Time Horizon

Time Horizon

The time horizon is the span of time over which an economic evaluation captures relevant differences in costs and outcomes among the options being compared. It determines when the analysis begins, when it ends and which consequences are included.

This page explains how analysts choose a suitable horizon, why trial follow-up and treatment duration are not necessarily the same as the analytical horizon, and when a lifetime horizon may be needed. It also explains extrapolation, stopping rules, discounting, cycle length, budget-impact horizons and the checks used to decide whether a model runs for long enough.

What the time horizon controls

The time horizon sets the outer boundary of an economic evaluation. Costs and outcomes occurring after that boundary are excluded, even when they would otherwise be relevant to the decision.

Costs and outcomes should normally be evaluated over the same horizon so that the alternatives are compared consistently. Choosing different horizons for different quantities requires a clear methodological justification.

• The time horizon determines which treatment costs enter the analysis.
• The time horizon determines which health benefits and harms enter the analysis.
• The time horizon determines whether later recurrence, progression or mortality is captured.
• The time horizon affects the amount of extrapolation and uncertainty in the model.
• The time horizon can change incremental costs, incremental outcomes, ICERs and net benefit.

Why choosing the horizon matters

Intervention costs and health effects often occur at different times. A short horizon may capture an intervention's early costs while excluding later health gains or cost offsets, making the intervention appear less favourable.

A longer horizon can capture more relevant consequences, but it may also depend more heavily on uncertain assumptions. The preferred horizon is therefore the shortest period that still captures all material differences caused by the decision.

How analysts select a suitable horizon

The horizon should follow the decision problem and the expected timing of consequences. It should not be selected simply because a trial ended, a dataset stops, a treatment course finishes or a convenient number of model cycles has been reached.

Analysts should consider:

• The condition's natural history and expected duration.
• The timing of intervention and implementation costs.
• The duration of treatment.
• The expected duration and possible waning of treatment effects.
• Delayed benefits, adverse effects and complications.
• Recurrence, progression, recovery and retreatment.
• Subsequent treatment pathways.
• Mortality and competing risks.
• Future healthcare use and cost offsets.
• Patient, caregiver or cross-sector consequences required by the perspective.
• The applicable institutional reference case.
• The consequences remaining at the proposed endpoint.

A longer horizon is not automatically more appropriate. Its projected clinical pathways, survival, costs and outcomes must remain plausible.

Common types of time horizon

Different decision problems can support different horizons. The label attached to a horizon is less important than whether the selected endpoint captures the relevant differences between alternatives.

The selected horizon should be described in years or another explicit unit. Terms such as “long term” are not sufficiently precise without a defined endpoint.

When a lifetime horizon is appropriate

A lifetime horizon is appropriate when important differences in survival, chronic disease progression, recurrence or continuing quality of life may persist across the population's remaining lifetime. It follows the relevant population until nearly all material differences between the alternatives have occurred.

A lifetime horizon is not an arbitrarily large number of years. Its endpoint depends on the population's age, mortality, condition, model inputs and stopping rule.

• A mortality effect commonly supports considering a lifetime horizon.
• A lasting morbidity or quality-of-life effect may support a lifetime horizon even without a mortality difference.
• A short horizon may be sufficient when all material consequences resolve within a clearly defined period.
• A lifetime horizon still requires plausible survival and treatment-effect assumptions.
• A lifetime horizon does not mean that treatment effects must last for life.

Trial follow-up is evidence, not necessarily the endpoint

Trial follow-up describes the period during which outcomes were observed. The analytical horizon describes the period over which consequences must be considered for the decision.

A two-year trial can inform an evaluation that needs a much longer horizon when treatment affects survival, recurrence, disability or later resource use. Extending beyond the observed period requires extrapolation and makes the additional uncertainty visible.

• The observed evidence period shows what was measured directly.
• The extrapolated period shows what is projected after observation ends.
• Treatment duration shows how long the intervention is given.
• Treatment-effect duration shows how long its effect is assumed to continue.
• The analytical horizon shows how long the complete analysis runs.
• The stopping point shows where the model ends.

