In a nutshell
This page discusses the evidence of effectiveness and cost-effectiveness of New Incentives’ program to provide conditional cash transfers to increase infant vaccination in North West Nigeria. (More information on New Incentives and our rationale for making it a top charity can be found on our New Incentives charity page.)
A GiveWell-funded randomized controlled trial, conducted by IDinsight, found strong evidence that New Incentives' program increases vaccinations (see below). Combined with evidence that a) vaccine-preventable diseases are likely a significant cause of child mortality in North West Nigeria (see below), and b) vaccines effectively prevent the diseases they target (see below), we believe New Incentives' program is likely to substantially reduce child mortality in North West Nigeria.
In our cost-effectiveness analysis, we also model benefits of New Incentives' program resulting from: 1) reductions in mortality from vaccine-preventable diseases in individuals older than five, 2) "development benefits" due to improvements in health, and 3) consumption benefits from the cash transfers themselves (see below). We combine the value of these additional benefits with the program's primary effect — reducing child mortality — to estimate the program's impact in terms of "outcome as good as averting the death of a child under 5” (see here).
Our best guess is that New Incentives' cost per outcome as good as averting the death of a child under 5 is $3,148 (see a high-level summary of our calculations here).
Published: November 2020
- Does it work?
- What are the effects on child mortality?
- What are the benefits beyond child mortality?
- Additional benefits and negative or offsetting impacts
- Downside adjustments
- How much does it cost?
- How cost-effective is it?
- Our process
- Next steps
Does it work?
The primary benefit of New Incentives' program is reduction in mortality for vaccinated infants until the age of five. This accounts for about 67% of the benefit of New Incentives' program in our model (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
Other benefits of New Incentives' program that we incorporate into our overall estimate of New Incentives' impact are:
- Reduction in mortality beyond the age of five. Reductions in mortality from vaccine-preventable disease also occur in vaccinated people after the age of five. This accounts for about 18% of the benefit of New Incentives' program in our model (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
- Development benefits. Early-life health improvements from vaccines might lead to increases in income when vaccinated children become adults. This accounts for about 11% of the benefit of New Incentives' program in our model (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
- Consumption benefits. Cash transfers received by caregivers may improve well-being through increased consumption. This accounts for about 3% of the benefit of New Incentives' program in our model (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
We have also attempted to account for:
- Additional benefits (e.g., morbidity effects, benefits of herd immunity) and negative or offsetting impacts (e.g., crowding out of New Incentives by other vaccination promotion programs, inflation). Rather than explicitly modeling these, we have applied percentage adjustments based on our best guesses. (More)
- Downside adjustments. We adjust for risk of wastage, the quality of monitoring and evaluation, and our level of confidence in funds being used for their intended purpose. (More)
Overall, our best guess is that New Incentives' program leads to the equivalent of about 9.2 child deaths averted for each 1,000 infants in areas where New Incentives operates. We estimate the program costs roughly $29,100 per cohort of 1,000 infants, leading to an estimated cost-effectiveness of approximately $3,148 per under-5 death equivalent averted.1
What are the effects on child mortality?
We view reducing child mortality by improving the uptake of vaccines that prevent diseases that cause mortality as the primary benefit of New Incentives' program. Reduction in child mortality accounts for 67% of the benefit of the program in our model (before accounting for additional benefits and offsetting impacts and downside adjustments).
We estimate that New Incentives' program averts 5.5 child deaths per 1,000 infants in areas where New Incentives operates, based on the following factors:2
- The effect of New Incentives' program on vaccination rates. We estimate the program leads to a 22 percentage point increase in the use of (directly and indirectly) incentivized vaccines in three states in North West Nigeria. This estimate is based on a GiveWell-funded randomized controlled trial (RCT) of New Incentives by IDinsight, to which we apply adjustments for partial vaccination and attenuation due to self-reporting bias. (More)
- The prevalence of vaccine-preventable disease and mortality in areas where New Incentives operates. We estimate that the probability of death for an unvaccinated child under 5 from diseases that could be prevented by one of the vaccines incentivized by New Incentives is 3.3%. This estimate is based primarily on data from the Institute for Health Metrics and Evaluation (IHME), to which we apply a number of adjustments. (More)
- The effect of increased vaccination on prevalence of vaccine-preventable disease. We estimate that the vaccines incentivized by New Incentives reduce recipients' risk of vaccine-preventable disease by 76%, based on meta-analyses of vaccines' effects on vaccine-preventable diseases and adjustments for a) concerns about potentially lower vaccine efficacy in low-income countries and b) "non-specific" effects on mortality from diseases other than those directly targeted by vaccines. (More)
We discuss each of these factors in more detail below.
Does New Incentives' program lead to increased vaccination?
IDinsight, funded by a GiveWell Incubation Grant, conducted an RCT to measure the impact of New Incentives' program on vaccination rates in three states in North West Nigeria.3 The RCT found that New Incentives' program had high uptake and led to significant increases in the use of several vaccines incentivized by the program. Households served by clinics where New Incentives was providing CCTs for vaccinations had substantially higher self-reported vaccination, compared to control clinics. We view the RCT as high-quality, and its findings are consistent with similar studies.4
The RCT finds a 14 to 21 percentage point increase in the study's three pre-specified primary outcomes — i.e. self-report of having received 1) the BCG vaccine, 2) the measles vaccine, and 3) any of the three doses of PENTA vaccine. Relative to control households, households in the catchment area for treatment clinics were:
- 16 percentage points more likely to report receiving the BCG vaccination (95% CI 0.12-0.21, 63% control group mean),5
- 21 percentage points more likely to report receiving any of the three doses of PENTA (95% CI 0.16-0.26, 54% control group mean),6 and
- 14 percentage points more likely to report any dose of measles (95% CI 0.10-0.18, control group mean 59%).7
We assume the effect for the rotavirus vaccine, which was indirectly incentivized, is the same as for PENTA, since the two doses of rotavirus are administered at the same time as the first two doses of PENTA. We average the effects on each of these outcomes to generate an overall effect size estimate of 17 percentage points.8
We then apply the following adjustments to this initial estimate:
- A roughly 20% upward adjustment (from 17 percentage points to 20 percentage points) to account for self-report bias (see below).
- A 12% upward adjustment to account for partial vaccination, which we believe leads to an increase in the effective number of infants achieving full vaccination (see below).
- A 4% downward adjustment toward our prior of a 16 percentage point increase in vaccination rates (see below).
With these adjustments incorporated, our overall best guess is that New Incentives' program increased vaccination rates in the treatment group by 22 percentage points.
Overall, we view the evidence for the main effect on vaccination rates as reasonably strong. Factors that give us confidence in this result include:
- GiveWell has followed this RCT closely and believes that it meets high quality standards.
- These findings are consistent with our pre-registered prior for the effect size, based on trials of similar incentives for vaccination programs (see below).
- These findings are from a pre-registered RCT.
- The code was checked by an external auditor.9
- The increase in treatment clinic vaccinations over and above increases in control clinic vaccinations occurs in clinic data as well as self-report data.
- We find the results from New Incentives' baseline and endline surveys consistent with the plausibility of immunization incentives as a driver of observed behavior change (more here).
Our main remaining concerns include:
- We remain uncertain about the appropriate adjustments for partial vaccination (see below) and self-reporting bias (see below), which have a substantial impact on our estimate of New Incentives' impact and cost-effectiveness.
- The self-reported data in the RCT show a large and surprising increase in vaccination rates in control clinics from baseline to endline. At baseline in 2017, vaccination coverage rates in control areas were roughly 15%-25% across vaccines; at endline, they were around 60%. While this increase in counterfactual coverage appears in self-reported data, clinic administrative records in control areas do not show any increase in vaccination volume since baseline. We are very unsure how to explain this discrepancy (though we do expect administrative recordkeeping to be somewhat better in treatment areas than in control areas).
- In treatment areas, the baseline report found much higher vaccination rates than expected (20%-30%) for primary outcome measures. If this is driven by other vaccination programs, then we might have concerns about whether those programs were more frequent in treatment areas than in control areas.10
- GiveWell recommended the funding for the IDinsight RCT, and New Incentives is a GiveWell Incubation Grant recipient. As a result, there may be an incentive for GiveWell to review the results favorably.
The primary outcome New Incentives measures is caregiver self-report of vaccinations received by infants. Caregivers might misreport infant vaccinations, potentially leading to both false negatives and false positives in the data that could cause us to either over- or under-estimate treatment effects (see this page for more information). Self-reporting bias may also affect our estimate of vaccination coverage in control areas, which is a key input for our cost analysis.
Our best guess is that self-report bias causes an overestimate of control group vaccination coverage in the data and, overall, leads to an underestimate of the effect of New Incentives' program on vaccination rates.
Comparing self-report data to BCG scarring data
In addition to measuring self-reported vaccination, IDinsight reports on the effect of the program on:11
- child health cards (CHCs),
- clinic immunization records, and
- BCG scarring.
Each of these measures has limitations. For the purposes of attempting to correct for self-reporting bias, we have chosen to focus on BCG scarring.
The BCG scar rate among the control group is 41%, and the treatment effect on scars is 22 percentage points, suggesting a treatment scar rate of 63%.12 We make adjustments to this result to account for 1) some individuals who are vaccinated not developing scars, and 2) enumerators potentially failing to identify all BCG scars.
We assume, based on scarring rates in the literature, that 90% of infants in the control group who received BCG developed a scar.13 Based on data we received from New Incentives, we assume that 97% of infants in the treatment group who received BCG developed a scar.14 We also assume that a) 90% of infants who receive BCG and develop a scar have their scar detected by IDinsight enumerators, and b) 95% of infants who do not develop a scar are correctly identified by enumerators.15 This implies a control group BCG rate of 48% and a treatment group rate of 69% — i.e. a treatment effect of roughly 20 percentage points. This treatment effect of a 20 percentage point increase in vaccination is roughly in line with the increases in self-reported vaccination for the study's three primary outcomes — i.e., receiving the BCG vaccine, the measles vaccine, and any dose of the PENTA vaccine. For comparison, the treatment effect on BCG vaccination based on self-report data is a 16 percentage point increase over a control vaccination rate of 63% (see calculations here).
The following table summarizes 1) self-report data across the RCT's three primary outcomes, 2) BCG scar data, and 3) our best guess of actual BCG vaccination rates.16
|Self-report, any dose of PENTA||0.54||0.78||0.21|
If our best guess for control group coverage (48%) is accurate, this implies that self-report of BCG is biased upward in both the control group (63% using self-reported BCG vs. 48% using best guess true BCG) and the treatment group (79% using self-reported BCG vs. 69% using best guess true BCG), but the bias in the treatment group is less.17
We expect control group self-report to bias coverage estimates upward because we expect a significant number of caregivers who did not vaccinate their infants to report that they did. In the treatment group, we expect this effect to cause less upward bias because, even if the percentage of caregivers who didn't vaccinate but falsely report vaccinating is the same (or higher) than in the control group, we expect the absolute number of false reports of vaccination to be smaller since a smaller proportion of infants were in fact not vaccinated.
