Ex ante modelling of interventions

Ex ante modelling of interventions estimates the potential effect of an intervention on an indicator before it is rolled out, whereas ex post evaluation analyses what happened once it was implemented and, if the design allows, estimates its observed effect. It applies when the decision to launch a health, nutrition, behavioural or policy intervention has not yet been made and you want to anticipate what change to expect. The difference lies in the timing and in the nature of the question. Ex ante modelling projects a scenario before resources are committed, integrating the available sources of information to estimate the effect the intervention could have. Ex post evaluation analyses data already collected on an intervention that has already taken place. They are sister services and complement each other: one informs the decision to roll out, the other evaluates the roll-out once it has happened. If the question is "what can we expect if we do this?", it is ex ante modelling; if it is "what effect did our intervention have?", it is ex post evaluation.

Yes: the core of the service is integrating whatever sources the case allows, such as national surveys, situation or behaviour reports, your own data and the scientific literature, as starting evidence for the estimate. It applies precisely to the usual situation, in which you rarely start from a single clean and complete dataset. Each source is brought in with a role that matches its quality, how recent it is and how close it is to the target population, so that the estimate makes use of what is available without treating all sources as equivalent. Extensive data of your own are not essential: a well-used national survey and effect sizes from the literature can support a modelled estimate with its assumptions made explicit. What determines the scope is not the number of sources but their quality and how close they are to the question. The less information you start from, the more weight the assumptions carry and the greater the uncertainty, and the result reflects that.

Not necessarily: checking two independent sources against each other is an option used when the case allows it, not the structure of the service. It applies when there is more than one independent source on the same effect. In that situation, the sources are checked against each other as evidence triangulation: convergence makes the estimate more plausible, and divergence shows which assumptions need revisiting. But many cases arrive with a single usable source, or with several that are not independent, and that does not prevent modelling: the estimate is still built from what is available, with the assumptions and the uncertainty made explicit. Triangulation is a valuable tool when there is material for it, not a requirement without which the work cannot go ahead. What defines the service is estimating the potential effect by integrating whatever there is. Checking independent sources against each other reinforces the estimate when the case offers it; it is not a condition of entry.

It is not enough: potential impact is not what is expected to happen with the intervention, but the difference from what could be expected without it, in the same population. This applies to all ex ante modelling, because projecting only the scenario with the intervention conflates its effect with the underlying trend that was already under way. Counterfactual estimation compares the expected scenario with the intervention against the expected scenario without it, under explicit assumptions, and so estimates the incremental change associated with the intervention under those assumptions. If an indicator was already improving on its own, attributing all of that improvement to the intervention overstates its effect; the counterfactual guards against that. One honest limit is worth stating: this counterfactual is a projection under assumptions, not the causal certainty of an experiment. Its credibility depends on how reasonable those assumptions are, and they are stated explicitly. What you receive is an expected incremental effect, separated from the underlying trend, not a promise of what will certainly happen.

Yes: the effect of an intervention depends not only on its nominal coverage but also on adherence, compliance and actual utilisation, and all of that is modelled. It applies to any intervention where what is planned is not what finally reaches the target population, which is the case for many interventions in practice. The losses between what is planned, what is rolled out and what actually reaches people and is used are modelled, so that the estimated impact reflects realistic implementation conditions rather than an optimistic assumption of full coverage. An estimate that assumes everyone receives and makes full use of the intervention produces an inflated figure that reality will not match. Building in those losses, which form the cascade between the planned intervention and the real effect, is what makes the estimate defensible to whoever will have to answer for the result. Estimating the effect under realistic coverage is not pessimism; it is honesty about how interventions work.

An estimate with its uncertainty, not a single number with no margin: bounding the uncertainty of the estimate is a central part of the service. This applies to all ex ante modelling, because a point estimate with no margin conveys false precision about something that, by definition, has not happened yet. Uncertainty is analysed explicitly, using Monte Carlo simulation or other uncertainty propagation approaches when the case requires it, so that the result shows a range of effects consistent with the assumptions and sources available, not just the central value. This also reveals which assumptions weigh most on the result and where more information is worth obtaining before implementation. An effect presented as an exact figure hides precisely what a funder or a decision-maker needs to know: how much it can vary and what it depends on. Delivering the uncertainty with clear bounds is what turns the estimate into an honest basis for a decision rather than a numerical promise.

Ex ante modelling of interventions is a high-level service that can draw on tools such as synthetic population generation, scenario simulation, time series analysis or stratification, but none of them is a required part of the service: they are brought in when the case calls for them. This is worth clarifying because it is sometimes assumed that ex ante modelling always means simulating many scenarios or generating a synthetic population, and that is not so. One case may be resolved with a single national survey and a point estimate under one scenario, with no need for a synthetic population or a battery of scenarios; another, more complex case may well require those tools. The choice depends on the case in hand, not on a fixed package of techniques. Several of these tools are also services in their own right and can be commissioned separately when that is the main need. What defines ex ante modelling is the question, the effect of the intervention on the indicator before roll-out, not a closed set of methods.

No: ex ante modelling of interventions estimates what can be expected under explicit assumptions. It does not prove that the intervention will work, nor does it replace evaluating the intervention once it has been implemented. This limit is worth understanding before using the estimate to decide, because it is an informed projection, not a certainty. The estimate is built on the available evidence and a set of stated assumptions. Its value lies in turning an expected impact into a defensible estimate, with its uncertainty, so that the decision rests on more than intuition. But the actual effect only becomes known once it can be observed after roll-out, which is the task of ex post evaluation. The client decides on and rolls out the intervention and commits the resources; the service estimates, models and quantifies the potential effect that informs that decision. What you receive is an estimate that helps you decide better before committing resources, not an assurance of the outcome.