We used our previously described FRED ABM of Allegheny County, Pennsylvania,^6–10 which comprises a total population of 1,164,880. A collaboration of investigators (P.C., S.T.B., W.D.W., and B.Y.L., currently at RTI, the Pittsburgh Supercomputing Center, and Johns Hopkins, respectively) developed the initial ABM in 2008, which was used during the 2009 H1N1 pandemic to help inform policy making at county through national levels. After which, this initial ABM was transcoded and served as the basis for the subsequently developed FRED system by investigators at the University of Pittsburgh, Carnegie Mellon University, and Pittsburgh Supercomputing Center.¹⁰ The ABM represents each individual in Allegheny County with a computational agent. Just like a person, each agent has a set of characteristics (eg, age, sex, and employment status) and is assigned to a particular household, school (if the agent is of school age), and workplace (if the agent is employed). During the simulation, the model proceeds in discrete 1-day time steps. Each simulated day, the agents move among virtual representations of the county's households, schools, workplaces, and communities, where agents interact with each other. Agents' ages and locations are based on 2010 US census data.

At any given time, an agent is in one of 4 mutually exclusive influenza states: susceptible (S), exposed (E), infectious (I), or recovered (R). An agent in the S state has a probability of contracting influenza when contacting an agent in the I state. The transmission probability depends on the agents' ages and location and the reproductive rate (R₀) of the epidemic (Appendix Table A1).^11–13 If the agent in the S state contracts influenza, he/she moves into the E state and remains for the duration of the latent period (distribution with a mean of 1.2 d). Once the latent period elapses, the agent transitions to the I state, in which he/she can transmit influenza to others and remain for the duration of the infectious period (distribution with a mean of 4.1 d). Once the infectious period elapses, the agent moves into the R state in which he/she is not infectious and immune to infection, and where they remained for the rest of the simulation. Vaccinating an agent in the S state had a probability (ie, the vaccine efficacy) of moving the agent to the R state 2 weeks after the vaccination occurred (to represent the time lag in the onset of immunity postvaccination). In addition, some agents could have preexisting immunity; these individuals started the simulation in the R state.

For this study, the primary output from the ABM was the number of influenza cases in each age group on each simulated day.

We developed FluEcon, our economic model, in Microsoft Excel (Microsoft, Redmond, WA) and utilized a Crystal Ball add-in (Oracle, Redwood City, CA), following our previous models,^14–19 to translate influenza cases from the ABM into health outcomes and corresponding costs from the third party payer and societal perspectives. Appendix Table A2 shows the model inputs. Each influenza case from the ABM had a probability of being symptomatic. Each symptomatic case then had probabilities of seeking ambulatory care, being hospitalized, or dying from influenza, each associated with a corresponding cost. Thus, the outcome of each case determined the costs accrued. All costs and probabilities were age specific when available.

Third party payer costs included all direct costs of illness (eg, ambulatory care visit, hospitalization) derived from nationally representative data sources. Societal costs included direct and indirect (ie, productivity losses due to absenteeism from work or school and mortality) costs. The median hourly wage for all occupations estimated productivity losses and assumed an 8-hour work day. A death resulted in accruing the net present value of that person's remaining lifetime earnings, based on their life expectancy.^20,21 A 3% discount rate converted all costs into 2013 $US.

Health effects were measured in quality-adjusted life-years (QALYs). Each influenza case accrued QALY values based on their age-dependent healthy QALY value attenuated by influenza's utility weight (either hospitalized or not) for the duration of their symptoms (if not hospitalized) or hospitalization (if hospitalized). For example, a child has a healthy QALY over the course of a year of 1, if he/she contracts influenza and are not hospitalized, we use the published utility weight for influenza without hospitalization (0.659) to attenuate the healthy value, resulting in 0.659 QALYs (1×0.659) for the 7-day symptom duration. By contrast, an adult has a healthy QALY value of 0.92 over a year, if he/she contracts influenza and is hospitalized, they accrue 0.473 QALYs (0.92×0.514, the published utility weight for influenza with hospitalization) for the hospitalization duration. We also considered QALYs from vaccination side effects.

When there is no vaccination, varying R₀ from 1.2 and 1.6 results in approximately 229,000 more cases generating an additional $21.0 million in direct costs, $52.0 million in productivity losses, and 807 additional QALYs (Table 1). As the amount of preexisting immunity in the population increases, the difference in the number of influenza cases and their related costs and health effects between R₀ 1.2 and 1.6 decrease (Table 1). This highlights the variability from season to season (as there is no one most common influenza season) and how these parameters relate to epidemiologic and economic outcomes.

Tables 1 and 2, and Figure 2 show how the current timing of vaccination reduces the epidemic curve and impacts economic outcomes. The impact of vaccination depends on the transmission dynamics of the influenza season, on R₀, amount of preexisting immunity, and season peak. Vaccination can lead to a 1.9- to 14-fold decrease in attack rate, depending on these transmission dynamics. The current vaccination timing averted fewer cases when the season peaked in December (29,187–252,106 cases averted), increasing with R₀ values and decreasing with higher levels of preexisting immunity.

