Should bed nets be free? Cohen & Dupas (2010), Free Distribution or Cost-Sharing

Standard public finance says health goods with positive externalities should be publicly funded, even subsidized beyond 100% when side effects or other private costs are high. That logic is clean for goods that work whatever the recipient does, such as vaccines or deworming pills. It is less clean for goods that only help if people use them, such as bed nets or latrines. For those, a small charge, called cost-sharing, might avoid wasting subsidies on people who would not use the product.

The paper lists three ways a positive price could raise usage among those who acquire a product:

Selection

A price screens out people who do not value the good, so it reaches those likely to use it.

Sunk costs

Having paid, people may feel they should use what they bought.

Signaling

A higher price may be read as a sign of higher quality.

Against these, a price can sharply reduce take-up, and there is evidence that zero is a special price. If poor people who cannot pay are also the sickest, a price screens out exactly those who need the product most. The case for subsidies therefore turns on four things: how demand responds to price, how usage responds to price, how price changes the neediness of whoever buys, and whether the health benefits have externalities or other nonlinearities. The paper estimates the first three and uses them to model the fourth.

Uptake fell by about 60 percentage points between free distribution and 40 Ksh, still a 90% subsidy and 10 Ksh below the price Kenya’s cost-sharing program charged pregnant women at the time. Usage among those who did take a net did not rise with the price. And women who paid were no sicker, on the measure of anemia, than the average prenatal client.

Bottom line. For a product people already value, a small price mostly reduces coverage without improving targeting or usage.

Suppose a net can be put to a health use (hung over a bed) or a nonhealth use (used for fishing, or left in its bag until an old net wears out). Only hung nets create the health externality. A household chooses h nets to hang and n not to hang, and pays the marginal cost C minus a subsidy T for each:

Raising the subsidy buys more hung nets, H, but it also subsidizes every unhung net, N. Setting the marginal cost of the subsidy equal to the marginal benefit of the externality gives

If the ratio is 1, every extra net bought because of the subsidy gets hung, and the optimal subsidy equals the externality, as in Pigou’s classic rule. If people hang fewer of the extra nets, the subsidy should be smaller. It could be larger than the externality if a low price signals to households that nets are worth using. A sunk-cost effect, where people who paid more are more reluctant to leave a net unhung, would push the other way.

Bottom line. What matters is how the number of hung nets responds to price compared with how total nets respond. The experiment is designed to measure both.

Insecticide-treated nets (ITNs) have been shown to cut overall child mortality by at least 20% where malaria is the leading cause of death among young children. They protect pregnant women and their babies from the harm of maternal malaria, and they save on treatment costs and lost learning and income. Yet at $5–$7 a net, most families cannot afford one, and in 2008 most children and pregnant women in sub-Saharan Africa did not sleep under one.

The paper frames the policy debate. Advocates of cost-sharing, such as the social-marketing organization Population Services International (PSI) and William Easterly, argue that a price screens out people who will not use a net, and that positive prices support a commercial market that can sustain supply. Advocates of free distribution, such as Jeffrey Sachs and the World Health Organization, point to the externality. In a village-level trial in western Kenya, nets cut child mortality, anemia and malaria as much in households within 300 meters of beneficiary villages as in the villages themselves. In Ghana, child mortality rose 6.7% with each 100 meters from the nearest household with a net. Strong community effects are thought to need at least 50% coverage, and no cost-sharing program was known to have reached that.

The externality has three sources: insecticide kills mosquitoes that touch the net, fewer people to bite means fewer infective mosquitoes, and fewer infected people means fewer parasites to pass on.

Bottom line. Nets are proven, expensive and underused, and their benefits spill over to neighbors. That makes the price question central.

