Attitudes Towards Meat Consumption; A Network
Perspective
L. Dorresteijna
aUniversity of Amsterdam
Attitudes can play an important role in understanding what motivates meat consumption and finding key indicators of behavioral change towards more sustainable dietary lifestyles. Attitude research has typically used the tripartite model, where an attitude acts as a latent variable. Recently, a Causal Attitude Network Model (CAN) was pro-posed, where an attitude is represented by the evaluative reactions and interactions between these reactions. The attitude tend to clus-ter and therefore adhere to a small-world structure. The objective of this study was, first, to identify the different dietary lifestyles. A questionnaire was composed, which was completed by 650 partici-pants. We applied a latent class analysis and the results indicated three different attitudes toward meat consumption. The second ob-jective of this study was to use the network approach to investigate the attitudes. The networks of the three attitudes were estimated, rel-ative importance of items was determined using centrality measures, stability of the networks was assessed and the networks were com-pared. We found that the attitudes do not cluster as suggested by the CAN model and that these attitudes seem to strive for consis-tent, as opposed to accurate, attitudes. Additionally, the attitudes differed on centrality: in the meat consumers and reducers network ideas about the environment influence the network, while in the meat avoiders network animal abuse influences the network. Targeting these nodes in the network of meat consumers may alter frequency of meat consumption. Furthermore, while the structure of the net-work seemed to be relatively similar, the attitudes of meat consumers were stronger connected than meat reducers and meat avoiders. The relatively higher connectivity could make this attitude more difficult to alter.
Attitude | Meat Consumption | Dietary lifestyle | Network Analysis
N
owadays, meat consumption still increases (1),even though meat consumption is linked to severe environmental- (2) and health problems (3)(4) and elicits se-vere animal abuse (5). In general, society does not realize that consumption of animal-based food products negatively effect the environment and reducing meat consumption is necessary for a sustainable future (6). Most commonly four dietary lifestyles are considered in society: omnivorous, flexitarian, vegetarian and vegan. These diets vary from non-sustainable (omnivorous) to sustainable (vegan) respectively and are based on frequency of consumption of meat and other animal-derived products. Omnivores do not restrict their diet and are cur-rently the most common in the Netherlands (>45%), while flexitarians (40%) consume meat a maximum of 4 times a week, vegetarians (4%) abstain from meat every day and veg-ans (<1%) abstain from all animal-based products including milk, cheese and honey. Not only do vegans abstain from con-suming animal-derived products, but they strive to exclude all forms of exploitation of, and cruelty to, animals for clothing or any other purpose (7). This division between dietary lifestyles is solely based on behavior and is not driven by data. There-fore, it may not reflect the actual differences between dietary
lifestyles. Recently, another categorization was proposed (8), where a division is made between meat consumers, meat reduc-ers and meat avoidreduc-ers. Again, in this division meat consumreduc-ers do not restrict their diet and meat reducers are defined as someone who reduces their meat intake to less than 4 times a week. In contrast to the other division; no distinction is made between vegetarians and vegans. While different divisions are proposed, no consensus has been reached and a division driven by data is needed.
Attitude.Attitudes can play an important role in
understand-ing what motivates meat consumption. Numerous studies show differences in attitudes towards meat consumption be-tween dietary lifestyles. Kenyon and Barker (1998) found that omnivores perceive meat as positive and associate meat with taste, luxury, social status and special occasions such as Christ-mas, in contrast to vegetarian girls, who have strong negative associations with meat (killing, cruelty, blood, poor health and visceral disgust). Similar attitude differences between dietary lifestyles were found in research using the Implicit Association Task (10) and the Extrinsic Affective Simon Task (11). Ruby (2012) found that attitudes differ within the vege-tarian group as well. Namely, vegevege-tarians with ethical motives make more associations with their dietary choices, explicitly integrating it with philosophical frameworks and will react with more disgust to meat consumption. It is not unusual for other reasons to be added in addition to the initial motive for reducing meat consumption (13). Thus, there is a sizable body of evidence that omnivores and vegetarians perceive meat in different terms. However, attitudes of flexitarians and vegans are less investigated.
Not only do omnivores perceive meat differently, but most experience conflicting thoughts. Omnivores enjoy meat, but are simultaneously aware of the meat-associated problems (14). Thus, consumers who have these conflicting thoughts may experience cognitive dissonance (15). Research suggest that these attitudes are less stable, more pliable and tend to result in more systematic information processing (16). Even the idea of commitment can lead to conflicted feeling and to considerable more negative affect. Consequently, this could motivate someone to reduce the dissonance between action and attitudes (17). Consumers can use several strategies to reduce cognitive dissonance: avoidance, dissociation, perceived behavioral change, denial of animal pain, denial of animal mind, pro-meat justification and reduced perceived choice (18).
Desirably, consumers reduce their cognitive dissonance by changing their behavior towards more sustainable dietary lifestyles. Recently, Graça et al. (2015) suggested that the willingness to transition from omnivorous towards a more
tainable lifestyle may depend on the affective connection with meat. They propose that this affective connection may be a continuum in which one end refers to disgust (i.e. negative affect and repulsion, related with moral internalization) and the other to attachment (i.e. feeling sad when considering reducing meat consumption, and high positive affect and de-pendence towards meat). The willingness to transition from vegetarian towards vegan, however, may depend on moral engagement with animal welfare. Jabs et al. (1998) found that ethical vegetarians are more concerned with animal wel-fare and, thus, are more likely to transition towards veganism compared to health vegetarians. In addition, Hamilton (2006) interviewed vegetarians and concluded that health vegetarians gradually left out animal-based products, while ethical vege-tarians made more sudden changes in their diet. These effects of motivation are also reflected within omnivores. Rothgerber (2015) concluded that conscious omnivores were more disloyal to their diet than vegetarians. He hypothesized that they have less affinity with animals and evaluate animals less favorably, which may cause them to eat meat more often and feel less guilty when doing so. This explanation is supported by ear-lier evidence that vegetarians share more empathy towards animals, believe animals to be more similar to humans and support animal rights more often (18).
