Video Summary - [그림으로 완벽정리 :: 통계학] 심슨의 역설이란?

Main Ideas & Concepts

Simpson’s Paradox (illustrated with graphs)

Simpson’s Paradox is a phenomenon where the relationship between two variables:

Within each subgroup (after splitting the data by a category) does not match—or is reversed compared to—
The relationship seen in the combined dataset (when groups are ignored).

Example used: Math vs. English grades (male vs. female students)

Male-only data: shows an upward-right trend → positive correlation (and a positive slope in a linear model).
Female-only data: also shows an upward-right trend → positive correlation.
All students combined: shows a downward-right trend → negative correlation (and a negative slope in a linear model).

Naming

Simpson’s Paradox is named after a British statistician who published a paper on the phenomenon.

Correlation vs. directionality

The video emphasizes:

Positive correlation: as one variable increases, the other increases (and vice versa).
Negative correlation: as one variable increases, the other decreases (and vice versa).

It also stresses that correlation alone cannot determine causal direction (which variable influences which), because variables may be jointly associated without proving cause.

Causality is said to be covered separately in another video.

Why Simpson’s Paradox happens (role of a mediating/communicator variable)

The explanation introduces a variable described as a mediating factor (called a “commutator/communicator” in the subtitles).

This categorical factor (e.g., gender, class) influences both variables being compared.
The key mechanism described:
- Even if there is an underlying positive causal relationship between the two variables,
- partitioning into subgroups (e.g., by gender) can make the combined pattern appear reversed.

Model reasoning (clustering effect)

The data are clustered by category:
- e.g., males cluster toward the upper-left
- e.g., females cluster toward the lower-right
When combined, this clustering effect can mask the original positive relationship and produce an apparent negative one.

When it would not be Simpson’s Paradox

The paradox would not occur if the confounding categorical effect (e.g., gender/mediator effect) is weakened—meaning the subgroup differences are not significantly driven by that category.

In that case, combining the data would still show the same direction (e.g., positive relationship).
Therefore, the paradox does not occur.

“Methodology” / Instructional Logic (Conceptual Checklist)

To identify Simpson’s Paradox

Examine the relationship within each subgroup (e.g., males only, females only).
Determine the direction of association in each subgroup (positive vs. negative trend).
Examine the relationship using the combined dataset (ignore subgroup labels).
If the combined trend is reversed or does not match the subgroup trends, label it Simpson’s Paradox.

To interpret the cause (as described in the video)

Consider a categorical mediator/confounder (“communicator/commutator,” such as gender or class) that affects both variables.
Analyze how subgroup clustering induced by that category can shift the overall combined trend.
Recognize that the reversing effect can occur even when an underlying positive causal relationship exists.

Speakers / Sources Featured

No specific individual speaker is named in the subtitles.
The source referenced is a British statistician who originally published a paper on Simpson’s Paradox (no name is given in the subtitles).

[그림으로 완벽정리 :: 통계학] 심슨의 역설이란?

Key takeaways