The study ran in 20 rural public health centers across four districts of western Kenya (Busia, Bungoma, Butere and Mumias), where malaria transmission is year-round with peaks in the rainy seasons. Nearby, pregnant women could receive as many as 230 infective bites over 40 weeks of pregnancy, and up to a third of infants were born premature, small for gestational age or underweight. The 20 clinics were chosen from 70 in the region for being public, their size, services and distance from one another, then randomly assigned:

Clinics earned a monthly bonus worth 5,000 Ksh (about $75) if spot checks found no leakage or mismanagement. Even so, 4 of the 11 clinics charging a positive price sold some program nets to people who were not prenatal clients; none changed the price charged to pregnant women. None of the 5 free clinics mismanaged the nets. The program started between March and May 2007 and ran at least three months, through the peak malaria season. Posters announced the price.

A second, surprise price

In clinics charging a positive price, on unannounced days, women who had put the money on the counter were invited to pick an envelope from a basket. Depending on the envelope, they paid the posted price, nothing, or (at 40 Ksh clinics) a lower price. Everyone offered agreed. Because these women had already shown they were willing to pay, any difference in later usage by the price they actually paid isolates a sunk-cost effect from selection. The lottery began at least five weeks into the program, no more than once a week, on varying days, so that women would not wait for it.

Data

Three sources: clinics’ own sales records; surveys of every pregnant woman attending on three or four random days per clinic, 545 women in all, with a hemoglobin test; and unannounced home visits to a random sample of 246 women who had acquired a net, three to ten weeks later. The visits, all within three weeks in July 2007, found 92% (226 women). Enumerators asked to see the net and checked whether it was out of its packaging and hanging. Neither the women, the clinic staff nor the baseline enumerators knew usage would be checked, and reported usage did not rise over the three weeks of visits.

Because price was randomized across clinics but outcomes are individual, standard errors are clustered by clinic. With only 16 treatment clinics, significance is judged against a t distribution with 14 degrees of freedom: critical values of 2.98, 2.14 and 1.76 at 1%, 5% and 10%. The confidence intervals in this guide use 2.14 accordingly. As a second approach that needs no distributional assumptions, the authors also use randomization inference (Section IV.A).

Bottom line. A clinic-level price experiment with a surprise second-stage discount, individual usage checks, and inference built for a small number of clusters.

Randomization inference asks: if price had no effect at any clinic, how unusual would the observed difference between groups be? Under that null hypothesis every clinic’s outcome is fixed, so one can compute the difference for every possible way the clinics could have been assigned to groups and see where the actual assignment falls. The test assumes nothing about the distribution of errors, but with few clinics there are few possible assignments.

The main result of the paper is in the last panel: effective coverage, the share of prenatal clients who acquired a program net and were found using it, is 54 percentage points lower at 40 Ksh than under free distribution. Uptake at 20 and 40 Ksh is significantly lower than under free distribution, and effective coverage is significantly lower at 40 Ksh. None of the differences in sales is significant; at 40 Ksh, with one clinic’s records missing, no difference could have been.

Bottom line. Even with only a handful of clinics per group and assumption-free tests, charging 40 Ksh clearly cuts both uptake and effective coverage.

Two independent sources give the same picture. Clinic records show free clinics handing out about 41 nets a week; each 10 Ksh on the price cuts weekly sales by about 8 nets, roughly 20%, and 40 Ksh cuts them by 80%, to about 9 a week. The individual survey data agree: each 10 Ksh cuts the probability of buying by about 15 percentage points, a price elasticity of −0.37 at the mean.

Two worries are addressed. Access to free nets from other sources, such as the July 2006 measles campaign that gave nets to mothers of young children, might have dampened demand; but the 63% of women who had not received a free net in the previous year are just as price-sensitive. And women might have switched clinics to get a free net, which would not happen under a nationwide price; restricting to women who stayed at their usual clinic gives nearly the same results.

The paper extrapolates the linear estimate to the prevailing 50 Ksh price: demand would be 75% lower than under free distribution, so of 100 women who take a free net, 25 would buy one. Note that 50 Ksh is outside the range of prices actually tested.