Network perspective.Historically, the most influential model
of attitudes has been the tripartite model of attitudes, where an attitude acts as a latent variable that cause affective, be-havioral and cognitive components. The specific responses to the attitude questions reflect the latent variable (23). Dalege et al. (2015) posit that this tripartite model has several limi-tations. First, the latent variable model is not able to offer a plausible representation of the structure of attitudes regarding formation and change in attitudes. Second, it is not able to integrate inconsistencies between attitude and behavior, be-cause the model assumes that behavior is part of the attitude. Therefore, Dalege et al. (2016) propose a new model; the Causal Attitude Network (CAN) model. In the CAN model an attitude is still composed out of the three domains; affect, behavior and cognition, but these domains are represented by the evaluative responses and interactions between these responses. This model argues that the attitudes tend to cluster and therefore adhere to a small-world structure. In contrast to the tripartite model, the CAN model can be fitted to actual data, while still maintaining explanatory power. This benefit makes it particularly fit to visualize formation, change and inconsistencies in attitudes. When the CAN model is used, findings can be integrated, indicators of behavioral change can be identified and inconsistencies between behavior and attitude can be overcome.
Aim.To guide behavioral change towards more sustainable lifestyles a strong foundation should be laid down to identify, on the long run, key indicators of this behavioral change. This study aims to identify these key indicators of behavioral change, however, before this is possible a more descriptive study is needed to build upon. This study has two objectives. First, it aims to identify the different dietary lifestyles. The literature proposes two divisions (7)(8), however there is no consensus about the categorization of dietary lifestyles. It is hypothesized that these categories are not only determined by self-reported behavior, but also by their beliefs and feelings
about meat consumption.
Second, when the different dietary lifestyles are identified, it is possible to visualize the network of the attitude for each group and investigate group differences. A network perspec-tive is needed, because existing literature is only able - due to methodological limitations - to show links and therefore fails to provide a coherent overview. It is hypothesized that the attitudes differ on several characteristics; meat eaters have more positive associations with consuming meat (9), have less elaborate (12) and less stable attitudes (16), show cogni-tive dissonance (14) and experience more negative affect (17). These characteristics will diminish when diets become more sustainable.
Methods & Materials
Participants.Participants were recruited through off- and on-line channels. Sixty participants were recruited at a vegan event (Veggieworld Utrecht), the rest of the participants were recruited through several Facebook groups (e.g., ’Vegan Ned-erland’, ’Vedisch vegetarisch Koken’). Additionally, the link was shared in several social networks of participants. Materials & Procedure.Eighty questions regarding attitude towards meat consumption were collected from the literature (e.g. Berndsen et al., 2004; (19)(25) and categorized in af-fect, behavioral and cognitive items. Thirty-five items were removed due to similarity. We examined the remaining items in a pilot study (N = 45) on their variance and correlation. Items with high variance and low correlations were selected, resulting in 6 affect, 10 behavior and 6 cognition items (Table 1). Affect and cognition were measured on a 6-point Likert-scale (strongly disagree to strongly agree). Additionally, the first four behavioral items were measured on a 7-point (days of the week) scale, the next four items on a 4-point scale (al-ways to never) and the last behavioral questions on a 2-point scale (yes or no). Additionally, demographic questions were asked (age, allergies, education, gender, perceived lifestyle and religion). Furthermore, internal consistency was assessed with
Table 1. The attitude towards meat consumption questionnaire
Domain Question
Affect I like the taste of meat
Meat reminds me of death and suffering of animals If I had to stop eating meat I would feel sad If I eat meat I feel guilty
If I eat meat I would feel anxiety If I eat meat I feel disgust
Behavior How many times a week do you eat meat & fish for dinner How many times a week do you eat meat or fish for lunch How many times a week do you eat meat or fish for breakfast How many times a week do you eat cheese, milk or eggs How often do you order meat or fish in a restaurant How often do you cook with meat or fish at home
How often do you eat meat or fish when someone else prepares it How often do you eat meat replacements
Do you watch animal derived e-numbers while shopping Are you consciously reducing your meat intake Cognition Eating meat is morally wrong
Meat contains important nutritions for your body The production of meat is harmful for the environment Animals are inferior to people
By consuming meat you contribute to animal suffering There should be a tax on meat
Cronbach’s alpha (– = 0.73) and exceeded the value of .70 that is traditionally used as a rule of thumb (26).
The questionnaire was compiled and administered through the online survey program ’Qualtrics’ (Qualtrics Provo, UT). Data Analysis.
Latent class analysis.All analyses were conducted in R (28),
for the latent class analysis the R package poLCA was used (29). In the latent class analysis, the responses to the attitude questionnaire items were dichotomized and analyzed with an exploratory latent class analysis. This analysis determined the optimal number of latent classes; thus, the number of different attitudes toward meat consumption. In this analysis, the number of latent classes were increased and the model with the best fit was selected, which was expressed by the Bayesian Information Criterion (BIC; (30)). The BIC was used to compare models that show an acceptable fit to the data. A small value characterizes a model that fits well and is parsimonious. The BIC will usually be more appropriate for basic latent class models, because of their relative simplicity (31)(32). Additionally, the likelihood ratio chi-square (LR) and the Pearson’s ‰2goodness of fit were considered during model selection (33). Generally, the goal is to select models that minimize ‰2and LR without estimating excessive numbers of parameters.
The model with the best fit was selected, whereafter poste-rior probabilities were calculated. These probabilities reflected the probability that a response pattern of a participant belongs to a given dietary lifestyle and were used to categorize the attitudes for the network analysis. Thus, a participant was assigned to the group associated with the largest posterior probability.
Network analysis.After estimating the posterior probabilities,
we conducted the network analysis in four steps: network estimation, network centrality, network stability and network comparisons. The R-package qgraph was used to visualize all networks (34), the R-package Bootnet (35) was used to estimate stability and the R-package NCT (36) was used to compare the networks.
