The correlation reported for Figure 2 (-.79) is probably an overestimate resulting from the first and last case in the data. Using this figure to estimate the raw data you get an r as low as -.37 using some robust methods, and if you bootstrap the confidence interval around r, it is very wide (-1 to about 0.6) implying that the population effect might be very different to that reported. I elaborate on these computations at http://discoveringstatist...

Of course the presence of two data points that might be affecting the r from Figure 2 is interesting in itself - it would be interesting to know what it is that made these cases different to the main cloud of data.