2021/06/28

The problem with p-values

Fundamentally, hypothesis testing is based on conditional probability. 

We base our thinking on these premises: 

 

1.    Null Hypothesis H0 --> no effect

2.    Alternative Hypothesis H-->  There is an effect

If we assume His true and the P-Value of 5% is. There is a chance of 5% that we would have gotten the test results given the Null Hypothesis is true. Since this is a very low probability, we are rejecting the Null Hypothesis. So usually, a high p-Value indicates that my test results are significant. 



HOWEVER, 

What does a p-value of p=.2 indicate? It means given the Null Hypothesis is true there is a 20% that we would have gotten these effects. 

However, this is the problem: We fail to reject the Null Hypothesis --> 20%, but we can also not accept it. We have absence for evidence for an effect but we don't have evidence for the absence of an effect. 

 In other words, the p-value does not tell us anything about how likely it is that a hypothesis is true. 


SOLUTION: 

Bayesian Hypothesis Testing 

deals with: Which of the hypotheses is better supported by the data? 


Answer: The model that predicted the data best ! 

The ratio of predictive performance is known as the Bayes Factor (over 10 is usually good)