Another variable we introduce is the real price of oil. A higher price of oil should reduce the use of (sulphur-containing) oil. In fact, we can identify such a relationship in both the fixed-effects and random-effects model. However, on theoretical grounds the effect of a higher oil price on pollution is not necessarily as straight-forward as the above argument implies. If a higher oil price leads to a substitution effect and a switching from oil to other fuel types, it is uncertain if this other fuel is “cleaner’” ndturalgrsol “diiider” ioal. Tho detd teemtm ^ges^alT therebstitutimiit towords cleaner fuel types. Many people want to payday loans similar to speedy cash whenever they feel like it, and this is very much possible now that you can apply for such a loan at this. We make sure the application process is quick and there is no long-term commitment. You just get your money and are free to spend it on whatever.

In another sensitivity test we replace the linear time trend by year dummies. Since we have an intercept in the model, we do not include dummies for the first two years (as there were very few observations for the very first year 1971). The result is surprisingly supportive of a linear time trend. The estimates for the year dummies (not shown in tables С.1 and С.2 in order to conserve space) trace out a remarkably stable linear path.

C.2 Dependent Variable

In a second set of sensitivity analyses we explore the choice of our dependent variable. We have argued before—basef wtr thtobsaISiationeexprepsed in figuresA. a addA.2 that alogarithmic transformation of the dependent variable is appropriate. However, there is a menu of different SO concentrations to choose from. We opted for the median SO concentration because is more “robust” with respect to outlier observations than the arithmetic mean. The U.S. Environmental Protection Agency kindly supplied us with a variety of concentration statistics. We explore all of them in tables С.3 and С.4 for our fixed-effects and random-effects baseline model. In addition to the median (“Base” ww mue pheeri“hmeticmean t“Mepn”) and the 0Pth,95th,and 9dth percentileoP f02 concentrations (“P90%”,“P95%”, and“099%”). All odthoE2meashPes wete tp“pr“emreh inio logarithms when they were used as a dependent variable.

The first observation is that the intercept term is increasing from left to right, as the higher percentiles have higher average SO concentrations. Comparing the mean with the median, we find a higher intercept for the mean. One way of reading this is that, adjusted for our regressors, the mean exceeds the median. This appears to be simply a result of the non-normal distribution of the (linear) SO cocentrations, which we saw in figure A. 1 to be highly-skewed to the left.