobarometer waves. We use the data from special Eurobarometers that concern the attitudes of Europeans toward science in 1991, 1993, 1996, 1999, 2002, 2005, and 2010. Approximately 1000 people per country were questioned using a random multiphase sampling process (in geographical layers (infranational)). The surveys cover the population from the age of 15 years residing in each member state of the EU, as well as the countries linked to the EU (like Norway, Switzerland, and Turkey).¹⁶ We obtain a rich questionnaire giving rise to different uses (relationship with biotechnologies, on the European vision of science, etc.; e.g. Gaskell et al., 2004, 2011; Priest et al., 2003; Rousselière and Rousselière, 2010).

The models are applied to four successive database subsets. Thus, we test the general attitude toward biotechnologies (since 1991) across the countries (subset A1) and on the countries present over the entire period (i.e. in all the Eurobarometer waves; subset A2 equivalent to a “balanced” version of A1). The dependent variable corresponds to the response people gave to the question: “Do you think that biotechnologies and genetic engineering will have a positive effect on the way we live in the next 20 years?” Then, we test the attitude toward GMOs (since 1996) across the countries (A3) and on the countries present over the entire period (subset A4 equivalent to a “balanced” version of A3¹⁷). The dependent variable corresponds to the reply to the question: “Do you agree with the proposal: ‘we should encourage the development of GM foods’?” Table 1 reports the descriptive statistics for the various subsets.

As an explanatory variable, we use age (from 15 to 99 years) as part of a quadratic function (age and age squared)¹⁸ as we suspect non-linear effects. For the different control variables, we test all the variables that are available in each wave, but there are a very few number of them: gender, political scale (four possibilities: left, center, right, and refusal¹⁹), level of education (four possibilities: left the educational system before 16 years, between 16 and 19 years, at 20 years or after, and still in the educational system), professional activity (eight possibilities: self-employed, manager, employee, worker, stay at home, jobseeker, pensioner, and student). These last variables (levels of education and professional activity) are only available for subsets A3 and A4.

The contextual effects are determined by country variables (33 possibilities for subsets A1 and A3 but only 14 for subsets A2 and A4), generation (seven possibilities: 1934 and before, 1935–1944, 1945–1954, 1955–1964, 1965–1974, 1975–1984, and 1985 and after) and period (5–7 possibilities each corresponding to a survey wave: 1991, 1993, 1996, 1999, 2002, 2005, and 2010 for subsets A1 and A2 and 1996, 1999, 2002, 2005, and 2010 for subsets A3 and A4).

We can see that there are “high jumps” in the optimism toward biotechnologies: 49.3% of the Europeans though it will have a positive effect in 1991, a proportion that falls to 39.7% in 1999 before growing to 63.6% in 2005 and decline to 54.8% in 2010. On the contrary, the proportion of Europeans that agreed to the question as to whether or not GMO should be encouraged decreases from 43% in 1996 to 21% in 2010. The reasons behind this change can reside in the fact that the respondents are older (mean age of 47 years) or that the EU has a greater number of member countries in 2010 compared to 1996. Note also that there is a slightly change in the educational level of the population. Therefore, a multivariate methodology is required in order to control for these potential confounders.

Seven different models (M1, M2, M2a, M3, M4, M5, and M6) were estimated. The first one (M1) corresponds to a simple logit using only age as independent variable. M2 includes gender and political position effects. M2a is the same as M2 but includes, for subsets A3 and A4 only, education and activity effects. M3 corresponds to a hierarchical logit with only cohort random effects, and M4 the same as M3 but with random period effects included. M5 corresponds to the model presented in equation (1) (Section 2) and M6 to the model presented in equation (2).

5. Results

According to the DIC, model M5 has to be selected for subsets A2 and A4 and M6 for subsets A1 and A3 (see Table 2). We note a strong improvement when we add period effects (going from M3 to M4) and country effects (going from M4 to M5). Finally, adding a random slope for political position at the country level dramatically improves the model (going from M5 to M6) for subsets A1 and A3.

Table 2.

DIC estimates for different model specifications.

