Nonresponse in Contingent Valuation-reducing ... - Springer Link

2 downloads 94 Views 371KB Size Report
the WTP question (see Mitchell and Carson, 1989 and Bishop and Boyle, 1985, ... Mitchell and Carson, 1989; Craig and McCann, 1978). Second, when usingĀ ...
Environmental and Resource Economics 8:119-124, 1996. 9 1996 Kluwer Academic Publishers. Printed in the Netherlands. Research Note

Nonresponse In Contingent ValuationReducing Uncertainty In Value Inference GORAN BOSTEDT and MATTIAS BOMAN Department of Forest Economics, Swedish University of Agricultural Sciences, S-901 83 Umed, Sweden

Accepted 8 September 1995 Abstract. This paper focuseson problems associatedwith nonresponsein ContingentValuation

surveys. The results from a telephone follow-up survey show that value inference can be considerably improvedby informationon nonrespondents' attitudes. Key words: nonresponse,value inference,contingentvaluation, follow-upsurveys

Introduction

A major problem for researchers using the Contingent Valuation Method (CVM) to estimate nonmarket benefits is the problem of uncertainty in value inference caused by sample nonresponse. As far as we know, there is no CVM survey with a 100% response rate. This paper focuses on unit nonresponse, i.e. when no questions are answered because the person fails to participate in the survey. When using Willingness-To-Pay (WTP) estimates from a sample to infer the benefits to the population as a whole, two extreme positions are normally taken. The first assumes that nonrespondents have the same WTP as respondents(e.g. Walsh et al. 1984; Kristr6m, 1990; Carson et al. 1992; Li, 1994), whereas the second assumes that the nonrespondents' WTP is zero. WTP for the population is then WTPpop -- a*WTPresp, where c~ is the response rate to the WTP question (see Mitchell and Carson, 1989 and Bishop and Boyle, 1985, for discussions). Since CVM survey response rates are usually between 50 and 70% (Kristr6m, 1990; Fredman, 1994; Li, 1994), the 'range of uncertainty' between WTPpo p = ~*WTPresp and WTPpo p = WTP~e~pmay be very wide. This paper aims to show that a follow-up study of nonrespondents can be used to predict the mean WTP for those nonrespondents who answered the followup survey and, thereby, reduce this 'range of uncertainty'. Winter 1993, a mail CVM survey concerning the benefits of preserving the wolf in Sweden was sent to a random sample of 2000 Swedes. The realized sample was 1221, corresponding to a response rate of 61%. There was no item nonresponse on the discrete choice (DC) WTP question. Results are presented in Boman and Bostedt (1995), where it was found that the realized

120

GORAN BOSTEDT AND MATrlAS BOMAN

sample was fairly demographically representative for the Swedish population as a whole.

Methodology After the main survey, a follow-up was conducted using a random sub-sample of 150 individuals from the original sample. This consisted of 100 nonrespondents and 50 respondents to the main survey. To obtain a high response rate, telephone interviews were used instead of mail questionnaires. Respondents to the follow-up were asked to consider two questions previously used in the main survey: attitudes towards environmental issues (QI), and the importance of the existence of wolves in Sweden (Q2). Respondents were asked to rank their relative importance on a scale from 1 to 5, where 1 means 'least important' and 5 means 'most important'. Nonrespondents were asked exactly the same questions, together with an additional question concerning their reasons for not answering the mail questionnaire. The nonrespondents were not asked a WTP-question. There are three important reasons for this. First, one can expect a low response rate to the WTP-question beforehand, as compared to attitude questions (cf. Mitchell and Carson, 1989; Craig and McCann, 1978). Second, when using the telephone mode, it is generally advisable to avoid complex questions (Dillman, 1978), such as the WTP-question. Finally, since the nonrespondent sample for practical reasons was limited to 100 persons, it was statistically questionable to ask a DC WTP-question, which was the question format used in the main survey. A total of 98 individuals answered the follow-up questions, corresponding to a response rate of 65%. Table I shows the distribution between respondents and nonrespondents to the follow-up survey. Table I. Distribution of respondents and nonrespondents in the follow-up survey. -up

Respondents Nonrespondents Total

Respondents a

Nonrespondents

Total

34 64 98

16 36 52

50 100 150

"An individual is considered to be a respondent if at least one of the questions in the followup was answered.

Analysis and Results One important conjecture in the following analysis is that people's attitudes affect their valuation (Mattsson and Li, 1994), implying that attitude variables can be used to predict WTP. First a structural relationship must be

121

N O N R E S P O N S E IN C O N T I N G E N T V A L U A T I O N

established between WTP and attitudes, using the main survey. To estimate this, based on the DC WTP question, we have used a modified B o x - C o x functional form to allow for negative valuations (Draper and Smith, 1981). This results in a valuation function where the Box-Cox transformation of WTP equals yc; + ~ and c; is a vector of explanatory variables for individual i, for instance, answers to various attitude questions; 7 is a vector of parameters to be estimated; and e is an error term, assumed to be normally distributed. In our estimations, answers to the questions Q1 and Q2 were included in c;, together with sex (the only demographic characteristic that was known for nonrespondents). The log-likelihood function can then be written as: Ln(L) = ~ {D,.Ln[q~[[yci - U(Bi)]Io]] ,=l + (1 - D,) -Ln[1 - qb[[yci - f(Bi)]/o]] },