Network estimation A network describes an attitude as a network of mutually interacting characteristics and con-tains nodes (observed variables) and edges (relation between variables). To obtain the attitude network polychoric correla-tions were estimated (37), where an edge indicates a nonzero partial correlation between two nodes. Thus, edges can be understood as conditional independent relations among items of an attitude; if two items are connected in the network, they are connected after controlling for all other items. If no edge emerges, it means they are conditionally independent. How-ever, these networks can exhibit spurious connections that may arise due to multiple testing (38). Thus, a lasso regularization is applied to cancel out small partial correlation coefficients based on the Extended Bayesian Information Criterion (EBIC; (39) (40)) to control for these spurious connections. This re-sults in a sparse and easy interpretable network, where the absence of paths in the network suggest that two variables do not directly interact. To get a parsimonious network we set the hyperparameter to 0.2 in the analysis. We used the Fruchterman-Reingold algorithm to specify the layout of the
nodes in the network; thus, nodes that are strongly associated with one another appear closer together in the graphs, allowing for inspection of clustering (34).
Network centrality In this analysis we computed cen-trality measures for the different attitudes to analyze which variables greatly influence the network. The centrality mea-sures used in this analysis were strength, betweenness and closeness. Node strength measures the weighted number of connections of a focal node (i.e., the sum of all edges of a given node to all other nodes) and thereby the degree to which that node is involved in the network. Betweenness quantifies how important a node is in the average path between two other nodes (i.e., the number of times a specific node lies between two other nodes on their shortest connecting edge) and can be thought of a measure of how much the node exerts control over information flow in the network (41). Closeness quantifies how well a node is indirectly connected to other nodes (i.e., the inverse of the summed length of all shortest edges between a node and all other nodes). For all measures, higher scores are indicative of greater centrality. While previous papers have often investigated these three measures of centrality, recent investigations have shown that betweenness and closeness are often not reliably estimated (35). This was also the case in these analyses, therefore we will mostly focus on node strength. Betweenness and closeness are discussed in the Supplementary Materials.
Network stability A potentially problematic issue in network analyses is that sample size can impact their accuracy and replicability. Therefore, recent research has focused on evaluating the reliability of the network model parameters (35). Epskamp et al. (2016) developed the R package ’bootnet’ to construct confidence intervals (CIs) around the edges in a network and to determine the stability of the centrality measures. To estimate the edge-weight CIs, nonparametric bootstrapping with 1.000 draws was used, where overlapping edge-weight CIs indicate that the weights do not differ and that interpreting their order should be done with care.
Centrality stability was determined by subsetting the data and correlating the original centrality indices with those from the subsamples. Epskamp et al. (2016) recommend a coeffi-cient of at least 0.25, whereas a stability coefficoeffi-cient of 0.5 is ideal, although they note that evidence for these guidelines is currently limited. The stability coefficients guide the inter-pretation of the most central attitude items. To estimate the stability coefficients, casewise bootstrapping with 2.500 draws was used.
Network comparison Finally, we compared the net-works on strength and structure. We tested whether the networks differed from each other in their network structures via the R-package NetworkComparisonTest (NCT)(36). First, all networks were reciprocally compared to investigate whether all edges were identical. Second, we used the NCT to test whether global strength estimates differed across networks. This difference is defined as the weighted sum of the connec-tions for each group. The NCT randomly regroups participants from the groups (1.000 times) and calculates the differences. This results in a distribution which is used to test whether the groups significantly differ against a significance level of 0.05.
Table 2. Discrepancy in categorization between perceived lifestyle and reported behavior
Lifestyle vs. reported Omnivores Flexitarians Vegetarians Vegans
Omnivores 106 3 0 0
Flexitarians 89 101 7 0
Vegetarians 3 34 148 6
Vegans 0 1 21 131
Results
Participant characteristics.A total of 791 subjects started the questionnaire, but from a total of 650 subjects data was com-plete. In this study, the mean age was 34.55 year (SD = 12.95), with 477 women (71%). Most participants (N = 499) received higher education (HBO/WO), 116 received moderate educa-tion (MBO) and 34 attended high school. Addieduca-tionally, most participants were non-religious (N = 325; Atheist/ Agnostic), 71 were Christians, 56 had a religion with dietary restrictions (Buddhism, Hinduism, Islam and Jew-ism), all others (N = 197) had uncategorized religions. Furthermore, 11 participants were allergic to meat (most commonly pork) and fish and 25 were allergic to lactose or eggs.
Dietary lifestyles.
Behavioral data.In this study, 109 participants reported to
per-ceive their dietary lifestyle as omnivorous, 197 as flexitarian, 191 as vegetarian and 153 as vegan. The National Institute of Nutrition (1997) showed that perceived dietary lifestyle is not an accurate reflection of someone’s behavior, because one perceives their dietary lifestyle as more sustainable. Therefore, more stringent criteria where imposed on the self-reported consumption of meat and fish; flexitarians consume meat max-imum 4 days a week during dinner, vegetarians never eat meat or fish and vegans never eat any animal-derived products (8). The response on behavioral question on meat and fish con-sumption were summed and distinguished between omnivores and flexitarians (consume meat or fish less than 4 days a week). Additionally, the response on behavioral questions on cheese, eggs and milk consumption were summed to distin-guish between vegetarians and vegans (never consume these products).
After imposing the more stringent criteria, the sample con-sisted of 198 omnivores, 139 flexitarians, 176 vegetarians and 137 vegans. As expected, there was a discrepancy in cate-gorization between perceived lifestyle and reported behavior (t(649) = 11.17, p < 0.05). A total of 164 participants per-ceived their dietary lifestyle differently. Most of them (N= 89) perceived themselves as flexitarians, but were categorized as omnivores based on their behavior. Additionally, 34 perceived themselves as vegetarians, but were categorized as flexitarians and 21 perceived themselves as vegans, but were categorized as vegetarians. Interestingly, 16 participants showed more sustainable behavior than their perceived lifestyle suggests (Table 2).
As can be seen in Figure 1, omnivores and flexitarians differ in the times a week they consume meat or fish on all times of the day (t(334) = 34.28, p < 0.05; t(294) = 14.16, p < 0.05; t(267) = 6.73, p < 0.05). The most prominent difference is at dinner; omnivores consume meat or fish 6.66 (SD = 1.12) times a week during dinner, while flexitarians only consume meat or fish 2.02 (SD = 0.08) times a week, which is far less
than expected.
