A mark and recapture perspective on vaccination touchpoints Authors: Niket Thakkar Published: 2026 ABSTRACT This paper considers large-scale vaccination campaigns, a major platform for vaccine access in a lot of the world, as a recapture estimate of the target population marked by routine immunization. Framing the campaign as a measurement, we learn about its properties, including the campaign’s coverage of the target population and some implied sampling properties of post-campaign coverage surveys (PCCSs), the current gold-standard in implementation quality measurement.We develop this idea in the context of the 2023 measles campaign in Kano State, Nigeria, where we have detailed implementation data collected by vaccination teams involved in that effort. Looking specifically at the teams’ tally sheets, the daily records of who they vaccinated, we find significant discrepancies between the recapture estimates and those from the corresponding PCCS. Exploring a variety of bias models applied to both the tally sheets and the PCCS helps clarify how anecdotal issues from the field relate to this discrepancy.Overall, we find that the tally sheets, despite being an unorthodox population sample, provide a tractable perspective on implementation and measurement, one that’s in principle available nearly instantly and at high resolution for any vaccination touchpoint. I. Introduction It’s generally recognized that a commitment to improving global health is also a commitment to measuring the properties of interventions. 1 Unfortunately, the consensus tends to disintegrate from there. What needs to be measured, to what precision, how to do it, who it’s for (and so who should pay for it) are all open questions in a broadly chaotic measurement environment. Measles vaccine delivery is a concrete example. In the highest-burden settings, routine vaccination services are inaccessible or unreliable, leaving gaps in immunity that have to be closed by dedicated vaccination campaigns every few years. 2 In those efforts, medical teams bring vaccines to underserved areas and try to immunize kids within a predetermined target population. Designing these efforts to better reach susceptible individuals before the virus does is a major global health initiative, 3 one that requires high-quality measures of intervention impact, comprehensiveness, and failure-modes. The accepted standard for measuring campaign quality is the post-campaign coverage survey. 2 , 4 PCCS is a 2-stage cluster survey where individuals in the target population are asked about their campaign experience, including if they received a vaccine at all. While useful in principle, in practice PCCS is underpowered at the granular scale relevant to implementation, it’s rarely conducted in a reasonable time frame, and often its findings are colloquially reduced to just a coverage number and then forgotten. It’s becoming clear that PCCS isn’t satisfying programmatic measurement needs. In fact, most countries that conducted a campaign since 2020 haven’t even bothered with a PCCS. 4 For a complex intervention like a vaccination campaign it seems unwise to rely so heavily on one measurement platform. Ideally, measurement methods offer light touch, timely, accurate, and interpretable insights into intervention details, which is a lot to ask of a single approach. As campaigns progress with no measurement plans at all, the toolbox is starting to feel pretty empty. Along those lines, the main idea in this paper is to make better use of the implementation records already collected in a vaccination campaign. We focus specifically on the tally sheets, 5 Figure 1 is a blank example, literal paper records of who vaccination teams encounter as they work. In exploring the data photographed and digitized during the 2023 measles campaign in Kano State, Nigeria, 6 we show that campaign coverage can be estimated directly from the tally sheets in a principled Bayesian inference scheme, making use of prior information on routine immunization coverage and target population size. The approach is analogous to mark and recapture methods common in