Central government's motivation to devolve property rights to experimentation

Why would central government devolve the right to experiment? We can identify three motivations for experimental devolution. The first derives from uneven knowledge distribution. The notion that each individual holds idiosyncratic bits of knowledge suggests that individuals can generate different bits of new knowledge if interaction is left to them (Hayek, Reference Hayek1948; Turgot [1775] Reference Turgot and Baker1987). The second is that complex problems usually push diverse individuals to pool resources together. Sharing the pain of policy experimentation entails sharing the gain: new knowledge (Kollman et al., Reference Kollman, Miller and Page2000). The third motivation concerns the containment of the costs of a possible policy failure.

The three motivations, as parts of our story shall also intimate, often operate together. However, we presuppose that the primary motivation for central government to grant rights to experiment to local governments is the third one. The reason is simple: policy failure is politically unpopular. A significant policy failure (like a perfect policy success) is rare, but it does occur, and usually correlates with bad policy design. Recall for instance the UK public expenditure programmes on the Concorde aircraft and on the Advanced Gas-Cooled Reactor (Henderson, Reference Henderson1977). More recently, we can think about the higher costs that the US would have incurred had the California electricity deregulation experiment of the early 2000s been done nationally; or, in a similar vein, about the costs that would have turned up if Texas's 1993 multiyear experiment with school finance equalisation (the so-called Robin Hood plan) had been national. In essence, our stance is that central government has an interest in devolving property rights to experiment primarily for political concerns.

Limited cognition and implicit conditionality

Suppose that the grantor writes down several grant conditions. This conditionality might be interpreted by the grantee (e.g. a Brazilian mayor, an Italian regional president, a US state governor) as a signal that the grantor views other aspects of the grant as relatively less important, leaving many degrees of freedom along the unwritten aspects, also in terms of not necessarily spending for experimentation. Under a totally unconditional grant this would be the case a fortiori. As a result, there could be cases of policy drift (McCubbins et al., Reference McCubbins, Noll and Weingast1997). A recent illustration is when in March 2015 the South African Treasury withheld one of its grants to 60 municipalities when the latter did not employ the unconditional transferred funds as expected.Footnote ⁸

In these cases, the first instinct is to try to tie the grantee's hands through more conditionality. A scenario that comes to mind is for the grantor to use grants as a way to direct subnational policy experiments toward the (perceived) highest expected value experiments. The idea is that local policymakers do not exercise their experimental potential in ways that are welfare maximising, as they may experiment only on policy options that make the grant programme a safer bet for them. Reasons for grant conditionality then may include differences in legal enforceability for third parties (e.g. local beneficiaries of the grant project), legislators wishing to constrain the discretion of bureaucracies that will administer the grants, the desire to bind subsequent national legislatures that might take a different view of the terms of the grant, attempts to decrease local corruption, and the like. Thus, the rationale would be that highly conditional grants, in which the central government closely directs the policy chosen by each sub-unit, may leave less discretion for each local policymaker, but produce more valuable experiments overall (Galle and Leahy Reference Galle and Leahy2009).

In this scenario, conditionality avoids the costs of directly monitoring the binding of the target experiment. We do not disagree with the view that more conditionality ex ante has the potential to offset ex post commitment and enforceability costs tied to limiting perverse incentives. But this scenario, where experimentation is planned from the grantor, would at least imply that the grantor to some extent suffers from fewer cognitive limitations than the grantee (Minkler, Reference Minkler1993).

A second scenario would be for the grantee to guess about the importance of all the terms ex post. In this case we would need to think of the grantee not just as a risk lover, but also as possessing particularly strong political intelligence. A scenario that is less aleatory and that would require a more mundane political intelligence is the one that lets the grantee negotiate (at some cost) the importance of all the terms ex post. But this third possibility does not necessarily guarantee reduction of perverse incentives after negotiations close (Williamson, Reference Williamson2005).

There is another ex ante scenario that can assist in assuring binding, however. We refer to implicit conditionality, the notion that there may be conditions from relational understandings that are outcomes of negotiation as well, albeit not explicitly spelled out in the grant per se. Since, as underlined at the outset, political relations are intertemporal rather than discrete, these implicit conditions reduce the probability of deviation from the stipulated terms even when a grant is explicitly unconditional (Brennan and Pincus, Reference Brennan and Pincus1990; Garzarelli, Reference Garzarelli2006).

