The AHEAD model simulates the course of AD according to the patient's characteristics at a given point in time 1322. This is done by forecasting the time until each patient with AD requires full-time care (FTC) – the consistent requirement for a significant amount of caregiving and supervision for the greater part of the day, regardless of the locus of care and who the caregiver is. These forecasts are derived by estimating the failure-time curves that result from the corresponding time-dependent hazards. The hazards are calculated using equations that depend on the presence of extrapyramidal symptoms (EPS), the presence of psychotic symptoms, age at onset, duration of illness and cognitive status. The model also forecasts survival, which depends on sex, EPS, duration of illness and cognitive status. The required equations were derived from published long-term follow up data of patients with AD 26.

The model considers four states: 1) not yet requiring FTC (pre-FTC), 2) requiring FTC but still living in the community (FTC community), 3) requiring FTC and institutionalized (FTC institution), or 4) dead. The model uses individual patient data as inputs into the predictive equations, and the individual results are then aggregated into an overall forecast for the entire cohort. Comparisons between treatment scenarios are achieved by re-computing the failure-time curves for the same patients under each treatment scenario.

The cognitive portion of the Alzheimer Disease Assessment Scale (ADAS-cog) 27 is used as the measure of cognition in this model as cognitive outcomes of trials are typically reported using this measure. Cognitive status after six months (C₆) in the absence of pharmacologic treatment is forecast using a regression equation based on baseline cognitive status (C₀). The equation, derived from patient-level data on untreated participants in three clinical trials 282930 was C₆ = -0.54+1.10 C₀ (the 95% confidence intervals for the coefficients were -1.94 to 0.97 for the intercept term, and 1.05 to 1.15 for the slope term). An increase in the score reflects deterioration in cognitive functioning.

Along with the duration of FTC and economic outcomes, the model calculates quality adjusted life years (QALY) remaining by assigning quality "weights" to each state. Values of 0.603 for pre-FTC, 0.338 for FTC regardless of location of care, and 0 for death were derived from published data 31. A single value was used for FTC as the study on which quality weights were based found that setting of care did not have an independent effect on outcomes 31.

If one treatment is estimated to increase health benefits but also the costs relative to the next most effective treatment, the model calculates the incremental cost-effectiveness ratio in terms of the net cost per QALY gained (both discounted). Pair-wise comparisons, even for treatments otherwise dominated, are also reported here since these may be relevant in settings where the dominant treatment cannot be used.

Results are forecast over ten years from the start of treatment because by that time all survivors require FTC. A rate of 3% per annum was used to discount costs and health benefits occurring after year one.

To provide a common population of patients for these analyses, the clinical and demographic characteristics of participants in a recent clinical trial 30 were used. Among these 918 patients, who should be broadly representative of those for whom the drugs are indicated, 64.9% were female and the mean age was 77.2 years. ADAS-Cog scores ranged from 10 to 62, with the mean ADAS-cog score at baseline being 28.9 and the first and third quartiles of the distribution being 21 and 36, respectively. Psychotic symptoms were exhibited in 32.4% of the patient population, 7.5% had EPS, 8.9% had early onset of disease, and the mean duration of illness was just over four years. In the absence of pharmacologic treatment, it was estimated that after six months, the average ADAS-cog score would deteriorate to 31.3.

Computations for each treatment were based on the cognitive effects reported in the Cochrane reviews – formal meta-analyses of the randomized clinical trials assessing efficacy and safety of these drugs. For each drug and dose, the reviews reported point estimates and 95% confidence intervals (CI) for the weighted difference in mean cognitive change comparing the treatment to placebo using intention-to-treat analysis with last observation carried forward (LOCF) methods used for patients dropping out of the trial (Table 1). While it is possible that the LOCF method may slightly overestimate the cognitive effects of treatment if withdrawals from the trial are higher with treatment than without (as patients cognition, on average, should decline between the time of withdrawal and the end of the study), these were the only data available for all three active treatments and the same methods were used for all.

These estimates were used to determine the change in cognition expected with each treatment and dose over six months for each patient in the model population. Two approaches were taken to make these predictions. For the main analyses, the most direct approach was used: the Cochrane estimated mean difference from placebo (ΔC) was applied to the intercept term in the regression equation that describes progression without treatment: C₆ = (-0.54+ ΔC) + 1.10 C₀. This implies that treatment has a constant absolute effect on cognition, regardless of baseline ADAS-cog – thus termed the 'constant effect' analysis.