ecology, 7 , 8 and so we call this the recapture coverage estimate. A blank tally sheet from Nigeria circa 2023. Vaccinators tick off the relevant bubbles as they work, compiling a histogram each campaign day. In the specific example considered, the recapture coverage estimate is in disagreement with the corresponding PCCS estimate. Two quiet but common narratives in global health are that tally sheets are at least partially fabricated to increase campaign scale and that PCCS sampling might be biased to increase coverage in response to political pressures on performance. We consider these hypotheses quantitatively in our example, and we show that an unrealistically structured bias in tallying would be necessary to corroborate the PCCS. This paper does not fully settle the issue of campaign coverage and target population size in 2023 in Kano. But it does demonstrate that implementation records offer timely and interpretable insight into campaign coverage. As our understanding develops, that could free the PCCS to focus on other implementation details, and we conclude with some general thoughts to that effect. II. Four confusing numbers The 2023 measles vaccination campaign in Kano State took place over 9 days in mid December. In the planning phase, the Gates Foundation funded GRID3 to support implementation in a variety of ways, including a pilot study where paper records were photographed and hand-digitized by GRID3 employees in 6 local government areas (LGAs, the Nigerian administrative level 2 unit, essentially districts) considered representative of the state [1]. The tally sheet photographs in particular were taken on a nightly basis in a direct interaction with vaccination team supervisors. 6 For our purposes, the implementation data is essentially a collection of counts by vaccination team every campaign day, taken from records like that in Figure 1. While the tallies are disaggregated by sex and age, we’ll focus specifically on the 12 to 23 month-olds across their caregiver reported status of having had measles vaccine before (“other dose”, OD) or not (“zero dose”, ZD). In Kano, across the 6 LGAs, there were 204137 1 year-old tallies in total, with 189321 reporting OD. Since the previous campaign in Kano was in 2021, these caregivers must have been referring to vaccines delivered through the routine immunization (RI) system. With that thought in mind, we might take tallied 1 year-olds as a sample of RI coverage and conclude that 92.7% of kids get measles vaccine through RI in Kano. In the same year though, the Demographics and Health Survey (DHS) found that RI coverage among 1 year-olds was 38.6% . 10 , 11 Meanwhile, for this campaign, the PCCS estimated that 89% of the target population was covered, with 11% of them getting their first ever measles vaccine from the intervention. 12 These numbers are collected in Table I with some sample size and design details. Table I. Routine, campaign, and post-campaign measles vaccine coverage data from Kano State Nigeria, circa 2023.SourceCoverage Sample sizeDHS a (12–23mo.).386531Tally sheets (12–23mo.).927204137PCCS b (9mo.–5yr.).89600PCCS with ZD at the campaign c .11600 a The 2023 DHS report 11 has design factor 1.411, so De=1.99 for this statistic. b Approximately 107 of the 600 samples are 1-year-olds, and De=8 for overall coverage. 12 We take estimates across ages as representative of 1 year-olds but with higher uncertainty. c De=2 . Estimates across all states suggest this could be ≈.15 for 1 year-olds, higher, but still within the Kano interval. 12 Routine, campaign, and post-campaign measles vaccine coverage data from Kano State Nigeria, circa 2023. The 2023 DHS report 11 has design factor 1.411, so De=1.99 for this statistic. Approximately 107 of the 600 samples are 1-year-olds, and De=8 for overall coverage. 12 We take estimates across ages as representative of 1 year-olds but with higher uncertainty. De=2 . Estimates across all states suggest this could be ≈.15 for 1 year-olds, higher, but still within the Kano interval. 