Implicit conditionality therefore can be seen as bringing in a ‘gap-filling’ (Ayres and Gertner, Reference Ayres and Gertner1989) dimension to mistake-ridden policy innovation. That is, this last scenario can be a means to decrease uncertainty by saving on ex post costs of political compromise through ex ante relational agreements about terms for experimentation. The ensuing pages accordingly consider that, on balance, implicit conditionality crowds in the incentives to experiment should explicit conditionality fail to do so; and that those cases where grants are not used as destined can be thought of as mostly exceptions resulting from weak exogenous political incentives, such as local and/or central leadership turnover and end of term in office with no prospect of re-election.

Let us be clear here. We are not suggesting that a policy success is a necessary precondition to receive future grants, promotions, etc. Rather, the precondition that sufficiently qualifies for success in our context is that grants are employed as destined, namely for policy experimentation. In a behavioural setting, in fact, optimal rationality is about a reasonable rule-following procedure in light of the situation faced, which does not necessarily coincide with an optimal result in the standard optimisation sense (Simon, Reference Simon1990).

Economising rules

Since policymaking is a very uncertain production activity, it usually proceeds in incremental or so-called piecemeal fashion, where one tries out different solutions to actively seek mistakes in order to extend knowledge and improve practice. For example, in the recent case of legalising cannabis in some US states, reflecting on

So the rational (but fallible) policymaker proceeds gradually by exploring the solution space by taking many small decisions over time on each policy under consideration and treasuring the lessons learned at each step. The lessons learned during this muddling through are what endogenously drive policymaking forward, because each moment of choice builds on the outcome from each previous step. The point being that the positive and negative lessons learned serve as rules of decision-making when each policy – each production activity – is considered separately in terms of its value-generating attributes (degree of externality internalisation, institutional compatibility, revenue stream generation, etc.), which is the basic intuition behind the dynamics of our simple heuristic. Namely, that over repeated trials through time, the frequency of the selection of a policy is proportional to the net welfare value of selecting that policy.Footnote ⁹

If we suppose, for simplicity, that there are a number of experiments on two competing policies, x ₁ and x ₂, and that for each experiment there is a corresponding welfare value function, v ₁(x ₁) and v ₂(x ₂), then we can more precisely render how the frequency of policy selection changes over time through a familiar replicator equation, such as:

(1) $$\begin{equation} \left. {\frac{{d{x_i}}}{{dt}}} \right|_{i = 1}^2 = {x_i}\left[ {{v_i}\left( {{x_i}} \right) - \bar{v}\left( {\bm{x}} \right)} \right]. \end{equation}$$

$[ {{v_i}( {{x_i}} ) - \bar{v}( {\bm{x}} )} ]\ $is the difference between the average value of the trials of the chosen policy, v_i(x_i), and the average of the values of all policy trials, $\bar{v}( {\bm{x}} ) = \mathop \sum_{i\ = \ 1}^{n\ = \ 2} {x_i}{v_i}( {{x_i}} )$. If ${v_i}( {{x_i}} ) > \bar{v}( {\bm{x}} )$, then x_i will continue to be selected; if instead ${v_i}( {{x_i}} ) < \bar{v}( {\bm{x}} )$, then x_i will not continue to be selected. There is indifference between policies when they have the same value.

The simple evolutionary dynamic from equation (1), explicitly or otherwise, describes behaviour under uncertainty in a variety of contexts and fields where repeated trials are involved (Fudenberg and Levine, Reference Fudenberg and Levine1998; Hofbauer and Sigmund, Reference Hofbauer and Sigmund1998; Sugden, Reference Sugden2005). For the purposes at hand, it informs us that a policymaker experiments – learns by trial and error – according to average rather than marginal welfare value per unit of investment (a money sum, knowledge, opportunity cost, time, etc.) at each step taken. More precisely, the muddling through occurs at the margin, but the reinforcement (negative or positive) to the muddling through is considered in terms of its average return. The heuristic therefore is that the higher the average welfare value from a policy in relation to alternatives, the more a policy will be pursued.Footnote ¹⁰ It resonates with the view that when uncertainty does not permit straightforward decision making, rules of behaviour are coarser rather than finer (Bookstaber and Langsam, Reference Bookstaber and Langsam1985) – less is more (Gigerenzer and Brighton, Reference Gigerenzer and Brighton2009).

