This study used a retrospective, bottom-up micro-costing cost-minimization evaluation methodology taking the TB program perspective.^13–15 Cost-minimization analyses identify and compare costs of interventions while assuming equivalent health outcomes between types,¹⁶ thereby focusing solely on cost differences to inform program evaluation.¹⁷ This methodology has previously been used in analyses of TB treatment activities.¹⁸ Costs were calculated for all DOT sessions conducted by the DOHMH TB program during September 2014–August 2015. The cost data from 2016 were applied to resource utilization in the study period due to the unavailability of cost information. Labor and non-labor cost inputs were obtained from the DOHMH TB program except for gasoline costs, which were obtained through the New York State Energy Research and Development Authority.¹⁹ The analysis focused on cost per DOT session, rather than per patient, since some patients receive DOT from non-DOHMH providers for part of their treatment.

Labor costs per minute were estimated by taking the annual salary plus fringe benefits for each DOT worker and dividing by the minutes in 261 paid 7-hour working days (including holidays). DOT staff, who are trained, non-licensed professionals, range in job titles and salaries; therefore, we calculated the cost per session for each of the 17 DOT workers by multiplying cost per minute with each DOT type’s observation time. The duration of one LVDOT session was previously observed to be 5 minutes.⁵ We assumed the same duration for CDOT and one half of the LVDOT observation time (2.5 minutes) for RVDOT, since no real-time interactions occur during RVDOT sessions. For FDOT, we assumed the same observation time as for LVDOT plus one-way travel time. The per-session labor cost was the average of all the workers’ cost per session for each DOT type.

To calculate the FDOT travel times, we reviewed vehicle travel logs completed as part of standard program practice. Only DOT-specific trips submitted by DOT workers using designated vehicles were included. We randomly selected one weekday every 2 weeks (excluding holidays) from January 2014 through August 2014 and from January 2015 through August 2015 to obtain a representative sample of travel. The miles traveled and the travel times between locations were calculated. We excluded trips where neither the arrival nor departure site was a DOT location (e.g., relocation of vehicles for parking). We also excluded non-DOT trips (e.g., maintenance, hospital visits) and incomplete records—missing or illegible mileage information or data entry errors (e.g., negative mileage). Logs with missing travel times were excluded from the average travel time calculation.

Non-labor cost inputs included resources used for one or multiple DOT types. Costs associated with multiple program activities, which could not be separated from DOT, were excluded (e.g., administrative costs for operating clinics). Additionally, only costs directly incurred by the TB program were included in the analysis, which excludes other DOHMH and city government expenses (e.g., insurance). Table 1 shows the non-labor costs.

Based on the mileage logs, the average daily miles traveled for FDOT were aggregated by month and divided by the number of working days in that month. Fuel economy for hybrid sedans was assumed to be 30 miles per gallon of gasoline. The mean daily travel cost was the average daily distance in miles per gallon multiplied by the 2016 monthly per-gallon gasoline costs for the NYC metropolitan area.¹⁹ The number of daily FDOT sessions was the total number of one-way trips in that month divided by the number of person-days. Travel cost per session was the mean daily travel cost divided by the average FDOT sessions per day.

To calculate the total per-session costs, we defined a steady state of DOT implementation comprising equal use of all DOT types (i.e., 25% RVDOT, 25% LVDOT, 25% FDOT, 25% CDOT), given the average daily DOT capacity, which was the sum of observed DOT sessions across all DOT types divided by 247 working days. For all non-labor assets (e.g., webcams, signature pads, headsets, and tablets), we determined a 3-year useful life and a depreciation rate of one third. This rate, multiplied by the total asset cost, derives the annual cost assuming zero salvage value. Cost per day was the total cost of each non-labor input divided by 247 working days. Cost per session was the cost per day divided by the number of steady state sessions per day. The total non-labor cost per session for each DOT type was the sum of their costs of inputs per session.

