Perceptions of emerging biotechnologies

An online survey was distributed in April 2017 to a representative sample (n = 521) of the United States by Qualtrics, with the intent to obtain a sample that was roughly 68% White, 15% Hispanic, 12% African American, and 5% Asian. The survey established basic demographics about respondents including age, gender, ethnicity, level of education, political ideology, and income. Previously established groups of questions were included to measure pro-environmentalism (New Environmental Paradigm: NEP) [15], personal values [16], and scientific reasoning skills [4]. In addition to these established scales, questions based on the willingness to sacrifice framework [17], about levels of trust in academic, government, and industry scientists, and about religious beliefs were also used to generate scales. The consistency of each scale was determined using Chronbach's α. While originally planned as a scale, two questions regarding religion were used as stand-alone variables in order to distinguish between religious fundamentalism (Q: I believe the bible is the literal word of God) and general religiosity (Q: Religion is important in my life).

To measure opinions on biotechnology, respondents were presented with five different blocks each containing a different biotechnology product presented in randomized order (table S1) and asked to score their agreement with the following statements in respect to each technology on a Likert Scale (1: Strongly disagree to 5: Strongly agree, 3: Neither agree nor disagree):

• 1.  
  I am familiar with this technology.
• 2.  
  I support the development of this technology.
• 3.  
  I believe this technology is beneficial.
• 4.  
  I believe this technology is risky.
• 5.  
  I believe this technology should be regulated.

A series of statistical analyses were conducted in this study. First in section 3.1, multiple linear regression models were used to explain the relationship between independent variables (e.g. demographics, personal values) and general opinions toward biotechnology (i.e. the dependent variables). To measure general opinions, scales were generated for each of the above questions using responses to all five biotechnology products (see table S2). Next in section 3.2, we used analysis of variance (ANOVA) tests to determine if there was an effect of biotechnology product (i.e. independent variable) on responses to each of the five questions (i.e. dependent variables). For questions with significant differences between products (ANOVA; p ≤ 0.05), post-hoc Tukey's Honest Significant Difference (HSD) test was used to determine which biotechnology products were different. Additionally, we used the standard deviation to determine the extent to which an individual's responses to the five questions differed across the five biotechnology products. All statistical analysis was conducted in R 3.3.3.

To identify common biotechnology opinion patterns [18], responders were clustered based on their support for and their beliefs about benefits, risks, and regulation of all five biotechnology products. Two clustering algorithms (k-means and Partitioning Around Medoids (pam)) were tested with the clustering number (k) ranging from 3 to 10. The best algorithm (k-means) and k (k = 5) were selected using the 'clValid' package in R using silhouette width as a validation metric (figure S1(A)). Because k-means clustering is not deterministic, clustering was performed once, and those cluster assignments (figure S1(B)) were used for the remainder of the study.

To determine how well independent features could classify individuals into their biotechnology opinion group we trained multiclass machine learning models using each feature individually and with all features combined [19]. We randomly generated 10 datasets that were down-sampled to include equal numbers of individuals from each opinion group (i.e. balanced datasets). Both the parameter sweep (used to select the best combination of parameter values, see table S3) and training were done within a 5-fold cross validation scheme to avoid overfitting. Briefly, for 5-fold cross-validation, individuals are divided into five groups, then groups 1–4 are used for the parameter sweep and training and group 5 for testing, next groups 1–3 and 5 are used for the parameter sweep and training and group 4 for testing, etc. Model performance was measured using the F-measure (F1), or the harmonic mean of precision (True Positive Rate (TPR)/TPR + False Positive Rate (FPR)) and recall (TPR/TPR + False Negative Rate (FNR)), where an F1 is generated for each opinion group and the macro F1 represents the mean F1 from all five opinion groups. Because individuals could be classified into 1 of 5 opinion groups, a machine learning model randomly guessing would have an F1 = 0.2, while a perfect model would have an F1 = 1. To ensure model performance scores were not artificially inflated [19], only the performance on the testing data, which is withheld from the model during training, was reported.

All single feature models were built using the Logistic Regression algorithm. Note that binary feature (e.g. Hispanic, lived on a farm) are likely to perform below random expectation when used individually to classify individuals as belonging to one of five opinion groups However, four additional algorithms were tested for the combined model: Support Vector Machine (SVM), SVM with a polynomial kernel (SVMpoly), Random Forest, and Gradient Boosting. All algorithms were implemented in Scikit-Learn using Python3.