Additionally, omnivores, flexitarians and vegetarians differ in the days a week they consume cheese, milk and eggs (F(1) = 12.45, p < 005). Omnivores, consume these products 2.97 (SD = 1.24) days a week, which is more often than flexitarians (2.65 (SD = 1.21) days a week; t(300) = 2.37, p < 0.05) and vegetarians (2.51 (SD = 1.32) days a week; t(360) = 3.44, p < 0.05). Interestingly, there is no difference between flexitarians and vegetarians (t(205) = 0.95, p = 0.34).
Furthermore, several demographic measures predicted the dietary lifestyle based on reported behavior of participants. Firstly, men seem to eat meat more often compared to women (‰2(3) = 22.46, p < 0.05), which is in line with other studies (43) (44). Secondly, older participants (> 46 year) tended to consume meat more often (‰2(9) = 23.27, p < 0.05). Thirdly, the higher educated tend to eat less or no meat (flexitarian, vegetarian or vegan) (‰2 (9) = 20.77, p < 0.05). Fourthly, flexitarians and vegans tend to be more non-religious, while vegetarians tend to be more religious (‰2(6) = 60.29, p < 0.05). However, this could be due to non-random sampling; participants were, among others, recruited in a ’Vedisch Vege-tarian Cooking’ Facebook group, were people tend to be more spiritual-oriented. Fifthly, allergies were equally distributed among dietary lifestyles and therefore did not seem to influence their behavior (‰2(3) = 4.47, p = 0.12), but this should be carefully interpreted due to small sample size (N = 36).
Latent class analysis.The first objective of this study was to
de-scribe the groups of dietary lifestyles. Therefore, the responses to the attitude questionnaire were recoded in the same direc-tion for easier interpretadirec-tion, dichotomized and analyzed with an exploratory latent class analyses. The exploratory latent class analysis was conducted to determine the optimal number of latent classes. Evidently, the latent class analysis does not function optimally when the sample size is too small in com-parison to the number of variables (45), therefore variables with the smallest variance in affect, behavior and cognition were dropped so that each group consisted of 3 variables. Thus, the first models consisted of 9 variables (dinner, restaurant, home cooking, death, taste, guilt, tax, nutrition and abuse), while different number of classes were considered. In these models the behavioral question dinner was dichotomized with the mean as cutoff point. Table 3 shows the goodness-of-fit statistics of these models. Based on the BIC value, the models with 3 and 4 classes had the best fit. However, the goodness-of-fit statistics of the model with 4 classes was not constant, indicating sparseness of data or parameterization. As can be seen in Figure 2, which shows the probabilities of each variable, there is a clear distinction between the classes. Class 1 shows a response pattern expected from meat consumers, class 2 as expected from meat reducers and class 3 as expected from meat avoiders.
This model was used to estimate the posterior probabilities that shows the probability that a particular response pattern belongs to a specific dietary lifestyle (Fig. 2). The model predicts that the first class contains 158 meat consumers , the second class 122 meat reducers and the last class 370 meat avoiders. Interestingly, the groups based on the posterior probabilities are distinctively different from the groups based on reported behavior ((Table 4); t(649) = 8.83, p < 0.05); indicating that attitudes are not solely reflected by behavior.
Fig. 1. Frequency of meat or fish consumption per week for omnivores and flexitarians during dinner, lunch and breakfast. For each time of
the day omnivores consumed significantly more meat of fish per week.
Table 3. Goodness-of-fit statistics for the models with 1 to 6 classes
Number of classes BIC df ‰2 LR
1 class 7245 502 65306 2701 2 classes 5172 492 2475 563 3 classes 5084 482 910 411 4 classes 5089 472 919 350 5 classes 5154 462 608 251 6 classes 5193 452 1273 225
Table 4. Discrepancy in categorization between classes and behav-ior
classes vs. behavior Omnivorous Flexitarian Vegetarian + Vegan
Meat consumers 143 15 0
Meat reducers 55 63 4
Meat avoiders 0 61 309
Network analysis.
Network estimation.The second objective of this study was to
de-scribe the networks of the attitude towards meat consumption for the different dietary lifestyles. Figure 3 shows the net-work of the combined sample (discussed in the Supplementary Materials), whereas Figure 4 shows the networks of the meat consumers (a), meat reducers (b) and meat avoiders (c). The questionnaire consisted of 22 questions, however in a network analysis numerous parameters need to be estimated, which makes the analysis less reliable when sample size is relatively small (24). Therefore, the items anxiety, breakfast, cheese, death, e-numbers, inferior, lunch, moral and other cooking were removed to reduce the number of parameters to be es-timated while still maintaining relatively similar number of
items per domain in the attitude. Thus, four affect items, five behavior items and four cognition items remained. To obtain the networks polychoric partial correlations were calculated of the remaining 13 variables; abuse, dinner, disgust, envi-ronment, guilt, home cooking, nutrition, reduce, replacement, restaurant, sad, taste and tax. As described in Dalege et al. (2016), an attitude consists of affect, cognitive and behavioral
and these variables cluster together. In the network nodes in the same domain are shown in similar colors; black, dark gray and light gray. The edges between nodes within a network correspond to polychoric partial correlations between items; where a stronger connection is shown as a thicker and more saturated edge. Positive and negative connections are denoted by green and red edges respectively. Dalege et al. (2016) explained that these reactions tend to cluster and therefore adhere to a small-world structure. However, when inspecting the networks, it visually shows that the networks do not adhere to the three domains of an attitude. The smallworldindex indi-cates that these networks do not cluster, because none of the three networks (1.01, 0.84, 1.04 respectively) exceeds the con-servative small-worldness criterion threshold of 3 (46). Others (47) use a threshold of 1 to conclude that there is clustering. Therefore, the Average Shortest Path Length (ASPL) - which is the average of all shortest path lengths between all nodes in the network, where a low L indicates high connectivity -and communities (48) - highly interconnected groups with few links to other groups - were estimated, because a smallworld index could also mean that a network has high clustering, but a small average shortest path length (24). We found ASPLs in the networks of 6.73, 15.99 and 10.57 respectively, indicating that especially the network of the meat consumers is strongly connected. The walktrap algorithm is able to detect commu-nities and showed that the meat consumers network contained
Fig. 2. Probability of finding a specific value for the three classes. Class 1 exhibits a pattern expected from omnivores (meat consumers),
class 2 from flexitarians (meat reducers) and class 3 from vegetarians and vegans combined (meat avoiders). no clusters. Furthermore, it detected two big clusters in the
meat reducers network; the first contained environment, taste, disgust and dinner and the second nutrition, abuse, tax, guilt, reduce. Lastly, it detected one cluster in the meat avoiders network; consisting of nutrition, abuse, tax, sad, guilt, dinner, home cooking, replacement, and reduce. Thus, none of the networks showed clustering or clustering according to the three domains combined with a small ASPL, which is in line with the results of the smallworldindex. Since no clustering was found, we investigated internal consistency within the attitude. All networks showed moderate internal consistency (.64, .55, .75 respectively), indicating that the evaluative reactions in the networks are aligned.