Based on the above considerations, our CEA assumes that a) the true control group vaccine coverage rate is 48% across vaccines (rather than 59% across vaccines), b) a 20 percentage point increase holds across all vaccines (not only BCG), c) the percentage of infants who would have been vaccinated in the absence of New Incentives' program is 48%, and d) that the total percentage of infants vaccinated is 69%.18
Our key remaining uncertainties about accounting for self-reporting bias include:
- For both treatment and control groups, we assume the coverage rates for all vaccines are similar — i.e., differences are due to either statistical error or different levels of measurement error across different vaccines. This seems like a plausible assumption and puts more weight on BCG estimates, for which we have a clearer understanding of the biases present. However, an alternative approach (which we have not used in our estimates) would be to make best guesses about the sensitivity (i.e. the degree to which self-report correctly identifies infants that have been vaccinated) and specificity (i.e. the degree to which self-report correctly identifies infants that have not been vaccinated) of self-report data for each vaccine independently and make adjustments accordingly.19
- It might be worthwhile to consider effects across vaccines in more detail. For example, we believe self-report sensitivity is probably lower for other vaccines than for BCG, which would imply a smaller downward adjustment in control coverage. BCG has the highest control coverage (63%, compared to 59% and 54% for measles and PENTA, respectively), which is consistent with our belief, though we do not know by how much sensitivity varies across vaccines.
- We are unsure about our guesses for the rates of BCG scarring and of enumerator error in identifying BCG scars.
- This 20 percentage point increase is higher than our prior (see below), but previous studies may not explicitly account for self-reporting bias.
The vaccination schedule for infants in the areas of Nigeria where New Incentives operates includes:
- three doses of the PENTA vaccine (which includes the DTP and HiB vaccines),
- three doses of PCV, and
- two doses for the rotavirus vaccine.
One of the primary metrics by which we evaluate the effect of New Incentives' program is whether individuals report received any dose of the PENTA vaccine (see above). We also use this data to infer whether those individuals received any dose of PCV and the rotavirus vaccine, since those vaccines are administered at the same time as PENTA. However, because we base our estimate of New Incentives' mortality effects on meta-analyses that focus on disease reduction resulting from increases in rates of full vaccination, we have attempted to translate this "any dose" metric into terms of an effect on full vaccination. Failing to correctly account for partial vaccination could lead us to over- or under-estimate the effect of New Incentives on full vaccination rates.
See here for our methods and calculations. Our best guess is that accounting for partial vaccination leads to an increase in effect of roughly 18% for PCV, DTP, and HiB20 and a 10% increase in effect for rotavirus,21 for an overall increase in effect on vaccination rates of 12%.22
It's possible that some caregivers in control clinic catchment areas went to clinics in treatment areas to receive incentives from New Incentives. This type of positive spillover effect would cause us to underestimate the effect of New Incentives, since those infants would be counted as vaccinated but are in the control group.
However, because IDinsight did not find any evidence of spillovers in the control group,23 we do not include any adjustment for spillover effects. (We do estimate there were substantial numbers of infants enrolled who were outside treatment clinic catchments but not in control clinic catchments. We discuss this further below.)
Evidence from other studies and adjustment toward skeptical prior
Before seeing the results from the RCT, we did a literature review of other programs providing incentives for vaccinations. Based on that review, we predicted that New Incentives' program would increase vaccination rates by 16 percentage points. (See our literature review and prior for program impact.)
This prior falls within the range of the reported effect sizes on primary outcomes in IDinsight's analysis (14 to 21 percentage point increases across vaccines).24 However, after accounting for self-report bias (see above), we estimate a treatment effect of 20 percentage points for New Incentives' program, which is roughly 25% higher than this prior.
To adjust toward this skeptical prior, we put 80% weight on the IDinsight RCT effect size and 20% weight on our prior. We put a higher weight on the RCT because a) it focuses on the same program in the same setting (whereas our prior was based on merely similar programs in different settings), and b) it has undergone an independent code audit and a larger amount of scrutiny from GiveWell staff. Putting 80% weight on the observed 20 percentage point effect and 20% on the prior of a 16 percentage point effect leads to a weighted effect of 19.6 percentage points, or a downward adjustment factor of 0.96 (see our calculations here).
Comparison to pre-analysis plan
Before receiving results of the RCT, we drafted a pre-analysis plan that described how we would interpret the RCT results in our cost-effectiveness analysis. We split these into internal and external validity considerations.
Internal validity. Regarding internal validity, we planned to focus primarily on bias from self-reports, along with several additional factors.25 A comparison of our pre-analysis plan and what we ended up considering is below:
- Self-report. In our cost-effectiveness analysis, we made adjustments for self-report bias, described here. As discussed in the pre-analysis plan, we conducted cross-checks with other measures and decided to focus primarily on BCG scars, since we have the most confidence in this as an alternative measure. We also made an adjustment to put some weight on our prior for the program (see here), as specified in the pre-analysis plan. We did not explore changes in variables related to social desirability or evidence of difficulty with caregiver recall, which were two additional considerations mentioned in the pre-analysis plan.
- Additional factors. Among the additional factors we planned to consider, we did consider treatment spillovers, alternative measles campaigns, and study quality. Factors specified in the pre-analysis plan that we did not prioritize review of or make adjustments for are: a) the plausibility of immunization incentives as a driver of observed behavior change, and b) analysis of partially treated areas.
External validity. Regarding external validity, we planned to consider several factors. We ended up not making adjustments for several of these factors, because our funding will support New Incentives' expansion in the same states in which it operated during the RCT (which reduces concerns about the RCT's external validity). However, we did adjust for several factors that might change over time within the same states between the RCT and implementation. A comparison of our pre-analysis plan and what we ended up considering is below:
- Baseline coverage. We include an adjustment for the potential "crowding out" of New Incentives due to counterfactual vaccination coverage changing over time (see below).
- Presence of competing vaccination campaigns. We would guess that some of the potential increase in counterfactual coverage and crowding out of New Incentives is due to other vaccination programs. We also account for some chance that another organization would have set up a CCT program similar to New Incentives in the absence of our funding New Incentives (see below).
- Introduction of additional vaccines. Our CEA incorporates benefits from rotavirus vaccine, which we guess will be introduced in 2021 but which was not available during the RCT.26
- Inflation-adjusted transfer size. We include an adjustment for the risk of inflation over time reducing the effectiveness of the current nominal cash transfer amount (see below).
- Other factors. We did not prioritize review of the following factors, since we guess that they are less likely to be important, given that New Incentives is scaling up in the same states in which it operated during the RCT: poverty rates, cultural acceptability, delivery of routine immunization services, geographic access, supply-side issues, program outreach capacity and intensity, catchment security, New Incentives’ program implementation, vaccine effectiveness, disease environment, and disease spillovers.
What is the prevalence of vaccine-preventable disease and mortality in areas where New Incentives operates?
Our best guess is that the probability of death for an unvaccinated child under five from a disease that could be prevented by one of the vaccines incentivized by New Incentives is 3.3%.27 This estimate is primarily based on the Institute for Health Metrics and Evaluation (IHME)'s data on child mortality from vaccine-preventable diseases in Nigeria.28 We make adjustments to account for:
- The proportion of deaths from vaccine-preventable diseases coming from specific pathogens (i.e. disease etiology) (More)
- Deaths occurring before vaccines are administered (More)
- The share of children in the mortality data who received vaccines (More)
- A higher mortality rate in North West Nigeria for unvaccinated children due to worse health overall (More)
Adjustment for disease etiology and pre-vaccination deaths
The diseases targeted by the vaccines that New Incentives directly incentivizes are:
|Lower respiratory infection caused by S. pneumoniae||PCV|
|Meningitis caused by S. pneumoniae||PCV|
|Lower respiratory infection caused by H. influenzae type b||HiB vaccine|
|Meningitis caused by H. influenzae type b||HiB vaccine|
|Diphtheria, tetanus, and pertussis||DTP vaccine|
|Hepatitis B||Hepatitis B vaccine|
New Incentives also indirectly incentivizes vaccines for the following diseases:
|Diarrhea caused by rotavirus||Rotavirus vaccine|
|Polio||Oral polio vaccine|
|Yellow fever||Yellow fever vaccine|
We take our initial estimate of the probability of death for children under 5 from vaccine-preventable diseases (LRTI, pneumococcal meningitis, whooping cough, diphtheria, tetanus, HiB meningitis, measles, and tuberculosis) from IHME data. We adjust this initial estimate as follows:
- We remove deaths that occur before vaccines are administered from the data (for example, we estimate that 21% of under-5 measles deaths occur before the age of 9 months, the age at which the measles vaccine is provided, so we apply an adjustment factor of 0.79).29
- We adjust the probability of death from specific diseases to account for the etiological fraction of different pathogens — i.e., the share of LRTI deaths due to S. pneumoniae and HiB and the share of diarrhea deaths due to rotavirus. We've decided not to use IHME's vaccine etiology data, but are highly uncertain about the appropriate adjustments to use. (See the cell notes here for further discussion.)
- We exclude child mortality from hepatitis B, polio, and yellow fever, since IHME data show a very low probability of death for children from these diseases in Nigeria.30
Summing across diseases, with these adjustments incorporated, we find a probability of death for children under 5 in Nigeria from vaccine-preventable diseases of 1.7%.31
Adjustment for proportion of vaccinated children in IHME data
Baseline mortality rates from IHME necessarily include some vaccinated and some unvaccinated infants. We adjust baseline mortality rates to reflect mortality among only unvaccinated infants as follows:
- We estimate that vaccination coverage in Nigeria is 43% across vaccines over the 5 years leading up to 2017. This is based on coverage of BCG, DTP, measles, HiB, PCV and rotavirus from UNICEF/WHO data.32
- We adjust these coverage estimates to account for partial vaccination (i.e., children receiving the first dose but not all three doses of PCV).33
Applying the formula in this footnote,34 we estimate the probability of death in Nigeria from vaccine-preventable disease for unvaccinated children is 2.5%.35 We have not reviewed the calculations behind UNICEF/WHO data, and it's possible that we would refine these estimates after further investigation.36
Mortality rates in North West Nigeria vs. Nigeria as a whole
We estimate that the probability of death for children under five due to vaccine-preventable disease is 30% higher in North West Nigeria, where New Incentives operates, than in Nigeria as a whole, due to factors such as overall health, prevalence of disease-causing pathogens, and access to healthcare.
To arrive at this estimate, we compared all-cause under-five mortality, diarrhea mortality, and lower respiratory tract infection (LRTI) mortality (accounting for differences in vaccination rates) in the three North West Nigerian states where New Incentives operates to rates in Nigeria as a whole:
- All-cause under-five mortality. We find all-cause under-5 mortality is roughly 50% higher in North West Nigeria than in Nigeria as a whole.37 All-cause under-5 mortality differences include both the effect of higher vaccination rates in Nigeria as a whole than in North West Nigeria, as well as any other effects (e.g., overall health and access to healthcare, which affects susceptibility to death from infection by vaccine-preventable disease). We view this as an upper bound on the difference between the probability of death for children under five due to vaccine-preventable disease in North West Nigeria vs. Nigeria as a whole.
- Diarrhea mortality. We find diarrhea mortality is roughly 40% higher in North West Nigeria than in Nigeria as a whole.38 The rotavirus vaccination (which protects against some forms of diarrhea) is not yet available in North West Nigeria, though we expect it to become available starting next year.
- LRTI mortality. We find that LRTI mortality, accounting for higher vaccination rates in Nigeria as a whole vs. North West Nigeria, is roughly 20% higher in North West Nigeria than in Nigeria as a whole.39 This is particularly relevant because deaths from LRTI prevented by vaccines account for the plurality of the effect of New Incentives' program.
