Decomposing local: a conjoint analysis of locally produced foods.

A total of 80 product profiles (5 price levels x 4 location descriptions x 2 producer identifications x 2 freshness levels) were generated in order to create the eight pairs of product profiles that are presented to each respondent. To make the design more efficient, product profile pairs were individually checked; we removed pairs that featured identical profiles or pairs that featured one dominating profile (e.g., lower price and a freshness guarantee). (4) The average values of these attributes are summarized in table 4 for the observations used in estimation.

Model Estimation

Each respondent was asked to cast eight decisions during the conjoint questioning for a potential of 4,240 usable responses. Due to nonreported demographic information, 107 conjoint choices were dropped from estimation. Furthermore, for 167 conjoint choices, individuals either failed to answer the choice question or answered that both choices were equally desirable; these responses were omitted from the analysis. This leaves 3,966 usable conjoint choices for analysis.

Statistical analysis of the model proceeds by estimating the utility difference model using a random effects probit estimator. This estimator allows for the correlation of errors across observations contributed by the same respondent. That is, we postulate that the error term from individual i on response t is [[epsilon].sup.i.sub.t] = [[eta].sub.i] + [e.sup.i.sub.t], where [e.sup.i.sub.t] is an independently and identically distributed error term for all i and t whereas [[eta].sub.i] is an individual-specific error component and E[[[epsilon].sup.i.sub.t][[epsilon].sub.i.sub.[tau]] = [rho] [for all] t [not equal to] [tau]. Estimation of this random effects panel model yields a statistically significant p coefficient. Failure to account for this within-subject error correlation would lead to inefficient estimates. Failure to implement such an estimator may lead to improper inferences concerning parameter estimates.

We model the probability that the respondent chooses the product shown on the left side of the conjoint graphic. The model's parameters are estimated via maximum likelihood procedures for a random effects probit model. Two interaction terms, one between two product attributes and one between a product attribute and respondent gender, did yield significant results and were included in the final model.

Most of the product attributes profiled in the conjoint instrument have well-defined expectations with respect to sign. All else equal, respondents are expected to prefer products that feature a lower price, provided a freshness guarantee, and are produced in a location that is closer to the location of the store or market where the interview takes place.

Results

Table 3 lists the product attribute variables included in the final model. Using the production location attribute, we decomposed the concept of "local" into two degrees of distance--within-state and within an undefined substate region ("nearby"). This allowed us to test whether consumers discriminated between products at a substate level, and effectively determine their definition of the term local. Both farm size (5) and freshness guarantee attribute variables were binary in nature, and thus were represented as dummy variables. The price variable measures the price differential for the two products, ranging from $-2.00 to $2.00

We also considered various higher-order effects representing interactions among the four product attributes. For example, in the case of a product with an unidentified location, it is possible that the presence of a freshness guarantee may partially offset the negative impacts of the first cue--that is, the signals may be substitutes. The combination of a guarantee and unidentified label, therefore, may have predictive value in the consumer's choice beyond what each could predict independently. The same might be true for the farm size cue and an unidentified or national label as well. Finally, we considered the effects of socioeconomic variables by interacting them with product attribute variables. Consumers with differing demographic or economic status may value the product attribute cues differently. In preliminary models, various hypothesized socioeconomic interaction effects were tested, though only one such interaction between gender and local production proved to be robust.

Our sample of consumers was drawn randomly, but from two very different market types. Because consumers select the type of market they attend, it is possible that these consumer groups will differ in important ways regarding food product preferences. To test the regularity of preferences between these subsamples, models were estimated and a likelihood ratio test was used to test for differences based on market type. Our results suggested a systematic difference in response between the groups, and thus we estimated separate models for these groups ([chi square](10) = 58.36, p < 0.001). Similar tests were conducted for group subsamples established on the basis of household income, education, support for local production, and other socioeconomic measures, but in each of these cases the likelihood ratio test statistic was not statistically significant at the 10% level.

We also tested for learning that may occur within the experiment. Because each respondent made eight consecutive choices, it is possible that the respondent took a few choices to clarify his or her preferences. If there is systematic variance between the early and late experimental results, we can infer that learning occurred during the experiments. A likelihood ratio test for stability of parameters between the first and last four choices failed to reject the null hypothesis (for the direct market sample, [chi square](10) = 0.69, p > 0.10; for the grocery sample, [chi square](10) = 14.25, p > 0.10).

Table 5 shows the results of the final models for the grocery and direct market subsamples whereas table 6 reports the deterministic and stochastic WTP for several key product attributes for each subsample. Table 5 displays two models for each subsample. The first model features the greatest articulation of the location of production with separate categories for "grown nearby" and "grown in Ohio." For both grocery and direct market respondents, we fail to reject the hypothesis that the values of these two parameters are equal ([chi square](1) = 0.01, p > 0.10 for the grocery sample, [chi square](1) = 1.92, p > 0.10 for the direct market sample). We find that our sample of respondents do not distinguish between "grown nearby" and "grown in Ohio" even though all respondents see both distinct claims multiple times during the course of the conjoint experiment. We consolidate the two locations into the singe "composite local" variable, which is a dummy variable indicating whether the product was marked "grown nearby" or "grown in Ohio." Henceforth, we shall refer to such products as "local."

The results from model 2 clearly show that both sets of consumers show preferences for locally grown over grown in the United States, and direct market shoppers displayed nearly twice the WTP for the closer location ($0.92 vs. $0.48 per basket for the stochastic WTP with no overlap in the 95% confidence intervals).

Grocery shoppers tended to penalize strawberries that did not identify the location of production compared to strawberries marked as produced in the United States, although direct market shoppers did not. So, grocery store customers discriminate between the three locational attributions of strawberries, with locally grown being the preferred, which is distinct from grown in the United States, which is distinctly preferred to unidentified production locations. Direct market shoppers in our sample distinguish only between locally produced and all other locational attribution. Note that, due to the design of the choice experiment, these values attributable to the location of production hold constant the degree of product freshness and the corporate affiliation of the production, giving us an estimate of the demand for location that is not confounded with the naturally correlated feature of product freshness or farm size/organizational structure attributes.

The presence of a freshness guarantee was significant and positive in sign for both groups of consumers, with a stochastic WTP for the freshness guarantee of $0.54 for grocery store shoppers and $0.73 for direct market shopper, though the 95% confidence intervals of the WTP for these two groups features some overlap. For both groups the WTP for a freshness guarantee is similar to the WTP for a guarantee of local production, suggesting that strawberries that feature either a locally grown label or a freshness guarantee can support similar price premiums in a market constituted of subjects such as the grocery store and direct market shoppers included in our sample. Furthermore, the fact that there was no statistically significant interaction term between locally grown and freshness (the interaction term was highly insignificant in preliminary modeling and dropped from the final model) suggests that the two features are valued independently by our sample of subjects. That is, freshness guarantees are valued equally for products produced locally and within the United States.