The Underlying Causes of Simpson's Paradox in Statistics

This article explores the mathematical and statistical mechanisms behind Simpson's Paradox, a phenomenon where a trend present in multiple individual groups of data completely disappears or reverses when those groups are combined. We will break down the primary causes of this statistical illusion—specifically focusing on confounding variables and unequal sample distributions—and demonstrate how these factors interact using a clear, practical example.

Simpson's Paradox is not a flaw in mathematics; rather, it is the result of how weighted averages work when applied to complex datasets. The reversal of trends occurs due to two main factors working in tandem: confounding variables and disproportionate sample sizes.

1. Confounding Variables

The primary driver of Simpson's Paradox is the presence of a confounding variable, sometimes called a lurking variable. This is an outside, often unmeasured factor that correlates with both the independent variable (the cause) and the dependent variable (the effect).

When data is divided into subgroups, the confounding variable is controlled for, allowing you to see the true relationship within that specific context. However, when you aggregate the data into a single pool, you erase the visibility of the confounding variable. Its hidden influence remains active in the math, skewing the combined outcome in the opposite direction.

2. Disproportionate Sample Sizes

A confounding variable alone is not enough to trigger the paradox; there must also be a significant imbalance in the size of the subgroups.

When you look at separated groups of data, you are usually comparing simple percentages or rates. When you combine the data, you are calculating a weighted average. If the sample sizes of the subgroups are vastly unequal, the largest subgroups will dominate the combined average. If one group is heavily weighted toward a specific condition (due to the confounding variable), the overall average will reflect the baseline reality of that condition rather than the trend you are actually trying to measure.

How the Causes Interact: A Clear Example

To understand how confounding variables and unequal sample sizes cause this reversal, consider a hypothetical comparison of two medical treatments—Treatment A and Treatment B—used on 2,000 patients with either mild or severe symptoms.

The Divided Data (Subgroups):

  • Mild Cases: Treatment A cures 90 out of 100 patients (90%). Treatment B cures 720 out of 900 patients (80%). Treatment A is better.
  • Severe Cases: Treatment A cures 540 out of 900 patients (60%). Treatment B cures 50 out of 100 patients (50%). Treatment A is better.

In both mild and severe cases, Treatment A has a higher success rate than Treatment B.

The Combined Data (Aggregated):

  • Treatment A Total: Cured 630 out of 1,000 patients (63%).
  • Treatment B Total: Cured 770 out of 1,000 patients (77%).

When combined, the trend reverses, making Treatment B look vastly superior.

Why It Happened:

  1. The Confounding Variable: Disease severity is the hidden factor. Naturally, severe cases have a much lower survival rate than mild cases, regardless of the treatment used.
  2. Disproportionate Sample Sizes: Doctors did not assign the treatments equally. They gave the better option (Treatment A) predominantly to the sickest patients (900 severe vs. 100 mild). Conversely, they gave Treatment B mostly to patients who were only mildly ill (900 mild vs. 100 severe).

When the data is combined, Treatment A's overall average is dragged down because 90% of its sample size consists of difficult, severe cases. Treatment B's overall average is artificially inflated because 90% of its sample size consists of easy, mild cases.

Summary of the Cause

Ultimately, Simpson's Paradox is caused by improper data aggregation. It happens when an analyst combines distinct groups of data without accounting for a baseline inequality (the confounding variable) that dictates how the data was distributed into those groups in the first place (the disproportionate sample sizes).