	A1		A2		A3		A4
	DIC	$Δ D I C (M_{i - 1} - M_{i})$	DIC	$Δ D I C (M_{i - 1} - M_{i})$	DIC	$Δ D I C (M_{i - 1} - M_{i})$	DIC	$Δ D I C (M_{i - 1} - M_{i})$
M1 (age)	192,916		152,238.6		86,121.01		73,175.71
M2 (M1 + female and political position)	191,699.5	1216.5	150,771.8	1466.72	85,467.27	653.74	72,720.22	455.49
M2a (M1 + female, political position, education, activity)	–		–		85,423.94	43.33	72,694.45	25.77
M3 (M2 or M2a + model with cohort effects)	190,346.8	1352.71	150,421.2	350.63	85,419.13	4.81	71,966.82	727.63
M4 (M3 + period effects)	187,385.8	2960.98	147,831.2	2590.06	83,561.09	1858.04	71,528.79	438.03
M5 (M4 + country effects)	183,949.1	3436.65	145,687.7	2143.49	81,384.28	2176.81	69,982	1546.79
M6 (M5 + political affiliation random effects at country level)	182,396.4	1552.72	–		79,963.21	1421.07	–

DIC: deviance information criterion.

The results of the different estimations are presented in Table 3. The first comment is that the parameters estimated on the general population or its “balanced” version are relatively similar (A1.M5 compared to A2.M5; A3.M5 compared to A4.M5). It is as if the new countries included in the EU behave on average like the old European countries. The main differences are for the age parameters between A1.M5 and A1.M6 when a random slope for political scale is added.

Table 3.

Results of the different estimations.

	A1.M5			A1.M6			A2.M5			A3.M5			A3.M6			A4.M5
	Coefficients	SE	OR	Coefficients	SE	OR	Coefficients	SE	OR	Coefficients	SE	OR	Coefficients	SE	OR	Coefficients	SE	OR
Intercept	0.237	0.223	1.267	0.668	0.063	1.950***	0.243	0.180	1.275	−0.492	0.315	0.611	−0.245	0.115	0.783**	−0.516	0.263	0.597**
Age	−0.007	0.003	0.993***	0.003	0.002	1.003**	−0.006	0.003	0.994*	−0.012	0.004	0.988***	−0.017	0.003	0.983***	−0.012	0.004	0.988***
Age2/100	−0.006	0.003	0.994**	−0.017	0.002	0.983***	−0.007	0.003	0.993**	0.005	0.004	1.005	0.012	0.004	1.012**	0.006	0.005	1.006
Left	−0.164	0.016	0.849***	−0.160	0.022	0.852***	−0.166	0.018	0.847***	−0.169	0.025	0.845	−0.190	0.037	0.827***	−0.172	0.026	0.842***
Center	−0.136	0.015	0.873***	−0.124	0.020	0.883***	−0.142	0.017	0.868***	−0.116	0.023	0.890	−0.123	0.030	0.884***	−0.099	0.025	0.906***
Politicalposition_refusal	−0.522	0.017	0.593***	−0.518	0.030	0.596***	−0.507	0.020	0.602***	−0.261	0.026	0.770	−0.349	0.050	0.705***	−0.219	0.029	0.803***
Female	−0.297	0.011	0.743****	−0.302	0.011	0.739	−0.320	0.012	0.726****	−0.354	0.018	0.702****	−0.363	0.018	0.696***	−0.344	0.019	0.709****
Educ16_19										0.078	0.025	1.081***	0.080	0.025	1.083***	0.069	0.026	1.071***
Educ20+										0.213	0.028	1.237****	0.203	0.029	1.225***	0.205	0.030	1.228****
Student										0.143	0.052	1.154***	0.195	0.054	1.215***	0.162	0.055	1.176***
Manager										0.054	0.041	1.055	0.070	0.041	1.073	0.063	0.043	1.065
Employed										−0.038	0.040	0.963	−0.010	0.040	0.990	−0.020	0.043	0.980
Worker										−0.049	0.035	0.952	−0.024	0.036	0.976	−0.028	0.038	0.972
Housewife/man										0.002	0.042	1.002	0.038	0.043	1.039	0.010	0.044	1.010
Jobseeker										−0.039	0.046	0.962	0.002	0.047	1.002	−0.001	0.050	0.999
Pensioner										−0.024	0.041	0.976	0.005	0.041	1.005	0.019	0.044	1.019
Random effects
Cohort	0.007	0.007		0.000	0.000		0.008	0.009		0.005	0.006		0.000	0.000		0.004	0.008
ICC (Cohort)	0.19%			0.00%			0.22%			0.13%			0.00%			0.11%
MOR (Cohort)	1.083			1.000			1.089			1.070			1.000			1.062
Period	0.207	0.192		0.134	0.056**		0.183	0.166		0.290	0.470		0.076	0.052		0.255	0.419
ICC (Period)	5.70%			3.65%			5.10%			7.47%			2.04%			6.88%
MOR (Period)	1.543			1.418			1.504			1.671			1.301			1.619
Country	Estimated	SE	Correlation	Estimated	SE	Correlation	Estimated	SE	Correlation	Estimated	SE	Correlation	Estimated	SE	Correlation	Estimated	SE	Correlation
Intercept	0.126	0.034	***	0.249	0.065	***	0.104	0.043	**	0.297	0.082	****	0.352	0.073	***	0.155	0.061	**
Left				0.035	0.008	***							0.063	0.021	***
Center				0.024	0.007	***							0.027	0.011	**
Politicalposition_refusal				0.086	0.016	***							0.172	0.037	***
Cov(Intercept, Left)				−0.066	0.016	−0.707***							−0.012	0.028	−0.081
Cov(Intercept, Center)				−0.042	0.014	−0.543***							−0.007	0.022	−0.072
Cov(Intercept, Refusal)				−0.103	0.022	−0.704***							−0.014	0.036	−0.057
Cov(Left, Center)				0.025	0.006	0.863***							0.031	0.013	0.752**
Cov(Left, Refusal)				0.041	0.010	0.747***							0.056	0.021	0.538***
Cov(Center, Refusal)				0.030	0.008	0.660***							0.049	0.017	0.719***
Between−country variance
Left				0.153									0.391
Center				0.190									0.364
Right				0.249									0.352
Refusal				0.130									0.495
ICC (Country)	3.47%			6.78%			2.90%			7.65%			9.47%			4.18%
MOR (Country)	1.403			1.610			1.360			1.682			1.761			1.456
Deviance	183,898			182,053			145,652			81,327			79,704			69,941
ICC (Cohort, Period, Country)	9.36%			10.43%			8.22%			15.52%			11.51%			11.17%
MOR (Cohort, Period)	1.555			1.418			1.517			1.679			1.301			1.625
MOR (Cohort, Country)	1.416			1.61			1.376			1.689			1.761			1.463
MOR (Period, Country)	1.734			1.805			1.667			2.077			1.866			1.842
MOR (Cohort, Period, Country)	1.744			1.805			1.679			2.083			1.866			1.847
Number of observations	140,119			140,119			110,519			69,674			69,674			57,659