When visually expecting the networks we find several simi-larities. The meat consumer and meat reducer networks show both stronger connections between the nodes home cooking -dinner - restaurant, which means that people who consume meat at dinner also consume meat at restaurants and at home and vice versa. Both networks also show stronger connections between guilt - disgust - taste - sad. Furthermore, the net-works of the meat reducers and meat avoiders show several similarities as well. Both networks show stronger connections between home — dinner - restaurant and abuse — environ-ment - reduce. As expected, the networks of meat consumers and meat avoiders show less similarities.
Network centrality.Network connectivity provides general
infor-mation on the dynamics of the network, but the centrality of a node provides information on how change in a particular node would affect the rest of the network. Figure 5 shows the centrality measures for all items per group. The first centrality measure considered was node strength, which measures the weighted number of connections of a focal node and thereby the degree to which that node is involved in the network. This measure only considers the local structure of the focal node. In the meat consumers network home cooking, dinner and en-vironment have a high strength, whereas enen-vironment, reduce
and taste are more prominent in the meat reducer network. Reduce, taste and environment have high strength in meat avoiders.
The weak nodes in the meat consumers network are nu-trition, tax and disgust, while the weak nodes in the meat reducers are replacement, nutrition and sad and in the meat avoiders are nutrition, tax and dinner. Betweenness and close-ness are discussed in the Supplementary Materials.
Stability.Next, we checked accuracy of the networks by
non-parametric bootstrapping. The results revealed sizable boot-strapped CIs in all three groups (Supplementary Materials), which indicates overlapping CIs for most edge weights. Over-lapping CIs indicate that many edge-weights likely do not significantly differ from one-another, which means that inter-preting the order of most edges in the network should be done with care. Nonetheless, it revealed that the strongest edges are substantially stronger than others.
Additionally, we investigated the centrality stability (CS)-coefficients for all centrality measures. The CS-coefficient for strength centrality for the three networks was 0.2, 0, 0.2 for networks 1 through 3 respectively. The meat consumer network shows moderate stability, while the other networks shows less stability. Other centrality measures showed no to small stability (betweenness: 0, 0.13, 0.05 and closeness: 0, 0.5, 0.05). Even though we found small stability coefficients, literature (35) suggests that strength is the most reliable measure. Details are available in the Supplementary Materials.
Network comparison.Network connectivity provides general
in-formation on the dynamics of the network. A common index of network connectivity is the Average Shortest Path Length (L; West, 1996), which is the average of all shortest path lengths between all nodes in the network. A low L indicates high connectivity. The connectivity of the meat consumers, meat reducers and meat avoiders networks are 6.73, 15.99 and 10.57 respectively, which shows that the network of meat consumers
Fig. 3. Network structure of meat consumers (N = 158), meat reducers (N = 122) and meat avoiders (N = 370) combined. The domain affect is depicted with light gray nodes,
behavior with black nodes and cognition with dark gray nodes. Green edges represent positive associations, whereas red edges represent negative association. Thicker edges represent stronger association. abu indicates animal abuse; dinner, consumption of meat during dinner; dis, feeling of disgust; env, harmful for environment, gui, feeling of guilt; hom, consuming meat at home; nut, meat has nutritional value; red, consciously reducing intake; rep, usage of meat replacements; res, meat consumption in restaurant; sad, feeling of sadness; tas, meat is tasteful; tax; meat tax is necessary
is more connected than the meat reducers, as well as that the network of the meat avoiders is more connected than the meat reducers.
Additionally, we used the NCT to compare all edge weights across the networks. When comparing the meat consumer and the meat reducer network one edge (replacement - reduce) sig-nificantly differed, which implies considerable similarities for the rest of the network. When comparing meat reducers and meat avoiders two edges significantly differed (guilt - restau-rant; environment -reduce) and when comparing the meat consumers and meat avoiders networks no edges significantly differed. It should be noted that the NCT has low power, especially when sample sizes across groups are unequal.
Furthermore, we tested whether the global strength esti-mates of the networks significantly differed. The NCT showed that the meat consumers and reducers do differ in strength - defined as the weighted sum of the absolute connections (”strength = 28, p < 0.05) as well from the meat avoiders
network (”strength = 10, p < 0.05). It also showed that the
strength of the meat reducers network does not differ from the strength of the meat avoiders network (”strength= 18, p
= 0.07). It should be noted that this does not rule out the existence of local differences in the network structure, as statis-tical power to detect local differences is limited. Nevertheless, care must be taken in interpreting the differences between the networks.
Discussion
Behavioral change towards more sustainable lifestyles is needed, thus a strong foundation should be laid down to identify key indicators of behavioral change.
