Signalling, screening and information

You are standing at a used-car lot outside Mombasa, or at a cattle auction in the Rift Valley, or across a desk from a loan applicant who has no payslip and no title deed. In every case one side of the trade knows something the other cannot verify. The seller knows whether the engine is sound; the borrower knows whether she will actually plant the loan or divert it. Asymmetric information is not a frictionless-market nicety you can assume away — in much of the African economy, where collateral is thin and formal records are sparse, it is the central obstacle to trade. This module gives you the apparatus to see it precisely, and the mechanisms markets invent to defeat it.

Two failures, not one

Draw the distinction sharply, because everything follows from it. Adverse selection is hidden information: the uninformed party cannot observe the other's type before the contract is signed — the car's true quality, the borrower's riskiness, the worker's ability. Moral hazard is hidden action: the type may be known, but after the contract the informed party takes an action the other cannot observe — how hard the borrower works, whether the insured driver still locks the car. The timing is the whole game: adverse selection is a problem of selection at the door, moral hazard a problem of behaviour once inside.

Akerlof's market for lemons: how a market unravels

Start with George Akerlof's 1970 model — the paper that helped found the field and win a Nobel. Take a market for used cars (a 'lemon' is American slang for a bad one), though a Sahelian cattle market or a Kampala boda-boda market works identically. Sellers know the quality of their own vehicle; buyers see only a row of machines that look alike. Crucially, gains from trade exist for every car — the buyer values it above its current owner. Under full information every car would change hands. Watch what hidden quality does to that benchmark.

Quality q = the car's value to its current owner (the seller).
   q is spread evenly (uniform) over [0 , 10,000], in some currency unit.
   Only the SELLER knows a given car's q. Buyers know only the spread.
   A buyer values a car of quality q at  k × q,  with k = 1.5
   (real gains from trade: 1.5q > q for every car).

Efficient benchmark: since 1.5q > q, EVERY car should trade.

Now a buyer offers price p. Who sells?
   A seller sells only if  p ≥ q  →  cars offered are those with q in [0 , p].
   Average quality of what is offered:  E[q | q ≤ p] = p / 2.
   A buyer therefore expects value  k × (p/2) = 1.5 × (p/2) = 0.75 p.

The buyer will pay p only if  0.75 p ≥ p  →  0.75 ≥ 1, which is FALSE for any p > 0.

The death spiral:
   offer p = 10,000 → offered avg 5,000 → true value 7,500 < 10,000  (overpaid) → cut
   offer p =  6,000 → offered avg 3,000 → true value 4,500 <  6,000  (overpaid) → cut
   ... every round the best cars withdraw, dragging the average down ...
   stable price → 0.  The market collapses to the single worst car (q = 0).

Follow the logic. When a buyer offers a price, only sellers whose cars are worth that price or less will part with them — the good cars are held back. So the pool on offer at price p contains only qualities up to p, average p/2. The buyer, valuing that pool at 1.5 × (p/2) = 0.75p, pays p for something worth 0.75p, and revises the offer down. But a lower price withdraws the next tier of cars, dragging the average down again. Each round of 'the best remaining cars exit' feeds the next, and the price chases a falling average to zero. This is adverse selection at full strength: the market does not shrink, it unravels.

Real used-car and livestock markets obviously do function, which tells you traders have invented ways to move private quality information across the counter. Warranties are the cleanest: a promise to cover repairs is cheap to keep if the car is sound and ruinous if it is a lemon — so only good-type sellers offer one, and the warranty itself certifies quality. Third-party inspection does the same by verification rather than incentive: the pre-export JEVIC certificate on cars shipped through Mombasa, or a trusted mechanic's once-over, reveals q directly. Brokers — the dalali in a Swahili-speaking market, the cattle middleman who knows every herder — stake a reputation built over many trades; because their future income depends on not vouching for junk, their word carries information a one-shot seller's cannot. Each device supplies a credible, differentially costly channel for the truth.

Moral hazard: hidden action after the contract

Now hold type fixed and hide the action. A microlender funds a tailoring business; whether the borrower buys a sewing machine or diverts the cash to a cousin's wedding is an action the lender cannot cheaply observe. An insurer covers a shop against fire; whether the owner still keeps an extinguisher is hidden. In each case the contract itself changes behaviour, because the informed party no longer bears the full consequence of the action. A quick test separates the two failures: if the problem would exist even if no contract were ever signed — bad types are simply out there — it is adverse selection; if the problem is created by the contract, arising only because behaviour changes once cover or credit is granted, it is moral hazard.