Bottom line. Demand barely moves between free and 10 Ksh but collapses by 40 Ksh, even though every price tested is a subsidy of 90% or more.

An important caveat first: because few women bought at high prices, the usage sample is small, especially at 40 Ksh, and the estimates are imprecise. On average, 62% of women visited at home said they were using their program net, and 57% had it visibly hanging. Of those who said they used it, 95% had it hanging.

There is no pattern. Women paying 10 or 20 Ksh were slightly less likely to use their nets than women who got them free; women paying 40 Ksh were somewhat more likely; none of the differences is significant. The linear estimate implies that 10 Ksh more raises usage by about 4 percentage points, with a 95% interval running from −4 to +12 points. Among women at their first prenatal visit, the group most relevant for a permanent program, usage was highest with free nets. Retention was above 90% at every price, and no second-hand market developed. Women who were not yet using their net most often said they were waiting for the baby to be born or for an old, untreated net to wear out.

Usage among owners is noisy, but effective coverage, the share of all prenatal clients using a program net, can be estimated precisely: 65% under free distribution against 15% at 40 Ksh (Table VI), a difference significant at 1%. Each 10 Ksh lowers it by about 12 points, an elasticity of −0.44. The drop in demand is not offset by any rise in usage. The authors attribute the high valuation even among those who paid nothing to years of advertising, the effort of travelling to the clinic, and counseling during the visit. They also note that the region is very poor, so many women may value nets but cannot pay.

Bottom line. Charging did not buy higher usage. It bought lower coverage.

The second-stage lottery holds willingness to pay fixed: every woman in this sample had already agreed to pay the posted price. Randomly varying what she actually paid therefore tests whether the act of paying, by itself, makes people use a product more.

This fits a similar test for a water purification product in Zambia (Ashraf, Berry and Shapiro), which also found no sunk-cost effect. The authors add that they lack data on time preferences, which might reveal subgroups prone to sunk-cost behavior, and on whether women mentally netted the discount against the payment.

Bottom line. There is no sign that paying makes people use their nets more.

Does a price at least steer nets toward women who need them more? The measure is hemoglobin: anemic women (with low hemoglobin) are likely to be those most exposed to malaria and least resistant to it, exactly the women a cost-sharing program would want to reach. The comparison is between women who took a net at each price and all prenatal clients at the control clinics.

Women paying 10, 20 or 40 Ksh look just like the average prenatal client. Surprisingly, women who got free nets were healthier. Table VIII suggests why: they were 12 percentage points more likely to be on a repeat visit, less likely to have walked, and paid about 3.5 Ksh more for transport. The authors read this as free nets drawing women back for an earlier revisit, before the benefits of their first visit wore off; prenatal care in Kenya includes free iron supplements and presumptive malaria treatment, both of which raise hemoglobin.

Two further points. In absolute numbers, far more anemic women are covered under free distribution: all of them take a net when it is free, but only about 40% at 40 Ksh, so with similar usage, effective coverage of anemic women is about 60% lower under cost-sharing. And the few 40 Ksh buyers were more likely to pay for transport, read Swahili, wear shoes and own animals, suggesting that cost-sharing selects partly on wealth rather than need. The paper’s text puts the share of moderately anemic women at 71%; Table VIII gives 69% at control clinics.

Bottom line. A price does not target the sick. If anything, it targets the better-off.

The authors combine their demand and usage estimates with the medical evidence to estimate child lives saved under each price. Two things are uncertain. First, how much a net protects its own user as a physical barrier. Second, how the protection of nonusers grows with the share of users, which they model as an S-shaped (logistic) function whose threshold may be low, medium or high. That gives nine scenarios. They assume a free net costs the planner exactly 40 Ksh more than one sold at 40 Ksh, ignoring the cost of handling money under cost-sharing, and that 65% of households have a pregnancy within five years and so become eligible.