The first part of this study described the different dietary lifestyles reflected in the attitudes. The literature proposes two kinds of divisions. The first division suggests that there are four dietary lifestyles; omnivorous, flexitarian, vegetarian
and vegan (7). While the other division suggests there are only three dietary lifestyles; meat consumers, meat reducers and meat avoiders (8) - no distinction is made between people who only restrict their meat and fish intake (vegetarians) and who avoid all animal-derived products (vegans). These divisions are solely based on behavior and do not take other relevant domains of an attitude into account. Our analysis shows that the latter division is the most probable and that a division solely based on behavior does not reflect the different attitudes towards meat consumption. Finding three groups is in line with several studies that show that vegetarians and vegans experience similar emotions, e.g. Filippi et al., (2010) showed that vegetarians and vegans do not differ in empathy, while omnivores score significantly less. It also contradicts the research that suggests vegans - compared to vegetarians - hold stronger beliefs about meat eating, animal welfare, and the environment (12). However, in this study we did not take motives (health vs. ethical (12)) into account, thus further research is needed to make this division more conclusive. Additionally, this study focused on attitudes towards meat consumption. It could be possible that a distinction between vegetarians and vegans will be found when other aspects of animal utilization (e.g. wool, leather) are considered.
The second part of this study investigated similarities and differences in the networks of meat consumers, meat reducers and meat avoiders, while addressing the considerable concern of replicability in the network literature (50) (35). Specifically, we estimated networks for the three different kinds of attitudes toward meat consumption. Our results can be summarized as follows.
First, while literature suggest that the attitudes consist of three domains that cluster together (24), no clustering according to the CAN model was found in these attitudes. Second, the attitudes have different nodes that are central in the network. Third, while the structures of the network
Fig. 4. Network structure of each group; (a) the network structure of the meat consumers (N = 158), (b) network structure of the meat reducers (N = 122) and (c)meat avoiders
(N = 370). The domain affect is depicted with light gray nodes, behavior with black nodes and cognition with dark gray nodes. Green edges represent positive associations, whereas red edges represent negative association. Thicker edges represent stronger association. abu indicates animal abuse; dinner, consumption of meat during dinner; dis, feeling of disgust; env, harmful for environment, gui, feeling of guilt; hom, consuming meat at home; nut, meat has nutritional value; red, consciously reducing intake; rep, usage of meat replacements; res, meat consumption in restaurant; sad, feeling of sadness; tas, meat is tasteful; tax; meat tax is necessary
Fig. 5. The three node centrality measures of meat avoiders (red), meat consumers (green) and meat reducers (blue): strength, closeness and betweenness. Centrality indices
are shown as relative scores. Abuse indicates animal abuse; dinner, consumption of meat during dinner; disgust, feeling of disgust; environment, harmful for environment, guilt, feeling of guilt; home cooking, consuming meat at home; nutrition, meat has nutritional value; reduce, consciously reducing intake; replacement, usage of meat replacements; restaurant, meat consumption in restaurant; sad, feeling of sadness; taste, meat is tasteful; tax; meat tax is necessary
seemed to be relatively similar, the networks differed in degree of strength.
In the next sections, we discuss our results in more detail, and conclude with strengths and limitations.
In contrast to our hypothesis, the attitudes do not cluster as suggested by the CAN model (24). In the CAN model an attitude is composed out of three domains; affect, behavior and cognition. These domains are represented by the evaluative reactions and interaction between these reaction. The model predicts that similar evaluative reaction form tight clusters, which are connected by few edges. According to this theory, attitude networks should strive for consistency. Obviously, individuals are motivated to hold at least some-what accurate attitudes, thus striving for a consistent attitude is limited by the motivation to have an accurate attitude. Striving for consistency would lead to aligned evaluative reactions, while striving for accuracy can lead to unaligned attitudes. Thus, to deal with trade-off between consistency and accuracy, attitudes are supposed to show clustering. The attitudes towards meat consumption do not show clustering and therefore do not show this trade-off between accuracy and consistency. It seems that overall, these attitudes strive for consistency. This is indeed what we found when estimating the internal consistency within the attitude, because all networks showed moderate consis-tency within the attitude. While this is inconsistent with the CAN model, it is consistent with the different coping mecha-nisms someone can apply. Rothgerber (2014) discusses several dissonance reducing mechanisms and some of them seem to be relevant here; perceived behavioral change, denial of animal mind and pro-meat justification. First, there was a discrep-ancy in categorization between perceived dietary lifestyle and reported behavior, which suggest that people substitute actual change by convincing themselves and others that they avoid meat consumption. Research has shown that people claim to
be vegetarian, while consuming red meat, chicken or fish (12). Second, meat consumers and meat reducers evaluate animals more inferior than meat avoiders, which they could use to justify their meat consumption. These people enlarge dissimi-larities between animals and humans (51). For example, the discrepancy between “I eat animals” and “I don’t like to hurt animals” seems less important when the capacity of animals is diminished. Third, Rothgerber (2012) found that the more consumers support pro-meat taste statements (e.g., “I enjoy eating meat too much to ever give it up”) the greater their reported meat consumption. This is in line with our results, because meat consumers show more sadness when asked to re-duce their meat intake. It seems that these coping mechanisms are reflected in their attitudes, and therefore show consistency. Additionally, meat avoiders do not need to apply any coping mechanisms, because their emotions and beliefs are already consistent with their behavior, as suggested by the relatively high internal consistency in the attitude. However, further research should investigate whether these coping mechanisms indeed affect the attitudes.
Conceptualizing attitudes as networks also has implications for attitude change; they can be changed by using centrality measures, because high centrality indicates substantial influ-ence on the network. In the meat avoiders network the node animal abuse had high centrality, which means that it has a substantial influence on the state of the network. When looking at the responses of participants on this item it shows that meat avoiders associate meat with animal abuse, which is in agreement with other research (13) . Additionally, the node associated with environmental problems had high centrality in the meat reducers network, indicating that this node had a substantial influence on the state of the meat reducer network. When looking at the responses of participants on this item, it shows that meat reducers moderately associate meat with
environmental problems. This is in line with other research, which showed that although flexitarians equally care about environmental issues compared to vegetarians (8), most people are only moderately aware of the fact that meat production causes a stress to the environment (52). Further research is needed to investigate how these nodes can be used to drive behavioral change towards more sustainable dietary lifestyles or are only indicative of the current dietary lifestyle. It should be noted that a node with high degree is more difficult to influence, but if they change state the consequences for the network will be more pronounced (24).
