Valuing lives saved from safer food--a cautionary tale revisited.

The possibility of reporting errors and biases is hard to avoid in survey studies and we were prepared for their presence here. Our goal was to qualify the survey responses with hard data for two purposes: (1) to confront the survey responses with factual data with the goal of removing some of the reporting errors and biases, and (2) to combine the survey data and the behavioral data (because neither data source contains all the relevant information) to identify each person's willingness to pay for mortality risk reduction. But both of these aims require a connection between the survey data and the behavioral data. We find it disconcerting that the picture emerging from table 2 seems to be this: there are internal correlations between the factually oriented indexes and between the more subjective indexes, but there is little correlation between the two groups of indexes. There seems to be a disconnection between exposures to health risk and (reported) health risk reduction activities.

Fourth, this disconnect story is strengthened by noting the significant correlation between household income and both HI and BMI. Again, we observe correlation between factual data, which appears to reflect socioeconomic patterns in health and obesity. Household income is also significantly related to the general risk reduction index, RI, which in part is based on specific factual questions. A weak relationship exists, not significant at the 1% level, between the FPI and SPEI indexes.

Fifth, the respondents' relativerisk is not significantly correlated with any of the indexes, and neither is the personalrisk. Interestingly, baserisk, the respondents' perception of the risk exposure of other individuals, is significantly correlated (at the 1% level) with the individual's health index, HI, and the household income with correlation coefficients of 0.108 and -0.123. The lack of correlation between personalrisk or relativerisk with any of the other variables confirms the picture of a disconnect in the survey responses. To us, respondents seem to have vague notions of their risk exposure; these notions are so vague that they correlate with health status, reported behavior, or effectiveness of risk reduction.

While the evidence so far indicates a disconnection in the survey responses, the last and perhaps most consequential result we report concerns a correlation between the responses to certain questions that seems "too high." Table 3 reports the correlations between the responses to the various components of Q43. The correlations are similar for Q44-46, and for Q48-51 (correlations are calculated treatment by treatment). The lowest correlation observed in any of these correlation matrixes is 0.773 while most correlations actually are above 0.9, in which the responses to Q48-51 fare worst with correlations ranging between 0.93 and 0.99 (with only one exception--a correlation of 0.88).

This strong correlation in the presence of the otherwise low to nonexistent correlations is baffling. We are at a loss to explain this finding, other than that the survey respondents seem to be guessing at numbers they perceive only vaguely, at best. Their guesses have a strong internal consistency between what they perceive to be similar questions, but there is no external consistency to several things we expected to be related: health status, reported behavior, reported self-protection activity, and worst of all, the responses to differently worded questions about similar underlying realities. The strong internal correlations between the responses about risks posed by beef, pork, and chicken suggest surrogate behavior. These results contrast the lack of correlation between similar responses summarized by SPEI and IRRI.

As a final check, we examine the correlations between the responses to matched pairs of questions that form part of SPEI and IRRI. Question 43 is matched with question 48, 44 with 49, etc., and the matching is done treatment by treatment, i.e., conditional on amount. The contrast still exists--too much correlation across different food groups, 0.7; too little correlation across the same questions asked in different ways, 0.177. Moreover, there is no tendency for matched sub-questions, like q43_beef and q48_beef, to be more correlated than nonmatched sub-questions, like q43_beef and q48_chicken. The strongest observed correlation, of -0.177 (negative due to differing treatments), is between q45_chicken and q50 beef.

This strong internal consistency, and poor near-internal and external consistency, means these data are unsuitable for either the proposed GMM approach or for any other method. The internal consistency means that not enough variation exists between a person's responses to the risks posed by beef, pork, and chicken that could be correlated to behavior. The lack of near-internal consistency causes us to doubt that the responses represent strongly felt perceptions. Rather, the responses appear to be "best guesses." Finally, the lack of external consistency indicates that the responses are near orthogonal to other facts and behaviors, making them uninformative and unable to shed further light on the value of risk reduction.

In summary, the evidence from Wave I of the UW survey revealed that responses contained poor information about how people perceive food safety risks and how they will respond. This implies these data cannot support reasonable VSL estimates. In the end, one can question whether this result is due to the theory of endogenous risk, the survey design, or the rationality of the respondents. We conclude on a constructive note. Future survey work that tries to collect data at the level of detail needed to run a GMM-style model will have to find a set of different behaviors that really are distinct in the average person's mind, even after accounting for endogenous risk. The challenge is to identify a framework of comparable activities that are sufficiently distinct to generate both different behavioral reactions and decided perceptions that are tied to the true nature of the activities. What we have learned is that the behaviors in the risk questions have to be sufficiently differentiated, almost a day and night distinction, to make inferences about the value of statistical life given multidimensional heterogeneity

Thanks to the USDA/ERS for supporting this research, and Tanya Roberts and Fred Kuchler for their support.

References

Bateman, I., A. Munro, B. Rhodes, C. Starmer, and R. Sugden. 1997. "Does Part-Whole Bias Exist? An Experimental Investigation." Economic Journal 107:322-32.

Hammitt, J., and J. Graham. 1999. "Willingness to Pay for Health Protection: Inadequate Sensitivity to Probability?" Journal of Risk and Uncertainty 18:33-62.

Hayes, D., J. Shogren, S. Shin, and J. Kliebenstein. 1995. "Valuing Food Safety in Experimental Auction Markets." American Journal of Agricultural Economics 77:40-53.

Kahneman, D., and J. Knetsch. 1992. "Valuing Public Goods: The Purchase of Moral Satisfaction." Journal of Environmental Economics and Management 22:57-70.

Roberts, T. 1986. "The Economic Losses Due to Selected Foodborne Diseases." Proceedings of the Ninetieth Annual Meeting of the United States Animal Health Association, October 1986, pp. 336-53.

Shogren, J., and T. Crocker. 1991. "Risk, Self-Protection, and Ex Ante Economic Value." Journal of Environmental Economics and Management 20:1-15.

Shogren, J., and T. Stamland. 2002. "Skill and the Value of Life." Journal of Political Economy 110:1168-97.

--. 2006. "Consistent Estimation of the Value of Statistical Life." Resource and Energy Economics 28:262-81.

(1) In GMM terms, the goal is to obtain a model in which all parameters are identified. This model then yields an asymptotically unbiased estimate of the key parameter--the value of statistical life. This approach has the potential of restoring greater accuracy to the estimates of the value of statistical life by avoiding the bias inherent in single-equation approaches. This advance can come at considerable cost. Keeping the model tractable and its data requirements reasonable requires some strong assumptions. But we believe these assumptions do not invalidate the main strength of the model, which is to strip away the presumption--implicit in previous estimation approaches--that all people are identical.

(2) Due to the fact that NSL participants primarily resided on the east coast and Midwest of the United States, only KN panelists in the following states were selected for the survey: Connecticut, New York, New Jersey, Pennsylvania, Ohio, Indiana, Michigan, Maryland, Washington DC, Virginia, North Carolina, South Carolina, and Florida.

(3) We use the 'frequency value' of the response in decimal notation, for instance q21 = 0.001 if the response was '1 in 1000 people.'

Jason F. Shogren is the Stroock Professor of Natural Resource Conservation and Management, University of Wyoming, and the King Carl XVI Gustaf Professor of Environmental Sciences, Umea University; Tommy Stamland is Associate Professor of Finance, Norwegian School of Economics and Business Administration.