Using nowcasting, forecasting and scenario analysis to understand epidemics

One of the key aims of epidemiological modelling is to try and predict how an epidemic might progress. For short-term predictions, modellers use a process known as forecasting. However, for medium- and long-term projections, predictions cannot be made with any certainty. A different approach, scenario analysis, is required to gain insights into what might happen to an epidemic over a longer time-frame and the impacts any interventions may have on the course of the epidemic.

This blog post outlines the differences between forecasting and scenario analysis and the various applications of the two approaches, using influenza as an example. 

Photo by Patrick Assalé on Unsplash

Every year, seasonal influenza is responsible for thousands of hospitalisations and deaths. It is important for policymakers to have information about how many cases of influenza there are likely to be each winter. This enables resources to be efficiently deployed and mitigations put in place to try and minimise the numbers of hospitalisations and deaths. Epidemiological modelling plays a crucial role in predicting how an infectious disease such as influenza is likely to spread through a population and what the impacts of mitigation measures may be.

Influenza surveillance data are important but there can be a lag in influenza cases being detected, diagnosed, reported to a surveillance system and then the relevant data being analysed, interpreted and published. To overcome this issue, models can be used to estimate levels of influenza in the population at the current time; this is a technique known as nowcasting.

Forecasting is used to make predictions a few days up to a few weeks into the future. However, once you move beyond a few weeks into the future, forecasting is no longer possible as there are too many uncertainties. Forecasting produces absolute numbers that are accompanied by a range known as prediction intervals. For example, the US Centers for Disease Control and Prevention (CDC) produce forecasts that predict the number of hospitalisations due to influenza in the next week. Other models can be used to forecast hospital admissions for influenza for up to two to four weeks ahead.

Whereas forecasting is used to estimate what ‘will’ happen in the near future, scenario analysis can be used to guide decision-making over the longer term using ‘what if’ scenarios. In this type of approach, a model is developed to compare scenarios involving different policy interventions and explore what changes a particular policy may result in.

For example, the seasonal influenza vaccine is developed based on predictions about the influenza strains that are likely to be circulating in the upcoming influenza season. However, there may be sudden, rapid changes in the predominant strain of influenza in circulation. Scenario analysis models can be used to explore what the implications of this may be if the vaccine remains highly effective against this strain or if the vaccine is ineffective against this strain. Other scenarios that could be compared include various values for the infectivity of the predominant influenza strain in circulation and the levels of prior immunity to that strain among the population.

The inputs used to model the scenarios are based on a series of assumptions. In some cases these assumptions are based on known, measured values, such as the hospitalisation rate for a given influenza strain. Some assumptions must be explored through scenario analysis, for example what proportion of infected individuals will avoid contact with others to avoid passing on their infection. There are other assumptions that must be changed during the modelling of scenarios, to take into account the changes that may occur as a result of policy interventions. For example, if people are asked to avoid meeting elderly or immunocompromised relatives, the impact this may have on influenza hospitalisation rates must be estimated as part of any scenario analysis. Behavioural changes in response to policy are particularly difficult to model.

As already mentioned, forecasting often depicts absolute numbers together with prediction intervals. Scenarios, however, are generally presented in the form of a graph showing the central estimate with a confidence interval that indicates the level of uncertainty around that central estimate. The wider the confidence interval, the greater the degree of uncertainty around the central estimate for a given scenario. Often, multiple scenarios are plotted on the same graph to aid comparison between them.

At Model Health Ltd we are experts in a wide range of epidemiological modelling techniques, including nowcasting, forecasting and scenario analysis – contact Lisa to find out more!