I can see where you're coming from, but I also see two problems. Firstly, you're generalizing towards the origin from a relatively high section of the graph.

I can see where you're coming from, but I also see two problems. Firstly, you're generalizing towards the origin from a relatively high section of the graph. Leaving aside the relative happiness, a 5% increase in income is actually much more significant in someone's life when they make $10,000 a year than when they make $100,000. There are thresholds established by absolute resource requirements. This makes it dangerous to reason from the 'flat' part of the curve downwards. Relevant analogy: transistor response curve. Secondly, insofar as income represents a portion of an absolute resource [ie to some extent], some information is obscured that would otherwise be intuitively and immediately available. It's remarkable that the money required to increase a person's happiness by one unit is, in every case, enough money that another person could share their initial happiness. A graph with a linear relation would demonstrate this as an (immediately obvious) diminishing return on a finite resource. A graph with a log relation presents a diminishing return as natural and efficient. In this way, the choice of scaling method significantly biases the meaning attributed to the data.