SE: standard error; OR: odds ratio; ICC: intraclass correlation coefficient; MOR: median odds ratio.

p value < .1; **p value < .05; ***p value < .01; ****p value < .001.

The low ICC and MOR calculated behave like any analysis using micro data to underline the importance of individual effects with regard to contextual effects. The addition of contextual effects provides an explanation for 8%–15.5% of total variance. For individuals belonging to different periods, countries, and generations, the MOR is 1.679 (A2.M5) and 2.083 (A3.M5). We can already underline the importance of period and country effects with regard to cohort effects. In fact, the MOR for individuals observed at different periods but belonging to the same generation in the same country varies between 1.301 (A3.M6) and 1.671 (A3.M5). It is only 1.000–1.089 for individuals of different generations but the same country and observed at the same period. The variance of the random intercept is not significantly different from zero for the cohort. It means that the generations cannot be differentiated after controlling for different covariables. The conclusion is the same for the period effect for A2, A3, and A4, and this effect is significant only in the A1.M6 case.

Although it is not the focus of our article, we can highlight the differentiated impact of the different control variables on the acceptability level for biotechnologies, by calculating the odds ratio (corresponding to the exponential of coefficients). The odds of accepting genetic engineering or biotechnologies are reduced for women by 25%–27% and the odds of accepting GMO applications by 29%–30%. A high level of education has a positive impact on the acceptability of GM food (between +23% and 24% if the educational system was left after 20 years of age and between 16 and 19 years. When the respondent is still in the educational system, this positive impact falls to 17%). These last effects are clearly in line with the literature on the consumer willingness to accept biotechnologies or GM food (e.g. see Rousselière and Rousselière, 2013, on ISSP data; Sturgis et al., 2005, on British Social Survey). Distinguishing between objective (tested) and subjective (self-rated, also known as perceived) knowledge (measured by education), House et al. (2004) show that people holding a college degree or more are more likely to accept GM foods. They specify that objective knowledge (partially an outcome of education) is not linked to acceptance, whereas subjective knowledge (also partially linked to education) is an important variable of GM food acceptance. As we don’t have the relevant variable, we cannot test these two dimensions of knowledge.

