This is the third in this year’s series of posts by PhD students on the job market.
Creating accurate policy registers is challenging when administrative capacity is low, whether it be lists of the poorest households for anti-poverty programs, or tax registers with individuals' incomes and assets. Most African governments only have reliable information for a very small subset of taxpayers. As a consequence, it is frequent for administrations to rely on discretionary decisions of bureaucrats to expand the tax net. This might have advantages if bureaucrats have specific knowledge on taxpayers, but could backfire if their incentives depart from the government's objective, or if they vary in their performance. Alternatively, administrations can use rule-based or algorithmic processes. This option is becoming increasingly available thanks to digitalization. In my job market paper, with Victor Pouliquen and Bassirou Sarr, I provide the first experimental comparison of bureaucrat discretion with an algorithm for the creation of a tax roll.
Senegal's First Digital Property Tax Census
We partnered with the national tax administration to embed our study in Senegal's first digital property tax census. In theory, owners are subject to an annual tax, the base of which is the value that is or could be obtained from the property rented at market prices. Before the program, to update the tax roll, bureaucrats were sent in the field to observe properties, discuss with occupants, and use their judgment to assign values. This discretionary approach is a reasonable option in a setting where there is a lack of information on real estate prices -- a common feature in low-income countries. This field work was inefficient and only occurred for a dozen properties a year: everything was done on paper and there was no harmonized addressing. As a result, before the program, only 16% of properties were registered.
To meet the administration's new objective of massively expanding the tax net, we collaborated with a developer to create a property tax software including geocoded plot identifiers. 268 bureaucrats deployed in Dakar used the application to generate the modernized tax roll. 96 neighborhoods were covered, spanning approximately a quarter of the region and 40,000 plots, 92% of which were registered.
We experimentally vary the extent to which bureaucrats have discretion in the valuation of the tax base. We randomly assigned half of the neighborhoods to the status quo fully discretionary valuation. In the other 48 neighborhoods, valuations are rule-based: bureaucrats enter 18 observable characteristics of the property (e.g., whether there is a balcony, the type of fence, ...), and an algorithm combining these with location and built area from satellite data generates a predicted value. A third option is a pure rule, an algorithm prediction computed remotely using only location and built area, without any bureaucrat input.
We developed the algorithms with the valuation department and international practitioners. We worked with licensed real estate assessors to create a database of benchmark market values that we used for our calibration.
Discretionary Bureaucrats Undervalue Properties and Generate a Regressive Tax Profile
We find that discretion harms both horizontal and vertical tax equity. Many properties of the same value end up with very different tax bills. This is driven in part by important differences across bureaucrats - we estimate that their heterogeneity accounts for 40% of the overall variation in the extent to which tax roll values depart from market values.
The second striking finding is that bureaucrats undervalue properties, with this undervaluation becoming more severe as the property value increases. The resulting tax profile is shown in black on Figure (1a), with the gray dashed line showing the expected tax profile when applying the tax code directly to benchmark market values. For the 20% cheapest properties, the tax rate observed under discretion (3.8%) is in the realm of the expected tax rate (4.4%). But for the 20% most expensive properties, the observed tax rate is 1.7% instead of an expected 8.6%.
The Algorithm is a Promising Way to Expand the Tax Net Equitably
The algorithm replicates market values relatively well based on property characteristics (especially when compared to similar settings). In contrast, bureaucrats' discretionary valuations are poorly explained by objective property features. Applying the algorithm to a test sample, we find a 0.90 R2 and that 60% of the predictions are within 30% of the market value (54% for the pure rule with a 0.87 R2). We find an R2 of 0.25 for the bureaucrats' implicit algorithm.
Figure (1a) plots the tax profile under the rule-based process (in blue) and the pure rule (in red) and leads to three take-aways. Overall, both rules bring the tax profile significantly closer to the government's objective when compared with discretion.
Second, while in our calibrations, the algorithm including the 18 property characteristics (the rule) is more precise than the pure rule based on satellite data only, this superiority disappears when the implementation is delegated to bureaucrats. This is due to erroneous entries of characteristics (for instance misclassification of the type of cladding), suggesting that even partial discretion affects policy outcomes.
Third, for the cheapest properties, while both rules are more accurate than discretion as measured in property-level regressions, they are also likely to overvalue -- a common drawback of property valuation models due to unobserved variables. Hence, for this segment of the real estate market, bureaucrats' discretion might offer some advantages over the algorithms.
Which Mechanisms drive Bureaucrats' Valuations?
We show that bureaucrats lack knowledge on how expensive high-end properties are. In a lab-in-the-field, bureaucrats were shown the picture of a high-end property and asked to provide their best estimate of the rental value. They massively undervalue, on average bringing the property down from the top to the third quintile of market values (Figure 2).
There are no possible gains for bureaucrats from undervaluing in the lab-in-the-field, and strikingly the tax rate implied by their responses is similar to what we observe in the census data for high-end properties. This provides evidence that collusion between bureaucrats and owners is not the main driver of the regressivity results. Additionally, we don't observe any differential undervaluation gradient for cases where bureaucrats met the owner compared to cases where there were no interactions.
Local information does help reduce the tax base gap, such as when bureaucrats meet the owner or when the property is rented, but only for the cheapest properties, restricting the benefits of discretion to this segment of the real estate market. Furthermore, bureaucrats are subject to biases related to what they perceive as fair. In the lab-in-the-field, those who are told an owner is retired (a status associated with vulnerability) provide a 37% lower value for the same picture.
Policy Implications
Our results highlight that under discretion, bureaucrats generate important inequities and losses for the government. The fact that this stems from their lack of knowledge rather than self-beneficial behavior reveals the importance of misconceptions of the wealth distribution held by the officials in charge of taxation.
The implications of for local revenues are immense in a country with 3.4 billion FCFA (US$ 5.5 million) of property tax collections. Tax liabilities amount to 8 billion FCFA (US$ 13 million) in the discretion arm against a potential of 19.6 billion (US$ 32 million) based on market values. Adopting the rule-based system (respectively, pure rule) increases liabilities to 11 billion FCFA or US$ 18 million (resp., 19.7 billion FCFA or US$ 32 million).
The administration has asked for support to expand rule-based methods. An optimal policy could be to predict which properties are the cheapest based on location and built area, use discretion for these, and apply the pure rule elsewhere (Figure 1b). More generally, our results pave the way for other data-driven strategies to increase fiscal capacity.
Justine Knebelmann is a Postdoctoral associate at MIT with J-PAL. Her research is at the intersection of development and public economics. She relies on field experiments with governments, administrative data, digitalization interventions, and behavioral insights to investigate how states can grow their capacity to tax and to deliver services in an equitable way. Twitter: @JustineKneb. Link to paper.
Join the Conversation