Received 25 February 2016; accepted 3 April 2016; published 6 April 2016
To address this question seven pork attributes were investigated. Animal welfare, environmental impacts, locally raised/farmed pigs, and locally processed pork were analyzed in order to incorporate some of the commonly discussed production attributes. These four attributes were studied in combination with the three product attributes identified by Lusk and Briggeman  as important food values, namely price, food safety, and taste.
2. Materials and Methods
2.1. Survey Instrument and Data
A total of 1004 completed responses were collected. In total 1924 individuals clicked on the survey link (entered the survey) at some point during the data collection process, although 1004 completed the survey in its entirety during the data collection time period. Thus, 52% of those respondents who clicked on the survey link to enter the survey completed it during the data collection time period. Table 1 contains the basic demographics of the survey respondents compared with the census statistics for age  , gender  , and income  and geographic location of residence  .
2.2. Maximum Difference Scaling
Maximum difference scaling, commonly referred to as best-worst scaling, was originally used by Finn and Louviere  . The best-worst scaling analysis was explained more thoroughly by Marley and Louviere  who clarified the underlying method. The method used by Lusk and Briggeman  , which utilized recent advances in this type of modeling, was adapted for this analysis to ultimately determine consumers’ shares of preference for each of the studied pork attributes. This method forces survey participants to make tradeoffs between attributes of a particular subject. In this analysis, participants were forced to make tradeoffs between seven pork attributes. The specific attributes involved in this study were: animal welfare, price, pork/food safety, taste, environmental impacts, locally raised/farmed pigs, and locally processed pork.
The best-worst scaling methodology has advantages over the commonly used Likert-scale-type questions or ranking-type questions which seek the same general insight. When participants are asked to rank attributes in relative importance the result is only the rank (also referred to as order). Ranking-type questions do not provide a way for understanding the magnitude of importance of the individual attributes relative to one another. It is easy for participants responding to Likert-scale types of questioning to classify all attributes as important or equally on the scale, both of which are pitfalls to the rating question type according to Lusk and Briggeman  . Also, the rating method scale can be interpreted differently by different individuals making the results unclear. According to Wolf and Tonsor (  , pg. 231) “unlike, for example, the Likert-scale-type questions, best-worst scaling is cardinal;” thus it allows the analyst to determine the ordering (or rank) of attributes in addition to establishing how important attributes are to consumers.
To determine both ordering and relative importance of the seven pork attributes for consumers, best-worst scaling was utilized. A balanced incomplete block design that optimized positional frequencies and allowed for question sets of equal size was used. Respondents were asked to indicate from a set of four pork attributes which attribute in the question (choice set) shown was the most important (best) and the attribute that was the least important (worst) to them when they purchased pork. Every respondent saw the same seven choice sets, where each choice set showed four attributes. Each attribute was shown four times within the best-worst design and all possible best-worst pairs had the potential to be made two times throughout the seven choice sets. Within each choice set respondents were unable to select the same attribute in a choice set as both the most and least important.
2.3. Best-Worst Analysis
The best-worst question presented 7 pork attributes (K) to consumers, thus for this analysis. There was a total of best-worst combination possibilities any individual consumer could make. Given that consumers were shown seven choice sets with four attributes each, consumers had the opportunity to make
Table 1. Demographics of survey respondents.
twelve combinations per question leading to a total of 84 combination possibilities (12 combinations per question × 7 questions = 84 combinations). Participants’ i’s latent, unobservable, level of importance of the attribute k was where λk was the location of k on the underlying scale of importance of pork attributes, and was the random error term. If are i.i.d. type 1 extreme value, then the probability of any specific one attribute chosen as most important and any other chosen specific attribute as least important was of the multinomial logit (MNL) form following Lusk and Briggeman  , shown in equation (1).
Then, λk was estimated with the maximum of the log-likelihood function. The pair of attributes takes on the value 1 for the pair that the individual chose as most and least important, and 0 for the remaining pairs not chosen by the participant. The estimated λk was thus the value of attribute k relative to the attribute that was normalized to zero in the estimation of the model.