Ending an analysis at the end of trial follow-up assumes that no relevant differences occur afterward. That assumption should be justified rather than treated as automatic.

Treatment duration and effect duration are separate

Treatment duration states how long an intervention is administered. Treatment-effect duration states how long its effect is expected to persist, while the time horizon states how long the complete analysis continues.

These periods may end at different times. The model should represent them separately so that a long analytical horizon does not silently become an assumption of permanent treatment benefit.

• The treatment effect may stop when treatment stops.
• The treatment effect may persist for a defined period.
• The treatment effect may gradually wane.
• The intervention and comparator effects may eventually converge.
• The treatment effect may persist for the population's remaining lifetime.

Persistence and waning assumptions should be supported by clinical evidence, biological plausibility or clearly identified expert judgement. Important alternatives should be tested through scenario analysis.

How a model decides when to stop

A model needs a stopping rule that explains why extending the analysis further would not materially change its conclusions. Stopping after a round number of years is not sufficient unless that endpoint follows from the decision problem.

A stopping rule may be based on:

• Death of nearly all members of the modelled cohort.
• A clinically justified maximum age.
• Resolution of the disease episode.
• Convergence of the alternatives.
• Negligible remaining incremental costs and outcomes.
• Negligible change in incremental net benefit.
• A justified fixed endpoint required by the decision problem.

The analysis should report the population remaining at the endpoint and any residual costs, outcomes or net benefit. Material remaining consequences indicate that the horizon may need to be extended or more clearly justified.

How to check whether the horizon is long enough

Horizon adequacy should be checked by examining how the underlying results develop over time. Looking only at the final ICER can conceal continuing changes in incremental cost or incremental outcome.

• Cumulative incremental cost has become stable or remaining changes are negligible.
• Cumulative incremental health outcomes have become stable or remaining changes are negligible.
• Incremental net benefit has converged under the stated threshold.
• Very little of the cohort remains alive or in consequential health states.
• Material treatment effects have ended or converged.
• Remaining undiscounted consequences are negligible.
• The proportion of the result arising from extrapolation is reported.

A stable ICER alone is not a sufficient stopping rule. Ratios can appear stable even while both their numerator and denominator continue changing.

A worked example

Suppose an intervention has substantial initial costs but produces health gains over several years. Results calculated at one year can look very different from results calculated over five years or the remaining lifetime.

The one-year result includes most of the early cost but little of the later health gain. The lifetime result is more informative only if its projected survival, continuing effects and later costs are plausible.

The analyst should report how much of the lifetime result occurs beyond observed evidence. A favourable long-term result supported mainly by extrapolation should be presented with corresponding uncertainty.

How discounting differs from the time horizon

The time horizon determines whether a future cost or outcome enters the analysis. Discounting determines the present value assigned to a future quantity that has already been included.

Discounting cannot recover consequences omitted by a short horizon. Extending the horizon can increase the importance of discounting because more costs and outcomes occur further in the future.

• The horizon controls inclusion.
• The discount rate controls present-value weighting.
• The reference time determines when discounting begins.
• The timing convention determines where within a period a quantity occurs.
• Horizon and discount-rate scenarios may need to be tested together.

How time horizon differs from cycle length

Time horizon and cycle length are separate model settings. The time horizon is the full period covered by the analysis, while cycle length is the interval at which a state-transition model updates events, states, costs and outcomes.

A lifetime model might use monthly, quarterly or annual cycles. Shortening the cycle can improve timing accuracy, but it does not extend the analytical endpoint.

• Time horizon answers, “How long does the analysis continue?”
• Cycle length answers, “How often does the model update?”
• Both settings require justification.
• Changing cycle length may require a timing correction.
• Changing cycle length must not unintentionally change the total horizon.

How timing conventions affect implementation

The same nominal horizon can produce different results when models handle time zero or partial periods differently. Clear timing rules prevent unintended differences between alternatives and make calculations reproducible.