Free distribution saves the most lives in all nine scenarios. On cost per life saved, it is the cheapest option in five of the nine; 40 Ksh is cheapest in three, when nets protect their users strongly or when the externality needs very high coverage, and 20 Ksh in one. Given the imprecise usage estimates, the differences in cost per life saved are mostly not statistically distinguishable from zero.

Bottom line. Cost-sharing is at best marginally more cost-effective, while free distribution saves many more lives.

William Easterly had written that selling nets for $0.50 through prenatal clinics “gets the nets to those who both value them and need them”. The experiment finds no support for that claim:

Demand

Falls about 60 percentage points from free to 40 Ksh ($0.60), still a 90% subsidy; the paper extrapolates a 75% fall at the prevailing 50 Ksh.

Usage

No higher among payers; no sunk-cost effect in the lottery.

Targeting

Payers are no sicker than the average prenatal client, and somewhat better off.

Effective coverage

65% free against 15% at 40 Ksh: the share of all prenatal clients sleeping under a program net.

Lives saved

Highest under free distribution in every scenario, at similar cost per life.

The authors are careful about scope. Nets were already well known and highly valued in Kenya after years of advertising, so a low price was unlikely to be read as a sign of low quality; the results may not carry over to unfamiliar products. Given how valuable nets are and how little price affected usage, they read the high price sensitivity as a sign that pregnant women were short of cash or credit. The experiment could not measure effects of free distribution on the retail market for nets, or on the behavior of health workers (though, strikingly, the only clinics that leaked nets were those charging a price), or general-equilibrium effects at scale. Their results also differ from the Zambian water-purification study, where paying more did raise reported use; that study did not include a zero price, and a bottle of disinfectant lasts a month rather than three years.

Bottom line. For a proven, well-known product with spillovers, free distribution reaches far more people at little or no extra cost per life saved.

Anemia

A low level of hemoglobin in the blood. In pregnancy it is a common result of repeated malaria infection; here, moderate anemia is below 11.5 g/dL and severe anemia 9 g/dL or below.

Cluster-robust standard errors

Standard errors that allow outcomes to be correlated within a group, here a clinic, because price was assigned to whole clinics.

Cost-sharing

Charging users a subsidized but positive price, so they bear part of the cost.

Effective coverage

The share of all eligible people who both acquired a program net and are using it.

Externality

An effect of one person’s action on others that the person does not take into account; here, one household’s net protects its neighbors.

Insecticide-treated net (ITN)

A bed net treated with insecticide. Long-lasting nets keep their insecticide for three to five years without retreatment.

Kolmogorov–Smirnov test

A test of whether two samples come from the same distribution, based on the largest gap between their cumulative distributions.

Pigouvian subsidy

A subsidy set to match the external benefit of an activity, so that private decisions account for it.

Price elasticity

The percentage change in a quantity, such as demand or usage, for a 1% change in price.

Randomization inference

Testing a treatment effect by comparing the observed difference with the differences under every possible random assignment, assuming no effect at all.

Selection effect

A change in who acquires a product as its price changes.

Social marketing

Selling health products at subsidized prices with brand advertising, as PSI did with nets.

Stochastic dominance

One distribution dominates another if its cumulative distribution lies at or below the other’s everywhere: its values are higher throughout.

Sunk-cost effect

The tendency to use something more because one has paid for it, even though the payment cannot be recovered.

This guide paraphrases the published paper and redraws its results from its tables. Values for Figures I, II and III were read from the vector drawings in the paper’s PDF, so they match the printed figures to within plotting precision. The optimal-subsidy calculator and the envelope game are illustrations, not results from the paper. Later research cited: Dupas (2014), “Short-Run Subsidies and Long-Run Adoption of New Health Products: Evidence from a Field Experiment”, Econometrica 82(1): 197–228; Tarozzi, Mahajan, Blackburn, Kopf, Krishnan and Yoong (2014), “Micro-loans, Insecticide-Treated Bednets, and Malaria: Evidence from a Randomized Controlled Trial in Orissa, India”, American Economic Review 104(7): 1909–41.