Besides its intuitive appeal, the heuristic is factually not unrealistic. Take a tangible policy case, which while not in the specific context of grants can be easily related. The Johannesburg Roads Agency introduced the freely downloadable JRA Find&Fix application for smartphones in 2014, which allows drivers to report in real time a pothole with GPS coordinates as well as a photograph.Footnote ¹¹ The app competes with more traditional policies for reporting a pothole, such as contacting the relevant authorities by telephone to report its approximate location and description.

If we think about these two competing policies as our x ₁ and x ₂, then it is patent that, over time and with other conditions remaining the same, v ₁(x ₁) trumps v ₂(x ₂). For, among other things, x ₂ first requires a search – usually after some time from the initial reporting, which would also include a description of the size of the pothole – for the precise pothole location so that, usually after some additional time, a repair team can fix it. x ₁ instead bypasses the description and search steps, implying that it exhibits transaction cost and time saving economies that on average allow for quicker pothole repair compared to x ₂. Hence, x ₁ will be preferred to x ₂.

As a matter of fact, thanks to JRA Find&Fix, ‘the average time to resolve’ a road infrastructure ‘service request’, such as ‘potholes, faulty traffic lights and other road-related defects’, reduced ‘from 32.4 days to less than a day’. The app ‘complements’ the Roads Agency's ‘ability to plan repairs . . . based on reports and data received, thereby enhancing efficiency and increasing productivity’.Footnote ¹²

Grants and policy experimentation

We are prepared now to illustrate how an intergovernmental grant aids policy experimentation in a federation. The illustration hinges on likening the dynamics of trial and error policy production, as introduced above, to experimenting within a grant-generated policy experimentation space, with the bottom line being that the conditionality from different grants creates different policy experimentation spaces.

In line with the discussion so far, we continue to reason in two dimensions, and take as our comparative institutional cases two grants of the same amount, but of different types: an unconditional lump-sum and a conditional closed matching. The reason for selecting these two grants is that the characterisation of most other grants can be seen as falling within these two. That is, we can think about unconditional lump-sum grants and conditional closed matching grants as the ideal-typical inclusive polar extremes of a possible grants spectrum.Footnote ¹³ More generally, the simplifications do not entail loss of generality (Oates, Reference Oates1972: 101–2). The first section of the Appendix contains a formal treatment.

Assume two levels of government, central and local, the absence of fiscal illusion and of local government borrowing, that the pre-grant local budget is, as usual, credibly fixed from the previous year, and that central and local taxes remain constant. In keeping with our previous example, we can keep thinking about experiments on two competing policies, x ₁ and x ₂, that try to solve the same problem. Following standard practice, the number of experiments on both policies is chosen so that the cost of one unit in terms of the other is unity. The pre-grant local budget constraint, raised from only local taxes, is AB, with slope τ. See Figure 1.

Figure 1.

Experimentation with an unconditional lump-sum grant.

Suppose that local government, in the form of our policymaker, is the recipient of an unconditional lump-sum grant. Further suppose that the Bradford–Oates ‘veil hypothesis’ holds, viz. that there is equivalence between the grant and a budget increase of the same size (Oates, Reference Oates1999: 1129–30). (We will characterise this supposition more precisely presently when discussing the flypaper effect.) In terms of Figure 1, the grant shifts the local public budget from AB outward to CD.

Our primary interest here is on the experimental effects from the grant. We know that the grant serves both as a constraint and as an option for policy experimentation. The constraint is the budget for conducting experiments on x ₁ and x ₂. It is equal to the money amount transferred from central to local government through the grant, entailing that experimentation is not unbound. Relatedly, recall that it is in the interest of the grantee-as-policymaker to employ the grant amount as destined.

