Study design
We conducted a cost-of-illness (COI) study aiming to estimate the total healthcare expenditures used to diagnose and treat people with GW. We used an activity-based (micro-costing) technique to estimate the cost of diagnosis and each treatment option from the provider’s perspective, the facilities under the purview of the Peruvian Ministry of Health (MoH).
The micro-costing technique decompounds each service (i.e., diagnosis and treatment options) into the inputs and quantity required to provide it. We then find the best price for each input and multiply it by the amount needed. The sum of all inputs provides an estimate of the cost per service. Since we used the provider’s perspective for the costing analysis, only direct medical costs (e.g. drugs, materials, equipment, and physicians’ and nurses’ wages) were included [20].
Since we used a prevalence-based approach, the COI is determined by the product of the prevalent cases and the average treatment cost [21, 22]. The prevalence was obtained from a previous study conducted by Garcia et al [6]. In the following sections we describe how each treatment’s technique cost was estimated, as well as how we arrived at the overall average treatment cost.
Materials
We leverage the results found by Garcia et al [6] regarding the prevalence of GW, providers’ preferences for GW diagnostic methods, and distribution of cases across gender and type of case. Cases were categorized by physicians into “new” – no history of previous diagnosis, “resistant” – episode lasting longer than six months despite treatment, and “recurrent” – new case that appears within 12 months of previous episode. That study included physicians from six specialties: primary care physicians (including general practitioners and family medicine doctors), gynecologists, urologists, dermatologists, and infectious disease specialists.
To identify resource use for a typical visit, we developed a flow map of key activities completed during a visit (Fig. 1). Then we conducted a review of national guidelines [23] to identify the materials used in each activity according to protocol. Additionally, we updated and improve this information with eight in-depth interviews with physicians that participated in Garcia’s study [6]. These interviews were used to get further information about treatment practices, preferences for specific treatments, materials and equipment used in each procedure, duration of each procedure, and validation of the treatment algorithms.
In 2016, we conducted primary costing-data collection in Lima (coastal city and capital of Peru), Ayacucho (Andean region), and Iquitos (Jungle). The selection of sites was purposive. Each site represents a major region in Peru and therefore it allowed us to collect the most heterogeneous costing data, resources utilization (i.e., quantity of the resources used), and clinical practices to create robust estimates. In addition, it is coherent with the study design of Garcia, et al., so it preserves internal consistency.
We interviewed a total of nine administrative and logistics officers that provided the unitary costs of all drugs, materials, medical supplies, and equipment used for each treatment option. From each interview we obtained purchasing data that contained, for each input, volume of purchase and price paid, or directly unitary cost. The unit cost of disposable inputs (e.g., cotton, needles) relies on the assumption that in every session the entire input is used (i.e., no partition for reuse). To estimate the unitary cost of durable inputs (e.g., equipment, medical instruments) we used the depreciation method [24]. From the interviews with administrative officers, we obtained the total cost of the good, enquire about the rotation period (e.g., how often an equipment is changed, or infrastructure renovated), from which we obtained the useful life, which we finally used to estimate the depreciation cost per minute. Thus, the unitary costs were in the same unit as the duration of each activity.
Regarding human resources, we used the opportunity cost of the paid-time of the healthcare workers(HCW) [25]. We obtained salary data from the National Registry of Healthcare Personnel (INFORHUS) from the MoH to improve the precision of the wage estimates. We used this information to estimate the cost per minute per type of HCW and the estimated time per activity reported by the interviewed physicians to calculate the attributable costs to each treatment. We obtained the costs for six key HCW: receptionist, file staff, cashier, nurse, physician, pharmacist, whose regular activities are fairly differentiated and therefore we minimized the risk of overlapping. We used the information from the validate flow map (Fig. 1) to match activities with the HCW that most likely will perform them.
