n.V1-9.09 | The Discretion Trade

01| The claim is easy to state and awkward to prove. A minister has discretion over a pot of money for small social projects. The projects are approved, announced and delivered. Their effects fade within a few years because nobody follows up. But the minister's network of obligated people grows, and that network has value long after the projects stop working. Nothing illegal occurs at any point.

02| This is not a suspicion in search of evidence. It is one of the better-studied phenomena in political science, under the names clientelism, patronage, and distributive politics. What researchers argue about is not whether discretionary spending is directed politically. It is which voters get targeted and why.

The case that matches the description exactly

03| The cleanest evidence comes from a programme built precisely as described: a fund for small social projects, disbursed at official discretion. Norbert Schady examined Peru's social fund, FONCODES, during Alberto Fujimori's first term — a compensatory fund, money for local projects, distributed by the centre.

04| Two findings, both robust across a large number of specifications and controls. Expenditures increased significantly in the run-up to national elections. And projects were directed to the provinces where the marginal political effect of the spending was likely to be largest.

05| The precision of the second finding matters. This is not a story about money disappearing. The projects were built. What was allocated politically was not stolen funds but the timing and the location of legitimate spending — which is exactly the all-very-legal quality of the original observation. No audit would find anything. The discretion was the instrument.

Why fading impact is politically irrelevant

06| There is a reason the durability of these projects does not constrain the behaviour, and the timing evidence points to it.

07| The political return on a social project is collected at the moment of delivery: the announcement, the visit, the ribbon, the identifiable benefit traceable to an identifiable patron. Whatever effect the project turns out to have arrives years later, is difficult to attribute, and is not observable at the ballot box in the meantime.

08| The clustering of spending before elections is itself the evidence. If political credit depended on outcomes, there would be no reason to concentrate disbursement in the months before a vote — outcomes lag disbursement by years, so election-timed spending could not produce them in time. Spending spikes before elections precisely because the announcement, not the result, is what is being bought.

09| That inference is ours rather than Schady's, and should be read as such. But it follows from the timing pattern he documents, and it explains why a project that works and one that does not produce the same political return at the only moment that counts.

10| One more caution. The literature contains two competing accounts of who gets the money. One says discretionary resources go to core supporters — the loyal base, cheap to reach, reliable to mobilise, easy to monitor. The other says they go to swing voters, where the marginal cost of an additional vote is lowest. Schady's finding is closer to the second. Which pattern prevails appears to depend on the country, the electoral system, and how well parties can observe how people voted. So the minister rewards his own network is a plausible version rather than the established one. What is established is narrower and still damning: discretionary social spending is allocated by electoral logic rather than need.

What happens when the discretion is removed

11| The interesting half of the question is whether design can defeat the mechanism, and there is a large case that says it can.

12| Brazil's Bolsa Família distributes cash to poor households on rule-based criteria, and researchers have looked hard for clientelistic capture. The constraint that appears to do the work is structural: municipal enrolment quotas, calculated in advance from national poverty surveys. A politician wanting to enrol additional supporters runs into a ceiling set elsewhere, by data he did not generate. Independent audits of fraud have found only a few thousand cases among many millions of enrollees. And in places where vote-buying is otherwise common, Brazilians report that this programme is not used for it.

13| The second-order finding is more striking. Beneficiaries are significantly more likely to reject hypothetical vote-buying offers than people on the waiting list in the same municipality — a comparison that holds location constant and differs only in whether the household already receives a benefit no local patron controls. Their voting behaviour does not differ from non-beneficiaries, so this is not a programme buying votes for the government that created it. What it changes is the recipient's exposure. A person with an income that does not depend on anyone's approval is harder to buy.

14| And a broader pattern that cuts against the cynical reading of large transfer programmes: as coverage expanded, both beneficiaries and non-beneficiaries turned against local incumbents. Removing the patron's leverage did not transfer the loyalty upward. It appears to have dissolved some of it.

The machine that decides who qualifies

15| The obvious next move is to remove the decision point with technology. Replace an official's judgement with a database, a biometric check and an automatic payment, and the patron has nothing left to trade. This has been built at enormous scale and studied rigorously enough to say what happens. The answer is not that it fails. It is that it works, that its cost is a different kind of failure, and that both were measured in the same country with the same technology.