Third, the structure of the networks seemed to be relatively similar (i.e., most edges were not significantly different), but they did differ in connectivity. In the CAN model, global connectivity of an attitude network is defined as attitude strength (24). In this study, the attitudes of meat consumers and avoiders showed relatively similar strength, while the attitude of meat reducers showed less strength. Dalege et.al. (2016) shows that highly connected attitudes more often align than weakly connected attitudes and that the aligned networks more often predict behavior. Firstly, this could mean that the attitude of meat reducers is less predictive of their behavior. Secondly, this could mean that attitude of meat reducers are relatively easy to change towards a more sustainable dietary lifestyle. This also means that attitudes of meat consumers -and in some degree, of meat avoiders - are relatively difficult to change. The catastrophe theory predicts that high involvement makes it more difficult to change the attitude (53). Both meat consumers and meat avoiders experience strong emotions (54), which support the idea that these attitudes have high global strength and would be difficult to change.
Strength and limitations.The particular strengths of this study are that it offers a plausible representation of the structure of attitudes and that it integrates inconsistencies between attitude and behavior, as the model assumes that behavior is part of the attitude. Importantly, it allows for application to actual data, while still maintaining explanatory power. When the CAN model is used, findings can be integrated, indicators of behavioral change can be identified and inconsistencies between behavior and attitude can be overcome. At the same time, we must acknowledge several limitations.
First, the network perspective has its limitation on its own. Currently the model can only be applied on group data while key indicators of behavioral change differ between individuals. However, no such method has been developed yet and steps need to be conducted to test whether this model is also utile in individual settings. In extension of this limitation, this study also combined a latent class analysis - which is used to explain underlying, unobservable categorical relationships by identifying correlations between the items - with a network analysis, which investigates correlations too. Therefore, it could be the case that the latent class analysis affected the network analysis and made it more difficult to interpret.
Second, the current replication crisis in psychology (55) stresses the importance of replicability, thus testing for ac-curacy and robustness of the networks is important. Our results showed that accuracy of the networks was low and interpreting the networks should be done with care. However, it should be noted that networks with larger sample sizes are estimated more accurately, because it is easier to detect dif-ferences between centrality measures and it increases stability
of the network (35). It is unclear how many observations are needed, thus poor accuracy in this study could be due to the limited sample size.
Third, the recent develop NetworkComparisonTest exhibits some limitations. Borkulo et. al. (2016) explained that in the network structure invariance and the edge strength invariance test power is higher for less densely-connected networks. It could be that - due to our densely-connected networks - power was low and the test was unable to show a difference. For the global strength invariance test, however, this is reversed: power is higher for more densely connected networks, which could explain why we did found a difference in global strength. Additionally, they suggest that this test does not perform optimal under unequal sample sizes, which was true for our dataset. Again, this could have altered the results.
The last limitation, is that although the network perspective looks promising, it does assume no latent variables while in some cases latent variables can be an accurate representation of an attitude (56).
Conclusion.In conclusion, this study found that 1) the
atti-tudes towards meat consumption do not cluster as suggested by the CAN model, 2) that the attitudes of meat consumers, meat reducers and meat avoiders differ in strength (connec-tivity) and 3) that the nodes that influence the rest of the attitude network differ per attitude. Additionally, it highlights the need for more research into individual changes in attitudes, subsequently key indicators of behavioral change towards more sustainable dietary lifestyles can be identified.
ACKNOWLEDGMENTS. We would like to thank H.L.J. van der
Maas from the University of Amsterdam for his suggestions and feedback during this project, and all participants who shared the questionnaire to collect data for this study.
References
1. Bruinsma J, et al. (2015) World agriculture: Towards 2015/2030: An fao perspective. fao, rome. World agriculture: towards 2030.
2. Steinfeld H, Gerber P, Wassenaar T, Castel V, de Haan C (2006) Livestock’s long shadow: environmental issues and options. (Food & Agriculture Org.).
3. Clonan A, Wilson P, Swift JA, Leibovici DG, Holdsworth M (2015) Red and processed meat consumption and purchasing behaviours and attitudes: impacts for human health, animal welfare and environmental sustainability. Public health nutrition 18(13):2446–2456. 4. Walker P, Rhubart-Berg P, McKenzie S, Kelling K, Lawrence RS (2005) Public health
implica-tions of meat production and consumption. Public health nutrition 8(04):348–356. 5. Grandin T (2014) Animal welfare and society concerns finding the missing link. Meat science
98(3):461–469.
6. Lang T, Barling D (2013) Nutrition and sustainability: an emerging food policy discourse. Proceedings of the Nutrition Society 72(01):1–12.
7. Dagevos H, Voordouw J (2013) Sustainability and meat consumption: is reduction realistic? Sustainability: Science, Practice, & Policy 9(2).
8. De Backer CJ, Hudders L (2014) From meatless mondays to meatless sundays: motivations for meat reduction among vegetarians and semi-vegetarians who mildly or significantly re-duce their meat intake. Ecology of food and nutrition 53(6):639–657.
9. Kenyon P, Barker M (1998) Attitudes towards meat-eating in vegetarian and non-vegetarian teenage girls in england—an ethnographic approach. Appetite 30(2):185–198.
10. Barnes-Holmes D, Murtagh L, Barnes-Holmes Y, Stewart I (2010) Using the implicit associa-tion test and the implicit relaassocia-tional assessment procedure to measure attitudes toward meat and vegetables in vegetarians and meat-eaters. The Psychological Record 60(2):287. 11. De Houwer J (2003) The extrinsic affective simon task. Experimental psychology 50(2):77. 12. Ruby MB (2012) Vegetarianism. a blossoming field of study. Appetite 58(1):141–150. 13. Beardsworth A, Keil T (1992) The vegetarian option: varieties, conversions, motives and
careers. The Sociological Review 40(2):253–293.
14. Povey R, Wellens B, Conner M (2001) Attitudes towards following meat, vegetarian and vegan diets: an examination of the role of ambivalence. Appetite 37(1):15–26.
15. Elliot AJ, Devine PG (1994) On the motivational nature of cognitive dissonance: Dissonance as psychological discomfort. Journal of personality and social psychology 67(3):382. 16. Berndsen M, Van der Pligt J (2004) Ambivalence towards meat. Appetite 42(1):71–78. 17. Van Harreveld F, Van der Pligt J, Yael N (2009) The agony of ambivalence and ways to resolve
it: Introducing the maid model. Personality and Social Psychology Review 13(1):45–61. 18. Rothgerber H (2014) A comparison of attitudes toward meat and animals among strict and
19. Graça J, Calheiros MM, Oliveira A (2015) Attached to meat?(un) willingness and intentions to adopt a more plant-based diet. Appetite 95:113–125.