Cost attribution analysis captures the changes in resource use as the parameters of the intervention are modified. We modified the proportions of DOT utilization and multiplied by the program’s total DOT capacity to identify the number of sessions for each DOT type. We multiplied these by the total per-session costs and examined the savings potential. We identified 35 implementation scenarios for the attribution analysis by adjusting the DOT proportions at quartile increments (0, 0.25, 0.50, 0.75, and 1) and choosing a proportion for each DOT type where the sum of the proportions equals one. We hypothesized that increased use of VDOT would significantly (P < 0.05) decrease the annual program cost.

We applied four bivariate generalized linear models to the scenario data to predict the average changes in cost outcome. We created four ordinal variables, one for each DOT type. Each variable indicated the quartile of utilization for that DOT type. For example, in a scenario where only RVDOT is used, the variable for RVDOT would indicate a value of ‘4,’ as 100% utilization describes the highest quartile. The variables for all other DOT types would equal ‘0.’ Three sensitivity checks were conducted on the re-attributed data set: 1) re-allocating non-labor inputs across VDOT types, 2) increasing RVDOT session time to match LVDOT, and 3) evaluating the economies of scale assumption.

Microsoft Excel (Microsoft Corp, Redmond, WA, USA) was used to generate random dates for assessing the DOT logs and to develop the data tables for the cost analysis. Regression analyses were implemented using Stata v14 (StataCorp, College Station, TX, USA).

This analysis was considered a public health program evaluation activity and did not require review by the DOHMH Institutional Review Board (IRB). This project was reviewed by the Centers for Disease Control (Atlanta, GA, USA) and determined to not be research requiring IRB review.

Of 34 randomized days, 33 had FDOT observations conducted and 2494 one-way trips were identified with 223 (8.9%) trips excluded due to incomplete records or identification as non-DOT trips. Of 2271 trips included in the mileage calculations, 1703 (75%) were retained to calculate the average FDOT travel time; the others were excluded due to incomplete or inaccurate times. The mean travel time was 27.43 minutes (interquartile range [IQR] 15–35). The average miles traveled per person-day was 38.4 miles and the average steady state per-session gasoline cost was $0.34.

The average annual salary was $47 353. Average labor costs per session were $2.91 for CDOT, $18.67 for FDOT, $2.91 for LVDOT, and $1.47 for RVDOT. Notably, $15.21 (81%) of the FDOT labor cost per session was attributed to travel time. Table 1 shows the non-labor cost components.

Table 3 summarizes the cost impacts in the attribution analysis. We started by assuming that one DOT type is unused and the others are held in equal proportions (33%). We then increased the use of that DOT type by 25% (one quartile) and estimated the new total program cost. The mean cost for DOT without using RVDOT was $422 938 (model 4). For each quartile increase in RVDOT utilization, we noted a 16% decrease (P < 0.001) in the total costs to $359 232 (95% confidence interval [CI] $332 743–$387 830), keeping the non-labor input allocations unchanged. By comparison, each quartile increase of LVDOT use produced an 11% decline (model 3; P = 0.01), and for FDOT, there was a 30% increase (model 2; P < 0.01). Using a uniform distribution of DOT delivery (25% for each DOT type, or 9510 sessions each), the total annual cost was $382 145 (or $10 using a per-session average basis), reflecting a 20% reduction in cost.

First, we re-distributed the patient phones equally across VDOT types, which did not statistically change the absolute cost outcome for RVDOT ($359 433, 95%CI $331 670–$389,521; P = 0.003). Second, we increased the per-session RVDOT observation time from 2.5 to 5 min, resulting in a smaller decrease, from 16% to 11% (P = 0.003), matching the LVDOT cost outcome. Finally, we reduced the program capacity for each DOT type by 10%, which increased the cost for a one quartile increase in RVDOT to $369 733 (an increase of 3%, or approximately $10 500). The sensitivity analyses demonstrated the robustness of RVDOT as a cost-minimizing DOT type.