We average the differences in mortality rates for diarrhea and LRTI to generate an estimate of overall difference in probability of death from vaccine-preventable disease in North West Nigeria vs. Nigeria as a whole of 30%, which is consistent with our treatment of the 50% difference in all-cause mortality as an upper bound (see here for calculations).
Incorporating this 30% adjustment, we find that the probability of death under five years of age due to vaccine-preventable diseases among unvaccinated children is 3.2%.40
We have fairly high confidence that vaccine-preventable diseases are a significant cause of mortality in North West Nigeria. However, we have some uncertainties about the exact magnitudes, given uncertainties about some of our assumptions:
- Concerns about cause of death data. Our calculations rely heavily on cause-of-death data, and we're highly uncertain about how accurate these data are at correctly attributing death to a particular disease.41 Our calculations also rely on etiological data to estimate, for example, the share of diarrheal disease deaths caused by rotavirus and the share of deaths from LRTI or meningitis caused by S. pneumoniae and HiB. We have not vetted the methods used to produce cause-of-death or etiological data. In particular, we're highly uncertain how much weight to put on the findings of the PERCH study, which finds much lower etiological fractions of pneumonia due to bacteria like S. pneumoniae and HiB.42
- Subnational adjustments for North West Nigeria. We're uncertain how well cause-of-death data describe the three states in North West Nigeria where New Incentives operates. It's possible the prevalence rates of specific diseases or etiologies are significantly different from national figures. It's also possible the overall child mortality rate is much different than we've estimated, including for diseases like hepatitis B, polio, and yellow fever, which we've estimated have a very low probability of death for children in settings where New Incentives operates.
- Baseline vaccination coverage calculations. We haven't reviewed the calculations underlying the estimates of vaccination rates in Nigeria and have not made adjustments to account for, for example, self-reporting bias in these data. If the coverage figure we use is an overestimate, then the probability of death for unvaccinated individuals will be lower than we've estimated.
Does increased vaccination lead to reductions in disease?
The extent to which New Incentives' program leads to reductions in mortality depends on the extent to which the vaccines incentivized by New Incentives lead to reductions in the diseases those vaccines target. We estimate that the vaccines incentivized (either directly or indirectly) by New Incentives reduce recipients' likelihood of contracting the diseases targeted by 76%.
This estimate is primarily based on meta-analyses of the effects of vaccines incentivized by New Incentives. A weighted average of vaccines' efficacy against targeted diseases and those diseases' contribution to child mortality in Nigeria yields an overall efficacy of 67%. (More)
We apply adjustments to this initial estimate to account for:
- All-cause mortality effects. We apply a 33% upward adjustment to account for observed "all-cause" mortality effects — i.e. it appears that vaccines have mortality benefits beyond their impacts on the targeted diseases. We believe the impacts on all-cause mortality are likely due to a) "non-specific" effects of vaccines — i.e., vaccines may make children healthier and less susceptible to other infections or death due to diseases other than those targeted by the vaccines themselves — and/or b) the proportion of deaths caused by vaccine-preventable diseases being higher than estimated. However, for simplicity, we've modeled these benefits as an increase in the efficacy of vaccines against vaccine-preventable disease. This adjustment corresponds to 0.5 deaths indirectly averted for each death directly averted. (More)
- Concerns about vaccine efficacy in low-income countries. We adjust estimates downward by 19% to account for concerns about vaccine efficacy in low-income countries and Nigeria in particular. (More)
- Imperfect coverage. We include a 95% coverage adjustment to account for less than 100% coverage of all doses of vaccines in trials. This is a rough guess.
In aggregate, these adjustments imply a vaccine efficacy of 76%.43
We have a moderate level of confidence in these estimates. Some of our concerns are:
- We have not thoroughly vetted the meta-analyses we use to quantify vaccine efficacy, and it's possible that further review would lead us to make adjustments to these values.
- We are highly uncertain whether the sizes of our adjustments for concerns about vaccine coverage, vaccine efficacy, and non-specific effects are appropriate.
Meta-analyses of vaccines' effects on disease for children under five
The vaccines we have included in our cost-effectiveness analysis, along with their target diseases, doses, and efficacy against targeted diseases are summarized in the following table.44
|Vaccine||Diseases targeted||Doses||Vaccine efficacy45|
|PCV||Protects against pneumococcal bacteria.46 Our CEA models effects on (i) LRTI associated with S. pneumoniae and (ii) pneumococcal meningitis.||3 doses: at 6 weeks, 10 weeks, and 14 weeks. Co-administered with PENTA.47||0.58 (95% CI 0.29-0.75) for all serotypes-IPD (invasive pneumococcal disease)48|
|DTP||DTP is a combination vaccine used against diphtheria, tetanus, and pertussis (whooping cough). It is also known as DTaP or DTwP.49 Our CEA models effects on diphtheria, tetanus, and pertussis.||3 doses: at 6 weeks, 10 weeks, and 14 weeks. Part of PENTA.50||0.84 (95% CI 0.81-0.87) for acellular vaccines and 0.94 (95% CI 0.88-0.97) for whole-cell vaccines for clinical case definition of pertussis, based on combinations of vaccines containing pertussis vaccines.51 We focus on the effect on pertussis, since it accounts for 85% of probability of death across diphtheria, tetanus, and pertussis.52|
|HiB||Protects against Haemophilus influenzae type b (HiB), a bacteria responsible for severe pneumonia, meningitis, and other invasive diseases almost exclusively in children aged less than 5 years.53 Our CEA models effects on (i) HiB-caused LRTI and (ii) HiB-caused meningitis.||3 doses: at 6 weeks, 10 weeks and 14 weeks. Part of PENTA.54||0.82 (95% CI 0.73-0.87) for invasive HiB disease.55|
|Measles||Measles||1 dose at 9 months.56||0.85 (95% CI 0.83-0.87) for measles.57|
|BCG||Tuberculosis58||1 dose (at birth, or as close as possible)||0.85 (95% CI 0.69-0.92) for meningeal and/or miliary tuberculosis,59 which are the life-threatening forms of tuberculosis.60|
|Rotavirus||Protects against rotavirus, which is one cause of diarrhea.61||Indirectly incentivized. 2 doses (6 weeks and 10 weeks), co-administered with the first 2 doses of PENTA and PCV||0.46 (95% CI 0.29-0.59) for severe rotavirus gastroenteritis in sub-Saharan Africa.62|
For our estimates of vaccine efficacy, we rely on meta-analyses from the Lives Saved Tool (LiST), except for BCG (which is not included in LiST) and HiB. For BCG, we use a meta-analysis by Mangtani et al. 2013. For HiB, we use the meta-analysis by Thumburu et al. 2015, which provides an intention-to-treat (ITT) analysis rather than a per-protocol analysis.63
Overall, vaccine efficacy in these meta-analyses ranges from 46% (for rotavirus) to 85% (for measles and BCG).64 For our cost-effectiveness model, we estimate an overall vaccine efficacy of 64%, produced by taking a weighted average of vaccine efficacies, weighted based on the estimated share of under-five mortality in Nigeria accounted for by the disease targeted by a given vaccine.65
We largely take these vaccine efficacies at face value and have not thoroughly vetted the meta-analyses or individual studies included in them. Our key remaining uncertainties related to these meta-analyses include:
- Internal validity. It is possible we should include downward adjustments to account for internal validity concerns about the effect sizes reported in the meta-analyses. We have not currently done so, because the effect sizes are almost all based on RCTs, and our impression is that the efficacy of vaccines is well-established. The effect size for measles is based on a combination of RCTs and observational studies;66 there is also other evidence consistent with the effect size reported from serological studies of measles.67 Our impression is also that measles has been reduced in many countries that have had good vaccination coverage.68
- External validity. Our impression is that the meta-analyses above include a combination of results from low- and non-low-income countries, and there may be reasons to expect smaller effects in low-income countries. It might be appropriate for us to weight studies in low-income countries more heavily, since those settings likely resemble the setting of New Incentives' program more closely. Additionally, many of the studies referenced above were conducted decades ago, and it's possible that the results' validity may not extend to current settings. We discuss how we account for external validity below.
Our other remaining questions include:
- How closely do recipients' age, dosage, and other features of the interventions included in the meta-analyses match New Incentives?
- Do the effect sizes incorporate practical challenges in implementing vaccine programs in the field (e.g., maintaining the cold chain)?
- How would incorporating any information on coverage of vaccines in ITT estimates influence vaccine efficacy? We include a 95% coverage adjustment in the CEA, but this is a rough guess.69
- How does efficacy vary with vaccine "type"? Our understanding is that vaccines can vary along various dimensions (e.g., "valence" for PCV, "acellular" or "whole-cell" for pertussis) and that efficacy may vary with these characteristics.
- Are there interactions between vaccines that would shift our cost-effectiveness?
- How important is serotype replacement?70
It's possible that further investigation of these questions would lead us to adjust our best guess on vaccine efficacy.
Vaccine efficacy in Nigeria
The meta-analyses above include studies from both low-income and non-low-income countries, and we're uncertain about the extent to which effects vary across settings. It's possible that vaccine efficacy may be lower in low-income countries. This may be due to issues with maintaining the cold chain, which might cause vaccines to be inactive by the time they're received by children, or other problems preventing immunity from occurring.
To estimate the size of this effect, we've mainly considered:
- Results from biomarkers studies of measles immunity in Nigeria (More)
- Seroconversion studies of measles immunity in Nigeria (More)
- Meta-analyses of vaccine effectiveness in low-income settings specifically (More)
Overall, these factors lead us to make an adjustment of 0.81 (i.e., a reduction of 19%) to the vaccine efficacies from the meta-analyses described above. This adjustment reduces our estimate of vaccine efficacy across vaccines from 64%, based on meta-analyses, to 52%.71 For measles alone, this implies an efficacy of 69% (down from 85% reported in the meta-analysis we use), which is roughly in line with seroconversion studies and a meta-analysis (Uzicanin and Zimmerman 2011) based on studies specifically from sub-Saharan Africa. However, we are highly uncertain about this adjustment. We view the biomarkers results in particular as somewhat concerning, and we are continuing to investigate this issue. We may decide to conduct follow-up research to understand biomarkers for infants vaccinated through New Incentives. We also have not deeply reviewed seroconversion studies or meta-analyses that separate effects by low-income vs. non-low-income countries, and we have not fully explored all potential explanations for low vaccine efficacy.
Biomarkers studies in Nigeria
IDinsight conducted two operational pilots to test the feasibility of using oral fluid biomarkers to validate caregiver-reported vaccination status. Both of these pilots found low agreement between the oral fluid test and caregiver-reported vaccination status as well as administrative records (e.g. child health cards).
We view these biomarkers results as a negative update on measles vaccine efficacy in our cost-effectiveness analysis. We are highly uncertain how much to adjust based on these results, and our best guess is that the appropriate discount factor is between 0.3 and 1 (i.e., a 0% to 70% reduction in the overall effect size of New Incentives' program). This range corresponds to the following scenarios:
- 0.3 (i.e., 70% reduction in efficacy): This adjustment assumes that the biomarkers result is explained entirely by low vaccine efficacy, and that we should expect the drivers of low vaccine efficacy to be present in New Incentives' implementation for all vaccines.