For the different models, following a Wald test as defined for multilevel models by Goldstein (2003), we can reject the null hypothesis of a conjoined nullity of age, political positioning, or education but not of occupational status. We can see that being on the left side of the political scale (rather than being on the right side) contributes to a reduction in the odds of accepting genetic engineering in general and GM food, approximately by 25%.

The political effect exists also to a lesser extent for the center and an even lesser extent for the refusal category. By adding a random slope in the models A1.M6 and A3.M6, we can show some contrasting results. For the “biotechnology and genetic engineering” model, between-country variance is higher for right-wing people, but for “GMO” model, the between-country variance is higher for left-wing people, with the highest level for nonpartisan or refusal category. It is as if there is stronger difference on the attitude toward biotechnology for left-wing sympathizers across Europe than for right-wing sympathizers. Moreover, in the A1.M6 model, countries with higher intercept have a flatter slope for left-wing people. It seems that there is a less negative impact of being left wing on acceptability in country with a high level of acceptability. Furthermore, a country with a higher level of acceptance for left-wing politics has a higher slope for center and refusal categories. The same remarks apply to the A3.M6 model.

Concerning our hypothesis H1, in all cases, we find a significant age effect. Figure 1 corresponds to the graphical representation of these effects on the acceptability of different biotechnology applications. These age effects are calculated at the mean, adjusted for all control variables, and averaged over all contextual effects (periods, cohorts, and countries). A decreasing (unconditional) age effect can be found. The probability of being optimist about biotechnologies and genetic engineering thus goes from 56% at 20 years of age to 51% at 40 years, 44% at 60 years, and 37% at 80 years. Concerning GM food, the probability of support goes from 33% at 20 years to 29% at 40 years, 27% at 60 years, and 25% at 80 years.

Figure 1.

Age effects on the probability of supporting biotechnologies. Note that age effect are adjusted for all control variables and averaged over all contextual effects.

In the multilevel random effects model, each country has a constant that represents the cohort effect net of individual effects and averaged over all contextual effects (Yang, 2008). We used empirical Bayes prediction for these random intercepts (see Skrondal and Rabe-Hesketh, 2004).

Figure 2 corresponds to the graphical representation of the “country effect” (net of individual effects and averaged over political scale effect and all other contextual effects). We note the very considerable differences of association between “encouragement of GM food” and “optimism regarding biotech” across countries. Although there is a wider difference on the encouragement scale than on the optimism scale, this loose correlation allows us to classify countries into four clusters. A total of 5 countries are unfavorable to GM food but optimist about biotechnologies and genetic engineering (France, Sweden, Cyprus, Estonia, and Norway), 10 countries are more favorable toward GM food and less optimist about biotechnologies and genetic engineering (Belgium, Finland, Portugal, the Netherlands, Great Britain and Northern Ireland, Poland, East Germany, Ireland, and Malta), 6 countries are more unfavorable toward GM food and optimist about biotechnologies and genetic engineering (Luxembourg, Slovakia, Hungary, Italy, Czech Republic, and Spain), and 12 countries more unfavorable toward GM food and less optimist about biotechnologies and genetic engineering (Turkey, Bulgaria, Austria, Greece, Latvia, Romania, Croatia, Lithuania, Switzerland, Denmark, Slovenia, West Germany and Poland).

Figure 2.

Country effects on the probability of supporting biotechnologies. Note that country effects net of individual effects and averaged over all contextual effects.

6. Discussion and conclusion

The analysis proposed in this article enables the separation of age, period, cohort, and country effects. Contrary to traditional approaches such as those used by Bonny (2008), the modeling allows the quantification of the relative importance of contextual effects. Relative to H1 (support for biotechnologies decreases with age), we underline a strong age effect corresponding to weaker and weaker support for biotechnologies and GM food. The effect of age is more marked for biotechnologies or genetic engineering in general: according to the risk–utility dilemma (see Gaskell et al., 2004; Klingeman and Hall, 2006), the perception of these technologies as risky is not compensated for older Europeans by an increased perceived utility. Finally, we must note that as the average age is increasing with the period, the effect of population aging may be complex and can have structural effects on European societies. The development of functional foods even with new biotechnologies, for example, may lead to a greater acceptance of middle-aged and elderly consumers (Siro et al., 2008).

