Should we use different discount rates for different projects?

The discount rate is one of the most important considerations of a cost-benefit analysis. It is a critical component that allows us to determine the value we place now on benefits that do not accrue until later. 

Essentially, the discount rate is how we measure uncertainty about the future value of our expected benefits. For example, $100 of benefits today is better than $100 worth of benefits in 50 years because the world in 50 years will likely be very different from the world today. Therefore, the benefits we earn (or costs we incur) tomorrow might not mean as much as benefits earned today.

In order to determine a policy that has benefits 50 years into the future is more efficient than one that has benefits today, we’d need to factor this uncertainty about future benefits into our cost-benefit analysis.

In policy analysis, the selection of the discount rate can dramatically change how we value different proposals. Higher discount rates mean we are less certain about future benefits, and therefore would prefer projects with short-term gains. If the discount rate is too high, we might be short sighted and not plan far enough ahead as a society, valuing present benefits inordinately over those accrued in the future.

Smaller discount rates often make it more likely that future benefits outweigh short term costs. From a policy perspective, this would lead to much more spending on big investment projects. However, if we don’t discount enough, we might be worse off in the short term for projects that might not pan out very well.

In the United States, the base discount rate recommended by Circular A-4 is 7%. The best practice in cost-benefit analysis is to see if your results are sensitive to changes in the discount rate, but for all projects that is the base number to use. For the vast majority of agencies in the cost-benefit analysis world, a single discount rate across projects is preferred.

However, there is one prominent exception to this rule. Analysts in France use variable discount rates depending on the riskiness of the project. For example, imagine we had to choose between investing in a new hospital or in new railroad infrastructure.

Both have massive short term costs and only begin creating benefits after their completion. The main difference is the riskiness of those benefits actually coming to fruition. The hospital will almost always be a useful investment. The “worst case” scenario for the hospital would be if in the future there were no major disasters or pandemics and those hospital beds remained empty. 

Conversely, the railroad infrastructure is much more likely to not realize its benefits. More railroads are important during strong economic times when there are lots of goods that need to be shipped around, but there are always periods of economic downturn. It is much more likely that we will eventually face a difficult time in the economy than we will suddenly not need hospital beds.

Given the intuitive understanding that the discount rate is a tool we have to measure the risk associated with long term benefits, this approach makes a lot of sense. If we can get an idea of how much risk is associated with each type of policy proposal then we would do a better job of efficiently allocating resources. We can do this by allowing for different discount rates between projects. 

There is little consensus among economists about how to choose discount rates for different policies. Because the discount rate can have a significant impact on the results of cost-benefit analysis, it is important that there is a well defined method in place. Allowing analysts to choose discount rates without any framework to back up their choice could lead to severely biased results, and would open the door to potential manipulation.

Currently, the French government adjusts discount rates by looking at price and income elasticities for the goods generated by a policy proposal. This seems to have some strong promise, but in order to fully understand it, more policy analysis needs to be done using it.