1
$\begingroup$

I am developing a classification model for covid19 symptoms (after being ill) and I don't understand statistical analysis importance (some parts of it)

1 Firstly: Basically we perform statystical analysis to learn about data. However what's the purpose of counting mean, standard deviation as shown here:

https://www.sciencedirect.com/science/article/pii/S0010482522000762#bib27

What insight will it give me?

2 Moreover: They perform statistical test like Chi-Square to find the statistically significant features. Suppose they have around 15 "blood parameteres" and the tests would tell that only 10 of them are statistically important. Does it mean those 5 won't be used in the training and can be removed?

3 If they can be removed: Would feature elimination prove the same? Suppose we used Recursive Feature Elimination / Random forest with 10-best features. Would results be the same?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Though not in the details, it looks like they took some of the continuous variables, ranked them, and then used Chi-square to determine feature set. No explanation given as to why they did that. Also regarding the features not found significant. You can certainly uses them in model. chi-square is a weak test, and there may be interactions found in the model which are meaningful.

In any case The statistical tests were exploratory. Then were not used for inference directly. It is always a good practice to perform basic statistical descriptive statistics before approaching any ML. For example they could have not performed the missing value imputation without first seeing how many there were. Also note that MVC variable has overlapping confidence intervals between COVID and non-COVID responses, which sometimes is a signal that there is not a significant difference due to that variables.

They selected four features: white blood cell count (WBC), monocyte count (MOT), age, and lymphocyte count (LYT) and they ran them through 8 machine learning algorithms to classify and they used a stacked ML model.

Improve this answer
$\endgroup$
6
  • $\begingroup$ So suppose I have 15 blood parameters. Should I perform those statistical analysis tests as they did and keep these where p_value is less than 0.05 or should I keep all of those parameters and later evaluate with random forest what is usefull? $\endgroup$
    – yras8
    Commented May 1, 2022 at 16:48
  • $\begingroup$ I understand for what they use Chi-square in feature selection. But I am confused with statistical use-case. $\endgroup$
    – yras8
    Commented May 1, 2022 at 16:50
  • $\begingroup$ i think they use statistical test not feature elimination ,because they want to determined the group of features that are most statistically significant in both groups COVID-19 and Non -COVID groups.If they use feature elimination they could finnished with different features for each group. $\endgroup$
    – mariq vlahova
    Commented May 1, 2022 at 19:50
  • 1
    $\begingroup$ I updated my answer. It is unclear what you are trying to do with this. Are you trying to replicate this? $\endgroup$
    – Ralph Winters
    Commented May 1, 2022 at 21:38
  • 1
    $\begingroup$ see updated answer regarding my opinion on why there were using the statistical analysis $\endgroup$
    – Ralph Winters
    Commented May 2, 2022 at 13:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.