Question of the model

[Wanqi-Wang]

I am wanqi, It's a excellent and interesting website!

Actually, I cannot understand why to use a logistic model in prediction.

Personally feeling, logistic models are most recently used when evaluate binary outcome (e.g. survival/ death; get a disease/ not). By its easy to get odds ratio, this model become popular in the epidemiology.

However, when logistic model is used in this website, I don't think it's a good choice. Because it seems that it didn't consider incubation period, difference of vulnerable among population, so do you think SIRS model will be a better choice?

👍 💯 🥇 🐕 🍰

[Wanqi-Wang]

@YiZou01

[YiZou01]

Dear Wangqi,

Thanks for the concern. In fact, the logistic regression contains a board series，from two-parameter to five parameters. In this case, it describes the "S" shape. And it is true in some case it can be use to predict the probability (i.e binomial distribution, Poisson distribution). If the growing pattern is "S" shape (i.e. resource restricted), then basically it will work. And as you may see, it works.

The advantage of this model is that it is a very simple one. Of course, other models, e.g. the SIRS you mentioned, may also work, or could be better, but in that case, you need to consider more parameters, which means more information is required.

[Wanqi-Wang]

@YiZou01 ,Thanks! Dr. Zou,

It's a good answer to my question,

I agree you! And logistic may be the most efficient one! It seems I relied too much on textbooks.😆

By the why, how you can test if the prediction work? And why logistic model has two-parameter to five parameters, why five, rather than 6 or 7?

Thank you very much! Wanqi

[YiZou01]

Maybe https://en.wikipedia.org/wiki/Logistic_function

[pzhaonet]

@Wanqi-Wang

Regarding how to test if the prediction works, you can use a subset of the data to parameterize the model, and compare the simulated values with ralated observations.

Think about it: we spend no money for getting the data, and the prediction agrees well with the reported data. It might be the model which has the highest cost- performance ratio, ever.

[Wanqi-Wang]

@Wanqi-Wang Regarding how to test if the prediction works, you can use a subset of the data to parameterize the model, and compare the simulated values with ralated observations. Think about it: we spend no money for getting the data, and the prediction agrees well with the reported data. It might be the model which has the highest cost- performance ratio, ever.

@pzhaonet Thanks, Dr. Zhao! Yes! I agree it is the model has the highest cost - performance ratio, ever! Actually, I don't understand "you can use a subset of the data to parameterize the model, and compare the simulated values with ralated observations." Why use a subset? Why not compare predicted value at a certain time with observed value at that time directly?

Thanks 🌷

[pzhaonet]

Why use a subset? Why not compare predicted value at a certain time with observed value at that time directly?

The same meaning.

[Wanqi-Wang]

Thank you all !