Human papillomavirus (HPV) vaccines

HPV is the primary cause of cervical cancer. Cervical cancer is the fourth leading cause of cancer burden among women worldwide, resulting in over eight million disability-adjusted life years (DALYs) globally in 2017 [4]. Licensed HPV vaccines include first generation bivalent and quadrivalent vaccines and a second generation nonavalent vaccine. Since the licensure of quadrivalent Gardasil in 2006, HPV vaccines have been introduced in 110 countries [5]. Universal vaccination could prevent 70–90% of HPV-related disease [6], yet coverage remains low in many low- and middle-income countries (LMICs) [7]. As of 2014, more than one-third of females ages 10–20 years had received the HPV vaccine in high-income countries, compared to only 2.7% in LMICs [7].

The World Health Organization (WHO) Director-General made a global call to action for scaling up cervical cancer prevention efforts in 2018 [8]. Subsequently, the WHO global strategy to eliminate cervical cancer was endorsed by the Seventy-third World Health Assembly in resolution WHA73.2 [9]. One target is 90% coverage of a full HPV vaccine sequence for girls, in addition to targets for high coverage of cervical cancer screening and treatment [10]. To achieve this high level of vaccine coverage, the strategy underscores the importance of a sufficient supply of affordable HPV vaccines, introduction of HPV vaccination in countries that have not yet adopted the vaccine, and increased quality and coverage of vaccine delivery. Gavi, the Vaccine Alliance approved a plan in 2017 to accelerate introduction of HPV vaccines into national vaccine programs. They aim to support the delivery of 25 to 35 million doses of HPV vaccines annually beginning in 2021 [11]. Given the strong potential to eliminate cervical cancer with HPV vaccination coupled with cervical cancer screening and treatment, HPV vaccines are a global health priority.

Data sources

Our data are from the Tufts University Center for the Evaluation of Value and Risk in Health registries through 2017 (Tufts registries). The Cost-Effectiveness Analysis Registry [12] contained 7287 studies that measure cost per quality-adjusted life year (QALY) gained. The Global CEA Registry [13] contained 621 studies that measure cost per disability-adjusted-life-year (DALY) averted (Fig 1). Between these two registries, there were 23,479 cost-effectiveness results across a wide range of interventions. Details on how these registries were compiled, including search strategies, data collection, and article review are published elsewhere [14,15]. This study complies with the Guidelines for Accurate and Transparent Health Estimates Reporting statement [16] [S1 Table].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flow diagram of study selection in logistic and meta-regression models.

https://doi.org/10.1371/journal.pone.0260808.g001

Data extraction, standardization, and mapping

Given that the Tufts registries are not compiled with the intention of being used for meta-regression analyses, additional data extraction, standardization, and mapping was necessary [S1 and S2 Appendices]. Missing data were extracted from articles in the Tufts CEA registries, including the age and sex of the target population, comparator and intervention descriptions, discount rates for costs and health outcomes, study time horizon, health outcomes targeted by the intervention, and study locations. Each ICER was mapped to at least one age group, sex, location, and cause that matched the categories modeled in the Global Burden of Diseases, Injuries, and Risk Factors 2017 (GBD 2017) study [3]. The age, sex, location, and cause were used to pull the total burden targeted for each intervention in the Tufts registries from DALY estimates from GBD 2017. The location and year were used to pull the gross domestic product (GDP) per capita for each intervention from GBD 2017. Each ICER for HPV vaccination was also mapped to one or more delivery platforms that were adapted from Jamison et al [17]. Five characteristics that consistently differentiated HPV vaccination interventions across articles were: vaccine cost, vaccine coverage, vaccine type (e.g. bivalent, quadrivalent), target population with respect to sex, and whether or not a booster dose was included in the vaccination schedule. When this information was not available in the Tufts registries, these data were extracted from the articles. The study currency and currency-year were used to convert all ICERs to 2017 United States dollars (US$).

The Tufts registries include a categorical variable that describes the comparator intervention, including placebo, no intervention, standard of care, and an “other” category. We defined the null comparator as no intervention (n = 1436, 79% of final sample), standard of care (n = 374, 20%), or placebo (n = 18, less than 1%). An exception was studies where the standard of care was screening for HPV infection (n = 86, 4.7% of ratios in the final sample); we defined a variable, “screening comparator” to estimate the effect of this higher standard of care for cervical cancer prevention. For this study, all ICERs compared to the “other” category were re-calculated such that they were compared to the null to facilitate comparisons between ICERs. To do so, we used data in the Tufts registries on total or per person costs and total or per person health benefits in the intervention and comparator groups. For ratios without these data reported in the registries, the necessary numerator and denominator data were extracted to re-calculate the ICERs.