- 1 (i.e., 0% reduction in effect): This adjustment assumes that the biomarkers result is explained solely by testing inaccuracy or that we expect any drivers of low vaccine efficacy to be resolved in New Incentives' implementation for all vaccines.
We're not sure what's driving the poor biomarkers results. Experts that GiveWell and IDinsight spoke to point to several possible explanations.72 One commonly-suggested explanation was that low vaccine efficacy can occur if the cold chain is not maintained. We are unsure whether this applies to just measles or all vaccines.
To produce an adjustment to account for these biomarkers results, we "pool" the following factors:
- Our best guesses as to what is driving biomarkers results for the measles vaccine and how this should affect our expectations across vaccines — in particular, whether the results are driven by either (i) low vaccine efficacy (and we should expect the drivers of low vaccine efficacy to be present in New Incentives' implementation for all vaccines), or (ii) testing inaccuracy (or that we should expect whatever is driving low vaccine efficacy to be resolved in New Incentives' implementation for all vaccines)73
- External studies of measles vaccine efficacy (including measles meta-analyses and biomarkers studies using blood tests)74
- Our guess about the extent to which poor measles vaccine efficacy would imply problems with other vaccines.75
Our calculations are described here. These guesses and probability weights are very rough and should be viewed as a first pass.
Some older studies from Nigeria (though not North West Nigeria) find somewhat low levels of antibodies following measles vaccination and suggest a vaccine efficacy of ~70%:
- In a sample of 286 children in Ilorin, Nigeria, Fowotade et al. 2015 find 68.6% of infants developed protective antibody titres. The authors attribute this to low vaccine potency, and find that three out of six vaccine vials tested had virus titers above WHO-recommended cutoffs. They attribute this to lapses in the cold chain, though they do not provide any formal tests or analysis to support this. We find this somewhat surprising, since the authors also note that the vaccines were maintained in cold chains after they received them from the State EPI Unit.
- We superficially reviewed Adu et al. 1992 and Omilabu et al. 1999. Both of these studies also find lower than expected antibody levels, attributed to low vaccine efficacy, and speculate that lapses in the cold chain are the cause (though they do not formally test this explanation).
Meta-analyses of vaccines' effectiveness in low-income countries
Our impression, based on a shallow review of meta-analyses for measles, PCV, and BCG, is that vaccine efficacy may be lower in low-income countries:
- A different meta-analysis of measles vaccines than that used by LiST (which we rely on for our main analysis of vaccine efficacy) suggests lower efficacy against measles in low-income settings.76
- A meta-analysis of the effects of BCG notes that BCG is less effective closer to the equator.77 It is not clear to us whether this result applies to BCG administered at birth and to BCG's effects on fatal forms of tuberculosis.78
- The two African trials in a meta-analysis of the effects of PCV have smaller effects than trials in other settings, though the difference does not appear to be statistically significant, and it is not clear what is causing this difference.79
All-cause mortality effects
As described above, we model the effect of vaccines by multiplying IHME estimates of the probability of death from vaccine-preventable diseases for unvaccinated children by the reduction in that disease from meta-analyses of vaccine efficacy, which suggests an all-cause mortality relative risk across all vaccines incentivized by New Incentives of roughly 0.88. However, two meta-analyses we looked at — Higgins et al. 2016 and Lucero et al. 2009 — find all-cause mortality effects from vaccines that are substantially larger than this estimate.80
In our CEA, we estimate that measles, BCG, and PCV each account for less than 10% of child deaths. However, measles, BCG, and PCV would need to account for 30%, 42%, and 19% of child deaths, respectively, in order for vaccinations against those diseases to lead to the all-cause mortality effects observed in the above meta-analyses.81 Combining the all-cause mortality effects of the measles vaccine, BCG vaccine, and PCV suggests an overall all-cause mortality risk reduction of 0.4682 across those diseases; this would be even lower if we incorporated other vaccines. A rough guess is that, overall, these meta-analyses suggest an all-cause mortality relative risk from vaccination of roughly 0.4.
Overall, this leads us to adjust our estimate of the effect size of vaccines upward. Our best guess is that vaccination prevents 0.5 deaths indirectly (i.e., via indirect effects on non-directly vaccine-preventable diseases) for every death prevented directly (i.e., for directly vaccine-preventable diseases). We model this as a 35% increase in vaccine efficacy. However, we have very limited data with which to estimate the magnitude of this effect, and our current approach is a rough first attempt. We believe it is highly likely that we would update this parameter based on further investigation.
Explanations for this observation that would lead us to put more weight on our estimates based on current disease prevalence and less weight on the all-cause mortality estimates include:
- All-cause mortality effect point estimates may be incorrect. The 95% confidence intervals in Higgins et al. 2016 and Lucero et al. 2009 include a risk reduction on all-cause mortality of 0%. This would mean we should put more weight on our estimates based on current disease prevalence.
- The proportion of child deaths caused by vaccine-preventable disease may have changed over time. When these studies were conducted, measles, tuberculosis, and pneumonia due to S. pneumoniae may have been responsible for a larger share of child deaths than they currently are. This would mean we should give more weight to our estimates based on current disease prevalence.
- Overall health may have improved over time. Access to healthcare, nutritional status, and access to prevention may be greater now than when the all-cause mortality trials were conducted, leading to smaller non-specific effects.
- Simply adding all-cause mortality effects likely "double counts" non-specific effects. For instance, presumably, some of the non-specific effects of BCG on all-cause mortality are due to BCG reducing risk of death from e.g. measles and pneumonia, so adding the effects of measles vaccine and PCV to that impact would count those beneficial effects twice.
- Characteristics of infants in trials may not be representative. Two of the BCG trials were limited to a sub-sample of low-birthweight infants, for whom we would guess this effect is larger than for the average infant in North West Nigeria today.
- There is high uncertainty about the mechanisms for non-specific effects. Our impression is that the mechanisms driving non-specific effects are poorly understood, which leads us to discount them slightly.
- There is evidence of a negative effect of DTP on mortality. The negative effect of DTP on mortality suggests that we ought to downweight non-specific effects, though this is based on observational studies with high risk of bias.83
Explanations that would lead us to put less weight on our estimates based on current disease prevalence and more weight on all-cause mortality estimates:
- IHME disease prevalence and etiology data may be incorrect. It's possible that the IHME data attribute an incorrect fraction of deaths to vaccine-preventable diseases.
- "Non-specific" effects of vaccines. Vaccines may have effects on diseases other than the ones they target by generally strengthening infants' health — i.e. by preventing infants from contracting an illness that would have increased their likelihood of dying from some other cause. Being vaccinated for, e.g., measles may lower the probability of death from malaria or other infections.
Our best guess is that vaccines cause an overall all-cause mortality risk reduction of 17% (i.e., between the 60% implied by taking all-cause mortality risk reduction point estimates at face value and the 12% implied by just looking at effect of vaccines on disease multiplied by diseases' contributions to death). This implies vaccines avert 0.5 deaths indirectly (i.e. deaths due to non directly vaccine-preventable diseases) for each vaccine-preventable disease death averted, which is in line with our CEA for seasonal malaria chemoprevention.84
There are two main lines of reasoning behind this estimate.
The first is based on benchmarking against SMC:
- Our SMC CEA also makes an adjustment for non-specific effects of preventing malaria.85
- We would guess that non-specific effects would be similar for vaccines because the age profile of the beneficiaries is the same (children under 5) and the mechanism seems similar (i.e. beneficial health effects reduce the likelihood of death from other causes).
- We'd guess these effects might be somewhat lower for vaccines, since all of these vaccines together are preventing a slightly larger share of overall deaths than SMC (and so there are fewer deaths "available" to prevent indirectly). However, vaccines also prevent a wider range of disease. As a result, an adjustment of 0.5 indirect deaths averted per vaccine-preventable disease death averted seems appropriate to us.
- This implies an all-cause mortality relative risk of 0.83.
The second is based on putting some weight on the 0.40 relative risk implied by all-cause mortality trials and some on the 0.88 relative risk implied by our current approach:
- We put some weight on the 0.40 relative risk, since a) we think it's plausible there are non-specific vaccine effects, and b) we have high uncertainty about the IHME data on cause of death.
- However, we put more weight on the 0.88 estimate, given the wide confidence interval on the all-cause mortality effect and concerns that a) the proportion of child deaths caused by vaccine-preventable disease may have changed over time, b) overall health may have improved over time since the all-cause mortality studies were conducted, and c) simply adding all-cause mortality effects would likely "double count" non-specific effects.
- We haven't assigned specific weights to these estimates, but view this line of reasoning as an indication that our 17% adjustment is reasonable, since, in light of these considerations, we think it's reasonable to expect all-cause mortality effects to be higher than what we estimate based solely on disease prevalence and vaccine efficacy against those diseases (12%) but likely significantly less than implied by taking all-cause mortality point estimates at face value (60%).
We have a moderate level of confidence in the above estimates. Our key areas of uncertainty are:
- How much should we discount for uncertainties about vaccine efficacy in Nigeria and low-income countries more broadly (for instance, the findings of the biomarkers study discussed above)? We currently apply a downward adjustment of approximately 20% for these uncertainties.
- How accurate are the adjustments made in our analysis for self-reporting bias, partial vaccination, and all-cause mortality effects? The values we have used for these parameters are speculative and could meaningfully affect our cost-effectiveness estimate if altered.
- How reliable are cause of death data, especially for deaths due to specific pathogens (e.g., pneumonia caused by S. pneumoniae)?
- We've chosen to use a relatively simple cost-effectiveness model of the benefits of increased vaccination compared to some other models used in the literature. It is possible our model leaves out key features or additional positive or negative effects that would substantially affect our cost-effectiveness estimate (e.g., disease transmission dynamics, herd immunity)? We incorporate some such factors in the "Inclusion/Exclusion" section of our cost-effectiveness analysis, but only as very rough guesses.
- How should we interpret the unexpected increase in control vaccination coverage observed from baseline to endline (see above) in the RCT? If this increase is driven by other vaccination promotion programs, those programs might "crowd out" New Incentives in the future.
- How does New Incentives' program affect caregivers outside of the catchment of treatment clinics?
- How would New Incentives' impact change if it were to scale up?
What are the benefits beyond child mortality?
In addition to effects on child mortality, we also model benefits of New Incentives' program occurring through:
- Reductions in mortality from vaccine-preventable disease in vaccinated people after the age of five. Our model estimates that this accounts for 18% of the benefit of New Incentives' program (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
- "Development benefits," i.e., increases in income later in life due to early-life health improvements. Our model estimates that this accounts for 11% of the benefit of New Incentives' program (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
- Consumption benefits from the cash transfers themselves. Our model estimates that this accounts for about 3% of the benefit of New Incentives' program (before accounting for additional benefits and offsetting impacts and downside adjustments). (More)
Effects on deaths above age five
Risk of vaccine-preventable diseases also occurs at ages beyond five. Vaccines may confer immunity beyond the age of five, reducing risk of mortality as vaccinated infants age.
We think it is reasonable to expect that vaccines would lead to reductions in mortality from vaccine-preventable disease at ages above five, but because most of the burden of vaccine-preventable disease above the age of five occurs at ages 50 or older, we have higher uncertainty about the size of this effect, given a) the lack of long-term data on vaccine effectiveness and b) uncertainty about what the prevalence of vaccine-preventable diseases will be when infants who are vaccinated now reach age 50. Consequently, we apply a large discount to these effects.