12 It’s not clear how the pieces fit together. For example, taking the PCCS and tally sheets at face value and assuming every missed child is ZD, we’d estimate RI coverage is 82.5% at a minimum, more than 14 standard deviations from the DHS. Meanwhile, accepted population estimates like that from GRID3 13 put approximately 144k 1 year-olds in the 6 LGAs, so we sometimes see reports that campaign coverage is an impressive ≈140% . The usual thing to do at this stage is to argue that the DHS and PCCS have rigorous sample designs while the tally sheets do not. Something must be wrong with the latter, and they should be ignored. As we’ll see, there is probably something wrong with the tally sheets. But they shouldn’t be ignored. III. Mark and recapture The size of the target population, that is the number of 1 year-olds in the 6 LGAs, is a fundamental uncertainty underlying the discordance above. Mark and recapture methods are an established approach to estimating population sizes based on relationships between distinct samples. They inspire a line of thinking for us in Kano. For the unfamiliar, the canonical 7 mark-recapture example is something like this: An ecologist wants to know how many fish are in a lake. She first goes to the lake, catches 10 fish, and marks them in some durable way. Then she puts them back into the water. When she returns the next day, she catches 20 fish, of which 4 are marked. Since the expected number of marked fish in 20 samples from a large population of N fish is 20×(10/N) , she might estimate that there are N=50 fish in the lake, the so-called Lincoln-Peterson 14 , 15 estimate. Our situation is a variant where we have some uncertainty in the number of individuals marked. Specifically, in Kano, the DHS tells us that 38.6% of the population are “marked” by vaccine through RI, and in the campaign, teams found a sample of 204 k children, of which 92.7% were marked. If we assume that the campaign exhausted the population covered by RI, that would imply there are 204k×(92.7/38.6)≈490 k 1 year-olds in the population, and campaign coverage was about 42% . This quick calculation is the basic idea in what follows. It turns out, with some care, we can leverage prior information to avoid assumptions about campaign coverage in OD children. Moreover, in framing the problem probabilistically, we can quantify uncertainty more thoughtfully throughout, giving us some perspective on the PCCS estimates in Table I as well. To get started, consider a population of N 1 year-olds. Conditional on N , we can define a length N binary vector R=(R1,R2,…,RN) representing each 1 year-old’s RI status just before the campaign, with Ri=1 if the i th child received one or more measles vaccines through RI. The population has some true RI coverage CN(R)=1N∑i=1NRi, (1) a population average based in this case on N samples, that is, a census for a given N . For our purposes, the DHS is an estimate of Eq. 1, and we can retrieve the corresponding average and variance directly from the survey report. But for completeness and clarity, we treat the DHS as another length N binary vector S0=(S10,S20,…,SN0) , where Si0=1 means a child was sampled in the DHS, and ∑iSi0=n0 is a known constraint called the sample size. The relevant estimate in our case is Cn0(R)=1n0∑i=1NSi0Wi0Ri, a weighted average where Wi0 is the i th child’s normalized survey weight. The DHS sample is constructed for these averages to be unbiased and E[Cn0(R)]=CN(R) , where the expected value is taken over p(S0|n0) , a joint distribution sometimes called the sample design. Meanwhile, for binary R , the variance V[Cn0(R)]≈(De0/n0)E[Cn0(R)](1−E[Cn0(R)]) , where De0 is called the design effect, quantifying variance inflation relative to a less economical, large N simple random sample. 