Review: (Holland and Wessells 1998) Predicting Consumer Preferences for Fresh Salmon: The Influence of Safety Inspection and Production Method Attributes

Daniel Holland and Cathy R. Wessells

Agricultural and Resource Economics Review 27:1–14

Objectives

- Questions to answer:

1) How important to consumers is an inspection that verifies seafood products are safe and of high quality?

2) Does it matter to consumers which agency of government is in charge of inspection program?

- major foci:

* determine whether production method (farmed or wild) will influence consumer preferences

* assess consumer perceptions of seafood quality and safety

* determine economic value of mandatory quality inspection program

Methods/Approach

- based on conjoint analysis to determine average relative importance and value of three product attributes for fresh salmon:
* seafood inspection

* production method

* price

^ The three price levels chosen were consistent with prevailing retail prices at the time of the survey. The mid-level price was equal to the average price in supermarkets from Narragansett, RI., Newark, Del., and Durham, NH. The price levels were $4.49, $4.99, and $5.49/lb.

- data set from mail survey of randomly selected households in northeastern and mid-Atlantic states

- rank preferences for fresh salmon products with different price, inspection and production attributes using conjoint experiment

- "The attribute combinations for the nine products represent a full profile, orthogonal
 Addleman design that eliminates multi collinearity and resulting bias in parameter estimate (Addleman 1962)
- "The estimation of the parameters that model these preferences can be done in a variety of ways, but it is commonly referred to as conjoint analysis (Green and Srinivasan, 1978)." (p. 4)

- Conjoint analysis

* used to find relative importance and value

* used to design products and services to maximize consumer appeal or used to estimate market share for competing products

* originally used in fields of psychometrics and marketing, now widely used to evaluate consumer preferences and demand for market goods

* used in market studies when new or hypothetical products or product attributes are assessed and market data does not exist

* when used in combination with demographic data and perception questions can suggest market segmentation and potential marketing strategies based on heterogeneity in preferences among consumers (Swallow et al. 1994)

REPRESENTATIVE CONSUMER MODEL vs. INDIVIDUAL-LEVEL

- a single utility function is estimated and assumed to be representative of all consumers

* componential segmentation = adding terms to model (demographic variables or responses to perception of preference questions) that capture interactions between attribute preferences and individually specific variables

* accounts for heterogeneity based on observable characteristics of individuals

* useful in identifying market segments that differ in preferences and consumption patterns

- also estimated individual preference models, used information to improve performance of aggregate models where entire data set is used

* proved to be critical in avoiding misinterpretations of results of aggregate model

* "Individual-level estimations have the disadvantage of requiring enough information collected from each individual to estimate his or her utility function. This can require a large number of observations when the number of attributes and attribute levels is large." (p.5)

* Tests of statistical significance of attribute parameters or dependent variable predictions are cumbersome (Elrod, Louviere, and Davey 1992), sometimes can't be calculated

- cluster analysis = separates sample into groups of respondents based on similar preferences, accounts for diversity in preferences that cannot be tied to observed personal characteristics

- Predictive validity: individual level models > cluster analysis > componential model (Moore 1980, Green and Helsen 1989)

- Individual models in commercial conjoint applications (Wittink and Cattin 1981; Green and Srinivasan 1990)

- This study compares group differentiation between: 1) "representative consumer" model, 2) componential model (with interactive terms between product attributes, demographic and perception variables) and 3) individual-level models

- Ordinal ranking, as opposed to a rating scale, was chosen because it is believed to be more reliable (Green and Srinivasan 1978), particularly in mail surveys

^ "A referendum or paired comparison format would have the same advantages but requires a greater number of questions to derive the same amount of information." (p. 5)

^ rank-ordered logit estimation of conjoint question is based on random utility theory

^ Because the ranks are ordinal rather than cardinal, and because the nine ranks given by each respondent are not independent, neither an OLS, ordered probit nor logit specification would provide consistent estimates.