Being revenue extraction from local taxes, AB does not incorporate the institutional nudge from grantor to grantee; and the space above the grant is not feasible. The option consequently manifests as a proportion between the simple policymaking heuristic from equation (1) and the area that, by difference, is created by the grant:

(2) $$\begin{equation} \left. {\frac{{d{x_i}}}{{dt}}} \right|_{i = 1}^2 \propto {\it ABCD}. \end{equation}$$

In different terms, relation (2) informs us that experimentation is commensurable to possible experiments. The constraint and the option jointly define the space for muddling through, which is essentially the budget set for experimentation. That is, the bundles of experiments lying inside and on the boundaries of region ABCD in Figure 1.

Policymaking from the grant proceeds along the following lines. As a starting point, the policymaker can pick any combination of policy experiments within region ABCD. Suppose experimentation starts at a locus where there is an equal number of experiments on x ₁ and x ₂, such as E. Then, depending on the experimental feedback, the policymaker will take the next step. Thus, the arrows emanating from E are not vectors, but indicative of direction.

If x ₂ has a relatively higher average welfare return, then the policymaker would move in a north-westerly direction, for example toward point F. A movement toward H reflects learning about a relatively higher average welfare return for x ₁. A diagonal movement between 0 and 90 degrees represents indifference: ${x_1}\sim\ {x_2} \Leftrightarrow [{v_1}( {{x_1}} ) = {v_2}( {{x_2}} )] \ge \bar{v}( {\bm{x}} )$.Footnote ¹⁴

The halt of experimentation is less evident. On the one hand, there is the case of funds from the grant being exhausted, which occurs anywhere on CD. However, lying on CD can imply either experimental failure or experimental success. That is to say that learning through the heuristic does not automatically guarantee policy success even when the entire grant amount is employed as expected. Indeed, recall that behavioural analysis is less about rational optimisation of outcomes than it is about adaptive rule-following procedure.

On the other hand, the experiment can halt before the grant funds are exhausted. Here too there can be experimental failure or success. Regardless of experimental outcome, the policymaker would be somewhere within region ABCD exclusive of the upper boundary CD – for example, at point I.

I represents failure when experimentation stops without stumbling upon a successful policy. In this case, we can think of I as, e.g. a case of policy drift, such as the episode of the South African municipalities reported earlier, which we hinted could originate from a variety of weak exogenous political incentives.

I is a success when experimentation stops in light of a successful policy. In this case, the residual funds can be invested in something related to the successful experiment until grant funds are exhausted. The Public Safety Department (which consists of both fire and police) of Fairfield, California, for instance, was allowed to keep savings from a lump-sum grant. This led to a change in strategy from only fighting fires to fire prevention, which lowered the running costs and improved the insurance rating of the department (Osborne and Gaebler, Reference Osborne and Gaebler1992: 222–6).

Notice the asymmetry. Experimentation stops when there is exhaustion of the grant funds, irrespective of success or failure; while the exhaustion of the funds is not necessary for experimental success or failure. The important detail to keep in mind, though, is that learning occurs on both sides of the asymmetry.

At this juncture the flypaper-effect theorist will point out that empirically there is non-equivalence in a change in spending between an unconditional lump-sum grant and a corresponding parallel shift in grantee budget (Hines and Thaler, Reference Hines and Thaler1995). For present purposes, a simple recasting of our reasoning offers one rationale about why the budget shift of Figure 1 might not be an anomaly.

The flypaper effect considers that local government spending may not always be in line with welfare maximisation. Rather, the grant stimulates local government spending beyond the welfare maximisation point (but still on the grant-aided budget constraint). Spending on policy experiments, however, is not per se welfare enhancing. In our approach, mistake-ridden learning also enters community welfare. Two considerations follow when also accounting for learning. First, a successful policy experiment may be either welfare enhancing (JRA Find&Fix improves pothole repair turnaround times) or welfare-neutral (JRA Find&Fix leaves pothole repair turnaround times unchanged). Neither of these cases contradicts flypaper-effect models (Bailey and Connelly, Reference Bailey and Connolly1998: 336–8). Second, policy failure can be seen as decreasing welfare from a financial viewpoint until learning offsets it. In this case, spending is not reflected as it is in flypaper models.Footnote ¹⁵ Thus, the approach proposed here accounts analytically for both variation in local government spending from the flypaper effect and spending on failed policy experiments – the policy experimentation space incorporates both.