16| Karthik Muralidharan, Paul Niehaus and Sandip Sukhtankar evaluated the introduction of biometrically authenticated payments — Smartcards — for India's rural employment guarantee and social pension programmes in Andhra Pradesh. The rollout was randomised across 157 subdistricts covering about 19 million people. This is not a successful pilot; it is a state-scale experiment with a genuine control group.

17| The findings were positive on every dimension that matters here. Payments arrived faster and more predictably. Leakage of funds fell. Enrolment rose rather than fell. Access was not adversely affected — the gains did not come at the cost of shutting people out. On cost, the time savings to beneficiaries alone were roughly equal to the cost of the whole intervention, before counting the reduction in leakage at all.

18| Then the same tools went into India's public distribution system — subsidised food rations, with Aadhaar-based biometric authentication required at the point of sale. Jean Drèze, Reetika Khera and Anmol Somanchi surveyed roughly 1,000 households across 32 villages in Jharkhand and estimated exclusion errors as high as twenty percent where biometric authentication was required for every transaction. One household in five, entitled to rations, not getting them.

19| There is also a live methodological dispute: the two teams disagree publicly about whether the leakage reductions were as large as reported, and the argument is unresolved. For our purposes the dispute is more useful than a verdict. It shows that even with good data, good design and capable researchers, whether a digital system improved matters can remain genuinely contested.

The principle underneath

20| The technology was the same in both cases. The difference — this is our reading, not a finding either team states — lies in what happens when authentication fails.

21| In a payment system, a failed fingerprint means the money arrives late. Annoying, sometimes seriously so, but recoverable; the entitlement persists and can be claimed next time. In a rationing system with monthly cycles and no carry-over, a failed fingerprint means no food this month. The entitlement does not accumulate. An elderly person with worn fingerprints, a labourer whose hands are damaged, a village with no mobile signal on delivery day — each becomes a household that does not eat.

22| Which points at the general principle, and it is not a technical one. Every mechanism that prevents an official from wrongly saying yes also prevents anyone from rightly saying yes. Discretion is what allows a system to accommodate the case the rules did not anticipate. Removing it converts corruption risk into exclusion risk. Neither is ever zero, and moving from one to the other is a choice about which error you would rather make, and about who bears it.

23| That choice is usually presented as a technical upgrade. It is a distributive decision. The gains from reduced leakage are diffuse and accrue to the public purse. The losses from exclusion are concentrated on specific households, who are by construction the people least able to appeal.

24| Procurement is where the same logic works much better, and the reason is structural rather than technological. Ukraine's ProZorro moved public tendering onto an open platform where bids, documents and prices are visible; alongside it sits a community of citizens and analysts examining the contracting data and flagging suspicious deals. The reported savings are large — the December 2021 US Strategy on Countering Corruption cited close to $6 billion since 2017 — and should be handled with care, since they come from official and advocacy sources rather than independent econometric evaluation, and "savings" in procurement depends on the counterfactual price you assume. The direction is well supported; the magnitude is a claim. But the structural point holds regardless. In procurement, the party excluded by a rule is a firm. A company that fails to qualify loses a contract. Nobody goes hungry. The exclusion cost that makes biometric rationing dangerous is largely absent, which is why transparency reforms in procurement have a better record than transparency reforms in benefits.

The failure mode nobody plans for

25| One more pattern, and it is why did the government digitise? is the wrong question. Isomorphic mimicry applies directly. A government under pressure to be transparent can build a portal that publishes contracts as scanned PDFs, unsearchable and unaggregatable. It can create a digital registry with a manual override senior officials retain. It can automate eligibility and keep a discretionary hardship fund large enough to restore the old relationship.

26| Each of these is digitisation. None of them moves the discretion. And each is substantially cheaper and less politically costly than the real thing, which is why the form spreads faster than the function. The diagnostic question is not whether a system was digitised, but whether anyone can still decide an individual case — and if so, who, and how large that residual is.

27| So: discretionary social spending is directed by electoral logic, measured robustly in at least one programme built exactly like the one described. The projects being real and the process being legal does not weaken the finding — it is the finding. The mechanism is defeatable, but not by better intentions. It was defeated in Brazil by removing the point at which a human being decides who qualifies, and in Andhra Pradesh by removing the human from the payment point. And each removal bought its result by trading one class of error for another.