To consider possible dependency of the treatment effect on the baseline disease severity 3233, a second, 'variable effect", approach was used in sensitivity analyses. For this approach, a term was added to the regression equation reflecting the interaction between treatment (T) and baseline ADAS-cog score: C₆ = -0.54+1.10 C₀+ β·T·C₀. The coefficient β was estimated for each treatment by solving the equation at the mean baseline ADAS-cog scores reported for each treatment arm in the trials 282930, 343536: β = ΔC/ (Table 1).

In both the constant and variable effect analyses, the treatment effect was applied only to patients who were predicted to complete the first six months of therapy. For those predicted to withdraw from therapy, the rates of cognitive change expected without treatment were applied. The likelihood of completing the first six months of treatment was predicted for each patient by adjusting the base withdrawal rate incorporated in AHEAD according to a treatment-specific factor derived from the Cochrane reviews by calculating the ratio of treatment to placebo withdrawals (Table 1). For example, if treatment were associated with a relative risk of withdrawing equal to 2, and the base withdrawal risk for a patient were 10%, then with treatment, the withdrawal risk would be 20% (2 × 10%). The overall base withdrawal rate in AHEAD was 16%.

Although there is evidence that these drugs may have an impact on psychotic symptoms 37, this aspect was not included in these analyses because those data were not reported in the Cochrane reviews for all treatments.

Evidence that treatment affects survival is not yet available so mortality was assumed to be equal for all therapeutic options. This equivalence was ensured by calculating mortality with all three treatments using the ADAS-cog scores had the patients remained untreated.

Costs are reported in 2001 Canadian dollars and are determined from the perspective somewhat broader than that of a comprehensive payer. For example, informal costs such as in-home personal care were included. Costs were estimated based primarily on resource use data from the clinical assessment portion of a follow-up survey of the Canadian Study of Health and Aging 38 and unit costs from the province of Quebec using methods described elsewhere 13. The estimates were $432 per patient per month before FTC is required, $903 monthly for FTC in the community and $3,931 for a month of FTC in an institution. Based on published data, it is expected that that almost three-quarters (73.5%) of patients requiring FTC will be institutionalized 313. No event cost was assigned to the transition to FTC or to death. The daily cost of each cholinesterase inhibitor was set at $4.59. Daily treatment cost was independent of dose, reflecting current pricing practices for cholinesterase inhibitors in Canada. Patients predicted to discontinue treatment were assumed to consume three months of drug. It was conservatively assumed that to maintain the cognitive benefit obtained, all those who do not withdraw in the first six months, require continued treatment until FTC is needed. Once this happens, pharmacologic treatment was assumed to end for all patients. Treatment discontinuation prior to needing FTC was not considered, although one might expect that in addition to no longer incurring treatment costs, benefits attained through treatment prior to discontinuation may be lost over time.

Sensitivity analyses were performed on key inputs and assumptions. To reflect the uncertainty in the relative cognitive effect of each treatment, univariate sensitivity analyses were run using the upper and lower 95% confidence interval bounds for each drug's treatment effect. Health utility and cost inputs were varied and a potential survival benefit based on the cognitive effect was explored. Sensitivity analyses were also conducted assuming lower discontinuation rates, as the Cochrane reviews present evidence of modified dose-escalation regimens for each of the three therapies leading to significantly fewer drop-outs over the trial periods. Finally, all analyses were rerun using discount rates of 0%, 5% and 10%.

Apart from univariate analyses of each factor, multivariate probabilistic analyses were also carried out. The ranges over which each factor was changed in the multivariate analyses are displayed in Table 2.

In the constant effect analysis, all treatments reduce the time FTC is required relative to no treatment, but only the higher dose of donepezil and both doses of galantamine reduce costs (Table 3). As these three treatment options also provide greater effects, they dominate the others (i.e., have greater health effects and lead to lower overall costs). Both doses of galantamine also dominate the higher dose of donepezil. The only cost-effectiveness scenario amongst active treatments exists with donepezil 5 mg, which has a superior health effect compared to rivastigmine 6–12 mg, but is more expensive, with incremental benefits attained at a cost of over $150,000 per QALY gained.

The confidence intervals of the withdrawal adjustment factors for the lower doses of all three cholinesterase inhibitors included a value of one (i.e., withdrawal no different than with placebo), and there is some evidence that slower titration regimens might result in lower discontinuation rates even for the higher doses of these drugs. An analysis was therefore undertaken assuming that the withdrawal factor was one for all treatments. The only resulting change in position of the treatments was that higher dose rivastigmine dominated the lower dose of donepezil under these conditions.