Week 3 Topic 1

[Wirichada Pan-ngum]

For topic discussion of this week I would like you to generate some discussions around the topic of “Bayesian VS Frequentist”. For example, you can talk about the similarities and differences or you can write what you have learned on this topic and share and discuss with other classmates.

[Karina Dian Lestari]

I am much more familiar with frequentist methods because it is what has been thought in my undergrad study. I have tried to understand the Bayesian method but having a hard time because I am just too used to the frequentist method. The biggest difference between the frequentist and Bayesian methods is that for Bayesian you need a prior knowledge for the parameter. How do I know my prior knowledge is true? I also do not know anyone that is proficient in the Bayesian method, so it is hard to ask for advice. However, I agree that we do not have to lean into just one side only. We should select the method that is suitable for our data and can help us to answer our question.

[Wirichada Pan-ngum]

Thanks for sharing your thought. I think we all familiar with Frequentist approach more. It is good to learn some Bayesian as it is becoming more acceptable. One of the main concepts is about prior and posterior distribution, Bayes’ theorem.

[Kansiri Apinantanakul]

Bayesian VS Frequentist:
Both of these concepts are inferential statistics.

Bayesian
– Rely on the prior knowledge or known data to infer what they would like to know, therefore if the prior data is biased it would affect the answer.
– The basic assumption: parameter is a random variable.
– Common jargon: Credible interval, Prior, Posterior, Marginal

Frequentist
– Rely on the repeated action/tests results of the sample group to infer what they would like to know (“population parameter”), therefore if the result from action/tests is biased it would affect the answer.
– The basic assumption: parameter is fixed (but we don’t know what it exactly is)
– Common jargon: Confident interval, P-value

From my personal experience, I’m quite familiar with Frequentist more than Bayesian statistics. However, I also have experience studying the clinical trial methodology using both concepts to answer the different research questions. For example, assuming I conduct a clinical trial to study the effect of drug X on the prevention of a rare disease. In each clinic visit, the subject would be tested using a test kit whether he/she were developed the rare disease or not. After testing subject will receive drug X to prevent this rare disease. The test result of this lab kit can be collected and analyzed in 2 ways
1) Study effect on disease prevention in sample group and infer to effect in the population.
2) Study the incidence rate of this disease in Thailand using the Bayesian approach.

[Wirichada Pan-ngum]

Thanks for sharing your thoughts and experiences.

[Sri Budi Fajariyan]

Bayesian and frequentist are two approaches in statistics that can both find answers in research. Bayesian uses data about history and prior knowledge in addition to models and probing.

[Auswin Rojanasumapong]

From what I have learned, Bayesian and Frequentist are different in perspective of the answers and methods of finding the answer. Like many examples provided, Frequentist tends to rely on current data from repeated test/action while Bayesian tends to rely on prior knowledge. From my experience, I am familiar with the way of Frequentist, but I can imagine why the concept of Bayesian can be used in some situations to predict the result. Since there is no right or wrong between the two, we should learn both approaches to understand the perspectives of both Frequentist and Bayesian.

[Yanin Pittayasathornthun]

I don’t have experience on clinical data and I have never used both Bayesian and frequentist statistics.

Therefore, I don’t have clear picture on these statistics. After read the paper, I have a few questions and would like to discuss with you.

First one is from the first page of the paper, nearly at the end of middle paragraph at the right side. He stated that “If the prior and posterior probabilities come from the same statistical distribution family, they are called…”
Compare Bayesian statistic with t-test and other common ones, we usually test the distribution of the input data which is equal to prior probability. Anyway, the outcomes from t-test would be just p-value, whereas Bayesian generates probability of probability. The questions are:
1) why we would like to know the distribution of posterior probability?
2) is it matter if the distribution of prior and posterior probability the same or not the same type?

Next questions are from the second page, the second last paragraph at the left side. He stated that there is no single, prescribed and well-defined method for choosing a prior. The consequence of that is people use different priors and thus obtain different posteriors and make different conclusions.
Here I question (3) how we know which prior is suitable for our study? or
(4) If we try different priors, how do we know the posterior is correct? Is there any method for validate the posterior?