From this information we estimated mean and standard deviation (SD) of the unitary cost of each input. In all cases outliers were excluded if a value was a at least 3 times bigger than its peers. While a few inputs have a large variability (see supplemental material sheet “Costing data”) we decided in favor of the mean, instead of quantile-based metrics such as the median, because it easier to communicate, it allows for a more straightforward implementation of the probability-based sensitivity analysis, and the impact of each individual input is too small to bias the results. The SD captures the variability of the unitary cost given regional differences, purchase preferences, and others.
Analysis
First, we estimated the cost per session for a diagnosis appointment and each treatment technique. This is the sum-product of the resources’ amount needed to provide a service and its unitary costs. To account for the variability of the unitary costs we performed a Monte Carlo simulation [26]. We used a random number generator to obtain 1,000 estimates of each unitary cost based on a gamma distribution – the recommended distribution for cost data [27] – parameterized using the mean and standard deviation. Each draw from the distribution for all parameters is a simulation of the costing data, and therefore we obtained 1,000 simulations of the data. We estimated the final average costs 1,000 times and we were able to obtain the 2.5 % and 97.5 % percentiles of the distribution. Some parameters had no sample variability when only one source of information was obtained. In those cases, we assumed a SD of 1e-9 to conduct the simulation. Following the same process, we decomposed the cost per session into categories of costs: human resources, infrastructure, equipment, drugs, medical instruments, disposable materials, and public services. We present both the cost per session and the cost per cost category.
Second, in our study, the GW are compartmentalized by the combination of four groups given by the biological gender of the patient (2 categories), the type of case (3 categories), the physician’s specialty (5 categories), and treatment technique received (6 categories), resulting in 180 possible combinations. Each combination is called a compartment. The probability across compartments is not homogeneous and hence we sought to find the specific distribution of cases for each one. We used the information reported by Garcia et al [6] to estimate the distribution of cases given the combination of patient’s gender, type of case, and physicians’ specialty. The probability of no receiving treatment care once the warts have been detected vary across specialties but in all cases is negligible (see Garcia et al. [6]). The in-depth interviews provided us with the probability of choosing each treatment by physician specialty. We use these probabilities to estimate the final probability of each compartment. We report all the probabilities used for this analysis.
Third, we estimated the annual cost of treatment per patient as the product of (a) the cost of each session plus the cost of the diagnosis appointment, (b) the number of times the treatment was applied to each patient, dependent upon patients’ characteristics, type of case (new, recurrent, or resistant), and physicians’ preferences, and (c) the number of episodes within a year for recurrent cases. Thus, we obtained the annual cost per patient in each compartment. We report the average number of sessions per treatment, the overall number of episodes in recurrent cases by gender, and the final cost by treatment.
Fourth, we obtained an estimate of the average cost of diagnosing and treating a typical patient in one year as the sum-product of the annual cost per patient in each compartment and the associated compartment’s probability. Hence, this corresponds to a weighted average, where the weights are the probabilities of observing each combination of patient’ characteristics and physicians’ preferences. Given the distribution estimated for the cost per session (first point), we can estimate 95 % confidence interval for the annual cost of treating a typical case of GW.
Fifth, the COI of diagnosis and treating GW, is the product of the average treatment cost and the prevalence of GW. We calculate the point-estimate and range of feasible values of the COI. The point-estimate is the product of the mean values of prevalence and the average treatment costs. The lower bound is the product of the 95 %CI’s lower-bound for both the prevalence of genital warts and the cost of treatment; conversely, the upper-bound uses the upper-bound of the 95 %CI for both metrics.
Finally, the number of GW cases is based on the most recent estimation of population by age and gender, in 2017 the population of Peruvians between 18 and 60 years old was 18.4 million, 9.3 of them are males and 9.1 females [28].
All the costing data was collected in 2016 Peruvian Soles (PEN), but the results are expressed in 2019 US Dollars (USD) using a fixed conservative exchange rate of 3.3 PEN for each USD, and a yearly inflation of 2.5 %.