Discover more from Overcoming Bias
Testing Tax Career Agents
Agents who are paid a larger fraction of their client’s income have a stronger incentive to promote and advise those clients. Thus the tax career agent idea makes more sense for governments that tax a larger fraction of citizen income. So they make less sense for city or state governments. But a national government may not be willing to try it without seeing results from a smaller experiment. So how could we make a smaller experiment to test the concept?
The big advantage of tax career agents is that they are basically free to create. As the government already sits in that role, all it has to do is transfer that role to someone else, at almost no net cost to anyone. But to test the idea of tax career agents, we don’t have to rely on the fact that such agents are free for governments to create; we can pay extra to create such agents privately, just for the test.
To create a test tax career agent regarding client C, we could hold an auction to see who is willing to pay the most to, every year for the next Y years, be paid an amount equal to what C pays that year in income taxes. If agent A were a real tax career agent, the money paid to A would come from what C actually pays their government in taxes. But for the purpose of an experiment, this money paid to A could instead from the budget of the experiment. This alternate payment source should not matter much for A and C behavior.
So auction winners would first give large auction win payments to the experiment, after which the experiment would commit to paying them back each year when their clients pay taxes. In the meantime, the experiment would hold and invest these assets. As the experiment should be able to invest as well as agents, auction competition should induce the net cost of this experiment to be mainly the time and effort costs that agents expect to make advising and promoting clients.
In a prior post, I estimated that an agent A who got 20% of client C’s wages would increase those wages by 1-3%. As agents wouldn’t on average put in more efforts than they get paid for those efforts, that gives us an upper bound to the financial size of agent efforts. Thus if N clients with average income I each get a test tax career agent for Y years, the auction revenue to be collected and invested would be ~20%*Y*N*I, and the cost to create these agents would be ~1-3%*Y*N*I. (For Y large, these amounts are lower due to time discounting.) Note that much of this “cost” is actually a transfer to clients, who we expect to enjoy higher incomes.
Of course we’d want to track a similar-sized control group of N clients who didn’t get test tax career agents. And if we wanted to give experimental subjects the choice of if to create such an agent, then if only a fraction V volunteer to get such an agent, we’d want to track ~N/V workers who were offered the choice, and another ~N/V control workers not offered the choice. Note that we’d also need funds to manage the experiment, to collect data on participants, and to analyze the results.
And that’s a simple outline of the experiment design, including a rough estimate of its cost. In the U.S., 3% of $31K median income over ten years is $9.3K, which for N=1000 comes to $9.3M. This cost would of course be less for lower-income workers. Any want to do an analysis of what size N we’d want given to see significant results given this expected effect size?
Added 21Oct: Christopher McDonald did a Sample Size Analysis for Tax Career Agent Experiment. Assuming tax career agents improve wages on average by 0.3% per year, he finds that you’d need 7000 subjects over a ten year experiment to get a true-positive probability of 80% and a false-positive probability of 5%. So applying the same estimates as above, that gives an upper-bound cost of ~$30M. In my next post on this, I’ll outline a cheaper experiment design.
Added 12Nov: I just realized that I’d previously mis-calculated the wage rise to be 1-3%, instead of 10-30%. A smaller experiment would of course be required to see such a larger effect.