1

Are F-test, P-test and t-test necessary in Simple Linear Regression and Multiple Linear Regression analysis, when we work with entire Population?

I have data of a particular field of last two years. I would like to use all of these data into Regression Analysis and want to predict what may happen in next stage when I will conduct the same type of work again.

In this case, do I need to use P-value, F-value or t-values?

Alexis
  • 26,219
  • 5
  • 78
  • 131
Nazmul Hassan
  • 119
  • 3
  • 1
    See https://stats.stackexchange.com/questions/2628/statistical-inference-when-the-sample-is-the-population but as you want to use your model for out-of-data predictions, I can't see why would you need either of the values? – Tim Nov 07 '17 at 15:03
  • 4
    It is extremely rare to be working with an entire population. Every question we have ever seen concerning that circumstance has been based on confusing a sample with a population. In particular, if you are going to predict something, you are using your data *as if it were a sample* representing the thing you are predicting. Therefore, it's unlikely any answer to your question will be appropriate for the circumstances you are in. – whuber Nov 07 '17 at 15:38

1 Answers1

2

If it is reasonable to assume that your data is a random sample from the population over the given time horizon including the nearby future (e.g. assuming stationarity over time), then it could still make sense to look at p values and friends in order to infer on properties of the full time horizon.

Quite similar is the following (not so unrealistic) setting: A pharmaceutical company performs a clinical study in Japan on some rare disease, testing some crazy new drug. Almost every such patient in Japan is participating. Nobody will say: "Don't use p values because you have basically accrued the full population". Why? Because the company also wants to sell the drug to future patients in Japan, and, more important, also to the rest of the world. From this point of view, the patient sample at hand is actually just a tiny fraction from all potential patients. A problematic aspect (which is almost always present in medical research) is that you cannot test the necessary stationarity assumptions required for the sample to be considered as "random".

There are also other reasons that could make you look at p values et al. if you possess the full population, e.g. if the values underwent any sort of "randomness". For instance if you ask every person in a company, how happy they are. It will make a difference if you ask them on a Monday or on a Friday... happyness is not 100% objective.

Michael M
  • 10,553
  • 5
  • 27
  • 43