Inclusion/Exclusion criteria

Our analysis focused on HPV vaccine interventions delivered at the health center platform, as this was the most common delivery platform represented in the published estimates (98% of all HPV vaccine estimates in the Tufts registries). We excluded articles that did not clearly state the discount rates, time horizon, intervention, or comparator. We excluded ratios if they were not calculated at the country level or if the ICER could not be re-calculated relative to the null comparator. We also excluded ratios if their corresponding intervention characteristics (vaccine type, target population, coverage, cost, and booster) were either missing or represented a category for which we had insufficient data. Of the 109 ratios we excluded (6% of final sample), the two most common reasons for exclusion were vaccines that were not bivalent or quadrivalent (n = 23, 21%), and discounting not specified (n = 16, 15%).

After applying the exclusion criteria, our sample included 616 ICERs from 76 articles in 182 countries (Fig 1), which we refer to as main ratios [S2 Table]. From these 76 articles, we extracted 1,218 one-way sensitivity analyses for five covariates: vaccine cost, vaccine coverage, vaccine for females only or both sexes, cost discount rate, and discount rate for health outcomes. The one-way sensitivity analyses were matched to a reference ratio, which was often a main ratio, where the ICER of the one-way sensitivity analysis differed from the ICER of the reference ratio by the value of only one parameter. Three additional ratios from one article were excluded, because the vaccine cost was not specified.

With a total of 1,828 ICERs, 1,692 had positive incremental costs and health outcomes, and 136 were cost-saving. Table 1 summarizes characteristics of the published studies. The majority of the 75 articles (62.7%) reported ICERs for the High income super-region, while only 8.0% and 6.7% of articles reported ICERs for Sub-Saharan Africa and South Asia, respectively. When the sensitivity analyses were included, 23.5% ICERs in the final sample were for the High income super-region, 28.1% for Sub-Saharan Africa, and 4.3% of ICERs for South Asia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics on sample of incremental cost-effectiveness ratios (ICERS) and articles reported in Tufts registries on human papillomavirus vaccines.

https://doi.org/10.1371/journal.pone.0260808.t001

Covariates

Three categories of covariates were used in the analysis: covariates that explain ICER variation across countries, covariates that explain ICER variation due to intervention characteristics, and covariates that explain bias in ICERs as a function of methods. Two covariates for true variation in ICERs across countries were: GDP per capita, and cervical cancer DALYs per person. GDP per capita measured the cost to the health-care system, among other things, including both the cost of the intervention and treatment costs saved when death or disability is averted. Five covariates explaining true ICER variation due to intervention characteristics were: vaccine cost measured as cost per dose multiplied by number of doses in full vaccine sequence, vaccine coverage, vaccine type (quadrivalent or bivalent), target sex (female or both males and females), vaccine coverage, and whether or not a booster dose was included in the vaccination schedule. Seven methods covariates were: perspective (health-care payer or societal/limited societal), cost discount rate, discount rate for health outcome, time horizon (lifetime or less than lifetime), outcome measure (DALYs averted or QALYs gained), comparator (screening or no intervention), and the proportion of model population with access to cervical cancer treatment (100% or less than 100%). DALYs and QALYs are not always comparable, because they are based on different methods [18,19].

Modeling approaches

Our statistical analysis can be divided into two components: predicting country-specific ICERs for HPV vaccination, and predicting the probability that HPV vaccination is cost-saving in each country. We combined both sets of results to predict adjusted ICER estimates that incorporated the cost-saving probabilities.

The statistical model and fitting procedures for the analysis of ICERs was conducted in five stages, and used a mixed-effects meta-regression framework (MR-BRT) [20]. This model included priors on all covariates and a study-specific random intercept. Each stage is described briefly below; for further information, see S3 Appendix.

In the first stage, we estimated priors for selected covariates by leveraging the fact that one-way sensitivity analyses differ in no unmeasured covariates from their reference analyses. Four covariates had a sufficient number of sensitivity analyses reported in published CEA to estimate priors using crosswalk models: vaccine cost, vaccine coverage, cost discount rate, and discount rate for health outcomes. We matched each sensitivity analysis with its corresponding reference analysis, and the crosswalk model estimated the difference in log-ICERs between sensitivity and reference analyses as a function of the difference between values of that covariate. We then constructed Gaussian priors for these covariates to use in all subsequent stages of the analysis with means and standard deviations equal to the crosswalk parameter estimates and standard errors from these crosswalk models.

In the second stage, we estimated a nonlinear response curve for log-GDP per capita by modeling the log-ICERs as a nonlinear function of log-GDP per capita. Log cervical cancer DALYs per capita was entered linearly into this model, in addition to the four covariates addressed in the first stage, and the priors calculated in the first stage were placed on the corresponding covariates. To make this stage more robust to model misspecification, we placed a spline ensemble on log GDP per capita. This model also used a robust statistical approach for outlier detection, and outliers trimmed at this stage were discarded from subsequent steps of the analysis. The nonlinear response curve estimated by this model was used to transform log-GDP per capita for use in subsequent stages of the analysis.