(1)

where r denotes the cumulative normal distribution function, Di denotes yes or no to the DC bid, B i (yes implies D~ = 1), while f(Bi) is the Box-Cox transformation s(Bi)*(((IBil + 1) ~ - 1)/~), where s(Bi) = sign(Bi) when Bi ~ O, otherwise s(B~) = 1. o is the standard deviation of the stochastic disturbance term e. When using responses given on a scale (such as the responses to Q1 and Q2) as explanatory variables, one has to determine whether to treat answers as being cardinally or ordinally given. We decided to use both, when estimating equation (1). In the cardinal approach, QI and Q2 are simply coded with integer

Table II. Valuation functions, treating answers to attitude questions on a cardinal, or ordinal, scale. Dependent variable: B o x - C o x transformation of bid. Variables

Cardinal approach

Ordinal approach

Constant Qj Q2 QI = 3 (dummy) Ql = 4 (dummy) Q2 = 5 (dummy) Q2 = 3 (dummy) Q2 = 4 (dummy) Sex (male = 1) ~. o Log-likelihood DF

-2.863 0.395 2.157 -0.126 0.022 2.201 -467.695 1200

1.577 0.491 0.905 1.216 2.948 4.634 -0.148 0.019 2.182 -469.463 1197

(-1.34) (1.80) (2.68)**

(-0.60) (0.35) (2.66)**

(2.12)*

(0.81) (1.41) (1.41) (2.54)* (2.64)** (-0.70) (0.29) (2.63)**

T-values are within parenthesis, ** = significant at the 1% level and * = significant at the 5% level. Log-likelihood is the value of the maximized log-likelihood function, while DF denotes degrees of freedom.

122

GORAN BOSTEDT AND MAq"rlAS BOMAN

numbers. For the ordinal approach, responses are recoded as dummies. Results are presented in Table II. Table II shows that attitudes towards the wolf are an important determinant of the DC response, independent of the approach. General attitudes towards the environment, however, cannot be interpreted so unambiguously. A likelihood ratio test shows that the null hypothesis that the coefficients for Q1, Qz and SEX are jointly equal to zero can be rejected at the 99.5% level. As an initial comparison of the differences in attitudes between the main and the follow-up survey, we calculated the means and confidence intervals of Q1 and Q2 (which implies a cardinal scale). This showed that the means stated by the respondent group in the follow-up were lower than the corresponding means from the main CVM survey, which contradicts the hypothesis that respondents in telephone surveys attempt to be more accommodating towards the interviewer. Second, the nonrespondent group in the follow-up had a significantly lower mean response in Q2, but not Q~. Thus, the nonrespondents to the CVM survey do not seem to have a less positive attitude to environmental issues in general, although they rank the existence of the wolf lower than the respondents. To further test the consistency in the responses across survey modes, we used the Pearson chi-square test (implying an ordinal scale). The resulting p-values for Q~ and Q2 were 0.688 and 0.288, respectively. In this case, we cannot conclude that attitudes differ significantly across survey modes. When using the Pearson test to examine the consistency of responses within the follow-up, the p-values for Q~ and Q2 were 0.664 and 0.470, respectively. In both cases, the null hypothesis of general association cannot be rejected, i.e., there is no significant difference between respondents and nonrespondents to the main CVM survey when using this test. This highlights the reason(s) for nonresponse. The additional question to the nonrespondents regarding reasons for not answering the questionnaire then becomes important for finding out whether nonresponses were due to survey or non survey specific reasons. Results show that about 48% had forgotten about the questionnaire or did not have time to fill it in. Only 4.8% mentioned a survey specific reason, namely 'not interested in endangered species'. Given these results, we can conclude that most nonresponses are due to general rather than survey specific reasons. At this point, one could possibly infer that the WTP for nonrespondents equals the WTP for respondents, and, hence, aggregate WTPresp across the total population. However, this approach is somewhat crude since the cardinal and ordinal comparisons produced divergent r-suits. It is important to note that even small differences in attitudes between respondents and nonrespondents may indicate large differences in WTP. To investigate this, we have attempted to predict the mean WTP of those nonrespondents who answered the followup, WTPnresp. This can be done, assuming that the estimation of the structural relationship in equation (1) also holds for the nonrespondents. The estimated mean WTP can then be written as (e.g. Johansson, 1993).