• State whether the initial time point is included.
• State whether costs and outcomes occur at the beginning, middle or end of a period.
• Explain how a partial final period is calculated.
• State when discounting begins.
• Explain how maximum age and background mortality affect the endpoint.
• Apply the same timing rules to all alternatives.

How alternative horizons should be tested

Alternative-horizon scenarios show whether the conclusion depends on a particular endpoint. They are especially important when the base-case result includes substantial extrapolation or when treatment-effect duration is uncertain.

• Test horizons when important benefits or harms occur far in the future.
• Test horizons when survival must be extrapolated beyond observed data.
• Test horizons when treatment-effect persistence or waning is uncertain.
• Test horizons when a material proportion of the cohort remains at the endpoint.
• Test horizons when incremental results have not converged.
• Test horizons when decision-makers require short- and long-term results.

Where possible, report several nested horizons using consistent assumptions. This allows readers to see how incremental cost, outcomes and net benefit accumulate rather than comparing unrelated model runs.

Extrapolation and uncertainty

A horizon longer than the evidence period requires assumptions about what happens after observation ends. These assumptions may concern survival, recurrence, disease progression, treatment effects, later treatments, resource use, prices and quality of life.

The analysis should identify which results are observed and which are extrapolated. It should also explain how projections were selected, tested and externally validated.

• Report the evidence cutoff and extrapolated period.
• Report the extrapolation method and its data source.
• Test clinically plausible alternative models.
• Test treatment-effect duration and waning assumptions.
• Compare predictions with relevant external evidence where possible.
• Report the proportion of key results produced during extrapolation.
• Treat structural uncertainty separately when parameter sampling cannot represent it adequately.

Why economic evaluation and budget-impact analysis may use different horizons

Economic evaluation and budget-impact analysis answer different questions. An economic evaluation generally seeks to capture all material differences in costs and outcomes, while a budget-impact analysis usually examines shorter-term expenditure and affordability for a specific budget holder.

The two analyses therefore do not have to use the same horizon. The shorter budget-planning horizon should not replace a longer economic-evaluation horizon when important health or cost consequences continue beyond it.

Using the time-horizon workbook

The Time-Horizon and Stopping-Rule Explorer allows users to compare results at one-year, five-year and lifetime horizons. It also separates treatment duration, treatment-effect duration and analytical horizon while displaying cumulative results, residual survival and the share arising from extrapolation.

The workbook includes checks designed to reveal whether the selected endpoint still excludes material consequences. The workbook supports exploration of the inputs, calculations, convergence checks and interpretation.

What transparent reporting should include

A transparent report explains both the selected horizon and why it is long enough. It also shows readers how much of the result depends on unobserved future events.

• Report the selected horizon and its unit.
• Report why the horizon is appropriate.
• Report the applicable institutional requirement.
• Distinguish the observed and extrapolated periods.
• Distinguish treatment duration and treatment-effect duration.
• Report the model stopping rule and population remaining at the endpoint.
• Report remaining incremental costs and outcomes.
• Report timing and partial-period conventions.
• Report discount rates and reference time.
• Report cumulative cost, outcome and net-benefit patterns.
• Report results at important alternative horizons.
• Report the share of key results arising during extrapolation.
• Report material consequences excluded after the endpoint.

Common mistakes

Time-horizon errors often arise when a convenient endpoint is substituted for an analytical justification. These mistakes can materially distort both the size and direction of an economic result.

• Automatically using trial follow-up as the time horizon can omit later consequences.
• Selecting a lifetime horizon without a stopping rule leaves the endpoint undefined.
• Assuming a lifetime treatment effect because the model uses a lifetime horizon confuses separate assumptions.
• Stopping while material differences remain produces an incomplete comparison.
• Using different horizons for costs and outcomes creates an inconsistent analytical boundary unless justified.
• Confusing horizon with cycle length changes the wrong model setting.
• Treating discounting as a substitute for future modelling excludes rather than reweights consequences.
• Testing adequacy using only the ICER can conceal continuing changes in incremental results.
• Using a budget-planning horizon for cost-effectiveness can omit long-term value.
• Selecting the horizon because it produces a preferred conclusion introduces bias.