The probability of death from vaccine-preventable diseases (unadjusted for vaccination) in Nigeria is 0.13% from age 5-14, 0.98% from age 15-49 and 4.7% from age 50-74, compared to 1.7% for ages 0-5.86 This would suggest that roughly 80% of the benefits of vaccination accrue to individuals older than 5, with most of those (roughly 60%) occurring after the age of 50.87
We discount impacts on mortality occurring at older ages for the following reasons:
- We expect health to improve over time, suggesting mortality rates from these diseases will likely be lower when vaccinated infants reach these age groups.
- Vaccination rates in infancy were likely lower for individuals who are now in older age groups, so the probability of death for unvaccinated individuals is probably a lower share than what would be implied by directly comparing the figures above.
- Vaccine efficacy may diminish over time for some vaccines. The strongest evidence of this effect we found was for BCG, which is also the largest driver of mortality in older age groups (15-49 and 50-74).
- We discount to account for the possibility that fundamental changes occur that render the program ineffective in the future.
We estimate that 23% of "discounted deaths" (i.e., deaths averted in a given cohort once the above discounting factors are incorporated) occur in individuals older than age 5 (see calculations here).
Overall, we estimate 1.5 child death equivalents per 1,000 infants are averted for vaccinated people after the age of five,88 accounting for 18% of the overall benefit of the program. However, this estimate involves speculative assumptions and we have a moderate level of uncertainty in our guess.
We would guess that vaccination in infancy may have "development benefits" — i.e., early-life health improvements from vaccines might lead to improvements in income when vaccinated children become adults. While we have not reviewed direct evidence for development benefits from vaccination in infancy, we have included these benefits to be consistent with our cost-effectiveness evaluations of other interventions that improve health early in life. We estimate the equivalent of 0.9 deaths per 1,000 infants are averted due to development benefits.89 This accounts for about 11% of the program's overall benefit (see calculations here). This estimate involves some speculative assumptions and we have a moderate level of uncertainty in our guess.
Cash transfers received by caregivers who vaccinate their infants may also improve well-being through increased consumption. We have not observed direct consumption effects of cash transfers; instead, we model these based on the size of the cash incentives provided by New Incentives and estimates of average household income among recipients. We haven't deeply vetted these estimates, since we think they're unlikely to substantially affect our cost-effectiveness estimates. We estimate the equivalent of 0.2 deaths per 1,000 are averted due to consumption benefits.90 This accounts for about 3% of the program's overall benefit (see calculations here).
Additional benefits and negative or offsetting impacts
There are several benefits, as well as negative or offsetting impacts, of New Incentives' program that we have not modeled explicitly in the main CEA but that we do incorporate into our final cost-effectiveness estimate as rough guesses. Rather than explicitly modeling these, we have applied percentage adjustments based on our best guesses. On net, these effects would increase our estimate of the cost-effectiveness of New Incentives by 19%.91
There are several potential benefits of vaccines that we've excluded from our main cost-effectiveness analysis. These are included in the Inclusion/Exclusion section of the cost-effectiveness analysis:
- Lower likelihood of recipients infecting others. People who are vaccinated may be less likely to spread infection to others, creating a positive spillover effect. The size of this effect will depend on 1) how much vaccination reduces the recipient's likelihood of acquiring the disease and spreading it (rather than merely reducing their likelihood of dying from the disease themselves), and 2) the proportion of the population that is unvaccinated (the higher this proportion, the greater the benefit would be). This effect is included in more complicated models of vaccine efficacy, and it is possible that ignoring this effect leads us to underestimates of vaccine efficacy, especially for diseases with high transmissibility (e.g. measles). We have not yet sought to model these effects, since doing so would add a substantial degree of complexity to our model, but we believe this factor may be worth further investigation, since it is a key component of vaccines' impact.92
- Herd immunity. High overall vaccination levels in treatment groups (close to 80%) suggest that it is possible that New Incentives' program could lead at least some areas to achieve herd immunity as the program rolls out over a 3-year period or in specific areas where total vaccination rates have moved above 85%-90%. We have not yet sought to model these effects, since doing so would add a substantial degree of complexity to our model, but we believe this factor may be worth further investigation, since it is a key component of vaccines' impact.
- Benefits of additional indirectly incentivized vaccines beyond rotavirus. We've excluded any effects of vaccination on polio, yellow fever, mumps, rubella, varicella, hepatitis B, and meningitis A. It is possible that incorporating these could lead to additional benefits, though our impression is that these benefits are likely to be small.93
- Morbidity effects. We have not explicitly modeled impacts on morbidity from vaccine-preventable diseases.94
- Effect during outbreaks. Vaccines could reduce the risk of outbreaks for, e.g., measles, which could increase cost-effectiveness.
- COVID-19 may make New Incentives more effective. It's possible COVID-19 could increase the benefits of cash transfer programs.95
- Treatment costs/economic losses averted from prevention. By preventing diseases, vaccines may also prevent caregivers from incurring costs of procuring treatment (e.g. lost wages, transportation costs, out-of-pocket medical expenditures) for these diseases.
- Improved timeliness of vaccination. New Incentives may also make it more likely that vaccines are provided at the correct ages, which can boost their efficacy.96
- Increased clinic utilization. New Incentives may encourage more clinic visits, which could have spillover effects (e.g., improved uptake of other interventions). IDinsight did not find large effects on clinic utilization, so we have set this effect to zero.97
- Investment of income increases. We have included this factor to be consistent with our CEAs for other interventions that increase consumption.
Potential negative or offsetting impacts
There are several potential negative or offsetting impacts of vaccines that we've excluded from our main cost-effectiveness analysis. These are included in the Inclusion/Exclusion section of the cost-effectiveness analysis:
- Crowding out of New Incentives. High levels of vaccination at the endline of the IDinsight RCT raise questions about whether New Incentives' program could be crowded out by other vaccine promotion efforts that push vaccination rates close to 100%. This possibility is partially accounted for by our leveraging and funging adjustment (see below), but if there are other programs that would be likely to cause vaccination rates to increase over the next year, this could diminish the impact of New Incentives.
- COVID-19 may reduce New Incentives' effectiveness. It seems possible to us that cash incentives might be less effective during the COVID-19 pandemic, if (for example) healthcare staff capacity is diverted from vaccinations, caregivers are less likely to respond to cash incentives due to concerns about infection, or vaccine supply chains are disrupted. It's also possible that the risk (to caregivers or infants) of COVID-19 exposure from traveling to clinics to receive vaccinations could offset the mortality benefits of vaccination to some extent.
- Vaccine-derived polio outbreaks. The oral polio vaccine may cause outbreaks of polio and lead to paralysis.98
- Inflation. If inflation occurs, New Incentives may need to raise the amount of its cash incentives to maintain the same effect size.
- Serotype replacement. Vaccination may lead to higher prevalence of non-vaccinated serotypes of the diseases targeted, which could offset some of the beneficial effects of increased vaccination.
Additionally, a meta-analysis by Higgins et al. 2016 finds the DTP vaccine increases all-cause mortality by 38%, though this is from observational studies that the authors rate as having a high risk of bias.99 We do not include this factor in the "Inclusion/Exclusion” sheet, since we incorporate it into our consideration of all-cause mortality effects (see above).
We also adjust for risk of wastage, the quality of monitoring and evaluation, and our level of confidence in funds being used for their intended purpose. On net, we guess these reduce the cost-effectiveness of New Incentives by 5%.100
How much does it cost?
The cost of New Incentives' program includes a) costs paid by New Incentives to implement the program (including staff, cash for incentives, etc.), as well as b) the cost of the vaccines, which are paid for by the Nigerian government and Gavi.101 Our cost analysis also attempts to take into account the counterfactual use of the additional funding for vaccines that New Incentives' program causes the government and Gavi to spend — i.e. what those funds would have gone towards in the absence of New Incentives' program.
We estimate that New Incentives' program costs $29,100 per 1,000 infants in areas where New Incentives operates.102 In brief:
- We estimate a cost to New Incentives of about $25,300 per 1,000 infants. (More)
- We estimate a cost of about $5,800 per 1,000 children to the Nigerian government and Gavi. (More)
- Finally, because we guess that in the absence of New Incentives, funding from government and Gavi would be spent on less cost-effective programs than New Incentives, we incorporate a leveraging and funging adjustment of 1.07, which suggests overall costs of roughly $29,100. (More)
Costs to New Incentives
We estimate the overall costs to New Incentives as follows:
- Cost per child enrolled. We estimate a cost to New Incentives of about $38 per child enrolled, based on dividing New Incentives' operating costs by the number of BCG disbursements made (the BCG vaccine is the first in the series of vaccines incentivized by New Incentives and a key requirement for receiving the remaining disbursements).103
- Proportion of in-catchment infants vaccinated. We estimate that in catchment areas for treatment clinics, 48% of infants would have been vaccinated regardless of New Incentives' program ("always-takers"), based on our adjusted estimate of BCG vaccination rates (described above). Adding the estimated 22 percentage point treatment effect of New Incentives' program (described above) yields an estimate of 70% vaccination coverage in areas where New Incentives operates.104
- Proportion of vaccinated infants enrolled in New Incentives. We estimate that 95% of vaccinated infants were enrolled in New Incentives' program. Infants who were vaccinated as part of periodic campaigns, rather than routine immunization activities, might not be enrolled in New Incentives, but we guess that this proportion of infants is low, based on IDinsight's analysis of the percentage of households that report receiving vaccinations through routine immunizations (which would have been part of New Incentives' program).105
Together, these estimates imply a cost to New Incentives of roughly $25,300 per 1,000 infants in areas where it operates.106
Our key remaining uncertainties regarding New Incentives' costs are:
- We have assumed that the ratio of caregivers induced to vaccinate by New Incentives' program vs. caregivers who would have vaccinated regardless of New Incentives ("always-takers") is the same among caregivers who are within-catchment of treatment clinics and caregivers who are out-of-catchment (i.e., not within catchment of treatment clinics or control clinics). However, we have not seen evidence specifically supporting this assumption and remain highly uncertain about it.
- We have chosen to base our cost analysis on New Incentives' costs during the last year of the period evaluated by the RCT, since we believe New Incentives' future activities will most closely resemble its activities during that time period. However, it is possible that using a longer time period for our analysis, which might help account for longer-term fluctuations, may be more appropriate. Using a longer timeline would decrease our cost estimate by ~15%.
- We've identified a discrepancy between the number of enrolled infants implied by the RCT and data from New Incentives. Our best guess is that most of this discrepancy is driven by out-of-catchment infants, which we do not expect to affect our cost-effectiveness estimate. We do not make adjustments for this in estimating the size of the treatment effect, since these caregivers are not included in the estimating sample. We also assume that the ratio of "always-takers" to "induced to vaccinate" is the same for out-of-catchment caregivers, which means that the costs of the program are not affected by this type of spillover.107 Additionally, while we believe most of this discrepancy is driven by out-of-catchment infants, we believe part of the discrepancy is likely explained by fraud, in particular repeated disbursements (i.e. caregivers enrolling the same infant more than once). At the moment, our best guess is that roughly 8% of BCG disbursements were repeated disbursements, though we are continuing to investigate this issue.108
- At scale, New Incentives' clinics are likely to be closer to one another than during the RCT. Will this decrease transport costs? If so, by how much? Will it decrease the number of infants per immunization day? If so, will this increase costs, and by how much? We provide rough best guesses, but we are unsure about our estimates.