16 Along similar lines, the campaign is a distinct sample, again a length N binary vector S1=(S11,S21,…,SN1) where Si1=1 means the i th child received a campaign vaccine and ∑iSi1=n1 is known from the tally sheets. The sample is unorthodox in the sense that it has unknown design. In other words, we need a reasonable and tractable model for statistics of p(S1|n1) . To make progress, we consider Ri as a correlate of the i th child’s campaign participation, and we can hypothesis test against a sampling model with two distinct sub-populations, those with and without routine measles vaccine. Specifically, let k1≡∑i=1NSi1Ri∼Binom{NCN(R),π1}, k0≡∑i=1NSi1(1−Ri)∼Binom{N(1−CN(R)),π0}, such that k0+k1=n1 and both π0 and π1 are unknown sampling probabilities. Standard manipulations give the more concise form p(k1|n1,R)∝(NCN(R)k1)(N(1−CN(R))n1−k1)ψk1, (2) where ψ≡π1(1−π0)/π0(1−π1) is the odds ratio of being sampled based on RI status, and conditionality on N is implicit in the conditionality on R . For general ψ≠1 , this construction is sometimes called Fisher’s noncentral hypergeometric distribution. 17 Its properties and large n1 asymptotic approximations to its mean and variance are discussed in Appendix A. In the context of Eq. 2, the tally sheet RI coverage estimate is Cn1(R)=k1/n1 . As a result, the observed difference, Δ≡Cn0(R)−Cn1(R) has statistics E[Δ]=E[Cn0(R)]−1n1E[k1], V[Δ]=V[Cn0(R)]+1n12V[k1], where in the second line we assume the two samples are independent. From Appendix A, V[k1]∼O(n1) , so the DHS variance dominates when n1>>n0 , and p(Δ|N,ψ) is Gaussian to a good approximation when we’re thinking about a scale where the DHS is well-powered. Under that assumption, Bayes’ theorem gives us an approachable, 2-dimensional inference problem characterizing the campaign, p(N,θ|Δ)=p(Δ|N,θ)p(N,θ)p(Δ)≈N(Δ|E[Δ],V[Δ])p(N)p(θ)p(Δ), (3) where N is a Gaussian distribution, and we’ve chosen to work in terms of θ≡lnψ to facilitate the numerical integration needed to estimate p(Δ) and solve the inference problem. Note that all the distributions in Eq. 3 are implicitly conditional on the sample sizes n0 and n1 . In Appendix B, we show that the maximum entropy principle 18 – 20 guides us to logical forms for the prior distributions p(N) and p(θ) assuming there’s some known prior average target population size, N^ , used for campaign planning, like the one from GRID3 13 mentioned previously. For completeness in this section, we take p(N)=1N^(1−1N^)N−n1 (4) p(θ)=12csch2(θ2)[θ2coth(θ2)−1], (5) on domains N≥n1 and θ∈(−∞,∞) . To whatever extent it’s helpful, Eq. 4 might be called a left-truncated geometric distribution. Eq. 5 doesn’t have a name to our knowledge but is an exact consequence of 0≤π0,π1≤1 and the definition of the log odds ratio. The joint prior distribution p(N)p(θ) is log-contour-plotted (black lines) in Figure 2’s first panel. The distribution is peaked at N=n1 , θ=0 , and then falls geometrically with increasing N and exponentially with θ . Intuitively, the prior encodes that the target population is at least n1 and otherwise finite. Meanwhile, the symmetry around θ=0 represents our ignorance on how Ri=1 is associated with campaign coverage in the absence of any data. The tally sheets at face value. (Left) The joint prior distribution p(N,θ) (log contours in black) gives the N,θ -space some basic structure. The likelihood (tan) highlights the ridge consistent with the data. (Right) Their combined contributions are evident in p(N,θ|Δ) (log contours at the same levels), which can be used to compute the relevant marginals (grey). The observed Δ=0.386−0.927 in Table I specifies the likelihood p(Δ|N,θ) , which we visualize as a tan shaded interval within 5% of the log-maximum in Figure 2. The likelihood defines a ridge of N , θ pairs that are consistent with Δ . As θ increases to infinity, OD children are infinitely more likely than ZD children to participate in the campaign, and we see the ridge approach the perfect OD coverage estimate, N≈490 k, discussed above. But with the machinery in Eq. 3 we can solve the inference problem more completely, as illustrated in Figure 2’s second panel. Log joint posterior contours show how geometric features of the inputs combine to highlight the region most probably consistent with the data and prior uncertainties. The intuition that both perfect OD coverage and extremely high N are unlikely leads us to the corner in the likelihood ridge. Marginal distributions p(N|Δ) and p(θ|Δ) , estimated via numerical integration and visualized