Relative to H2 (the effect of the cohort would correspond to the succession of generations that are ever more favorable to the use of biotechnologies), our modeling underlines no decisive generation effect. For GM food, it should be noted that no generation effect can be reported, whereas empirical findings coming for studies on meat consumption (Fukuda, 2011), fresh fish consumption (Mori and Saegusa, 2010), or eating habits in general (Recours and Hébel, 2007) usually underscore important cohort effects. However, we cannot predict that future generations born in a European society who are less and less favorable toward GM food and more aware of environmental aspects, but faced with an increasing supply of GM products in the other parts of the world, will have the same attitude.

Relative to P3 (the succession of periods is characterized by considerable effects on the social acceptability of biotechnologies), our modeling underlines “important jumps” in optimism about biotechnologies and genetic engineering in model A1.M6, but the effects are not significant in the other models. It seems to appear that although Europeans can perceive the utility of certain biotechnologies and, therefore, accept their use (Gaskell et al., 2004), they are more and more skeptical regarding the application of these techniques to human food. But that skepticism is largely country specific due to a “polarization effect.” This is the confirmation of the meta-regression on 51 studies reporting 114 GM food valuation estimates by Dannenberg (2009): “while the aversion to GM food is increasing steeply in Europe, it is only gently increasing in America and even decreasing in the rest of the world” (p. 2189). As not all new technologies are affected, the widespread perception of a “crisis of trust” in the relationships between science, politics, commerce, and society need to be nuanced (see Marris and Rose, 2010).

More than the validation of P4 (country effect is characterized by the existence of strong heterogeneity between different European countries), we find that the newly admitted countries have a tendency to behave like the old European countries. Here, we have some divergence with the analysis based on univariate statistics of Gaskell et al. (2006). But as stated by the same authors, although the new EU group is characterized by more optimistic view, “such views can be subject to dramatic changes.” Therefore, this has to be confirmed in further works, in which we can try to estimate the net impact of the EU adhesion on the acceptability of biotechnologies. However, four groups of countries can be highlighted following the lesser or greater probability of their population being favorable to biotechnologies in general or GM food. We also show that the effect of political affiliation is largely country specific. Although leftists are more likely to be hostile to GMO, there are strong divergences between countries. In a country with a high acceptability, left-wing sympathizers tend to be less opposed to GMO. This finding is in line with the results of Costa-Font et al. (2008) who show that in the United Kingdom, “Conservative and Labour supporters were more likely to express positive feelings about GM Food” (p. 231), while Veltri and Suerdem (2011) show that left-wing people in Turkey share the view that “GMOs can be a Pandora’s Box hiding unknown cultural, ecological, ethical, and economic risks” (p. 151). This highlights the importance of the national structuring of the public debate on GMO. Our work leaves the theoretical explanation of the heterogeneity of European attitudes toward biotechnologies largely open. We suspect that the difference between regions can be explained not only by the chosen governance regime as developed by Kvakkestad and Vatn (2011) but also by the national stakeholder competition for public trust (Aerni and Bernauer, 2005). According to different works, the acceptability difference first stems from a “trust gap” between countries highlighted by Priest et al. (2003). While controlling the level of knowledge, trust in scientists (Canavari and Nayga, 2009), public authorities, or manufacturers (Rousselière and Rousselière, 2010) has a positive impact on the acceptability of GM foods. On the contrary, trust in environmental associations (Huffman et al., 2004; Marques et al., 2015) reduces its acceptability. The “trust gap” explains the difference in acceptability of GMOs in Europe and the United States by the fact that Europeans have a greater trust in consumer and environmental protection associations, whereas in the United States, people have a greater trust in the “biotechnological system.” Using samples representative of the European population, this work needs to be completed, notably through the use of micro-simulation techniques enabling the impact of a context’s evolution (notably via a public policy or population aging) on the individual behavior, to be tested (Stefani, 2008).

Footnotes

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

Author biographies

Damien Rousselière is Professor of economics at Agrocampus Ouest (Public University of Agricultural Sciences in Angers and Rennes) (France) and visiting professor at Université du Québec à Montréal (Canada). He was also an external consultant for the Environment Directorate of the OECD. He serves on the scientific councils of Coop de France (Federation of agricultural cooperatives) and Végépolys (Cluster on horticulture, seeds and urban landscaping).

Samira Rousselière is Associate Professor of Management at Oniris (Nantes Atlantic College of Veterinary Medicine, Food Science and Engineering), Department of Management, Statistics and Communication (France). Founding member of the French national consortium on Food Eco-innovation (IDEFI Ecotrophelia), her current interest includes functional foods, food eco-innovation and public understanding of new technologies.