In the third stage, we selected additional covariates to include in the final meta-regression using a generalized Lasso approach for linear mixed effects models. The spline-transformed log-GDP per capita, log cervical cancer DALYs per capita, and the four crosswalk covariates, were pre-selected covariates at this stage, and the priors estimated for the crosswalk covariates were placed on those covariates. This process selected from eight additional candidate covariates: vaccine type, target sex, booster dose, perspective, time horizon, outcome measure, comparator, and the proportion of model population with access to cervical cancer treatment. Only one of these covariates, the booster dose, was not selected for inclusion in the final model.

In the fourth stage we selected the standard deviation of a Gaussian prior to apply to all regression parameters other than the intercept and the parameters for the four crosswalk covariates. To select a standard deviation, we fit a mixed effects meta-regression models with random intercepts by study, and priors on crosswalk covariates as calculated in the first stage. We normalized all other covariates and included Gaussian priors on those covariates, centered at zero and with a standard deviation that was constant across covariates. We varied this standard deviation using a grid-search and used 10-fold cross-validation to select the standard deviation that minimized the mean squared error for predicting values in the holdout set. We then converted the prior standard deviation back to the unstandardized scale for each covariate.

In the fifth stage, we fit a mixed effects model with a random intercept and priors on covariates determined in the first and fourth stages. This model included priors on covariates calculated in the first and fourth stages and the transformed version of log-GDP per capita, and random intercepts by study.

The probability that HPV vaccination was cost-saving was analyzed using a mixed effect logistic regression model. We sought to use the same covariates as the meta-regression model. The model included two country-level covariates, four intervention-level covariates, and four of seven method-level covariates: cost discount rate, discount rate for health outcomes, time horizon, and the proportion of model population with access to cervical cancer treatment. To account for between study heterogeneity, data were grouped by article, and a random intercept was calculated for each article [S4 Appendix].

The estimated meta-regression and logistic regression models were used to predict the ICER and probability that the vaccine was cost-saving, respectively, for each country as function of the two country-level covariates, and vaccine cost. The HPV vaccine does not have a single global market price. We used the cost of all required doses based on the 2017 cost per dose as reported to the WHO’s Market Information for Access to Vaccines (MI4A) [21] and aggregated by Linksbridge [22]. We used the United Nations Childrens’ Emergency Fund (UNICEF) price for the 57 Gavi-eligible countries [23], and Pan American Health Organization for 22 countries eligible for their Revolving Fund for supported countries and vaccines [S5 Appendix]. The three other intervention characteristics were held constant for our predictions: vaccine coverage of 70% (median across all studies), a bivalent vaccine, and target sex of females only. Our country predictions used a health sector payor perspective, 3% discount rate for costs and health outcomes, lifetime time horizon, DALYs averted as the health outcome measure, null comparator, and less than 100% access to cervical cancer treatment.

Adjusted ICER predictions were combined predictions from meta-regression and logistic regression models. The logistic regression provided the probability that HPV vaccines were cost-saving in each country. We subtracted each probability from one, and multiplied the resulting value by the mean predicted ICERs, lower bound of predicted ICERs (2.5^th percentile), and upper bound of predicted ICERs (97.5^th percentile) from the meta-regression analysis. This ensured that for countries with the highest probabilities of being cost-saving, ICERs were adjusted downwards (i.e. towards 0) more than for countries with lower cost-saving probabilities.

To place the results in the context of each country’s economy, we also report the ICERs as one-half, one, and three times GBD per capita. ICERs can contribute to a country-specific process for interpreting results and deciding whether or not to adopt the intervention [24,25]. In the absence of this process, one times GDP per capita, and three times GDP per capita are frequently cited thresholds for categorizing interventions are very cost-effective or cost-effective, respectively, in LMIC [26]. The opportunity cost of health-care expenditures is an alternative threshold, which corresponds to one-half times GDP per capita or less in the countries where it has been estimated [27,28]. To account for uncertainty, we compare the upper bound of the 95% uncertainty interval (UI) to these threshholds.

The meta-regression analysis was performed with an open source mixed effects package https://github.com/zhengp0/limetr [20]. ICERs for 195 countries with adjustments were predicted with Python version 3.0 (Python Software Foundation, available at http://www.python.org). The logistic regression analysis was performed using the open source software lme4 in R version 4.0.5 (Comprehensive R Archive Network, available at https://cran.r-project.org/bin/windows/base/). The maps are original content that was created with open source software ggplot2 in R package 4.0.5, and map boundaries from DAGM (Database of Global Administrative Areas, available at https://gadm.org/data.html).