along the corresponding axes, highlight the specific θ and N values most consistent with the first two rows in Table I. Overall, in Kano in 2023, if we believe the tally sheets, we also must believe in a large, significantly less accessible ZD population that was missed by the campaign. IV. Recapture coverage Getting from Figure 2 to a coverage estimate is straight-forward since campaign coverage is n1/N , a one-to-one function of N . In other words, we can take the marginal distribution p(N|Δ) and compute p(n1/N|Δ) directly. The result (black) is visualized alongside the corresponding PCCS estimate (grey) in Figure 3. Recapture coverage. Taking the tally sheets as they are, we get a campaign coverage estimate (black) that’s inconsistent with the corresponding post-campaign survey (grey). Based on the tally sheets, our best guess for campaign coverage is n1/N∗≈39% , with a hard stop at the 42% , ψ→∞ limiting estimate. There is essentially no chance the 89% PCCS coverage estimate ( 95% interval shaded) is consistent with the recapture approach. To add some texture to this result, we can calculate an individual-level contingency table p(Ri,Si1|Δ) encapsulating the implications consistent with the tally sheets. Specifically, if we take the marginal probabilities ∑Si1p(Ri=1,Si1|Δ)=p(Ri=1|Δ)≈Cn0(R),∑Rip(Ri,Si1=1|Δ)=p(Si1=1|Δ)≈n1N∗, we find Ri=0Ri=1p(Ri,Si1|Δ)=Si1=0Si1=1{0.5590.0500.0550.336} (6) making use of the most probable log odds ratio in Figure 2 and noting that πj≡p(Si1=1|Ri=j) for j=0,1 in this new notation. At the most probable parameter values, an OD child is more than 9 times as likely to be reached by the campaign than a ZD child. Along those lines, we estimate that the campaign covers 87% of the OD population but only 9% of the ZD population. The probability of a child getting their first vaccine from the campaign is p(Ri=0|Si1=1)≈14.2% . Meanwhile, in this picture of the 6 LGAs, we find N∗≈522 k, more than 3.5 times the GRID3 estimate. Eq. 6 paints a surprising picture, particularly in light of the campaign’s stated 95% coverage goal. 11 We should and will be skeptical of it. But before that, the PCCS results in Table I are still lingering, and it’s interesting that surveyed coverage among ZD was 11% , in reasonable agreement with the recapture estimate. It’s helpful to think of participation in the PCCS as another length N binary vector, S2=(S12,S22,…,SN2) with Si2=1 indicating the i th child was surveyed and ∑iSi2=n2 . While the PCCS has well-defined survey design, and is unbiased in principle like the DHS, Figure 3 at least suggests we should test for potential issues. With that motivation, we might have two competing hypotheses for its sampling distribution p(S2|S1,R) in the post-campaign context. The more diplomatic hypothesis is that p(S2|S1,R)≈p(S2|R) . In words, PCCS is conditionally independent of the campaign, and sampling bias is explained by health care access. Surveyed campaign coverage estimates are spuriously inflated because the campaign shares a correlation with Ri=1 . In this case, as we did for S1 , we would test against ℓ1≡∑i=1NSi2Ri∼Binom{NCN(R),π˜1},ℓ0≡∑i=1NSi2(1−Ri)∼Binom{N(1−CN(R)),π˜0}, such that ℓ0+ℓ1=n2 and the sampling probabilities π˜0 and π˜1 reduce to one unknown as in Eq. 2. On the other hand, the more cynical p(S2|S1,R)≈p(S2|S1) has campaign participation itself as the dominant bias, maybe because incentive structures pressure programs to have high coverage campaigns. In that case we might test m1≡∑i=1NSi2Si1∼Binom{∑i=1NSi1,π^1},m0≡∑i=1NSi2(1−Si1)∼Binom{N−∑i=1NSi1,π^0}, a comparable model, with m0+m1=n2 , but with the population structured by campaign participation instead of RI coverage, and so with distinct sampling probabilities, π^0 and π^1 , in general. In the first case, campaign coverage is E[1n2∑Si2Si1]=E[ℓ1]n2p(Si1=1|Ri=1) +(1−E[ℓ1]n2)p(Si1=1|Ri=0), (7) a weighted average of OD and ZD coverage. Eq. 7 can be compared to the overall estimate in Table I to specify E[ℓ1]/n2 consistent with Eq. 6. Meanwhile, in the second case, surveyed campaign coverage is E[m1]/n2 directly. In other words, both models can