References

Aerni

Bernauer

(2005) Stakeholder attitudes toward GMOs in the Philippines, Mexico, and South Africa: The issue of public trust. World Development 34(3): 557–575.

Akaike

(1974) A new look at the statistical model identification. IEEE Transactions on Automatic Control 19(6): 716–723.

Alakasen

Myhr

(2007) The worth of a wildflower: Precautionary perspectives on the environmental risk of GMOs. Ecological Economics 60: 489–497.

Alston

Bradford

Kalaitzandonakes

(2006) The economics of horticultural biotechnology. Journal of Crop Improvement 18(1): 413–431.

Amin

Azad

MAK

Gausmian

Zulkifli

(2014) Determinants of public attitudes to genetically modified salmon. PLoS ONE 9(1): E86174.

Auer

(2008) Ecological risk assessment and regulation for genetically-modified ornamental plants. Critical Reviews in Plant Sciences 27: 255–271.

Austin

(2010) Estimating multilevel logistic regression models when the number of clusters is low. The International Journal of Biostatistics 6(1): Article 16.

Barnett

Cooper

Senior

(2007) Belief in public efficacy, trust, and attitudes towards modern genetic science. Risk Analysis 27(4): 921–933.

Bell

Jones

(2014) Another “futile quest”? A simulation study of Yang and Land’s hierarchical age-period-cohort model. Demographic Research 30(11): 333–360.

10.

Bonny

(2008) How have opinions about GMOs changed over time? The situation in the European Union and the USA. CAB reviews: Perspectives in agriculture, veterinary science. Nutrition and Natural Resources 3(93): 1–17.

11.

Bourdieu

(1990) Sociology in Question. London: SAGE.

12.

Browne

(2009) MCMC Estimation in MLwiN, Centre for Multilevel Modelling. Bristol: University of Bristol.

13.

Browne

Draper

(2006) A comparison of Bayesian and likelihood-based methods for fitting multilevel models. Bayesian Analysis 1(3): 473–514.

14.

Browne

Goldstein

Rasbash

(2001) Multiple membership multiple classification (MMMC) models. Statistical Modelling 1(2): 103–124.

15.

Burnham

Anderson

(2004) Multimodal inference: Understanding AIC and BIC in model selection. Sociological Methods and Research 33(2): 261–304.

16.

Cameron

Trivedi

(2005) Microeconometrics: Methods and Applications. New York: Cambridge University Press.

17.

Canavari

Nayga

(2009) On consumers’ willingness to purchase nutritionally enhanced genetically modified food. Applied Economics 41(1): 125–137.

18.

Chauvel

Schröder

(2015) The impact of cohort membership on disposable incomes in West Germany, France, and the United States. European Sociological Review 31(3): 298–311.

19.

Chen

Wong

Lee

(2001) Age, period, and cohort effects on life insurance purchases in the US. The Journal of Risk and Insurance 68(2): 303–327.

20.

Clements

(2014) Political party supporters’ attitudes towards and involvement with green issues in Britain. Politics 34(4): 362–377.

21.

Costa-Font

Mossialos

Rudisill

(2008) Are feelings of genetically modified food politically driven? Risk Management 10(3): 218–234.

22.

Dannenberg

(2009) The dispersion and development of consumer preferences for genetically modified food—A meta-analysis. Ecological Economics 68: 2182–2192.

23.

DeRosier

Sulemana

James

Jr Valdivia

Folk

Smith

(2015) A comparative analysis of media reporting of perceived risks and benefits of genetically modified crops and foods in Kenyan and international newspapers. Public Understanding of Science 24(5): 563–581.

24.

Durant

Legge

(2005) Public opinion, risk perceptions, and genetically modified food regulatory policy. European Union Politics 6(2): 181–200.

25.

Einsiedel

Medlock

(2005) A public consultation on plant molecular farming. AgBioForum 8(1): 26–32.

26.

Fiechter

(ed.) (2000) History of Modern Biotechnology I. Dordrecht: Springer.

27.

Freire

(2008) Party polarization and citizens’ left-right orientations. Party Politics 14(2): 189–209.

28.

Frewer

Miles

Marsh

(2002) The media and genetically modified foods: Evidence in support of social amplification of risk. Risk Analysis 22(4): 701–711.

29.

Fukuda

(2006) Age-period-cohort decomposition of aggregate data. Journal of Applied Econometrics 21: 981–998.

30.

Fukuda

(2011) A comparative study of U.S. and Japanese meat consumption using age-period-cohort decomposition. Journal of International Food & Agribusiness Marketing 23(2): 151–166.