The mean total vaccination cost was $3,451,555 (SD = $325,879) for Centers for Disease Control and Prevention vaccine prices and $4,810,976 (SD = $473,959) for private sector vaccine prices. When combined with the direct costs, third party payer costs can be determined for vaccine price points. Societal costs for both vaccine price points can be calculated as the respective vaccination costs plus direct costs plus productivity losses. For example, given the current timing, with private sector vaccine costs, the third party payer costs were $16.0 million and societal costs were $52.3 million (December peak, R₀ 1.2).

When the influenza season peaks in February, shifting vaccination for all ages earlier can reduce the attack rate from 13.7% (current schedule) to 12.8% regardless of when vaccination was complete (ie, attack rate same for shifts to September, October, or November) with R₀ 1.2 and no preexisting immunity. Earlier vaccination could avert ~34 QALYs, $0.6 million in direct costs, and $2.1 million in productivity losses (Table 1).

Table 1 can be used to determine the costs and QALYs averted by shifting the timing of vaccination from the current schedule and to determine the cost per person countywide (divide by 1,164,880). For example, completing vaccination by the end of October can avert 10,093 cases, $0.7 million in direct costs (saving $0.6 per person countywide), $2.2 million in productivity losses (saving $2 per person countywide), and 37 QALYs (R₀ 1.2; no preexisting immunity). These saving thresholds represent the investment ceilings that can be made to promote earlier vaccination (ie, shift vaccination) and remain cost-neutral.

When the influenza season peaks earlier (eg, December), it is more important to vaccinate earlier. Shifting the timing of vaccination had a greater impact with a December peak. A September shift provided the greatest gains, decreasing the attack rate from 15.6% to 12.8%, averting an additional 33,139 cases, $2.2 million in direct costs, $7.2 million in productivity losses, and 120 QALYs (R₀ 1.2, no preexisting immunity). This decreased with the amount of preexisting immunity in the population so that at 40%, shifting vaccination timing had little impact. A September shift yielded the greatest change, averting 73 cases, $5295 in direct costs, and $17,167 in productivity losses (December peak, R₀ 1.2). At an R₀ of 1.6, shifting vaccination could avert ≤11,687 cases and save ≤$680,627 in direct costs, and ≤$2,415,913 in productivity losses.

Figure 2 shows how earlier vaccination impacts the epidemic curves for both peaks and R₀ values with no preexisting immunity. Earlier vaccination shifts the curve later in the influenza season and reduces the peak; this is most prominent with a December peak and R₀ 1.2, whereas earlier vaccination has little impact with a February peak and R₀ 1.6 as the curves are stacked on top of each other. Table 2 provides the age breakdown for these epidemic curves for R₀ 1.2 and no preexisting immunity, showing the month for completing vaccination had little impact with a late peak (February).

Figure 3 shows the cost-savings per vaccinated person shifted earlier for both peaks and R₀ values. These savings represent the amount that could be invested per person to promote earlier vaccination and remain cost-neutral. Vaccinating everyone by the end of September (51,461 earlier vaccinations) saved $12–$72 and $41–$208 per shifted person in direct costs and productivity losses, respectively.

When only vaccinating children by the end of September (everyone else remains at current schedule), there were 147,670 total cases (11,366 fewer cases than the current schedule), saving $0.6 million in direct costs, $2.3 million in productivity losses, and 42 QALYs (February peak, R₀ 1.2, no preexisting immunity). Compared with the other vaccination shifts, it would be better to vaccinate only children by the end of September with a February peak, but vaccinating everyone by the end of September or October would be better than only vaccinating children by the end of September with a December peak (R₀ 1.2). However, at R₀ 1.6, vaccinating everyone earlier, regardless of when, resulted in fewer cases and incurred lower costs than only vaccinating children earlier (December and February peaks).

Table 2 shows the epidemiologic and economic impact of increasing vaccination coverage compared with the current timing and earlier vaccination. Increasing coverage could result in attack rates of 12.8%–27.2% (5% coverage increase) and 11.6%–26.1% (10% coverage increase) for R₀ 1.2–1.6 (February peak). With a February peak (R₀ 1.2), a 5% increase in coverage among all age groups would be comparable with earlier vaccination (similar attack rates and number of cases), however, a 10% vaccination coverage increase would be better than any of the shifts (Table 2). Results were similar for R₀ 1.6. With a December peak, vaccinating everyone earlier would be better than a 5% or 10% increase in coverage (R₀ 1.2 and R₀ 1.6). Vaccinating children by the end of September would be better than increasing vaccination coverage by 5% at an R₀ of 1.2 (regardless of the season peak) and R₀ of 1.6 with a December peak. However, vaccinating children by the end of September would only be better than a 10% increase in coverage at R₀ 1.2 (regardless of peak).