- Will New Incentives staff work at fuller capacity at scale? If so, will this decrease program costs?
Costs to other actors
By increasing the number of vaccinations administered in areas where it operates, New Incentives' program causes an increase in costs borne by the Nigerian government and Gavi, which provide support for vaccines. We estimate that New Incentives' program leads to $5,800 in additional costs to these other actors per 1,000 infants in a given cohort:
- New Incentives leads to roughly 219 additional vaccinations per 1,000-infant cohort.109 (New Incentives only causes additional costs to other actors for infants who would not have been vaccinated in the absence of New Incentives' program.)
- We estimate the costs to other actors are about $26.50 per vaccination (about $20 in government costs and $6.50 in Gavi costs).110
- In total, this leads us to expect about $5,800 in additional costs to other actors.111
Leveraging and funging
We would guess that in the absence of New Incentives' program, funding from the government and Gavi would be spent on less cost-effective programs than New Incentives. We incorporate this factor as a leveraging and funging adjustment in our cost-effectiveness analysis. This adjustment is 1.07 and suggests overall costs of $29,100 per 1,000 infants served.112
Our key remaining uncertainties about this approach are:
- What are the appropriate values for counterfactual uses of funding by the Nigerian government and GAVI? We've chosen our current estimates in part to maintain consistency with our other CEAs.
- What is the likelihood that the Nigerian government would have funded a different program like New Incentives' in the absence of New Incentives itself? We currently assume a probability of 10%, but this is speculative and uncertain. If this probability were higher, it would decrease our estimate of New Incentives' cost-effectiveness.
How cost-effective is it?
We estimate that New Incentives' program results in an outcome as good as averting the death of a child under 5 for each $3,148 spent. This is based on our estimates of roughly 9.2 child death equivalents averted per 1,000 infants immunized and a cost of $29,100 per 1,000 infants immunized.
Our cost-effectiveness model shows our calculations in more detail.
Note that there are limitations to this kind of cost-effectiveness analysis, and we believe that cost-effectiveness estimates such as these should not be taken literally, due to the significant uncertainty around them. We provide these estimates (a) for comparative purposes and (b) because working on them helps us ensure that we are thinking through as many of the relevant issues as possible.
Since our cost-effectiveness calculations form an important part of our decision making process when considering whether to recommend an organization as a top charity, several research staff members reviewed and provided feedback on a draft of the model. Their written reviews are much more informal than the content we typically publish, but we are sharing them below to give a window into an important part of our research process.
See the following links for informal reviews of our cost-effectiveness analysis by:
This work represents a more in-depth review process than what we have generally used in the past. We intend to use a similar approach going forward, spending proportionally more research time vetting conclusions that a) are more likely to impact our funding recommendations and b) are more likely to change as a result of being reviewed by additional researchers.
Research ethics inquiry into IDinsight paper on operational pilot study to determine optimal incentive sizes for measles vaccinations
On June 29, 2020, IDinsight informed us of concerns regarding its paper on a pilot study it conducted of New Incentives' immunization program. The pilot study, also supported by GiveWell Incubation Grant funding, is separate from the main RCT discussed here. The inquiry into the paper is limited to IDinsight's practices and to the pilot study. We take these concerns about IDinsight's research ethics seriously, but we do not believe they should impact our assessment of New Incentives as an organization. Additional details are in this footnote.113
We may update our assessment over the next two to three years, based on the following information:
- Changes in baseline vaccination rates in Nigeria. In particular, this could update our "crowd out" adjustment in the Inclusion/Exclusion section of our CEA.114
- Follow-up research on measles biomarkers. This could update our "adjustment for vaccine efficacy in Nigeria."115
- Changes in the program's effectiveness with scale-up.
See this spreadsheet, "CEA - Summary" sheet for a high-level summary of our calculations.
1,000 (people in cohort) x 0.033 (probability of death without vaccination) x 0.76 (reduction in risk of vaccine-preventable disease) x 0.22 (percentage point increase in vaccine uptake) ≈ 5.5 deaths averted. See the list below.
See IDinsight, Impact Evaluation of New Incentives, Final Report for the results of this RCT.
IDinsight, Impact Evaluation of New Incentives, Final Report, Table 3, pg. 23
- IDinsight, Impact Evaluation of New Incentives, Final Report, Table 3, pg. 23.
- This is similar to effects found for any dose of PCV (22 percentage point increase, 95% CI 0.18-0.27, 50% control mean), which is administered at the same time as PENTA. See IDinsight, Impact Evaluation of New Incentives, Final Report, Table 6, pg. 27.
IDinsight, Impact Evaluation of New Incentives, Final Report, Table 3, pg. 23.
An alternative method would be to treat the treatment effects for each vaccine as separate estimates. One rationale for this would be that there might be reasons to expect effects of New Incentives to differ across vaccines. We decided to use an average for a few reasons:
- It seems reasonable to view the observed effect sizes for different vaccines as "noisy" estimates of the same underlying effect.
- Applying the same effect across vaccines simplifies our model, and we would guess that separating the effect by vaccine would lead to similar results.
- In adjusting for self-report bias, we have used the observed effects on BCG scarring (see below). Translating this adjustment to other vaccines introduces additional complications.
From the independent review by the International Initiative for Impact Evaluation (3ie): "Overall, I found that IDinsight was receptive to suggestions, made major adjustments between baseline and endline that substantially improved the design, was diligent in documenting the Stata code, and their work and assessment are a valid assessment of the New Incentives Conditional Cash Transfer program. While there are remaining questions about the change in coverage (greatly improved) in control areas, and the poor accuracy of the biomarkers and its implications for the efficacy of the measles vaccine in Nigeria, we do not believe that these questions change the interpretation and conclusions of the study. The New Incentives program substantially and significantly increased immunizations rates." 3ie, Quality Assurance of IDinsight's Evaluation of New Incentives, p. 3
Note from IDinsight: "We found campaigns (as reported by clinic staff) to be equally frequent in treatment and control areas (and to not affect coverage that much). It is likely that some RI [routine immunization] activities were more intense in treatment than control areas but we see a good argument for considering those to be part of the treatment effect and likely to recur at scale."
"[W]e calculated the impact estimate for each outcome using vaccination card data instead of self-reported data. [Footnote 79: We also coded outcomes based on BCG scars, Child Immunization Registers, and various combinations of measures.]" IDinsight, Impact Evaluation of New Incentives, Final Report, p. 41
See IDinsight, Impact Evaluation of New Incentives, Final Report, Table 6, pg. 27.
We have not done a thorough review of the rate of scarring for newborns who receive BCG vaccine in settings like those where New Incentives operates. Our best guess of 90% is based on a superficial review of the literature.
Dhanawade et al. 2015 report a scar rate of approximately 90%, based on a simple of Indian infants: "Sixty-four (91.4%) had a visible scar at 12 weeks post vaccination representing a scar failure rate of 8.6%."
They note this rate is comparable to other studies in India but that studies in other countries have found varying scar rates: “In the present study, scar failure rate was 8.6%. This was comparable to other Indian studies on term infants by Rani (10%) and Lakhar (6.1%). Higher scar failure rate (55%) has been reported in low birth weight babies from India. A study from Pakistan reported scar failure rate of 19.6%, whereas the studies from LIMA Peru, Nigeria, and Brazil showed much lower scar failure rate of 1.4%, 3.7% and 3.1%, respectively.”
We don’t have evidence concerning the drivers of higher-than-expected scar rates among infants enrolled in New Incentives. We believe this higher-than-expected rate of scarring in the treatment group may be due to some infants in the treatment group receiving BCG more than once, either due to errors (e.g. unclear child health cards leading clinic staff to believe that an infant had not yet received BCG when they in fact had) or repeat enrollment, though this is speculative (see more discussion on our top charity page for New Incentives).
We think it is possible that enumerators mis-identify scars that aren't present, though we expect this to be less common than enumerators failing to identify scars that are present.
- For self-report data on the RCT's three primary outcomes, see IDinsight, Impact Evaluation of New Incentives, Final Report, Table 3, pg. 23.
- For BCG scar data, see IDinsight, Impact Evaluation of New Incentives, Final Report, Table 6, pg. 27.
We define the rate in the treatment group as the rate in the control group plus the treatment effect, which is the treatment coefficient from a multivariate regression controlling for covariates.
See our calculations here.
- We've compared this approach to simulations provided by IDinsight that estimate the effect of New Incentives on BCG using IDinsight's staff's best guesses on specificity and sensitivity of BCG self-report in control and treatment groups.
- First, we estimated what our adjustments to BCG scarring, combined with comparisons of BCG self-report and scars, would imply for self-report sensitivity and specificity. We find that our control and treatment "true" BCG coverage estimates, based on adjusting BCG scar data, are able to match reasonable values of self-report sensitivity and specificity. In particular, under what we view as reasonable assumptions, we calculate a self-report sensitivity of 0.90 in the control group and 0.92 in the treatment group, and a specificity of 0.62 in the control group and 0.58 in the treatment group. See our calculations in this spreadsheet.
- Second, we compare these estimates to estimates provided by IDinsight. IDinsight staff generated their best guesses for ranges of BCG self-report sensitivity and specificity, based on comparing self-report data to CHCs, clinic records, and BCG scars. Based on these ranges, they estimate a range of effects on "true” BCG coverage from 12 to 41 percentage points. See this spreadsheet, cells R28:R29. Our best guess of 20 percentage points falls slightly below the middle of that range.
See this spreadsheet, cell B19.
See this spreadsheet, cell B38.
We take a weighted average across vaccines, weighting the increase in effect due to partial vaccination (18% for PCV, DTP, and HiB, and 10% for rotavirus) by the percent of all vaccine-preventable disease targeted by the vaccine. See our calculations here.
"There is no substantial evidence of children in control areas receiving the incentive. We found a negligible number of control households that had either a CHC with a NI-ABAE stamp (n = 2, 0.1%) or the caregiver reported receiving cash incentives for vaccination (n = 9, 0.4%). We did not find any control caregivers with a NI-ABAE card. Accordingly, by these three measures, there is no evidence that a meaningful number of infants from control areas were enrolled in the program." IDinsight, Impact Evaluation of New Incentives, Final Report, p. 47.
See this spreadsheet, cells R28:R29.
"GiveWell believes that inaccuracy in self-reports of infant vaccination is likely to be the most important internal validity concern for this program. Inaccurate self-reports could ultimately lead to either upward or downward bias in RCT impact estimates, which GiveWell would like to account for in the internal validity adjustment for this program. While it is likely that self-reports are imperfect, on balance GiveWell has chosen self-reports of vaccination status as the primary outcome measure for this study because we believe that self-reports have fewer weaknesses than alternative impact measures. GiveWell will consider several pieces of evidence when evaluating the final accuracy of self-reported impact estimates. These additional validation exercises, while imperfect, are similar to those conducted in past work, and our impression is that past work did not find serious discrepancies between self-reports and other measures of vaccination status." GiveWell's Pre-Analysis Plan for the New Incentives RCT: Addendum on Internal and External Validity
See this spreadsheet, "Probability of death" sheet, cells D24:E24.
We discuss the prevalence of vaccine-preventable diseases among people older than five below.
See our summary of this data here.
See more detail here.
See IHME data here.
See this spreadsheet, "Probability of death" sheet, cell H15.
See our calculations here, based on data from WHO and UNICEF estimates of immunization coverage, Nigeria.