Pointing out that the approach analytically can account for the flypaper effect as a special case is not the same as simultaneously accounting for the economics of why we can consider the grantee-as-policymaker as having an incentive to span the full experimental space. Economists of diverse persuasions (Brennan and Pincus, Reference Brennan and Pincus1990; Inman, Reference Inman, Durlauf and Blume2009; Roemer and Silvestre, Reference Roemer and Silvestre2002) share common ground when emphasising the importance of political considerations for grantee behaviour, especially in the face of unconditional lump-sum grants. The common ground is that grant relationships are characterised by equilibria among many political actors, implying that the translation of budget effects from single-consumer to collective decision-making contexts is off the mark.

Our political scenario concerns repeated exchange. In repeated exchange, an initial large number of local governments involved in negotiation with the central government changes into a bilateral, non-anonymous grantor–grantee exchange relation when the grant is implemented (and when renegotiation occurs). The grantee–grantor relation is then a type of sunk investment in reciprocal political capital that can be changed only at significant sacrifice. As repeated exchange increases the value of the investment in reciprocal political capital, small number exchange increasingly dominates.Footnote ¹⁶ This means that the constraint on the feasible set tightens for the grantee's situation as such as well as for the repercussions of grantee–grantor interaction. The repercussions from repeated exchange create a strong ‘coercive force’ that renders the grant binding endogenous (Langlois and Csontos, Reference Langlois, Csontos, Mäki, Gustafsson and Knudsen1993: 117–20). It follows that we can heuristically consider the grantee-as-policymaker as having a vertical top-down incentive to employ an unconditional lump-sum grant for policy experimentation in its widest sense, rather than let it crowd out own-source funds on a dollar-for-dollar basis, not just because there is an exchange of commitments embedded in a grant's institutional interstices, but also because the exchange recurs bilaterally.

Let us now turn to the conditional closed matching grant of the same amount. Since the policymaking procedure is behaviourally fully analogous to case of the unconditional lump-sum grant, we can focus directly on the comparative institutional analysis of the two grants in terms of the respective experimental spaces.

Let the grant direct more funding to x ₁, as in Figure 2. The local public budget shifts from AB to AJC. m pivots the budget about point A to AJ. The slope of segment AJ is (1 − m)τ. The grant reaches its maximum amount (which, recall, is the same of the previous grant) at J. The remaining budget constraint JC has the same slope as the unconditional lump-sum grant, which is also the slope of the original budget (τ).

Figure 2.

Experimentation with a conditional closed matching grant.

The conditions involved are two: the matching ratio, m ∈ (0,1), and the maximum amount at which the grant closes. m represents the fraction of the cost of the experiment subsidised by the grant (e.g. if the donor funds 50 cents on every dollar spent on experimenting on x_i, then m = 0.5).

Conditionality here assures that, as a targeted grant, the closed-ended matching grant focuses on a specific experimentation area, and therefore would not allow policy trials outside the constrained area. Otherwise, if the closed grant is fully used – namely, in the range of constrained funding – then the grant's (equally valued) funds become fungible and resources can be allocated more generally within the budget – like the lump-sum grant. (The reverse is less problematic. If the closed end matching grant is not fully used, then we are in the range where the price incentive of the grant is still binding, even if in that range the local government will receive less dollar support from the targeted grant than from the lump-sum grant.)

By inspecting Figures 1 and 2 we notice that the constraints and the options for experimentation differ. The constraint of the conditional closed matching grant is twofold: the amount of the grant as well as the matching ratio. The grantee receives the full grant amount only by experimenting on x ₁ to the grantor's satisfaction, which is at least the horizontal distance BC. The option is:

(3) $$\begin{equation} \left. {\frac{{d{x_i}}}{{dt}}} \right|_{i = 1}^2 \propto {\it ABCJ}. \end{equation}$$

Accordingly, the space for muddling through in Figure 2 is represented by all the bundles of experiments lying inside and on the boundaries of the region ABCJ.