Spence's signalling: making the invisible visible

Michael Spence's 1973 model asks how the informed party can credibly prove its type when talk is cheap. His answer: take a costly action that the good type can bear more easily than the bad type. The classic instance is education. Suppose — provocatively — that schooling adds nothing to productivity; it is pure signal. High-ability workers still find studying less costly (in effort, time, repeated exams) than low-ability workers do. That cost difference is the entire mechanism. If firms come to believe that anyone holding a certain credential is high-ability, can that belief be self-confirming? Work the numbers.

Two worker types, paid their identified value in a competitive labour market:
   High-type value  w_H = 200,000 / year
   Low-type value   w_L = 100,000 / year      →  wage gap  w_H − w_L = 100,000

Education level e (years). Education adds NO productivity. Marginal cost per year:
   High type:  c_H = 20,000 / year   (studying is easier — LOWER cost)
   Low type:   c_L = 40,000 / year   (studying is harder — HIGHER cost)
                                    ← single-crossing: c_H < c_L

Firms adopt the belief: pay w_H if e ≥ e*, pay w_L if e < e*.

Condition A — the credential must be COSTLY ENOUGH to deter the low type.
   Low type prefers e = 0 (take w_L) over faking e* (get w_H):
       w_L ≥ w_H − c_L·e*      →      c_L·e* ≥ w_H − w_L
       40,000·e* ≥ 100,000     →      e* ≥ 2.5 years

Condition B — the credential must be CHEAP ENOUGH that the high type still buys it.
   High type prefers e* (get w_H) over e = 0 (be seen as low):
       w_H − c_H·e* ≥ w_L      →      c_H·e* ≤ w_H − w_L
       20,000·e* ≤ 100,000     →      e* ≤ 5 years

Separating band:  e* ∈ [ 2.5 , 5 ] years.
   Any diploma requiring between 2.5 and 5 years of (productivity-irrelevant) study
   cleanly sorts the two types.

The two inequalities are the heart of it. Condition A makes the credential expensive enough that a low-ability worker, offered the high wage for faking it, would rather stay uneducated on the low wage — the signal must cost the impostor more than the wage gain it buys. Condition B makes it cheap enough that a high-ability worker still acquires it rather than pool with the low types. Because the low type's marginal cost (40,000) exceeds the high type's (20,000), room opens between the bounds: any study length between 2.5 and 5 years satisfies both. Inside that band the firms' belief is self-confirming — high types educate, low types do not, and paying by credential is rational. That is a separating equilibrium: the two types take different actions and thereby reveal themselves.

Two cautions. First, separation is not the only outcome: if the credential falls outside the band — too cheap, so low types also acquire it, or too dear, so high types skip it — the market pools, everyone is paid the average, and the signal conveys nothing. Second, and more unsettling, in the pure model the years of study produce no skill; the whole expenditure is a sorting cost society pays to sift workers it could not otherwise tell apart. Read the African graduate labour market through this lens. Where many qualified applicants chase each vacancy, employers lean on the degree as a filter for the discipline and baseline ability that completing it signals, and requirements ratchet up — a diploma, then a master's — as each credential becomes common enough to stop separating. That ratchet is degree inflation, and the model says when it bites: as the cost gap between types narrows, the band shrinks and the old credential no longer sorts. The honest qualification is that real schooling also builds human capital, so the truth lies between pure signal and pure skill — but the signalling logic is unmistakably part of what a certificate does.

Screening: the uninformed party designs the menu

Signalling has the informed party move first and burn resources to reveal itself. Screening flips the order: the uninformed party moves first, offering a menu of contracts engineered so that different types choose different items and sort themselves. An insurer offering a low-premium, high-deductible policy alongside a high-premium, low-deductible one lets the low-risk pick the first and the high-risk the second — self-selection classifies them. The sharpest recent African instance is mobile digital credit. Products such as M-Shwari and Fuliza in Kenya cannot see a new borrower's type and cannot ask for collateral, so they screen with data: they read your mobile-money inflows, airtime purchases and repayment record, start you at a small limit, and grow it as your behaviour reveals your reliability. The menu is a schedule of limits and prices indexed to observed conduct — a screen that manufactures the credit history the formal system never recorded for you.

