Balancing Competing Principles of Environmental Equity
7.3 Methodological Framework
7.4.1 Experiment
Table 7.2 illustrates the results for the dummy variable model. The first column contains the estimates of the basic model represented by expression (7.5). Clearly, responsibility, i.e. ‘pollutes’, is the most important attribute, an outcome that is consistent with the predominance of the ‘polluter-pays-principle’ in environmental policy discussions. However, the results also indicate marked evidence of a trade-off between responsibility and other attributes in that the coefficients on ‘benefits’ and ‘income’ are highly significant. Indeed, the sum of the coefficients on the ‘benefits’ and ‘income’ variables (i.e. 0.76 + 1.73) is greater than the coefficient on the responsibility variable (i.e. 2.09). This would appear to suggest that an individual who both benefits from the programme and has high income should pay more, other things being equal, than someone who is responsible for the environmental problem but has low income and does not benefit in any way from the policy change, at least in the context o f this experiment.
Table 7.2 Dummy Variable Model Results (t-ratios in parenthesis) Variable Basic Model Model Without
Non-Response Individual Interactions Model Benefits 0.7583 0.70286 0.24257 (11.263) (7.539) (0.565) Income 1.7268 1.79034 2.01896 (26.725) (20.118) (13.541) Pollutes 2.0919 2.07676 2.20495 (32.824) (23.661) (16.855) Monthly Petrol*Pollutes -0.000008 (-1.266)
Health Status *Benefits 0.15854
(1.108)
Monthly Eamings*Income -0.000001
(-1.819)
N 198 105 105
Notes:
1. In these models the choice attributes are dummy variables: ‘Benefits’ - 1 if benefits from the programme; 0 otherwise; ‘Income’ - I if high; 0 if low; ‘Pollutes’ - 1 if is responsible for pollution; 0 otherwise.
2. The individual specific variables are coded as following: ‘ Monthly petrol’ - individual monthly expenditure in petrol; ‘Health status’ - 1 to 5, with I if very good and 5 if very bad; ‘Monthly earnings’ - monthly individual earnings.
An interesting question is the extent to which the preferences o f each respondent for burden sharing allocations are independent of his or her particular position in society. For example, does the knowledge of a given respondent that he or she has relatively high income bias his or her response away from options that assign burdens to high income individuals? The answer depends on whether selfish motives influence responses or whether the choice between alternative allocations is made as i f
respondents were under Rawls’ ‘veil of ignorance’ (Rawls, 1972). We specified variables to capture this aspect of responses using survey data on respondents’ monthly petrol expenditure (Petrol,), health status (Health^) and monthly earnings
(Earnings,). The impact of socio-economic factors can be gauged by interacting socio economic variables with the attributes of the alternatives and examining the signs and significance of the resulting interacted variables.^ The empirical model with interactions estimated can be specified as:
Uij = a I Pij + a2 Bij+ aj Yy+ a4 Petrol i*Py+ as Healthi*Bij+ a^ Earnings i*Yy (7.8)
The third column of Table 7.2 shows that allowing for individual interactions does not appear to add much to our discussion. The signs on the coefficients on Petrol*P and
Earnings *Y are negative as we would expect. However, the former is insignificant and the latter only significant at the 10% level. The coefficient on Health*B is positive but insignificant. Owing to the considerable number of non-responses, particularly with regards to respondents’ monthly earnings, we re-estimated the basic model on the sub-sample of respondents who had answered all questions. This provided a check on whether the comparison between the basic model and the model
^ Owing to the statistical specification of the logit model it is not possible to incorporate socio economic regressors directly into the utility index given by expression (7.5). The random utility framework is based on differencing the values of attributes across different alternatives and any attribute that does not vary across alternatives would simply drop out of the model as a result of this process. Interacting socio-economic characteristics circumvents the problem because the interacted
with interactions might be biased due to the fact that each model was estimated on a different sample. The second column o f results in Table 7.2 reveals that removing non-responses from the basic specification does not alter the conclusions drawn from the basic model.
Table 7.3 investigates the trade-off between attributes further by analysing how many respondents behaved in a way possibly consistent with lexicographic preferences. Lexicographic ranking is defined as a tendency for respondents to rank questions solely with reference to one o f the attributes, ignoring all other differences between options. A more complete form of lexicographic ranking may arise when respondents also refer to a second attribute to discriminate between those options that ‘tie’ in terms of their attribute o f prime interest. Such behaviour may either reflect the use of a simplifying heuristic to aid the ranking process or may be a manifestation of underlying preferences that are truly lexicographic. This is important because it is entirely possible that, for example, our findings compatible with say, the PPP whereby respondents always rank according to responsibility relying on other attributes only in the event of ties.
Table 7.3 Respondents with Possible Lexicographic Rankings
Ranking % Respondents (n=198) p,y,b 19% A B , Y 2% y,p,b 11% Y , B , P 2% B , Y , P 0 B , P , Y 1% Total 35%
Note: P - Pollutes; Y - Income; B - Benefits. For example, "P,Y,B' means that respondents first ranked burden-sharing allocations according to responsibility for pollution; then, in case of ties, ranked according to income; and finally decided upon any remaining ties according to the benefits from the programme.
variable both contains individual-specific socio-economic information and varies across alternatives in the choice set.