[Wirichada Pan-ngum]

To answer the first question, we sometime want to know the posterior distribution of the parameter estimated because we can learn much more about the parameters. The prior and posterior distribution may be similar or not, depending on how much we already know some information about that parameter and also the data observed.

[Wirichada Pan-ngum]

Fir q 3 and 4, sometimes you will see people try more than one prior distributions. If your data is strong, it doesn’t matter what you choose, it will tend to convert to the same value for the parameter estimate. Prior distribution may be coming from pilot study, literature, guess or people use uniform distribution for many unknowns.

[TARO KITA]

The concept of Bayesian is very new to me, and many aspects have already been mentioned here.
According to what I have learnt, one of the characteristics of Bayesian is its flexibility in incorporating new data. Bayesian allows data to be incorporated and analyzed in a sequential manner, while in Frequentist, the data must be analyzed again from the beginning each time more data is added. In terms of analyzing while incorporating new data, it is well suited to machine learning using big data, that is why it has been gaining attention in recent years.

[Arwin Jerome Manalo Onda]

At this point in time, I can say I am a frequentist since those are the methods I was taught in college. Hence, I was conditioned to be a frequentist. Perhaps, this is to standardize “answers” since Bayesian depends on your opinion – to which, opinion may vary from one person to another.

But after discussions on Bayesian, I think there will be instances that one method is “better” than the other. I might be wrong on this position but this is how I currently see this debacle.

I agree with Auswin that exploring each approach will help us understand the differences.

[Napisa Freya Sawamiphak]

I would say that I am more familiar with Frequentist than Bayesian because I’ve learned it before, and it was used in several clinical data and publications. I also heard about Bayesian from epidemiological research which is interesting to me. However, I haven’t used it myself. It seems like the probability is assigned ahead to a hypothesis in the Bayesian view but not for a frequentist view. I also agree with Auswin and Arwin that Bayesian tends to rely on prior knowledge, and we should learn both approaches to understand the difference.

[Tossapol Prapassaro]

From what I have learned in the past, I think I am a frequentist. I have some experience in clinical research and I use frequentist inference. Bayesian statistics is another type of statistical method that uses Bayes’ theorem. The difference between these two methods would be Frequentists use likelihood to calculate probabilities while Bayesian has to use the previous knowledge of the conditions that associated with the event to calculate the probabilities. The Bayesian seems to be more accurate in the statistic for diagnosis of the disease, however, there are some limitations to using it such as how to determine the prior knowledge and the calculation is far more complex than frequentist. Interestingly, the other utility of using Bayesian statistics is for machine learning algorithms which might be an important thing to learn in the near future.

[Wirichada Pan-ngum]

Famous Monty Hall Problem here https://youtu.be/obi_C1YbbIg
It is related to the Bayes Theorem, explain in here https://towardsdatascience.com/solving-the-monty-hall-problem-with-bayes-theorem-893289953e16

[Ashaya.i]

I am much more familiar with Frequentist than Bayesian. Considering between Frequentist vs Bayesian depends on the main objective of the study so we need to completely understand what we need from the study. The frequentist is relied on frequencies of events, meanwhile the Bayesian is relied on our knowledge of events. I think the concept of Bayesian is interesting because it is very new to me.

[Mayada Mohammed Abdalla]

Like most of you, I am more familiar With Frequentist than Bayesian and from what I understood Bayesian methods determine only the probability of an event based on the prior no matter the design was Applied. Frequentist methods can makes predictions using only data from current experiment while in Bayesian past Knowledge of similar experiment combined with Current experiment data to make conclusions.In my opinion Bayesian is more complicated and if I have to answer the question: does it matter which we use? I will say it depends on the data set.
I would like to discuss your opinions about the question {does it matter which we use?}.

[Hazem Abouelfetouh]

Both methods have advantages and disadvantages. I think choosing the statistics method depends on the hypothesis we are testing and the data set. The bayesian method depends on prior knowledge and on the probability of the observed data with relevant information necessary to make an inference which makes it more convenient and much faster for data processing in clinical research and machine learning.

[Nopporn Apiwattanakul]

I think both Bayesian and frequentist statistics can be useful in different settings. If we have a background knowledge or information, Bayesian one is more appropriate. If we are looking for something new, frequentist would give us new answers or knowledge.

[Author]

Posts