Further inspection of the two figures shows that ABCJ < ABCD. That is to say that the experimentation space is smaller for a conditional closed matching grant than for an unconditional lump-sum grant. The intuition is that the matching ratio reduces the experimental space by favouring a specific experiment (in this case, x ₁). There is therefore less experimentation potential under the conditional closed matching grant than under the unconditional lump-sum grant. The second section of the Appendix presents a proof of this result.

Some factual considerations

Consider US federalism. In 2015, welfare transfers to state and local governments amounted to 3.4% of GDP. Approximately 76% of these transfers (US$16 billion) are for the Temporary Assistance for Needy Families (TANF) programme, of which lump-sum grants for cash assistance comprise approximately 25% of total spending (about US$4 billion) (NASBO, 2016; OMB, 2017). The right to these non-trivial figures from less conditional grants was gained over the years by significant intergovernmental fiscal interaction, and was itself the fruit of mistake-ridden learning.

The Aid to Families with Dependent Children (AFDC) was a New Deal welfare programme in support of families without fathers that ran from 1935 to 1996. It provided states with federal funds in the form of open-ended matching grants for cash assistance, and allowed states to determine the level of monthly benefits. Under the AFCD, states could experiment with welfare policy from 1962 by applying for a statutory waiver from the federal government, hence creating leeway for ‘experimental, pilot, or demonstration’ projects aligned with the socioeconomic areas of the programme.Footnote ¹⁷ The federal government had discretion to approve the waiver in full or in part, and could change the scope of the experiment. By the end of the AFDC programme, almost all 50 states had received waivers and experimented with eligibility rules, work participation requirements, time limits of benefits and payment policies. However, among other issues, waiver applications were seen as a form of red tape that delayed local policy initiatives. Perhaps more significantly, states resented the interference of the federal government in their freedom to experiment, blaming the failure of many local experiments on central intervention. These central–local tensions around limiting discretion generated a phase of imbalance between levels of government that unintentionally led to institutional change (Bulman-Pozen and Gerken, Reference Bulman-Pozen and Gerken2009: 1274–6).

It is during the imbalance phase that, in the attempt to more precisely tailor welfare policies to their idiosyncratic needs, states induced significant reform in federal law toward more experimental autonomy. Eventually, the amount of vertical negotiation among levels of government during the phase led to a new welfare programme for cash assistance to families with children: the TANF, which, in 1997, replaced the more conditional AFDC (Van Wiggeren, Reference Van Wiggeren1997).

The TANF does not require the condition of an absent father for disbursement, and delegates to states the authority to design their own welfare programmes, subject to a few federally mandated conditions regarding work participation targets and time limits for cash assistance. States also have the flexibility to allow counties to decide the rules and cover preferred. Upon implementation of the TANF, states made numerous changes to their cash assistance programmes in terms of eligibility rules, work participation requirements, time limits for assistance, and payment policies (Zedlewski et al., Reference Zedlewski, Holcomb and Duke1998). Moreover, the experimental flexibility under the TANF allows states to make small or incremental changes to cash assistance programmes. Additionally, under the AFDC, matching grant states would bear the full cost of any failed experiments, but only retain a percentage of savings from the successes. The TANF, however, allows states to keep all savings from successful cost reduction experiments. Thus there is an indirect positive incentive to reduce welfare dependency that was not present under the open-ended matching grant of the AFDC.

From the AFDC to the TANF the US as a federal laboratory moved from a single cash assistance programme to more than 50 different programmes that have the potential to evolve more freely through local experimentation by means of less conditional grants. More generally, the policy shift testifies the iterative learning process that took place on two levels, which we can term macro and micro (using language somewhat informally). The macro level concerns iterations between central and local governments that eventually led to the TANF; the micro level concerns local iterations through grants. The iterations on the two levels were synergetic because there was bipartisan political will at the federal level both to concede rights to experiment and to learn from the state experiments, which is something that the US Congress currently seems to be missing when it comes to reforming healthcare.Footnote ¹⁸