- The scarcity of observations per individual (9 observations per individual and 7 parameters, including an intercept dropping one level of each attribute to avoid perfect multicollinearity) limits our ability to construct individual-level preference models from our data

* could estimate individual OLS models, but with only two degrees of freedom and violations of assumptions of cardinal data and independence of observations (statistical reliability unclear)

* however, did uncover over 46% of individuals in sample appeared to prefer a salmon product priced at 4.99/lb (med price) over 4.49 (low price) when all other attributes were identical, somewhat smaller proportion still preferred higher-priced product when prices were 5.49/lb and 4.49/lb (irrational)

^ may be due to artificiality of data-gathering

^ may represent actual consumer behavior

^ may be that important quality attribute was missing from product descriptions and some respondents assumed this attribute was represented at least partially by price (respondents told appearance was similar, but many assumed quality correlated to price)

^ frequent seafood consumers more likely to fall in "irrational" category supports that an important attribute that is related to price may be missing (in future ask if consumers think price is an indicator of quality)

* chose to rely on aggregate componential models

- Rank-ordered logit

* developed by in Beggs, Cardell and Hausman (1981)

* accounts for ordinal nature of data and lack of independence between observations for each respondent

* maxmimum likelihood estimates are those with maximum probability of resulting in observed sets of ranks

* rationale: individuals compare all choices, select most preferred - independent of ranking of remaining choices- then make next choices out of remaining subset of choices

- Sample:

* over 1,500 consumers in north-eastern and mid-Atlantic states (30% response rate)

* provided information about: preferences and consumption patterns for fresh fish, beliefs about safety and quality of fish, individual demographic characteristics

* 5,000 surveys mailed to randomly selected households according to each state's share of total population for region

* asked individual chiefly responsible for retail purchase of seafood to answer questionnaire

* 64% were male

* 97% were consumers of fresh seafood (seafood consumers more likely to return survey

* 756 (out of 1,529) answered conjoint question and all demographic and perception questions; 968 answered all demographic and perception questions

- Survey administration: 

* between April and August 1993

* began with personalized, hand-signed letter with information on impending arrival of survey and importance of responding and thanked for cooperation

* survey and letter followed initial letter by one week

* one week later postcard was sent to remind household =s to complete and return survey

* two weeks later follow-up survey was sent followed by another postcard reminder

- Willingness to Pay (WTP) for changes in attribute level can be estimated for representative consumer

* inappropriate if respondents assume implied relationship between cost and quality of market product or public goods

Results and Conclusions

- market segmentations and potential marketing strategies based on heterogeneity in preferences among consumers

- When presented with two choices (for example farmed, USDA-inspected salmon priced at $4.49/lb. versus wild, USDA-inspected salmon priced at $4.49/lb.), 81% of the people within our sample are predicted to choose the farmed product over the wild product. Also, 63% of the people within our sample are predicted to choose fresh salmon that is FDA inspected and priced at $5.49/lb. (but has no production method information) over farmed fresh salmon that is priced at $4.49/lb. (but has no inspection information). Finally, 97% of the people within our sample are predicted to choose the inspected product with no information on production method priced at $4.99/lb. over the farmed salmon product that has the same price but no inspection information.

- According to this sample of seafood consumers, the strongest preferences are related to seafood inspection, products indicating an inspection by either the USDA or the FDA are preferred to products with no inspection.

* FDA is preferred to the USDA as an inspection agency

* However, when these coefficients are modified by the interaction terms of the attributes with demographics and beliefs, the total effect is that the USDA is the preferred agency over the FDA

- wild salmon is preferred to farmed, modifying these effects by the coefficients of the dummy interactions with farmed and wild results in farmed being the preferred method of production

* The salmon industry’s strategy of labeling product as to what method of production is used, farmed or wild, seems to favor farmed salmon more than wild salmon in the northeastern and mid-Atlantic states. This may be the result of the perception that farmed salmon is higher quality and safer than wild, possibly because a farmed product is connected with a product over which the harvester has some degree of control, unlike wild salmon

- Conjoint analysis is useful for evaluating marketing strategies for seafood

- Public policies (namely an inspection program) can provide benefits to seafood consumers

- Componential models, which include interactive terms that capture the heterogeneity attributable to observable demographic characteristics, greatly improve model performance. Further improvements are made by including information derived from questions about individual perceptions. (However, much of heterogeneity is not correlated with observable characteristics)

- Individual-level analysis typically results in higher predictive ability and can identify important heterogeneity in preferences not revealed by aggregate model, but require large amount of data per individual and not applicable outside of sample

- many individuals apparently preferred to pay more rather than less for a given piece of fish, suggesting that they may associate price with quality (not apparent with aggregate model)