The RPL model assumes that is equal to one. The random error term, however, can vary between persons and thus it is possible that the mean of the parameter estimates of λk may be confounded with differences in scale. Thus the results from the parameter estimates in the RPL model are not easily interpretable. Following Lusk and Briggeman  to address the potential confound in scale the shares of preference, S, (the forecasted probability that the attribute k is chosen as best) were calculated following Equation (2) for each pork attribute examined.
The shares of preference for each of the seven attributes sum to one by design. Individual consumer-specific shares of preference for each of the seven attributes as well as the mean preference share for the entire dataset are estimated1. Shares of preference “can be analyzed to reveal cardinal rankings and respondent characteristics associated with those rankings” (  , pg. 222). Individual-specific shares of preference facilitate analysis of relationship between shares; relationships were evaluated utilizing correlations completed in SPSS.
2.4. Seemingly Unrelated Regression Analysis
To gain deeper insight into the potential relationships between the size of the share of preference for the pork attributes and consumer demographics, a seemingly unrelated regression was used. The intended goal was to determine which key pieces of demographic information were statistically significant determinants of the size of the share of preference for each of the pork attributes studied. When attempting to analyze determinants of the shares of preference for each attribute, it is possible that the standard errors could be correlated between the models. Seemingly unrelated regressions allow for simultaneous estimation of the set of models while at the same time accounting for correlated errors  , thus the seemingly unrelated regression was used.
The specific models used in this analysis were specified such that the dependent variables were the size of the share of preference for each attribute and the set of independent variables in each regression were the same set of key demographics along with the result of a simple validation test result (pass vs. fail)2. The set of specific demographic factors studied included gender, age (using 65 and older as the base category), income (with a base category of greater than $150,000), college education, pet ownership, having a source for animal welfare, diet type and pork purchasing behavior. Following the seemingly unrelated regression, Wald tests were employed such that the null hypothesis was that the coefficient estimate for an independent variable across the set of equations was concurrently equal to zero. The seemingly unrelated regression analysis and related testing was completed in Stata 12.
3. Results and Discussions
To gain deeper insight into potential relationships among shares of preference for the seven attributes studied, correlations amongst individual-specific preference shares were analyzed. The correlations (and their significance) between shares of preference are shown in Table 3. The size of the share of preference for animal welfare was statistically significant and negatively correlated with the size of the share of preference for the attributes, price, safety, and taste; meanwhile, it was statistically significant and positively correlated with the size of the share of preference for environmental impact, locally raised/farmed pigs and locally processed pork. The size of the share of preference for price was statistically significant and negatively correlated with the size of the share for all other attributes, excluding taste; this is reassuring in that participants that had larger preference shares for price tended to trade it off with the other attributes, as would be expected.
To gain deeper insight into the potential relationships between the size of the shares of preference for the pork attributes and consumer demographics, the seemingly unrelated regression was completed. Coefficient estimates,
Table 2. RPL results and shares of preference for pork attributes.
Note: Individuals made 7 choices and there were 1004 individuals, thus there were 7028 observations. ***, **, * indicate statistical significance at 1%, 5%, 10% level respectively.
Table 3. Correlations of shares of preference between pork attributes (n = 1004).
Note: Statistical significance (2-tailed) at the 5% and 1% level is represented by * and ** respectively.
“R-squared”, and chi2-statistic information for the seemingly unrelated regression are displayed in Table 4. Each model has a chi2 statistic which is statistically significant at the 0.1% statistical significance indicating that the set of independent variables in each model are statistically significant. Immediately following the regression analysis, the set of tests was completed to determine if each independent variable was statistically different from zero across all models. The null hypothesis was that the coefficient estimate for an independent variable across the set of equations is concurrently equal to zero. The resulting p-values from these tests are presented in Table 5. If the value is less than 0.05 then it can be interpreted as the coefficient estimates associated with the particular variable examined was statistically different from 0 across all of the seven models.
