[This Article is written on 6th May 2020, with data up to 5th May. Updates will be made as data becomes available ]
In my last post, I answered some important questions about Covid-19 in Pakistan, after introducing statistical uncertainties in the data and running 1000 simulation on it. However, for the end date of the epidemic, I did not give exact figures. Here, in this post, I am fitting a new kind of model that can predict the down trend of an epidemic based off of new 500 simulations.
However, before I give you the prediction, I wan’t you to understand that models are built from data and assumptions. If underlying assumptions start changing or if data starts’ behaving’ in a different manner, model will change. And given that this date is so far ahead in time, this can very well happen. Things like, change in lockdown, availability of tests, human behavior, virus’ behavior etc. can all affect it. However, even with their given uncertainties and shortcomings, models are still far superior than guess work, trends and biased judgement of media pundits (which we see in the news all the time).
So here is the prediction; Based on the data so far, The epidemic will start declining somewhere between 200 to 550 days from the day of first reported case. Which means:
We can expect it to start declining from anywhere between November of this year to September of next year. It will take a long time to be near to complete end. Roughly from 1 year to 3 years.
Following is the predicted trend(the data on y-axes here are daily cases, unlike cumulative cases in the previous model).
Note the thickness of the green area. This is the uncertainty and the region of all possible scenarios based on 500 simulations.
Differences from Previous Model
3 main differences from the original model are set in place to make these predictions,
Firstly,the derivative of the model below is fitted
Secondly, more data points have become available since 1st May (time of write-up of last post)and these are also used.
Thirdly, Poisson error imputations are now done directly in the data rather than from the fitted model to data.
Comments