- looking at group and individual heterogeneity can improve performance of models in predictions as well as provide valuable added information to industry or policymakers

- amount and type of data collected (ordinal ranks vs. scaled rating) may limit acceptable estimation procedures

Errors in experimental design, statistical analyses or analytical approaches?
Assumptions

- Consumers unaware of regulatory changes in seafood safety, communicating changes is important so seafood industry can increase demand and raise prices to cover increased regulation costs

- salmon industry experienced 97% increase in world production from 1986 to 1995 (Johnson 1996), in large part due to increased production of farmed salmon 

- "Orthogonal Addleman design assumes that the attributes are completely separable. This could be problematic if, for example, an inspection has relatively more value to consumers when they know fish is wild as opposed to farmed" (p. 4)

- sample may be representative of seafood consumers in the northeastern and mid-Atlantic

United States, we suspect that results might vary considerably elsewhere (preference for wild versus farmed may be reversed on the West Coast given the dominance of the wild fisheries in the Pacific)

Opinion

According to Google Scholar this study has been cited 79 times since 1998 which is not a lot. Many of the articles referred to eco-labeling and food safety. I was mainly looking at this article to understand their methodology. This has been helpful as I have read the Elrod (1992) article and decided we may not have a large enough sample size for choice-based. Combining this with information from Sawtooth (Orme 2009) I think we can do a CVA analysis where cards are ranked given our small sample. Sawtooth has a CVA/HB module for $1000 which allows for a better estimation of an individual’s part worths by borrowing from the population. This is useful especially since individual-level classification is not very important to us.

Useful References

Elrod, T., J.J. Louviere, and K,S, Davey. 1992. “An Empirical Comparison of Ratings-Based and Choice-Based Conjoint Models.” Journal of Marketing Research 29:368-77.

Addleman, S. 1962, “Orthogonal Main-Effect Plans for Asymmetrical Factorial Experiments.” Technometrics 4(1):21-46.

Beggs, S., S. Cardell, and J. Hausman. 1981. “Assessing the Potential Demand for Electric Cars.” Journal of Econometrics 17:1-19.

Swallow, S.K., T.W. Weaver, J.J. Opaluch, and T.S. Michelman. 1994. “Heterogeneous Preferences and Aggregation in Environmental Policy Analysis: A Landfill Siting Case.” American Journal of Agricultural Economics 76:431-3.

Johnson, H.M. 1996.1996 Annual Report on the United States Seafood Industry. Bellevue, Wash.

Green, P.E,, and K. Helsen. 1989. “Cross-Validation Assessment to Individual-Level Conjoint Analysis: A Case Study.” Journal of Marketing Research 26:346-50.

Green, P.E., and V. Srinivasan. 1978. “Conjoint Analysis in Consumer Research: Issues and Outlook.” Journal of Consumer Research 5:103-23.

Green, P.E. 1990. “Conjoint Analysis in Marketing: New Developments with Implications for Research and Practice.” Journal of Marketing 54(4):3–19.

Manalo, A,B, 1990, “Assessing the Importance of Apple Attributes: An Agricultural Application of Conjoint Analysis,” Northern Journal of Agricultural and Resource Economics 

19:118–124.

Anderson, J. L., and S.U, Bettencourt. 1993. “A Conjoint Approach to Model Product Preferences: The New England Market for Fresh and Frozen Salmon. ” Marine Resource Economics 8:3149.

Stoker, T.M. 1993. “Empirical Approaches to the Problem of Aggregation Over Individuals.” Journal of Economic Literature 31: 1827–74.

Moore, W.L. 1980. “Levels of Aggregation in Conjoint Analysis: An Empirical Comparison.” Journal of Marketing Research 17:5 16-23.

Green, P.E,, and K. Helsen. 1989. “Cross-Validation Assessment to Individual-Level Conjoint Analysis: A Case Study.” Journal of Marketing Research 26:346-50.

Wittink, D.R., and P. Cattin. 1981. “Alternative Estimation Methods for Conjoint Analysis: A Monte Carlo Study.” Journal of Marketing Research 18:101-6.

Louviere, J.J. 1988. “Conjoint Analysis Modeling of Stated Preferences.” Journal of Transport Economics and Policy 22 (1):93-119.

Vatn, A., and D. Bromley. 1994, “Choices without Prices without Apologies,” Journal of Environmental Economics and Management 26(2): 129-48.