31.

Gaskell

Allansdottir

Allum

Castro

Esmer

Fischler

. (2010) Europeans and biotechnology in 2010, winds of change? A report to the European Commission’s Directorate-General for Research, London School of Economics, London, October 2010.

32.

Gaskell

Allansdottir

Allum

Castro

Esmer

Fischler

. (2011) The 2010 Eurobarometer on the life sciences. Nature Biotechnology 29(2): 113–114.

33.

Gaskell

Allum

Wagner

Kronberger

Toergersen

Hampel

. (2004) GM foods and the misperception of risk perception. Risk Analysis 24(1): 185–194.

34.

Gaskell

Stares

Allansdottir

Allum

Corchero

Fischler

. (2006) Europeans and biotechnology in 2005. Report to the European Commission’s Directorate-General for Research, London School of Economics, London, May 2006.

35.

Goldstein

(2003) Multilevel Statistical Models, 3rd edn. London: Edward Arnold.

36.

Green-Pedersen

(2007) The growing importance of issue competition: The changing nature of party competition in Western Europe. Political Studies 55: 607–628.

37.

Hall

Mairesse

Turner

(2007) Identifying age, cohort, and period effects in scientific research productivity. Economics of Innovation and New Technology 16(2): 159–177.

38.

Holmgreen

(2008) Biotech as “biothreat”? Metaphorical constructions in discourse. Discourse & Society 19(1): 99–110.

39.

House

Lusk

Jaeger

Traill

Moore

Valli

. (2004) Objective and subjective knowledge: Impacts on consumers’ demand for genetically modified foods in the United States and European Union. AgBioForum 7(3): 113–123.

40.

Huffman

Rousu

Shogren

Tegene

(2004) Who do consumers trust for information: The case of genetically modified foods? American Journal of Agricultural Economics 86(5): 1222–1229.

41.

Inglehart

(1990) Culture Shift in Advanced Industrial Society. Princeton, NJ: Princeton University Press.

42.

Joly

P-B

Marris

(2003) Les Américains ont-ils accepté les OGM? [Have the Americans accepted GMOs?]. Cahiers d’économie et Sociologie Rurales 68–69: 12–45.

43.

Kahan

Jenkins-Smith

Braman

(2011) Cultural cognition of scientific consensus. Journal of Risk Research 14(2): 147–174.

44.

Kass

Raftery

(1995) Bayes factors. Journal of the American Statistical Association 90(430): 773–795.

45.

Klingeman

Hall

(2006) Risk, trust and consumer acceptance of plant biotechnology. Journal of Crop Improvement 18(1): 451–486.

46.

Knight

Barnett

(2010) Perceived efficacy and attitudes towards genetic science and science governance. Public Understanding of Science 19(4): 386–402.

47.

Kvakkestad

Vatn

(2011) Governing uncertain and unknown effects of genetically modified crops. Ecological Economics 70: 524–532.

48.

Larsen

Petersen

Budtz-Jørgensen

Endhal

(2000) Interpreting parameters in the logistic regression model with random effects. Biometrics 56(3): 909–914.

49.

Luo

(2013) Assessing validity and application scope of the intrinsic estimator approach to the age-period-cohort problem. Demography 50: 1945–1967.

50.

Marques

Critchley

Walshe

(2015) Attitudes to genetically modified food over time: How trust in organizations and the media cycle predict support. Public Understanding of Science 24(5): 601–618.

51.

Marris

(2001) Public views on GMOs: Deconstructing the myths. European Molecular Biology Organization Reports 2(7): 545–548.

52.

Marris

Rose

(2010) Open engagement: Exploring public participation in the Biosciences. PLoS Biology 8(11): 1–2.

53.

Mason

Winsborough

Poole

(1973) Some methodological issues in cohort analysis of archival data. American Sociological Review 38(2): 242–258.

54.

Mori

Saegusa

(2010) Cohort effects in food consumption: What they are and how they are formed. Evolutionary and Institutional Economics Review 7(1): 43–63.

55.

Nakamura

(1986) Bayesian cohort models for general cohort table analyses. Annals of the Institute of Statistical Mathematics 38(1): 353–370.

56.

ESW

Carpenter

Goldstein

Rasbash

(2006) Estimation in generalised linear mixed models with binary outcomes by simulated maximum likelihood. Statistical Modelling 6: 23–42.

57.

Nielsen

(1997) Public perception: Behind the color code of no. Nature Biotechnology 15: 1320–1321.