The extended case: joint liability and group microfinance

Bring the whole apparatus to bear on the problem that defines development finance. A poor borrower lacks precisely the two instruments that normally defeat asymmetric information: collateral, which gives the lender a stake that disciplines hidden action, and a credit history, which lets the lender screen hidden type. Without them, a conventional bank facing a pooled crowd of applicants must either price for the average risk — driving the safe borrowers out, an Akerlof unravelling in the credit market — or refuse to lend at all. Grameen-style group lending, born in Bangladesh and adapted across Kenya, Uganda, Ethiopia and much of the continent, manufactures substitutes for both out of the one resource the poor do have in abundance: knowledge of, and leverage over, their neighbours. The device is joint liability — a group collectively answerable for each member's loan — and it attacks adverse selection and moral hazard at once.

Peer selection defeats adverse selection

Two borrower types, each takes a loan of 10,000 and repays R = 12,000 if the project
SUCCEEDS, nothing if it FAILS (limited liability — the poor have no collateral to seize).
   Safe borrower:  succeeds with probability  p_S = 0.9
   Risky borrower: succeeds with probability  p_R = 0.5
   (The bank cannot tell them apart; the neighbours can.)

Joint liability: if YOUR project succeeds but your PARTNER's fails, you pay a
surcharge c = 3,000 toward the partner's loan.
   Expected surcharge = c × P(you succeed) × P(partner fails)
                      = c × p_you × (1 − p_partner).

Take a SAFE borrower (p_you = 0.9) choosing a partner:
   Partner SAFE  (p = 0.9):  3,000 × 0.9 × (1 − 0.9) = 3,000 × 0.9 × 0.1 =   270
   Partner RISKY (p = 0.5):  3,000 × 0.9 × (1 − 0.5) = 3,000 × 0.9 × 0.5 = 1,350
   Penalty for accepting a risky partner:  1,350 − 270 = 1,080 per loan cycle.

→ A safe borrower strictly refuses a risky partner. Safe pairs with safe; risky is
   left with risky. Groups form ASSORTATIVELY — and the bank, which never observed
   the types, has induced the village to sort itself.

Read the numbers as a screening result. Under joint liability the cost of your loan includes the surcharge you pay when your partner defaults, and that surcharge is larger the riskier your partner. A safe borrower pays 1,350 beside a risky partner against 270 beside a safe one — a penalty of 1,080 every cycle for keeping bad company. So safe borrowers refuse risky partners and group only with each other; risky borrowers, shunned, pair among themselves and bear the full weight of their own riskiness. Groups form assortatively, and the lender who cannot observe a single type has deputised the village to screen for it. Notice the effect on the unravelling: because risky types now internalise their risk instead of being cross-subsidised by the safe, the safe borrowers no longer flee, expected losses fall, the interest rate falls with them, and the market that would have collapsed under pooling stays open.

Peer monitoring defeats moral hazard

Selection handles the type; monitoring handles the action. Once the loan is disbursed, whether a borrower works the project or diverts the funds is hidden from a distant lender but plain to the neighbour who lives beside her and now stands to pay her surcharge. Joint liability thus enlists the group as monitor and enforcer, backed by social sanctions the bank could never wield — reputation in a tight community is collateral the poor genuinely possess. Repetition sharpens it: loans are small and sequential, and default forfeits the whole group's place on the growing ladder of future credit, so the pull of the relationship disciplines behaviour cycle after cycle. The same logic animates indigenous African institutions — the SACCO, the chama or merry-go-round, the village savings-and-loan association. The mechanism is not costless: peer pressure can turn coercive, members can collude against the lender, and covering a default can fray the very ties it relies on. Tellingly, mature programmes (Grameen's own 'Grameen II') moved toward individual liability while keeping the dynamic incentives and frequent repayment — evidence that monitoring and repetition, as much as joint liability, do the work.

Cheap talk versus costly signals

Stand back and ask why some claims are believed and others ignored. A seller's 'this car is excellent', a worker's 'I am able', a borrower's 'I will repay' are cheap talk — costless statements anyone can make regardless of the truth, so when every type has the same incentive to voice them they carry no information. What separates the good type is never the message but the cost of sending it: a signal informs only if it is costly and differentially costly, cheaper for the honest type than for the impostor — single-crossing again. A warranty is dear only to a seller expecting returns; collateral only to a borrower expecting to default; a degree only to the less able. Each converts an unverifiable claim into a costly act the wrong type will not imitate. Yet cheap talk is not always empty: when interests are aligned it transmits real information — a salaried extension officer with no stake in the sale can credibly advise which seed to plant. The question is always whether the speaker's incentives align with the listener's, and if not, what cost can make the truth the cheaper option.