explain the overall PCCS estimate with different underlying mechanisms. But with each sampling distribution specified by the 89% number in Table I, we can calculate the statistic F=∑iSi2Si1(1−Ri)/n2 , the surveyed estimate that a child gets their first measles vaccine from the campaign. In the p(S2|R) model, this is just the second term in Eq. 7, and we would estimate F<2% almost certainly. Meanwhile, in the p(S2|S1) model, we find F=(E[m1]/n2)×p(Ri=0|Si1=1)≈12.6% , and highly likely bounded between 9 and 15% , an order of magnitude higher. Reflecting on the two mechanisms, if the campaign and PCCS share an underlying correlation with Ri=1 , surveying a ZD child who was also vaccinated in the campaign is rare twice over. But if the PCCS is correlated with Si1=1 directly, surveying a ZD child is much less of a hindrance. In other words, if we believe the tally sheets from 2023, we must also believe that the corresponding PCCS was biased towards high coverage directly. V. Some asymptotic intuition So far we’ve discussed one explanation for the discordance in Table I, one that takes the DHS and the tally sheets as the two most reliable pieces of data. It’s healthy to ask how potential recording errors or biases might influence our results. With that motivation in this section, we derive a coarse but interpretable approximation to n1/N∗ , which highlights how the recapture estimates change with distortions to the data. In Appendix A, where we discuss some known statistical properties of the distribution in Eq. 2, we show that its mode, k1∗ , is related to θ through k1∗n1=A(θ)2n1(eθ−1)[1−B(θ)A(θ)], (8) where A(θ)=N+(eθ−1)(n1+NCN(R)),B(θ)2=A(θ)2−4eθ(eθ−1)n1NCN(R). In our case, reflecting on Figure 2, we’re in a situation where campaigns are much better at reaching OD children than ZD children. Mathematically then, we’re concerned with inference stability in the large θ limit, which means characterizing Eq. 8 to leading order in e−θ=1/ψ and calculating the downstream properties of Eq. 3. To start, we expand the terms outside and inside the brackets in Eq. 8 separately, and find A(θ)2n1(eθ−1)=12(1+CN(R)z+O(e−θ)), where we’ve introduced the notation z≡n1/N , and B(θ)A(θ)=1−4zCN(R)(z+CN(R))2+O(e−θ). Combining the two expressions and collecting terms exposes the mode’s leading order behavior, k1∗n1=CN(R)z+O(e−θ), recovering the ψ→∞ result discussed in the last two sections, that k1∗→NCN(R) , the full OD population. When n1 is large, Eq. 2 is sharp, and the mean is close enough to the mode for our purposes here. Taking E[k1]/n1≈CN(R)/z , Eq. 3 reduces to lnp(z,θ|Δ)≈const.+n1zln(1−1N^) −12V[Δ](Δ−CN(R)[1−1z])2, (9) where the constant depends on θ but not on z . Then, in the same large n1 limit, V[Δ]≈V[Cn0(R)] , and Eq. 9 can be maximized to find z∗ , the most probable campaign coverage estimate. Carrying out the calculation, with ln(1−1/N^)≈−1/N^ , we find 1z∗=N∗n1≈1−ΔCN(R)−n1N^V[Cn0(R)]CN(R)2. (10) Eq. 10 bounds coverage in the ψ→∞ limit and helps scaffold our thinking. The ratio V[Cn0(R)]/CN(R)2 is an index of dispersion, quantifying our uncertainty in RI coverage. It decays as 1/n0 for the DHS. As a result, when our RI coverage estimate is precise and the prior population estimate is O(n1) , the final term in Eq. 10 is small. In 2023 in Kano for example, it’s −0.008 if the DHS is genuinely unbiased so Cn0(R)≈CN(R) . If the DHS were biased, a possible error neglected in the previous two sections, we might more generally write Cn0(R)=(1+α)CN(R) with α≠0 characterizing the degree. In that case, a precise but shifted DHS estimate would modify z∗ like so, z∗≈Cn0(R)Cn1(R)[1−α(1−Cn0(R)Cn1(R))], to leading order in α . Realistically, if it’s biased at all, the DHS likely overestimates CN(R) , so α>0 . In the large θ regime we’re in, that would decrease z∗ , making the discrepancy in Figure 3 worse. DHS bias doesn’t seem like a viable explanation for the issues in Table I. Meanwhile, when it comes to tally sheet recording, Eq. 10 tells us that the recapture coverage estimate depends mainly on the ratio of RI coverage to the tally sheet average. In the Kano example, an error