Analysis o f the data reveals that 35% of respondents systematically ranked options with reference to only one of the three criteria, referring to a second attribute only where particular pairs of options tied on the basis of the first criterion. This total can be broken down as follows.^ About 21% of respondents seemed to be primarily concerned with the pollution (P) attribute, 13% with income (Y) and 1% with benefits
(B). The most prevalent lexicographic ranking is by P then by Y and then by B. In all, 19% of respondents answered in this way. While, at first blush, these results appear to indicate that a proportion o f respondents may have been using a lexicographic rule as a response algorithm, it can be shown that these rankings can be explained in terms of a standard utility function that accords a relatively high weight to the attribute concerned (Foster and Mourato, 2000). In the absence of a more sophisticated test to distinguish between these two hypotheses we argue that it is reasonable to speculate that respondents are willing to trade-off between competing principles. Clearly, however, it would be desirable to seek to reduce this ambiguity in future work.
7.4.2 Experiment 2
To the extent that there may be non-linearities in respondents’ preferences the dummy variable model with only two levels of each attribute will not capture this. For example, either of the following two scenarios might be plausible. On one hand, we might speculate that individuals believe that the burden assigned should be increasing in the level o f an attribute. On the other, the burden assigned might be a decreasing function of the attribute level as would be the case where respondents find it difficult to distinguish between relatively high levels of a given attribute. In order to evaluate how robust our initial conclusions are. Table 7.4 illustrates results from the continuous variable model, where each attribute was assigned three levels. As noted above, given that assigning 3 levels to 3 attributes would significantly complicate the ranking task that respondents have to perform, in our second experiment, we only asked them to consider two attributes at a time (i.e. PIB and PIY) therefore holding the omitted attribute (respectively, Y and B) constant.
’ Note that the probability of getting a (strong) lexicographic ranking amongst the 720 possible ranking combinations would be only 0.83% had the rankings been random.
Table 7.4 Continuous Variable Model Results (t-ratios in parenthesis) Variable Basic Model A Quadratic Model A Individual Interactions Model Basic Model B Quadratic Model B Individual Interactions Model Benefits 0.05683 (16.315) 0.25693 (12.338) 0.25835 (12.351) Income 0.00860 (52.922) 0.01029 (9.733) 0.01003 (6.533) Pollutes 0.27220 0.80704 0.79688 0.22404 0.45553 0.46350 (57.693) (32.197) (31.368) (43.563) (16.058) (11.328) (Benefits) ^ -0.00813 (-8.087) -0.00812 (-8.042) (Income) ^ -0.000005 (-1.646) -0.000003 (-0.622) (Pollutes)^ -0.02804 (-17.024) -0.02822 (-17.088) -0.01445 (-8.147) -0.01504 (-5.944) (Monthly petrol *Pollutes) 0.000001 (4.228) 0.000002 (4.294) (Health status *Benefits) -0.01290 (-0.668) (Monthly earnings *Income) -0.000000003 (-2.237) N 232 232 231 226 226 117 Log- Likelihood -1085.749 -936.5785 -929.6205 -1155.473 -1114.612 -550.3748
The presence of three levels for each attribute permits the estimation of quadratic specifications that allow for non-linear preferences for levels of an attribute:
Uij = W/ Py + W2 By + W3 y + WjB^y
Uy = Zi Py + Z2 Y y + ZjP^y + y
(7.9)
The second column of results in Table 7.4 indicates that the coefficients on both the ‘pollutes square’ and ‘benefits square’ variables are negative and significant. Similarly, in column 5, the coefficients on both the ‘pollutes square’ and ‘income square’ variables are negative (with the latter only significant at the 10% level). This indicates, to some extent, the presence of non-linearities in respondents’ preferences. In general, the results o f the quadratic models A and B seem to confirm the most important conclusions drawn from previous specifications, therefore indicating a high degree of consistency between models.
Table 7.4 also contributes to our earlier discussion regarding selfish preferences. Results are illustrated in the individual interactions models in columns 3 and 6, corresponding to the following empirical utility specifications:
Uij = wy Pij + W2 By + wj P^ij + 'W4B^ij + ws Petroh*Pij + weHealthi^By (7.11)
Uy = zi Py + Z2 Yy + + Z4 y + Z5Petroli*Py+ Zg EarningSi*Yy (7.12)
Again, the effect of a respondent experiencing bad health appears to be an insignificant factor in assigning a burden. However, the coefficient on the Petrolt*Py
variable is surprisingly positive and significant in both models. This would apparently suggest that respondents consider that they themselves should face a higher burden if they are responsible for pollution. However, in both models this effect is rather small in that the coefficient is relatively low (0.000001 and 0.000002 respectively). The interaction of the ‘Income’ attribute with respondents’ earnings in the final column indicates a negative and significant coefficient as expected. This offers further evidence that respondents with high income were somewhat less inclined to assign burdens on the basis of the income attribute. Overall, considering both experimental treatments, the evidence of the impact of selfish preferences in burden allocations is mixed.