See this spreadsheet, "Baseline vaccination coverage" sheet, rows 13-23, particularly the note in cell E13.
We use the following formula for each vaccine:
- Probability of death = (Percent vaccinated x RR x Probability of death without vaccine) + (Percent unvaccinated x Probability of death without vaccine)
- Probability of death without vaccine = Probability of death / ( (Percent vaccinated x RR) + Percent unvaccinated) where RR = 1 - vaccine effectiveness.
Relative risk for vaccines is calculated based on the meta-analyses for vaccines discussed below.
- See GiveWell, New Incentives CEA, "CEA - Main" sheet, cell B14.
- We have some remaining uncertainty about this estimate. We assume the 2017 probability of death given in IHME's data is for any child who was under 5 in 2017, which would reflect individuals receiving vaccinations in the past 5 years. If this probability of death is for children born in 2017, then it might be more appropriate for us to use 2017 vaccination rates only. We also do not know the exact method IHME uses to estimate these figures.
See this spreadsheet.
See this spreadsheet, "Main" sheet, cells E4:G4.
See this spreadsheet, "Main" sheet, cells E9:G9.
See this spreadsheet, "Main" sheet, cells E25:G25.
2.5% (probability of death from vaccine-preventable disease for unvaccinated children in Nigeria) * 1.3 (adjustment for higher mortality in North West Nigeria) ≈ 3.2%
We've compared data from IHME to data from the World Health Organization (WHO) and found the total share of vaccine-preventable deaths is roughly similar, though there are differences when looking at specific diseases.
See this write-up for further discussion of how we ought to weight the findings of the PERCH study in our estimates of the proportion of pneumonia deaths that are caused by S. pneumoniae and HiB.
0.67 (weighted average of vaccine efficacy and disease mortality contribution) x 1.33 (adjustment for non-specific effects) x 0.81 (adjustment for vaccine efficacy concerns) / 0.95 (coverage adjustment) = 0.76
We exclude hepatitis B (directly incentivized), yellow fever (indirectly incentivized) and oral polio vaccine (indirectly incentivized) because IHME data show a very low probability of death from these diseases in Nigeria. See IHME data here.
These efficacy numbers correspond to 1 minus the relative risk.
"Pooled vaccine efficacy (VE) for VT‐IPD was 80% (95% confidence interval (CI) 58% to 90%, P < 0.0001); all serotypes‐IPD, 58% (95% CI 29% to 75%, P = 0.001); World Health Organization X‐ray defined pneumonia was 27% (95% CI 15% to 36%, P < 0.0001); clinical pneumonia, 6% (95% CI 2% to 9%, P = 0.0006); and all‐cause mortality, 11% (95% CI ‐1% to 21%, P = 0.08). Analysis involving HIV‐1 positive children had similar findings." "Summary of findings for the main comparison" Lucero et al. 2009
- Fulton et al. 2016:
- "Study outcomes were required to be based on the current WHO definition of (1) "typical" pertussis (>14 days of cough with at least one of the following: paroxysmal cough, inspiratory whoop, or posttussive vomiting, in addition to laboratory confirmation); or (2) "severe" pertussis (>21 days of paroxysmal cough with laboratory confirmation of Bordetella pertussis infection, or epidemiological linkage). Studies using less stringent clinical criteria were also included if their laboratory criteria provided a high level of confidence for pertussis infection (eg, positive culture or polymerase chain reaction assay for B. pertussis)."
- "Regional Databases, with no date restrictions, for English-language studies using the following search terms: pertussis, whooping cough, DTwP, DTaP, vaccine, efficacy, morbidity, and mortality." Figure 2, Figure 4.
- We use the effect on aP in the cost-effectiveness analysis, since it is based on 2 RCTs and is the effect used in LiST; see Table 3.
- WHO does not seem to take a stance on using acellular or whole-cell, though we have not reviewed in depth; see WHO, Summary of WHO Position Papers, Recommended Routine Immunizations for Children.
- Fulton et al. 2016:
See calculations here.
"Nine randomized studies were included in the analysis. Pooled vaccine efficacy using a fixed effects model against confirmed invasive Hib disease following the 3, 2 and 1 primary dose schedule were 82% [95% confidence interval (CI) 73-87], 79% (95% CI 54–90) and 65% (95% CI 23–84), respectively, and the overall efficacy was 80% (95% CI 72–85)." Thumburu et al. 2015
See Figure 1, Sudfeld et al. 2010
- "Tuberculosis (TB) kills or debilitates more adults aged between 15 and 59 years than any other disease in the world. TB is caused by Mycobacterium tuberculosis, an intracellular pathogen of the genus Mycobacterium that includes some 55 species, half of which may cause disease in humans." WHO, Biologicals, BCG (Tuberculosis)
- "Diseases prevented: Disseminated disease and meningitis caused by M. tuberculosis." DCP-3, Methodologies Used for Impact of EPI Vaccines and New Gavi-Supported Vaccines, p. 4
Mangtani et al. 2013, Figure 5, p. 24
"Two forms of TB are life threatening: disseminated or miliar disease, and meningitis." Plotkin's Vaccines (2018), p. 1098.
"Diarrhea remains the second leading cause of death around the world for children under 5 years of age. Because the majority of diarrhea deaths occur in the low and middle-income countries, the etiologic agents responsible for diarrhea deaths among young children are unknown. Using hospitalization data as a best estimate of severe diarrheal disease and a proxy for diarrhea mortality, it has been estimated that rotavirus may be responsible for up to 39% of child deaths, the majority of which occur in low and middle income countries." Walker and Black 2011
See Lamberti et al. 2016, Table 1.
- An "intention-to-treat" (ITT) analysis includes all RCT subjects originally allocated to receive the intervention, whereas a "per-protocol" analysis is a comparison of treatment groups that includes only subjects who in fact completed the intervention.
- The meta-analyses referenced in LiST include a combination of ITT and per-protocol effects. Where possible, we've used ITT effects.
Note that some vaccines target specific serotypes of bacteria: for example, PCV is protective against some serotypes of S. pneumoniae, and we use the effect of PCV on all serotypes (not just those targeted by PCV), since we use deaths attributed to all cases of SP-caused LRTI (not only those caused by vaccine serotypes).
See here for our calculations.
A Cochrane review finds an effect of 95% for a single dose at age 9 months to 15 years, though this is also based on cohort studies and includes a much broader age range.
For example: "The proportion of children who develop protective antibody levels following measles vaccination depends on the presence of inhibitory maternal antibodies and the immunologic maturity of the vaccine recipient, as well as the dose and strain of vaccine virus (Figure 3, Table 1). Frequently cited figures are that approximately 85% of children develop protective antibody levels when given one dose of measles vaccine at nine months of age, and 90% to 95% respond when vaccinated at 12 months of age (17). Among the 44 studies listed in Table 1 in which children were vaccinated between 8 and 9 months of age, the media proportion of children responding was 89.6%. Among the 24 studies listed in Table 1 in which children were vaccinated between 9 and 10 months of age, the median proportion of children responding was 92.2% (mean 88.2; minimum 59; maximum 100; IQR 84, 96)." WHO, Immunological Basis for Immunization Series, Measles
In our CEA, a downward adjustment in measles effect size would also be offset by a weaker downward adjustment for biomarkers. (If measles' effect size were lower, that would bring it closer to what was observed in the biomarkers study.)
The study we use for HiB efficacy (Thumburu et al. 2015) uses ITT effects. However, other studies use a combination of ITT and per-protocol effects. The appropriate adjustment would also take into account efficacy achieved through partial vaccination.
"Vaccination with the protein-polysaccharide conjugate vaccine, PCV7, has significantly reduced
the burden of pneumococcal disease in populations where it is in widespread use and has had an important public health benefit. This vaccine targets only 7 of the more than 92 pneumococcal serotypes, and there have been concerns that the non-vaccine serotypes (NVTs) could increase in prevalence and reduce the benefits of vaccination." Weinberger et al. 2011, p. 1
0.64 x 0.81 ≈ 0.52
These possible explanations include:
- The test failed to detect immunity, potentially due to a) children exhibiting a different immune response, b) misreporting on cards, c) the test's accuracy being lower than reported in literature due to publication bias, d) children being vaccinated too early, or e) the test being administered too soon after vaccination.
- Vaccines failed to provide immunity, either because a) the vaccines lost their potency because they got too hot (i.e. the cold chain was compromised), b) the vaccines expired, c) the vaccines were fake or poorly manufactured, or d) the vaccines were incorrectly administered.
Our best guess is that there is roughly a 50% chance of (i) and a 50% chance of (ii).
When pooling studies, we put 1) 40% weight on the biomarkers pilot, 2) 40% weight on meta-analyses of the effect of measles vaccines on measles, and 3) 20% weight on biomarkers studies from Nigeria using more reliable seroconversion tests.
We have assumed 75% of the adjustment to measles applies to other vaccines, but this is a rough guess.
- Uzicanin and Zimmerman 2011 find a median vaccine efficacy for individuals who were 9-11 months old at age of first dose of 77%, with vaccine efficacy ranging from 73% in the WHO Africa region to 96% in the WHO European region (Table 2, p. 145).
- "Results of this literature review suggest that the VE of MCV1 administered at 9–11 months of age is generally lower than 85%, which is the usual expected rate of immune response after vaccination at that age." Uzicanin and Zimmerman 2011, p. 135
- "Generally, the reasons related to low VE estimates can be grouped into 3 broad categories, including (1) issues related to study methods; (2) program-related factors, such as appropriate vaccine storage, handling, and administration; and (3) host-related factors, most notably, age at vaccination." Uzicanin and Zimmerman 2011, p. 144
"Reviewed studies showed that BCG vaccination protects against pulmonary and extrapulmonary tuberculosis for up to 10 years. Most studies either did not follow up participants for long enough or had very few cases after 15 years. This should not be taken to indicate an absence of effect: five studies (one trial and four observational studies) provided evidence of measurable protection at least 15 years after vaccination. Efficacy declined with time. The rate of decline was variable, with faster decline in latitudes further from the equator and in situations where BCG vaccination was given to tuberculin-sensitive participants after stringent tuberculin testing." Pimpin et al. 2013, abstract.
"Because of the evidence that BCG protects against miliary and meningeal tuberculosis, in developing countries BCG vaccination is recommended at birth (or first contact with health services), taking into account HIV status. Our systematic review suggests that BCG also confers protection against pulmonary disease, the greatest burden from tuberculosis, when administered both in infancy and at school age, providing that children are not already infected with M tuberculosis or sensitised to other mycobacterial infections. Protection against pulmonary disease was seen in the Bombay Infants trial suggesting that, even close to the equator, if BCG is administered prior to exposure to tuberculosis and environmental mycobacteria it can provide significant protection. Further evidence of protection in populations close to the equator from BCG given before infection would strengthen these findings." Mangtani et al. 2013
The meta-analysis from Lucero et al. 2009 includes two African trials:
- Cutts et al. 2005 in Gambia: RR of 0.57 [0.38, 0.85].
- Klugman et al. 2003 in South Africa: 0.58 [0.26, 1.31].