Results from the seemingly unrelated regression revealed that being male and having purchased pork in the previous year were statistically significant determinants of the size of the share of animal welfare and had negative coefficient estimates. This finding supports those by Vanhonacker et al.  who found that men attached less importance to animal welfare relative to other production characteristics of livestock products than did women. Vanhonacker et al.  also found evidence of differences between consumption of meat products and concern for animal welfare which indicated that heavy meat consumers cared less about animal welfare relative to other production characteristics when compared with individuals reporting moderate or low meat consumption. Statistically significant determinants of the size of the shares of animal welfare preference with positive coefficient estimates included individuals in the age category 25 - 44 years old, individuals selecting income $50,000 to $74,000, having a source for animal welfare information, owning a dog, and owning a cat.
It is expected in primary data collection that the researcher make efforts to obtain the best data quality possible. Both survey design and data collection techniques are important elements commonly discussed regarding obtaining quality data. Guidelines for survey development, design, and administration so as to improve data quality have been discussed by many, including Dillman  . Various validation methods have been investigated in order to help identify participants who are not responding sincerely. The use of different data improvement methods and enhancements to these methods continue to be analyzed for the impacts in data quality improvement. The recent focus on attribute non-attendance in choice experiments   and the study of relatively simple validation questions  is evidence of interest in enhanced data quality controls. Validation questions are simple questions embedded within the survey or data collecting instrument to identify participants who are alert, reading, and understanding survey questions. These questions are simple to incorporate, do not take long for participants to answer, and the results (passing or failing the validation question test) can be applied to multiple questions or aspects of the survey. In their study of consumers’ willingness to pay for different animal products, Gao, House, and Bi  found, when examining the impact of a validation test, that there were differences between those who passed and those who did not pass the validation test. Gao, House, and Bi  suggested that the inclusion of a validation question would be a way to “detect careless respondents in the survey and improve data quality”. Thus, one question in the survey asked respondents to select the number six from the options one through seven. Those who correctly selected the number six were considered to have “passed” the validation test. Those who did not select six were considered to have “failed” the test. Eighty-five percent of
Table 4. Seemingly unrelated regression results.
Note: ***, **, * indicate statistical significance at 1%, 5%, 10% level respectively.
respondents passed the validation test question.
In addition to demographics for the entire sample, Table 1 displays the summary of those who passed the validation question. Statistically significant correlations existed between passing the validation test and income (positive correlation relationship), age (18 - 44 was negatively correlated with passing, age 65 years and older was positively correlated with passing the validation test), as was higher education at the 5% significance level. This finding is similar to Gao, House, and Bi  who found that respondents who failed the validation test tended to be younger, less educated, and had lower incomes.
Overall, passing the validation test question was a statistically significant determinant in each of the models evaluated. Thus this implies that passing the validation test is a statistically significant determinant when predicting the shares of preference for each of these seven pork attributes and implies that making use of a simple validation question within a best-worst analysis has the possibility of impacting the results.
While some people choose not to eat livestock-derived products, or meat products, there are many different reasons why individuals may choose to not purchase pork specifically. For example, religious beliefs may prohibit the consumption of pork products, individuals may not like or enjoy pork (relative to other meats), or consumers may be opposed to eating pigs (but find eating other species acceptable). Thus, pork consumption and purchasing was studied specifically by asking respondents to indicate if they had purchased pork in the past year.
Table 5. Test of the hypothesis that the coefficient for each independent variable is zero for all 7 outcome variables.
Eighty-three percent of respondents indicated they had purchased pork within the last year. Results indicate that having purchased pork in the last year was a statistically significant determinant of the size of the share of preference for the attribute animal welfare and had a negative coefficient estimate.
In McKendree, Croney and Widmar  pet owners (owners of dogs and/or cats) were found to be statistically more concerned for livestock animal welfare than non-dog and/or cat owners. They hypothesized that owning a pet may then also be related to seeking out animal welfare information given that pet owners more frequently reported having a source for animal welfare information. In addition, they examined relationships between reported concern for animal welfare and primary sources of animal welfare information. Their study found that an individual’s level of concern for animal welfare was related to whether or not they had a source for animal welfare information, but no statistical differences were found between which sources was used  . In our analysis, 46% of the participants indicated that they had a primary source for animal welfare information. Given the potential impact of having a source for animal welfare information and the importance of the different attributes, the multivariate multiple regression included a predictor variable which was called “Has a source for animal welfare information” and was given a 1 if the individual indicated they had a source for animal welfare information, and 0 otherwise. The variable of having a source for animal welfare information was a statistically significant determinant of the size of the share of preference for the attribute animal welfare and had a positive coefficient estimate, which supports the previous findings. In addition, having a source for animal welfare information was also a statistically significant determinant in models predicting the size of the share of preference for the attributes environmental impacts, locally raised/farmed pigs, and locally processed pork. In each case the variable had a positive coefficient estimate. Having a source for animal welfare information was a statistically significant predictor of the size of the share of price, but had a negative coefficient estimate.
