key: cord-0719048-irj05llp
authors: Moss, Robert
title: Commentary on “Transparent modeling of influenza incidence”: Because the model said so
date: 2021-02-26
journal: Int J Forecast
DOI: 10.1016/j.ijforecast.2021.01.028
sha: 534190959a36aa122055181ad9a8ce211e40b298
doc_id: 719048
cord_uid: irj05llp

nan

With limited data and prior experience to guide national responses to emerging infectious diseases such as COVID-19, models provide a means of assessing the likely impact of unmitigated outbreaks and exploring the relative merits of different response options. This is but one example of how models are being used with increasing frequency to inform policy decisions that affect us all. If we are to trust such decisions, the justification should be ''the model said so, because . . . '', not ''because the model said so''. For example, having a clear rationale to justify measures put in place in Australia to mitigate COVID-19 was crucial in maintaining public trust and promoting compliance with these measures (Leask & Hooker, 2020; Seale et al., 2020) .

The growing popularity of ''big data'' and machine learning has resulted in increasingly complex models, which presents a very real challenge when seeking to explain why a model produces the results that it does. This has stimulated the development of ''explainable artificial intelligence'' (XAI) methods, which aim to allow humans to understand these models and to determine when they should, and should not, be trusted (Samek et al., 2017) . In contrast, simple models are much easier for a human to understand, and are capable of matching, or even outperforming, complex models (Sherden, 1997) . Katsikopoulos et al. (2022) advocate using simple heuristics as a baseline against which the performance of complex models should be evaluated, and argue that using simple rules can yield algorithms that are both accurate and understandable. They propose a simple ''recency DOI 

What quantities do we need to predict, and when is a prediction good enough to inform our decision? A model should only be one part of a decision support system that recognises the broader social and political context . A perfect prediction is rarely (if ever) required, let alone possible, and we may instead ask: how much better/worse is each model than the other candidates? A simple model can then be a useful baseline against which to evaluate other models, as Katsikopoulos et al. (2022) demonstrate here, and as used in the US CDC FluSight competition (Lutz et al., 2019) . Decisions may also involve actions that modify the very process we are trying to predict, such as deciding whether to make facemasks compulsory to help reduce the spread of COVID-19. This requires a model that can make predictions for two different scenarios: one where facemasks are worn by most people, and another where they are not.

Is the model capable of providing useful predictions that perform well against our chosen targets and can support decision-making? In the context of infectious disease outbreaks, the prediction lead time is critical to informing preparedness and response activities. The oneweek lead time of the recency heuristic is too short; a more complex model is required and, as Katsikopoulos et al. indicate, longer-term predictions of influenza activity are much more difficult (Wilke & Bergstrom, 2020) . One alternative is the SIR model family, which is also based on simple rules. SIR-type models can be used to make predictions with a sufficient lead time to inform infectious disease responses (Kramer et al., 2020; Moss et al., 2019; Yang et al., 2015) and to evaluate the impact of different decisions . However, fitting SIR-type models to available data can be challenging. In the absence of a ''perfect'' model, we can capitalise on the relative strengths of different models by combining them into an ensemble that can outperform each individual model (Chowell et al., 2020; Reich et al., 2019) .

Can we examine the model and identify under what circumstances we should trust it? We need models that make good predictions for the right reasons. As model complexity increases, models can become harder to understand and may ''over-fit'' to irrelevant features of the data on which they are trained; as a result, they may perform poorly on new data. But simple, easy-to-understand rationales can also lead us to draw erroneous conclusions! Consider the paradox of moderate infection control: for pathogens that cause more severe disease in older persons, limiting transmission can increase the burden of disease (Cohen & Lipsitch, 2008) . Combining models in an ensemble can mitigate the limitations of each individual model, but may also reduce the transparency of the results.

When decisions are influenced by predictive models, we must understand the strengths and limitations of those models. To be useful in this context, a complex model should outperform simple models and be transparent in doing so. To paraphrase George Box, all models are wrong but some are useful if we can verify that they yield sufficiently good predictions, and we can understand why. A model that cannot be explained should not inform decisions that require justification.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Real-time forecasting of epidemic trajectories using computational dynamic ensembles

Too little of a good thing: A paradox of moderate infection control

Transparent modeling of influenza incidence: Big data or a single data point from psychological theory?

Forecasting influenza in Europe using a metapopulation model incorporating cross-border commuting and air travel

How risk communication could have reduced controversy about school closures in Australia during the COVID-19 pandemic

Applying infectious disease forecasting to public health: a path forward using influenza forecasting examples

Coronavirus disease model to inform transmission reducing measures and health system preparedness

Accounting for healthcare-seeking behaviours and testing practices in real-time influenza forecasts

Explainable artificial intelligence: Understanding, visualizing and interpreting deep learning models

COVID-19 is rapidly changing: Examining public perceptions and behaviors in response to this evolving pandemic

Infectious disease pandemic planning and response: Incorporating decision analysis

The fortune sellers: The big business of buying and selling predictions

Predicting an epidemic trajectory is difficult

Forecasting influenza epidemics in Hong Kong