58.

Nielsen

Jelsøe

Öhman

(2002) Traditional blue and modern green resistance. In: Bauer

Gaskell

(eds) Biotechnology: The Making of a Global Controversy. Cambridge: Cambridge University Press, pp. 179–202.

59.

O’Brien

Hudson

Stockard

(2008) A mixed model estimation of age, period, and cohort effects. Sociological Methods & Research 36(3): 402–428.

60.

Pardo

Engelhard

Hagen

Jorgensen

Rehbinder

Schnieke

. (2009) The role of means and goals in technology acceptance. EMBO Reports 10: 1069–1075.

61.

Pawar

(2009) Theory Building for Hypothesis Specification in Organizational Studies. London: SAGE.

62.

Priest

Bonfadelli

Rusanen

(2003) The “trust gap” hypothesis. Risk Analysis 23(4): 751–766.

63.

Raftery

Lewis

(1992) How many iterations in the Gibbs sampler? In: Bernardo

Berfer

Dawid

Smith

AFM

(eds) Bayesian Statistics 4. Oxford: Oxford University Press, pp. 765–776.

64.

Ramani

(2001) Introduction to the special issue on biotechnology. International Journal of Biotechnology 3(1–2): 1–3.

65.

Recours

Hébel

(2007) Evolution des comportements alimentaires: Le rôle des générations [Evolution of eating behaviours: Generational effects]. Cahiers d’économie et sociologie rurales. 82–83: 79–108.

66.

Reither

Land

Jeon

Powers

Masters

Zheng

. (2015) Clarifying hierarchical age-period-cohort models: A rejoinder to Bell and Jones. Social Science & Medicine 145: 125–128.

67.

Rodgers

(1982) Reply to comment by Smith, Mason and Fienberg. American Sociological Review 47(6): 787–793.

68.

Rousselière

(2010) On the impact of trust on willingness to purchase GM food: Evidence from a European survey. Review of Agricultural and Environmental Studies 91(1): 5–26.

69.

Rousselière

(2013) Assiste-t-on (réellement) à une polarisation du débat sur les OGM? Une perspective internationale sur la période 2000–2010 [Is there (really) a polarization of attitudes toward GMO? An international inquiry on the 2000–2010 period]. Revue d’économie politique 123(4): 593–622.

70.

Schaninger

Danko

(1993) A conceptual and empirical comparison of alternative household life cycle models. Journal of Consumer Research 19(4): 580–594.

71.

Scholz

Zuell

(2012) Item non-response in open-ended questions: Who does not answer on the meaning of left and right? Social Science Research 41: 1415–1428.

72.

Siro

Kapolna

Lugasi

(2008) Functional food: Product development, marketing and consumer acceptance, a Review. Appetite 51: 456–467.

73.

Skrondal

Rabe-Hesketh

(2004) Generalized Latent Variable Modeling: Multilevel, Longitudinal and Structural Equation Models. Boca Raton, FL: CRC Press.

74.

Spiegelhalter

Best

Carlin

Van Der Linde

(2002) Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society B 64: 583–639.

75.

Stefani

(2008) Micro-simulation of households: A new tool to assess the impact on society of food safety policies. Innovation 21(2): 171–176.

76.

Sturgis

Cooper

Fife-Schaw

(2005) Attitudes to biotechnology: Estimating the opinions of a better-informed public. New Genetics and Society 24(1): 35–58.

77.

Veltri

Suerdem

(2011) Worldviews and discursive construction of GMO-related risk perceptions in Turkey. Public Understanding of Science 22(2): 137–154.

78.

Verbeek

(2008) Pseudo-panels and repeated cross-sections. In: Matyas

Sevestre

(eds) The Econometrics of Panel Data. Berlin: Springer-Verlag, pp. 369–383.

79.

Yang

(2008) Social inequalities in happiness in the United States, 1972 to 2004: An age-period-cohort analysis. American Sociological Review 73(2): 204–226.

80.

Yang

Land

(2008) Age-period-cohort analysis of repeated cross-section surveys: Fixed or random effects? Sociological Methods and Research 36(3): 297–326.

81.

Yang

Land

(2013) Age-Period-Cohort Analysis: New Models, Methods and Empirical Applications. Boca Raton, FL: CRC Press.

82.

Zepeda

Douthitt

You

(2003) Consumer risk perceptions toward agricultural biotechnology, self-protection, and food demand: The case of milk in the United States. Risk Analysis 23(5): 973–984.