Evidence of the Cost-Efficiency of Scale as seen in Polio Vaccination and Surveillance Costs

This analysis examined how polio program costs vary with scale for vaccination and disease surveillance, based on historical budget data published by the Global Polio Eradication Initiative (GPEI) from 2005 to 2018. We applied a linear mixed effects regression model in order to understand the cost structure of the historical GPEI budgets, with the goal that lessons learned from polio may be extended to other global disease elimination programs. Our findings demonstrate that there are economies of scale for vaccine delivery operations and for disease surveillance, which means that larger programs can leverage fixed costs and achieve better cost-efficiency as they scale. This finding should enable decision makers to create more reliable budgets, which support fundraising and optimal resource allocation. They also provide insight into how cost effectiveness changes as programs scale up during progressive disease control and elimination, as well as what level of resources are needed to sustain a program that is scaling back post-eradication and through to certification.


21
In earlier studies (1)(2)(3), vaccine delivery costs are reported on an average per-unit basis, such as per-22 vaccine or per-child. Averages are calculated as the total of many costs including categories such as 23 salaries, transportation, vaccine acquisition, and supply chain, divided by the number of units (e.g. 24 population targeted or the number of vaccines administered), such as in Walker et al. (2004) and Page | 2 25 Griffiths et al. (2016). Some analyses (6) include a sensitivity analysis and demonstrate that the average 26 cost may vary from the reported value, but this only emphasizes the uncertainty in the point estimate, 27 not that the drivers of cost may change structurally under different conditions.

28
In addition, Geng et al. (2017) and Schütte et al. (2015) have demonstrated that program costs do not 29 scale linearly at a facility level and Ahanhanzo et al. (2015) demonstrated that the size of a clinic is an 30 important cost driver. In this paper, we examine whether programmatic costs for supplementary 31 immunization activities (SIAs) demonstrate efficiencies of scale. If so, budgeting based on unit cost 32 would drive inappropriate decision making, as shown by Claxton et al. (2016) and suboptimal budget 33 allocation, as previously demonstrated by Fitzpatrick and Bauch (2011).

34
Previous studies (Walker et al. 2004;Wolfson et al. 2008;Portnoy et al. 2015) have recognized the 35 importance of reporting both fixed and variable costs, as well as differences between subgroups of 36 countries, for example World Bank income class and World Health Organization (WHO) region.

37
However, these studies generally only provide an average per-unit cost without discussion of whether 38 this value is applicable across a range of activity levels, for example moving from pilot program to broad 39 deployment, or when building on an already successful platform. The insufficient understanding of the 40 drivers of cost may be one of the reasons that budgets are often quite far off from actual expenditures 41 (13), and Gandhi and Lydon (2014)

112
The LMER model has the form log(y it ) = β i0 + β 0 + β il * log(x it ) + β 1 * log(x it ) where y it is the cost and x it is 113 the number of units for country in year and both values have been transformed to the natural-log i t 114 scale. The β i0 and β i1 are mean-zero random intercepts and random slopes, respectively and describe 115 the country-specific deviations. The resulting coefficients β 0 and β 1 were used to determine the rate at which costs scaled with program size. With this model, a β 1 that is less than 1.0 means that an increase 117 in x it (e.g., number of vaccine doses distributed) is associated with a decreasing marginal increase in y it 118 (e.g., vaccination delivery cost), implying an efficiency of scale. That is, if the mean-zero random effects 119 are removed, log(y it ) = β 0 + β 1 * log(x it ) and 1, then dy/dx decreases as x it increases.

121
All model fits were done in the software package R (18) version 3.4.2. We used the lmer() function from 122 the lme4 library (Bates et al. 2014) to fit the LMER model.

124
We first examined whether a natural-log scale provided a good basis for analysis. Figure  we also calculated the correlation on the natural log scale, with the resulting correlation statistic of 135 0.935. From this, we concluded that using the historical budgeted data was adequate for our purposes.

136
The remainder of the results in this paper are based on natural log transformed values and utilize only 137 the budget dataset from the GPEI's FRRs.
Page | 7 138 Next, we examined the relationship between several hypothesized cost drivers and subsets of costs.

139
The cost categories of "surge" and "other" are not consistently reported over time or for every country, 140 so they were included in the total cost numbers but were not considered independently.

150
The LMER model was fit for all four scale-driver pairs. The total cost per capita is provided for 151 informational purposes, since this is how budgets for many programs are currently reported. The OPV 152 vaccine prices are negotiated between UNICEF and vaccine manufacturers on a per-vial basis, this 153 provides a natural control, which should have a β 1 close to 1.0. Indeed, we find that the confidence 154 interval for β 1 includes the value 1.0, which provides us with confidence in the model (see Table 1).

159
In the remainder of this paper we focus on the two costs over which GPEI has decision making control: 160 vaccination delivery and surveillance. Since a value of less than 1.0 implies an efficiency of scale, the 161 results demonstrate that there are economies of scale for these two costs.

162
In Figure 197 Our results are consistent with findings that there are indeed economies of scale in fields as diverse as 198 banking (Din et al. 1996;Wheelock and Wilson, 2017), education (Glass et al., 1995;Abbott and 199 Doucouliagos, 2003), real estate (24), and agriculture (25-27) -although Wilson and Carey (2004) found 200 that optimal scale may be locally-dependent and there is a risk of diseconomies of scale in certain Page | 10 201 situations (29). There may also be efficiencies of scope by deploying interventions for multiple diseases 202 in an integrated fashion, but we have not assessed that in this paper.

203
It is interesting to note that while large endemic and outbreak-prone countries have more predictable 204 budgets, this is not always the case for small and low-risk countries, particularly for surveillance costs.

205
The resulting flat-or negatively-sloped model fits do not lend themselves to obvious interpretation and 206 may be indicative that it is more difficult to estimate budgetary needs in low-intensity settings.

207
While it may be tempting to interpret these results directly, since these models are built on the natural-208 log scale we cannot draw definitive conclusions about how much cost is fixed vs. variable. This is 209 because the relationship in real terms is of the form . As a result of where α i = e β 0 + β i0 210 this structure, the slope is dependent on the intercept terms and , which implies that the rate of α i β 0 β i0 211 change in costs is not consistent across the range of x-values being evaluated. The intercept is also 212 outside the range of our dataset, so direct interpretation of the and is not appropriate.
With that in mind, the variation between countries described by their random slopes and intercepts 214 imply that there are important country-specific differences in how costs scale, in addition to the sub-215 linear scaling up of costs implied by the general model fits. The differences between sub-categories of 216 costs (OPV vs. vaccination delivery vs. surveillance) is also informative and suggest that multiple costs 217 components contribute to scaling efficiencies, which should be further investigated at lower levels of 218 detail than is currently possible from the aggregated budget data that GPEI publishes.

219
Overall, this analysis suggests a need to move beyond reporting "average per capita" or "average per 220 vaccine" as a standard measure, since they imply a linear growth in costs, without considering whether releasing them to be spent on other priorities. This is imperative for elimination programs that are in the 227 process of scaling up, such as regional measles elimination programs and the Central America malaria 228 elimination program.

229
It is also critical for GPEI to understand these dynamics so that they do not underbudget support for the 230 polio program as it winds down during the certifcation phase after eradication has been achieved.

249
Additionally, the GPEI's FRR documents are focused on the amount of external funding needed to run 250 the program, so they do not include in-kind costs provided by host countries nor self-funded programs.

251
To address this, we opted to exclude any country-year of data where there was a known bias, but this 252 may result in some selection bias and there may be some remaining unknown variance between GPEI's 253 budget and total costs.