that inflates n1 by 10 times would modify recapture coverage by only 0.1% if it maintains Cn0(R)/Cn1(R) . That said, those same types of errors effect N∗ directly, and we should interpret recapture target population estimates with that in mind. VI. A hypothetical tally Eq. 10 suggests that the recapture coverage estimate would align with Kano’s PCCS if Cn1(R) was roughly halved. In fact, it seems like the recapture estimate is robust in the sense that no other change will do the job. Tabling discussions of plausibility for now, let’s assume half of the OD tallies were misrecorded and were actually ZD. In that case Cn1(R) goes from 92.7% to 46.4% , and we can calculate the consequences from Eq. 3 onward. The new coverage estimate is visualized in Figure 4, and it agrees with the PCCS as expected. Revisiting the contingency table, p(Ri,Si1|Δ) , at the most probable estimates, we find Ri=0Ri=1p(Ri,Si1|Δ)=Si1=0Si1=1{0.1370.0490.4770.337}. (11) Figure 4. Hypothetical tally sheets in Kano.Recapture coverage (black) and the post-campaign survey (grey) agree if half of the OD tallies are actually ZD. Hypothetical tally sheets in Kano. Recapture coverage (black) and the post-campaign survey (grey) agree if half of the OD tallies are actually ZD. Now, an OD child is 1.1 times as likely as a ZD child to be reached by the campaign, the campaign covers 87.2% of the OD population and 77.8% of the ZD population, and the probability of a child getting their first vaccine from the campaign is 59% . The estimate of target population size changes as well, N∗≈251 k, a potentially more reasonable 21 1.7 times the GRID3 estimate. Our previous thoughts on PCCS sampling basically reverse. Specifically, to explain overall coverage surveyed around 89% , many more ZD children are required to have been vaccinated in the campaign. Under the p(S2|S1) sampling model specified by PCCS coverage, we’d estimate that F>42% of children surveyed got their first measles vaccine from the campaign. Meanwhile, in the competing p(S2|R) model, F is unconstrained, covering the 11% number in Table I. So in this hypothetical situation, the campaign and the PCCS sharing an underlying correlation with Ri=1 seems necessary. VII. Some structured error models Figure 4 demonstrates that it’s possible for the recapture and PCCS estimates to agree. Realizing that agreement in Kano would require a systematic recording error where about half of OD tallies are actually ZD. It seems unlikely that a net 50% false positive rate is realistic, but it’s worth ruling out more thoroughly. One potential mechanism is caregiver recall – maybe children get other vaccines and their parents get them mixed up when asked about measles. Literature on this topic 22 doesn’t suggest a large, one directional effect however. In most studies, people are roughly as likely to forget doses they received, falsely reporting ZD, as they are to invent doses they haven’t. Recall doesn’t seem affected by the campaign environment either. Looking at serological data from the 2020 measles campaign in Zambia, 23 differences in seropositivity between those with documented vaccine and those with recalled vaccine suggests that OD memory is accurate ≈86% of time. That number is in good agreement with more focused recall studies. 24 Moving on then to recording itself, we might guess that some vaccination teams didn’t bother asking about dose history. Across the 409 teams in the 6 LGAs, 100 never tally a ZD in any age group. If we remove those tally sheets from the dataset however, Cn1(R)→0.906 . Not even close to what we’d need. We can try more sophisticated misrecording indicators too. Since the number of children who show up on a given day is a random process, we shouldn’t expect any preference for round numbers. Looking at the work across the teams, with 10 numbers per sheet, we get a sample of up to 90 nonzero numbers per team across the campaign. We can test for example if the last digit in each nonzero entry is uniformly drawn from 0 to 9 using a standard Pearson’s χ2 test applied to each team’s tallies. There’s definitely evidence of non-uniform