The survey questions and code used to process, perform statistical analyses, and cluster the survey results can be downloaded from GitHub at https://github.com/azodichr/Biotech_Survey. The python code (ML_classification.py) used to train and test machine learning models is also available on GitHub at https://github.com/ShiuLab/ML-Pipeline. Any data that support the findings of this study are included within the article. See independent (Data S1) and dependent variables (Data S2).

Our respondents ranged from age 18–86 (median = 48). Of our respondents 33% were male, 72% were white, 15% were Hispanic, and 9% reported they were vegetarian/vegan (table S4).

Across the five biotechnology products, support was positively correlated with beliefs about the benefits of each technology (Pearson's Correlation Coefficient (PCC): 0.57–0.86) and negatively correlated with beliefs about the risks of each technology, although to a lesser degree (PCC −0.14 ∼ −0.27) (table S5). We used multiple linear regression to assess the association between independent variables and beliefs and opinions about biotechnology in general by averaging response over the five biotechnology products. Overall support for biotechnology ('I support the development of this technology') was most strongly associated with self-reported familiarity with biotechnology (p < 0.001), familiarity with agriculture (p = 0.005), and with gender (p < 0.001), with men being more supportive than women [20] (table 1). Because research on risk perception has shown what is referred to as the 'white-male' effect, where white men tend to judge risk as lower [21], we tested for an interaction between the male and white variables. We found a small white-male effect (effect size = 0.36; p = 0.04) on the belief that the biotechnology products are beneficial ('I believe this technology is beneficial'), where white men tended to believe more strongly in the benefits. However, the white-male effect was not present for beliefs about risks, benefits, or the need for regulation. Taken together, this suggests the 'white-male' effect was minimal in our survey. Two personal values were also significantly associated with beliefs about the benefits of biotechnology, with openness being associated with the belief that biotechnology is beneficial (p = 0.02) and biospheric altruism (i.e. altruism toward other species or the environment in general) being associated with the belief that biotechnology is not beneficial (p = 0.02) (table 1).

It has been demonstrated that extreme opposition to GM foods is associated with low objective knowledge about science and genetics [3]. While we found perceived riskiness and the need to regulate biotechnology were most strongly associated with the ability to think scientifically (p = 0.002 and 0.001, respectively; table 1), we found that those with a greater ability to think scientifically were more likely to believe biotechnology was risky and needed to be regulated (figure 1(A)). This suggests that individuals with generally negative opinions about biotechnology tend to have stronger scientific reasoning skills. Furthermore, self-reported trust in academic, government, and industry scientists was also associated with the perception of high risks for biotechnology products (p = 0.003), which counterintuitively suggests that individuals who trust scientists are more likely to believe that these products, developed by scientists, are risky. This is at odds with previous work that found trust in scientists lowered perceived risks of gene technologies [22] and with work that found trust in industry scientists was more strongly associated with support for biotechnology than knowledge about genetics [23]. Finally, and in line with previous findings [24, 25], we found that religious fundamentalism was strongly associated with the beliefs in the risks and need to regulate biotechnology.

Zoom In Zoom Out Reset image size

Next, we assessed the differences in beliefs and opinions about each of the five biotechnology products using ANOVA followed by Tukey's HSD tests when the ANOVA was significant (p < 0.05). We found that products related to agriculture were believed to be the least beneficial and had the least support, with support for hornless cattle significantly lower than the other products (table S6; figure 1(B)), while biopharmaceuticals were seen as the most beneficial and had the most support (HSD: p < 0.05). While these differences were significant, we found that 27% of respondents reported the same degree of support across all products (standard deviation (sd) = 0), and an additional 60% had minor differences in support (sd < 1) (figure 1(C)) between products. Furthermore, the perceived riskiness of biotechnology varied only moderately by product, with gene drives perceived as riskier and biopharmaceuticals perceived as less risky (p = 0.047; figure 1(B); table S6) and the no difference in opinions about the need for regulation across products (p = 0.23). Surprisingly, 50% (n = 261) of individuals agreed or strongly agreed that all five biotechnology products should be regulated, with an additional 14% (n = 71) believing four of the five products should be regulated and this response was not associated with political affiliation (p = 0.25; table 1). Taken together, this suggests most US adults support regulating biotechnology and have similar beliefs and opinions about all biotechnology products regardless of their unique benefits and risks.