123

NONRESPONSE IN CONTINGENT VALUATION

c

f(.,)]o d.

with the circumflexes denoting maximum likelihood estimates. Equation (2) was estimated by numerical integration for respondents, holding c equal to sample means. For nonrespondents, the mean WTP was then predicted using the corresponding means from the follow-up survey. To account for possible differences due to the assumption of cardinality or ordinality, both approaches were tried. We also calculated two 'ranges of uncertainty' for the mean estimates. Without a follow-up survey, the range of uncertainty - as previously shown - is ~*WTPresp ~ WTPresp. Using the predicted WTP for the nonrespondents, this is reduced to [(1 - (t)*~*WTPr~esp + ~ , W T P r e s p ] ~ [(1 - tx),WTPnres p + ( l , W T P r e s p ] , where ct is the response rate in the main survey, while [3 is the share of nonrespondents who answered the follow-up survey. The lower bound of the latter range of uncertainty then implies that those nonrespondents who were impossible to reach have no WTP at all. The upper bound implies that their WTP equals WTPnresp. This method has been suggested in Dalenius (1957). Results are shown in Table III. Quite surprisingly, the predicted mean WTP of the nonrespondents differs quite substantially in numerical terms from that of the respondents. Note, furthermore, that the results are quite similar irrespective of how responses to Q1 and Q2 are regarded, in terms of cardinality and ordinality. Table III. Mean WTP for a viable Swedish wolf population for respondents, and predicted mean WTP for nonrespondents answering the follow-up survey, with corresponding 'ranges of uncertainty'.

Mean WTP - Respondents Predicted mean WTP - Nonrespondents Range of uncertainty without follow-up Range of uncertainty with follow-up

Cardinal approach

Ordinal approach

713 344 435-713 521-569

728 280 444-728 514-553

Discussion Results reported in this paper show that a follow-up survey of nonrespondents to a CVM survey can provide very useful information. We argue that it is not necessary to adopt either the extreme position that the mean WTP for nonrespondents equals zero o r that it equals the mean WTP for respondents. The analysis can be made more elaborate. Given that the investigated attitudes affect WTP, we show that mean WTP for nonrespondents can be predicted based on information from follow-up studies, implying that the range of uncertainty about the population mean WTP can be reduced to more moderate levels.

124

GORAN BOSTEDT AND MATI'IAS BOMAN

Acknowledgments We would like to thank Leif Mattsson, Bengt Kristr6m, Hans Nyquist and Peter Fredman at the Swedish University of Agricultural Sciences (SUAS), and Glenn Harrison, University of South Carolina, for scrutinizing our report at different stages. The research was made possible by financial support from the research programmes 'Natural Resource and Environmental Economics Applied to Forestry and Agriculture' and 'Production-Nature-Environment' financed by the Swedish National Council for Forestry and Agricultural Research (SJFR) and by the Faculty of Forestry, SUAS, respectively. References Bishop, R., and K. Boyle (1985), The Economic Value of lllinois Beach State Nature Preserve, Final Report to Illinois Department of Conservation, University of Wisconsin, Madison. Boman, M., and G. Bostedt (1995), 'Valuing the Wolf in Sweden: Are Benefits Contingent Upon the Supply?', in M. Boman, R. Br~innlund and B. Kristr/Sm (eds.), Topics in Environmental Economics: Proceedings from the Second International Conference on Environmental Economics, held at Ulv6n, June 2-5, 1994, Submitted Book Manuscript. Carson, R. T., R. C. Mitchell, W. M. Hanemann, R. J. Kopp, S. Presser and P. A. Ruud (1992), A Contingent Valuation Study of Lost Passive Use Values Resulting from the Exxon Valdez Oil Spill, Report to the Attorney General of the State of Alaska, Anchorage. Craig, C. S. and McCann, J. M. (1978). 'Item Nonresponse in Mail Surveys: Extent and Correlates', Journal of Marketing Research XV, 285-289. Dalenius, T. (1957), Sampling in Sweden: Contributions to the Methods and Theories of Sample Survey Practice, Almqvist and Wiksell, Stockholm. Dillman, D. A. (1978), Mail and Telephone Surveys: The Total Design Method, John Wiley and Sons, New York. Draper, N. R. and H. Smith (1981), Applied Regression Analysis, John Wiley and Sons, New York. Fredman, P. (1994), A Test of Nonresponse Bias in a Mail Contingent Valuation Survey, Working Report No. 201, Department of Forest Economics, Swedish University of Agricultural Sciences, Ume~. Johansson, P-O. (1993), Cost-Benefit Analysis of Environmental Change, Cambridge University Press, Cambridge. Kristr6m, B. (1990), Valuing Environmental Benefits Using the Contingent Valuation Method - An Econometric Analysis, Ph.D. thesis, UmJt Economic Studies No. 219, University of Ume~, UmeL Li, C-Z. (1994), Welfare Evaluations in Contingent Valuation - A n Econometric Analysis, Ph.D. thesis, Ume~ Economic Studies No. 341, University of UmeL UmeL Mattsson, L. and C-Z. Li (1994), 'Sample Nonresponse in a Mail Contingent Valuation Survey: An Empirical test of the Effect on Value Inference', Journal of Leisure Research 26, 182-188. Mitchell, R. C. and R. T. Carson (1989), Using Surveys to Value Public Goods: The Contingent Valuation Method, Resources For the Future, Washington D.C. Walsh, R., J. B. Loomis and R. Gillman (1984), 'Valuing Option, Existence and Bequest Demands for Wilderness', Land Economics 60, 14-29.