20. Jabs J, Devine CM, Sobal J (1998) Model of the process of adopting vegetarian diets: Health vegetarians and ethical vegetarians. Journal of Nutrition Education 30(4):196–202. 21. Hamilton M (2006) Eating death: vegetarians, meat and violence. Food, Culture & Society
9(2):155–177.
22. Rothgerber H (2015) Can you have your meat and eat it too? conscientious omnivores, vegetarians, and adherence to diet. Appetite 84:196–203.
23. Bagozzi RP, Tybout AM, Craig CS, Sternthal B (1979) The construct validity of the tripartite classification of attitudes. Journal of Marketing Research pp. 88–95.
24. Dalege J, et al. (2016) Toward a formalized account of attitudes: The causal attitude network (can) model. Psychological review 123(1):2.
25. Herzog Jr HA, Betchart NS, Pittman RB (1991) Gender, sex role orientation, and attitudes toward animals. Anthrozoös 4(3):184–191.
26. Bernstein IH, Nunnally J (1994) Psychometric theory. New York: McGraw-Hill. Oliva, TA, Oliver, RL, & MacMillan, IC (1992). A catastrophe model for developing service satisfaction strategies. Journal of Marketing 56:83–95.
27. Qualtrics (2016) Qualtrics (Qualtrics, Provo, Utah, USA).
28. R Core Team (2016) R: A Language and Environment for Statistical Computing (R Foundation for Statistical Computing, Vienna, Austria).
29. Linzer DA, Lewis JB (2011) poLCA: An R package for polytomous variable latent class anal-ysis. Journal of Statistical Software 42(10):1–29.
30. Schwarz G, , et al. (1978) Estimating the dimension of a model. The annals of statistics 6(2):461–464.
31. Lin TH, Dayton CM (1997) Model selection information criteria for non-nested latent class models. Journal of Educational and Behavioral Statistics 22(3):249–264.
32. Forster MR (2000) Key concepts in model selection: Performance and generalizability. Jour-nal of mathematical psychology 44(1):205–231.
33. Goodman LA (1970) The multivariate analysis of qualitative data: Interactions among multiple classifications. Journal of the American Statistical Association 65(329):226–256. 34. Epskamp S, et al. (2012) qgraph: Network visualizations of relationships in psychometric
data. Journal of Statistical Software 48(4):1–18.
35. Epskamp S, Borsboom D, Fried EI (2016) Estimating psychological networks and their accu-racy: a tutorial paper. arXiv preprint arXiv:1604.08462.
36. Borkulo CD, Waldorp KJ, Boschloo L, Schoevers RA, Borsboom D (2015) Statistical compar-ison of two networks with respect to global strength.
37. Lauritzen SL, Wermuth N (1989) Graphical models for associations between variables, some of which are qualitative and some quantitative. The annals of Statistics pp. 31–57. 38. Friedman J, Hastie T, Tibshirani R (2008) Sparse inverse covariance estimation with the
graphical lasso. Biostatistics 9(3):432–441.
39. Tibshirani R (1996) Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B (Methodological) pp. 267–288.
40. Foygel R, Drton M (2010) Extended bayesian information criteria for gaussian graphical mod-els in Advances in neural information processing systems. pp. 604–612.
41. Opsahl T, Agneessens F, Skvoretz J (2010) Node centrality in weighted networks: Generaliz-ing degree and shortest paths. Social networks 32(3):245–251.
42. Facts C (1997) National institute of nutrition. Tracking nutrition trends 1989–1994–1997. 43. Prättälä R, et al. (2007) Gender differences in the consumption of meat, fruit and
vegeta-bles are similar in finland and the baltic countries. The European Journal of Public Health 17(5):520–525.
44. Gossard MH, York R (2003) Social structural influences on meat consumption. Human Ecol-ogy Review 10(1):1–9.
45. Nylund KL, Asparouhov T, Muthén BO (2007) Deciding on the number of classes in latent class analysis and growth mixture modeling: A monte carlo simulation study. Structural equa-tion modeling 14(4):535–569.
46. Humphries MD, Gurney K (2008) Network ‘small-world-ness’: a quantitative method for de-termining canonical network equivalence. PloS one 3(4):e0002051.
47. Watts DJ, Strogatz SH (1998) Collective dynamics of ‘small-world’networks. nature 393(6684):440–442.
48. Pons P, Latapy M (2005) Computing communities in large networks using random walks in ISCIS. Vol. 3733, pp. 284–293.
49. Filippi M, et al. (2010) The brain functional networks associated to human and animal suffer-ing differ among omnivores, vegetarians and vegans. PLoS One 5(5):e10847.
50. Forbes MK, Wright AG, Markon K, Krueger R (2017) Evidence that psychopathology symp-tom networks have limited replicability.
51. Bilewicz M, Imhoff R, Drogosz M (2011) The humanity of what we eat: Conceptions of hu-man uniqueness among vegetarians and omnivores. European Journal of Social Psychology 41(2):201–209.
52. Vanhonacker F, Van Loo EJ, Gellynck X, Verbeke W (2013) Flemish consumer attitudes to-wards more sustainable food choices. Appetite 62:7–16.
53. Van der Maas HL, Kolstein R, Van Der Pligt J (2003) Sudden transitions in attitudes. Socio-logical Methods & Research 32(2):125–152.
54. Rozin P, Markwith M, Stoess C (1997) Moralization and becoming a vegetarian: The trans-formation of preferences into values and the recruitment of disgust. Psychological Science 8(2):67–73.
55. Collaboration OS, , et al. (2015) Estimating the reproducibility of psychological science. Sci-ence 349(6251):aac4716.
56. Epskamp S, Rhemtulla M, Borsboom D (2016) Generalized network psychometrics: Combin-ing network and latent variable models. arXiv preprint arXiv:1605.09288.