To better understand common patterns of opinions and beliefs about the different biotechnology products, we used k-means clustering to sort respondents into five opinion groups (see Methods). Three of these opinion groups (i.e. A, D, E) had similar opinions about biotechnology across the five products (figure 1(D)). For example, individuals in opinion group A (n = 43) had strong negative opinions toward all biotechnology products, specifically they did not support or believe any product was beneficial (dark green), On the other hand, individuals in opinion group E (n = 119) strongly supported all biotechnology products and viewed them as beneficial (see dark purple). Interestingly, individuals in both of these response groups believed all biotechnology products were at least moderately risky and needed to be regulated. Individuals in two opinion groups, B (n = 68) and C (n = 131), had different opinions about different biotechnology products. While moderately supportive of some products, group B respondents were opposed to biofortified crops and hornless cattle, believing those products were risky and not beneficial. Of note, unknown long-term human health effects were listed as possible risks in the description of those two products, suggesting group B individuals were driven by their fear of negative impacts on human health (see table S1). Conversely, group C respondents believed the risks were similar across all products, however, only believed the products that explicitly addressed human health (i.e. biopharmaceuticals, biofortified crops, and gene drives) were beneficial. This suggests group C individuals were driven by their hope for products that improve human welfare.

No single demographic factor clearly distinguished individuals from the five opinion groups (table S4). Therefore, to better understand the social basis for these opinion groups, we evaluated 29 independent features derived from the survey data (see table S7) for their ability to predict an individual's opinion group using logistic regression. The logistic regression models were trained, and their performance was tested using a balanced, replicated, 5-fold cross validation scheme (see Methods 2.4). We used the F-measure (F1), or the harmonic mean of precision and recall, as a metric to measure predictive performance, where F1 = 1 would indicate the model correctly predicted the opinion group of every respondent and an F1 = 0.2 would indicate the model performed no better than random guessing (i.e. if the model randomly assigned each individual to one of the five opinion groups, the model would be correct 20% of the time). While seven features were able to distinguish between opinion groups better than random expectation (F1 > 0.2) (figure 2(A)), the features differed in which opinion groups they were able to distinguish (figure 2(B)). For example, across all five opinion groups familiarity with biotechnology had a performance (referred to as F1_macro) of F1_macro = 0.27 (± 0.04 sd), making it the best individual predictor of opinion groups. While it performed exceptionally well at identifying individuals from opinion group E (F1_E = 0.53 ± 0.04), with 73% of opinion group E individuals correctly predicted as belonging to group E (figure 2(B)), it was less useful for distinguishing between individuals belonging to the less supportive opinion groups (A–D). Biospheric altruism, on the other hand, performed just above random expectation overall (F1_macro = 0.224 ± 0.022), but performed well at classifying individuals from opinion group D (F1_D = 0.34 ± 0.10), with 51% of opinion group D individuals correctly predicted as belonging to group D. Such differences indicated a need to jointly consider all features in one model.

Zoom In Zoom Out Reset image size

To do so we used a machine learning approach (see Methods), that resulted in an integrative model that was more accurate (F1_macro = 0.37 ± 0.02) than any individual feature model (gray bar; figure 2(A); table S7). As expected from the overall better performance of individual feature models at classifying the more supportive opinion groups, the integrative model performed best at classifying individuals from opinion groups D and E (F1_D = 0.43 ± 0.06; F1_E = 0.54 ± 0.04; figure 2(B), S1). However, the model was less able to distinguish between non-supportive (opinion group A) and conditionally supportive (opinion group B and C) respondents than individuals from groups D and E (figure S2). While disappointing, this difficulty in itself suggests there is not a single social basis for different patterns of negative opinions about biotechnology, but rather that individuals with a broad range of personal values, religious and political beliefs, and educational backgrounds can have negative opinions about all or some biotechnology products.

Finally, to determine the degree and direction of the relationship between each factor and the opinion groups, we compared the distribution of responses for each factor across all five opinion groups. While differences were observed for a number of factors (figure S2), we chose to focus on those that were most predictive in the individual models. While individuals from the non-supportive (A) and conditionally supportive (B and C) opinion groups did not tend to differ for these factors, we found notable differences between the moderately (D) and strongly (E) supportive opinion groups (figure 2(C)). For example, opinion group E individuals had the highest median self-reported familiarity with the biotechnology products and were more willing to sacrifice for biospheric and animal welfare causes. While, individuals from the moderately supportive group (D) were less willing to sacrifice for these causes and were less likely believe in the New Ecological Paradigm (NEP) [15], a scale used to measure general pro-environmental beliefs. This highlighted the large gap in the social basis associated with moderate compared to strong support for biotechnology, where strong supporters tend to support, and moderate supporters tend to have neutral opinions regarding environmental issues.