Thread by @t0nyyates

@Samfr @iainmartin1 @DPJHodges There’s nothing wrong with what SAGE did, but that tweet is not really correct as it stands.

@Samfr @iainmartin1 @DPJHodges Forget the 'complexity' for a moment. That is a side issue. Think of the model, for simplicity, as y=y_1+x+shocks. y is infections, '-1' denotes one period before, x is a policy [like permitted contacts], shocks are good and bad luck.

@Samfr @iainmartin1 @DPJHodges SAGE are asking: what will happen to y, for different x's, ie different policy settings for permitted contacts. Try some different x's, and you get different time paths for y. When you see a time path for y you like, then you note down the x's and legislate for them.

@Samfr @iainmartin1 @DPJHodges But the answers are only interesting and to be taken seriously if the model can 'predict the future'. That is, if you feed in the y's as they are recorded in the data, and feed in the x's and the shocks, you should get the correct time path for infections.

@Samfr @iainmartin1 @DPJHodges If you don't - and this is an exercise you can do for the recent past by pretending that you haven't seen say the last year's data - then you know there is something wrong with your model, and to distrust the x's that you worked out in the very first exercise.

@Samfr @iainmartin1 @DPJHodges The amount of detail in the models SAGE are using tells you that they are trying very hard to make them realistic along some dimensions at least [save the economics]. And that they do want the quantitative nature of the lockdown measures simulated to be taken seriously.

@Samfr @iainmartin1 @DPJHodges Obvs because of necessary abstractions for a model to be tractable, coarseness in the data, imperfections in the data, you are going to make mistakes. There will be a cloud of uncertainty around both its policy advice and predictions.

@Samfr @iainmartin1 @DPJHodges The model outputs we have seen are not 'forecasts', by which strictly we mean not 'unconditional forecasts'. Which means forecast distributions for deaths, say. They are 'conditional forecasts' or what-if statements.

@Samfr @iainmartin1 @DPJHodges 'What is the probability distribution for deaths if we do x and if the omicron has severity z?'

@Samfr @iainmartin1 @DPJHodges But, neither have the what ifs been chosen entirely at random. They were what-ifs agreed along the chain of command from policy maker to scientific advisor to modeller. They were chosen because they lived in a relevant realm of possibility.

@Samfr @iainmartin1 @DPJHodges No-one has simulated 'what if omicron is like ebola?' because we know for sure that it isn't. Because the scenarios live in ranges that have enough probability mass to make them relevant for policy, they are in fact forecast like.