recording. Some examples are striking: One team recorded 33 non-zero numbers across the campaign, 30 of which ended in 0. More broadly, if we remove the sheets from teams with less than 10% probability of drawing from a uniform last digit distribution, we lose the work from 60% of the vaccinators. Doses delivered in the campaign decline substantially, n1→77139 , but the OD fraction remains essentially the same, Cn1(R)→0.923 . As a result, as expected from Eq. 10, the recapture coverage estimate is unchanged, but N∗→228 k, larger than the GRID3 estimate by 1.6 times. Returning to the motivating goal, while we could probably find more reasons to question subsets of the tally sheets, it’s difficult to imagine a mechanism for systematic false OD tallies at such a large scale. From an estimate stability perspective then, the recapture coverage approach seems generally reliable. VIII. Conclusion At this stage, reflecting on Table I, we’ve considered the data from multiple angles, which are summarized in Table II. The tally sheets give compelling evidence that Kano’s 2023 campaign had lower coverage in the 6 LGAs than the PCCS’s state-wide 89% . At the same time, the tally sheets imply there are more than 3 times as many 1 year-olds as people think, so it also seems likely that the number of tallies is inflated in some way. To what extent is unclear. Table II. 3 perspectives on the data in Table I.Data taken as reliable Implication92.7% of tallies are OD and campaign coverage is 89%RI coverage is at least 82.5% a RI coverage is 38.6% and campaign coverage is 89%About 46% of tallies are OD b RI coverage is 38.6% and 92.7% of tallies are ODCampaign coverage is at most 42% c a This was discussed briefly at the end of Section II. See also the comments on the implied DHS bias in the thinking below Eq. 10. b This was developed in Section VI, motivated by Eq. 10, and with implications encapsulated in Eq. 11. We also found N∗≈251 k in this case, and we discussed the plausibility of this recording bias in Section VII. c This is the result from Section IV, with implications encapsulated in Eq. 6, and with N∗≈522 k. Mechanisms for the implied PCCS bias were discussed in the text around Eq. 7. 3 perspectives on the data in Table I. This was discussed briefly at the end of Section II. See also the comments on the implied DHS bias in the thinking below Eq. 10. This was developed in Section VI, motivated by Eq. 10, and with implications encapsulated in Eq. 11. We also found N∗≈251 k in this case, and we discussed the plausibility of this recording bias in Section VII. This is the result from Section IV, with implications encapsulated in Eq. 6, and with N∗≈522 k. Mechanisms for the implied PCCS bias were discussed in the text around Eq. 7. More to the point though, we have shown that implementation data, despite being collected outside a traditional survey platform, can power provably robust inferences about campaign quality, target population size, and measurement. The approach illustrated in Figure 2 is applicable instantly after tallies are totaled, Eq. 10 suggests that resolution is limited mainly by precision in RI coverage, and results like Eq. 6 meaningfully highlight campaign failure modes. Analysis like this costs very little. It should probably be a part of all campaigns. If it were, it would inspire more thoughtful tally sheet design and more intentional data collection. We might reconsider PCCS’s role in that context. The Kano example suggests that PCCS is less suited to measuring coverage, and is better for measuring other campaign details, like the characteristics of children conditional on their campaign participation. It seems worth some study to design post campaign surveys with that narrower focus in mind, giving vaccination programs more fit-for-purpose measurement tools. Manuscript history A previous version of this article appeared as a preprint on arXiv. 25 Ethics statement The author did not obtain ethical approval for this research since all data used is routinely collected, predicated on voluntary participation in an immunization activity, and anonymous.