The overall effect is 0.42 [95% CI 0.25, 0.71]. Effects are lower in the American Indian trial (O'Brien 2003) (0.48), Gambia (Cutts 2005), South Africa (Klugman 2003), and the Philippines (Lucero 2009) (1.00), and highest in Finland (Eskola 2001 and Klipi 2003) and the US (Black 2000) (0.11-0.33). (See Figure 4, p. 15). It is not clear to us why these effects vary. The commentary does not explicitly mention changes varying by distance from equator or lower impact in LMICs.
- Higgins et al. 2016 find:
- BCG decreases all-cause mortality by 30% in clinical trials: "BCG vaccine was associated with a reduction in all cause mortality: the average relative risks were 0.70 (95% confidence interval 0.49 to 1.01) from five clinical trials." Pg. 1
- The measles vaccine decreases all-cause mortality 26% in clinical trials: "Receipt of standard titre MCV was associated with a reduction in all cause mortality (relative risks 0.74, [CI] 0.51 to 1.07) from four clinical trials." Pg. 1
- The DTP vaccine increases all-cause mortality by 38%, though this is from observational studies that the authors rate as having a high risk of bias: "Receipt of DTP (almost always with oral polio vaccine) was associated with a possible increase in all cause mortality on average (relative risk 1.38, [CI] 0.92 to 2.08) from 10 studies at high risk of bias."
- Lucero et al. 2009 finds PCV decreases all-cause mortality by 11% (95% CI -1%-21%): "Pooled vaccine efficacy (VE) for VT‐IPD was 80% (95% confidence interval (CI) 58% to 90%, P < 0.0001); all serotypes‐IPD, 58% (95% CI 29% to 75%, P = 0.001); World Health Organization X‐ray defined pneumonia was 27% (95% CI 15% to 36%, P < 0.0001); clinical pneumonia, 6% (95% CI 2% to 9%, P = 0.0006); and all‐cause mortality, 11% (95% CI ‐1% to 21%, P = 0.08). Analysis involving HIV‐1 positive children had similar findings." Pg. 2
- Higgins et al. 2016 find:
See our rough calculations here.
0.7 x 0.74 x 0.89 = 0.46
"Most studies indicated that receipt of DTP was associated with higher mortality, and three individual results had 95% confidence intervals that excluded no effect (one lower mortality, two higher mortality). The average relative risk was 1.38 (0.92 to 2.08) among these 10 studies, all assessed as being at high risk of bias." Higgins et al. 2016, pg. 4
See our calculations here. Even though we guess this effect is driven by non-specific effects (or higher than estimated proportion of deaths caused by vaccine-preventable diseases), we've modeled this as an increase in efficacy of vaccines against vaccine-preventable disease for simplicity.
In that CEA, we state: "Empirical research suggests that malaria control interventions often have a larger effect on all-cause mortality than would be expected exclusively from declines in malaria-specific mortality. We're highly uncertain about the exact value of this input, but we have spoken with malaria experts who told us that it is widely accepted there are roughly 0.5-1 indirect malaria deaths for every direct malaria death."
See this spreadsheet, "Probability of death" sheet, cells H20:K20.
See calculations here.
See this spreadsheet, "CEA - Main" sheet, cell B30.
- See this spreadsheet, "CEA - Main" sheet, cell B33.
- We convert development benefits to an equivalent benefit in terms of child deaths averted using "moral weights," which we use to make cost-effectiveness comparisons between interventions achieving different types of outcomes. See this spreadsheet, "Moral weights" sheet, for the moral weights used in our analysis of New Incentives. We consider these subjective and uncertain; see this page for more discussion of how we use moral weights.
We convert consumption benefits to an equivalent benefit in terms of child deaths averted using "moral weights," which we use to make cost-effectiveness comparisons between interventions achieving different types of outcomes. See this spreadsheet, "Moral weights" sheet, for the moral weights used in our analysis of New Incentives. We consider these subjective and uncertain; see this page for more discussion of how we use moral weights.
See this spreadsheet, "Inclusion/Exclusion" sheet, cell B17.
"Vaccination confers both direct and indirect effects. The direct effect implies protection against disease in vaccinated individuals. Indirect protection is when susceptible individuals avoid infection because the people who surround them are immunized. The magnitude of indirect effects is a function of transmissibility of the infectious agent, population mixing patterns, distribution of vaccine, and distribution of immunity in the population. 'Herd immunity' refers to population-scale immunity." Nymark et al. 2017
- Yellow fever and acute hepatitis B have a much lower DALY burden for children under five (53 and 69 per 100,000, respectively), relative to other diseases prevented by vaccines (see this spreadsheet, "Percent of effect from YLDs" sheet, cells J3, and J12.
- For polio, meningitis A, mumps, rubella, and varicella, our impression is that the number of cases are low, but we haven’t attempted to quantify the burden. In our cost-effectiveness analysis, we include a rough guess for additional benefits from effects on these diseases.
We believe the excluded morbidity effects are likely to be small (no more than 5% of the total DALY burden). This is based on the finding that only a small percent of DALYs is due to years lost to disability (YLDs) for children under 5 (less than 2% of DALYs across all diseases). Rates are higher for older age groups, but the absolute number of DALYs due to YLDs remains relatively low. Excluding these morbidity effects seems consistent with our decision to exclude morbidity from our CEAs of other interventions primarily aimed at reducing mortality (see here for further analysis).
"As the costs associated with seeking care increase and incomes fall, countries may consider using cash transfer programs as a vehicle to promote and expand vaccination coverage. In addition to the significant body of evaluation on cash transfer programs, microtransfers only for full vaccination coverage demand also seem worthy of further consideration (see New Incentives in northern Nigeria with a forthcoming evaluation showing that small demand-side incentives can play a major role in achieving full immunization coverage in very poor settings)." Ismail et al. 2020
"The incentive had the largest impact on the timeliness of the Measles 1 vaccine. Children in the treatment group who received the Measles 1 vaccine were 33 percentage points (95% CI: 28, 38; p-value < 0.001) more likely to have received the vaccination within one month of the recommended age (9 months) compared to those who received the Measles1 vaccine in the control group. Measles 1 vaccinations were also more likely to be timely in the treatment group than in the control group when using a cutoff of within 2 weeks of the recommended age." IDinsight, Impact Evaluation of New Incentives, Final Report, p. 30
See IDinsight, Impact Evaluation of New Incentives, Final Report, Table 13, p. 36: "Outcome: Ever visited clinic… Adjusted OLS results: 0.05 ([95% CI] 0.03, 0.08)"
- See here and Ismail et al. 2020.
- New Incentives may offset this risk through its effect on inactivated polio vaccine (IPV), which may lower risk of outbreak from oral polio vaccines: "Circulating VDPVs occur when routine or supplementary immunization activities (SIAs) are poorly conducted and a population is left susceptible to poliovirus, whether from vaccine-derived or wild poliovirus. Hence, the problem is not with the vaccine itself, but low vaccination coverage. If a population is fully immunized, they will be protected against both vaccine-derived and wild polioviruses." WHO, Poliomyelitis: Vaccine derived polio
"Receipt of DTP (almost always with oral polio vaccine) was associated with a possible increase in all cause mortality on average (relative risk 1.38, 0.92 to 2.08) from 10 studies at high risk of bias." Higgins et al. 2016, pg. 1
See this spreadsheet, "CEA - Summary" sheet, cell B13.
For full discussion, see our cost analysis here.
Note that we frame our cost estimates in terms of the costs per 1,000 infants in areas where New Incentives operates, even though not all of those infants will in fact receive vaccinations — i.e. costs are estimated per 1,000 eligible infants, rather than per 1,000 infants who are in fact vaccinated.
For full discussion, see our cost analysis here.
We add our estimate of the proportion of "always-takers" to the additional proportion of infants whose caregivers were induced to vaccinate by New Incentives' program because New Incentives disburses CCTs for all infants enrolled in its program — i.e., not just those who would not have been vaccinated in the absence of the program.
“The increase in coverage in control areas since baseline is not fully explained by methodological changes or by campaigns. Coverage for all three primary outcomes appears to have increased via routine immunization activities. For BCG and Penta vaccines, campaigns do not account for any meaningful proportion of vaccinations in control areas. Accordingly, we would assume that nearly all of the vaccinations recorded in the control group for BCG vaccine and Penta vaccine represent always-vaccinators who receive vaccinations through RI activities. If we assume that the NI-ABAE CCTs for RI Program pays incentives for all RI-delivered vaccinations, then control coverage is a reasonable proxy for the number of incentives paid to always-vaccinators for BCG and Penta vaccines.
“For the Measles vaccine, we expect that control coverage included a non-trivial proportion of campaign-vaccinators to whom the program may not always pay incentives. There remains uncertainty about the precise proportion, and our data only allows us to generate approximations. However, our best-guess is that RI Measles vaccine coverage in control is about 48% (or 11 percentage points lower than unadjusted self-reports coverage). This value falls in the middle of our various RI-only coverage estimates, which range from 44% to 53% depending on the adjustments applied. If GiveWell assumes that a large proportion of campaign-delivered vaccinations do not lead to program enrollment, then GiveWell should make a downward adjustment to control Measles vaccine coverage before using it as an estimate of enrolled always-vaccinators.” IDinsight, Impact Evaluation of New Incentives, Final Report, p. 62
1,000 infants x 70% infants vaccinated x 95% of vaccinated infants enrolled in New Incentives' program x $38.01 per enrolled infant ≈ $25,285. See GiveWell, New Incentives CEA, "CEA - Main" sheet, cells A43:B50.
However, we are highly uncertain about this ratio in out-of-catchment vs. within-catchment caregivers, and large differences in the ratio between out-of-catchment vs. within-catchment caregivers could lead to significant changes in costs and cost-effectiveness. If, for example, all out-of-catchment infants vaccinated at treatment clinics would not have been vaccinated otherwise, this would lead to an upward adjustment in New Incentives' cost-effectiveness. If, on the other hand, all out-of-catchment infants vaccinated at treatment clinics would have been vaccinated regardless (i.e., were "always-takers"), this would increase costs to New Incentives without increasing vaccination rates, which would lower its cost-effectiveness.
See the “Repeat enrollments” section here for our reasoning behind this estimate.
1,000 infants x ~22 percentage point increase in vaccinations (see above) ≈ 219 additional vaccinations
For full discussion, see our costing analysis here.
Note that this cost analysis does not include the rotavirus vaccine, which New Incentives plans to add to its schedule of incentivized vaccinations.
$26.49 per vaccination x ~219 vaccinations ≈ $5,811. See this spreadsheet, "CEA - Main" sheet, cells B51:B52.
See GiveWell, New Incentives CEA, "CEA - Main" sheet, cells B54:B55.
We retain a high degree of trust in New Incentives and the RCT results that inform our recommendation.
The questions relate to IDinsight, an independent organization. The inquiry from the National Health Research Ethics Committee, Nigeria (NHREC) centers on questions about whether IDinsight had proper ethical approvals prior to starting the pilot study and whether IDinsight appropriately acknowledged its local collaborators. The NHREC has not yet issued a formal response.
The inquiry does not relate to any aspect of the main RCT that informs our understanding of the program's impact. IDinsight told us that the main RCT received full ethical approvals and permissions in advance of data collection, and it shared the signed approvals it received from the NHREC. The questions do not relate to New Incentives' operation of the program. We thus do not believe that these concerns should impact our assessment of New Incentives.
Details on what happened and steps we and IDinsight are taking are here.
See this spreadsheet, "Inclusion/Exclusion" sheet, cell B15.
See this spreadsheet, "CEA - Main" sheet, cell B6.