Bias detection module

[Zethson]

Description of feature

Pre-requisite: #637

This serves as a meta issue to collect ideas for a bias detection and mitigation module:

• Implement Little MCAR #645
• Missingness per feature -> missingno stuff
• IQR outliers -> Add outlier detection #648 for numerical data @Zethson
• value counts for categorical counts + warnings? Some things are already in tableone @Zethson
• sensitivity analysis - easy with ehrapy because the API is so easy @Zethson
• Imbalanced subsampling through https://imbalanced-learn.org/stable/index.html ? -> only the most important sampling features in ehrapy. Not the generation stuff. @eroell
• Propensity score matching @Lilly-May
• standardized mean differences for numerical data @Lilly-May (Add bias detection to preprocessing #690)
• Correlations between variables @Lilly-May (Add bias detection to preprocessing #690)
• Feature importance -> requires target variable (only run if set) @Lilly-May (Add bias detection to preprocessing #690)
• generate_bias_report that runs a lot of methods subsequently and summarizes things?!? @Lilly-May (Add bias detection to preprocessing #690)

[eroell]

Comment regarding balanced sampling from imbalanced-learn, I started off something small a while ago due to a discussion of how to sample things for scanpy:
https://github.com/eroell/scimb

Anything like that could be helpful indeed.

[Lilly-May]

Regarding propensity score matching, we already have that implemented for causal inference with dowhy here. If we implement a causal discovery module (#619), implementing a propensity score matching algorithm would be important, but until then, I don't really see the need. However, please feel free to disagree!
Should we close this issue then?

[Zethson]

Yup! Let's close May 4, 2024 12:18:43 PM Lilly May ***@***.***>:

…

Regarding propensity score matching, we already have that implemented for causal inference with dowhy here[https://github.com/theislab/ehrapy/blob/1b5165d925f87304948fffe9b76cc68669ee6ab2/ehrapy/tools/causal/_dowhy.py#L37]. If we implement a causal discovery module (#619[#619]), implementing a propensity score matching algorithm would be important, but until then, I don't really see the need. However, please feel free to disagree! Should we close this issue then? — Reply to this email directly, view it on GitHub[#647 (comment)], or unsubscribe[https://github.com/notifications/unsubscribe-auth/AFHQA2DLHMCVC4RXT7X73RLZASYYDAVCNFSM6AAAAABCUZSVTWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAOJUGEYDQNJSGI]. You are receiving this because you were mentioned. [Tracking image][https://github.com/notifications/beacon/AFHQA2AQRISA2GSRSJRFIFTZASYYDA5CNFSM6AAAAABCUZSVTWWGG33NNVSW45C7OR4XAZNMJFZXG5LFINXW23LFNZ2KUY3PNVWWK3TUL5UWJTT42GHWU.gif]