The size of the share of preference for locally processed pork had the most similar set of statistically significant determinants to the share of animal welfare. In addition to having a source for animal welfare, owning a dog or cat, was also statistically significant with positive coefficient estimates determinants of the size of the share of locally processed pork. Similar to determinants of the size of the share for animal welfare, having purchased pork in the last year was also a determinant of the size of the share for locally processed pork (and the size of the share of environmental impacts) with a negative coefficient estimate. Being vegan was a statistically significant predictor of the size of the share of environmental impacts, locally raised/farmed pigs and locally processed pork and in each of these three models had a positive coefficient estimate. Indicating that one was vegetarian was also a statistically significant determinant of the size of the share of environmental impacts and had a positive coefficient estimate.
The attribute of price ranked fourth most important following pork/food safety, taste, and animal welfare, but preceding environmental impacts, locally raised/farmed pigs and locally processed pork. The statistically significant determinants with positive coefficients for the size of the share of preference for price were male, 45 - 64 years old compared to 65 years or older, and lower pre-tax income (compared with income of $150,000 or more). Determinants of the size of the share of preference for price with negative coefficient estimates were having a source for animal welfare information and passing the validation test. The coefficient estimate sign was generally opposite of the sign for each attribute when predicting shares of preference for the other attributes, which makes sense given that the size of the share of preference for price was found to be negatively correlated with the majority of the other shares of preference for the different attributes.
Taste was also ranked as one of the most important attributes. Seemingly unrelated regression results revealed that being male and having purchased pork in the previous year were statistically significant determinants of the size of the share of taste and had negative coefficient estimates. Independent variables associated with four income categories (less than $25,000, $25,000 to $34,999, $50,000 to $74,999, and $75,000 to $99,999) (compared to the income category $150,000 or more) as well as having passed the validation test were found to be statistically significant determinants of the size of the share of preference for taste.
This examination of the shares of preference for pork attributes in the U.S. found that determinants in explaining the size of the share of preference for these different attributes included different consumer’s demographics, pet ownership and sources of animal welfare information. These findings support previous studies which found a relationship between pet ownership and concern for livestock animal welfare. The majority of U.S. consumers did not have a primary source for animal welfare information, but having a source for animal welfare information was found to be a determinant with a positive coefficient estimate for each of the following attributes: animal welfare, environmental impacts, locally raised/farmed pigs and locally processed pork. Overall, consumers’ relative level of importance placed on animal welfare was predicted by gender, age, having a source for animal welfare information and having purchased pork in the past year.
Researchers at Purdue University developed the concept for this study in collaboration with Fair Oaks Farms, Belstra Milling Co., Indiana Pork and Indiana Soybean Alliance. Indiana Pork and Indiana Soybean Alliance provided the funding for the study. Researchers at Purdue University conducted the study and analysis without input, collaboration, sharing of survey design or participation in data collection by the funders, Fair Oaks Farms or Belstra Milling Co. in order to avoid biases or the perception of biases arising from working with industry groups.
MNL = multinomial logit.
RPL = random parameters logit.
1There are two ways to calculate the mean share of preference for each attribute. One way is to calculate the mean of the individual calculated preference shares for each attribute (using individual-specific coefficient estimates obtained in NLOGIT 5.0). The second option is to use the mean parameter estimates from the RPL model. In this work the latter is used to calculate mean shares of preference for each sample as utilizing the individual-specific preference shares facilitates calculation of correlations between preference shares and demographics for individual respondents.
2Survey respondents were asked to select the number six from the options one through seven. Those who correctly select the number six are considered to have “passed” and all others were considered to have “failed” the validation question.