The Square Root of the Sum of Squares: Why Comparing Two Polls Requires a Larger Margin of Error

On March 6, 2018, Stanford University researchers published a stark warning for data journalists and political analysts: "The margin of error is bigger than you think." The mathematical reality of survey science dictates that when a reader looks at a poll showing a candidate leading 48 percent to 46 percent, the standard error printed at the bottom of the page does not apply to the gap between them. Instead, measuring the distance between two independent survey figures requires a completely different calculation, one that expands the uncertainty bound by roughly 41 percent.[1][3][5]

The root of this statistical illusion lies in how polling variance is reported to the public. A standard survey of 1,000 respondents carries a baseline margin of error of roughly 3.1 percentage points at a 95 percent confidence interval. That figure represents the uncertainty of a single isolated proportion—for instance, the percentage of voters supporting Candidate A.[4]

However, elections and public opinion debates are rarely about a single isolated number. They are about the difference between two numbers. When an analyst compares Candidate A's 48 percent to Candidate B's 46 percent, they are no longer looking at one variable, but two. Because both measurements contain their own independent sampling noise, the total uncertainty of the gap between them is the combination of both errors.[6]

In statistics, standard errors cannot simply be added together. Instead, variances add. To find the margin of error for the difference between two independent polls, statisticians must calculate the square root of the sum of their squared margins of error. This geometric relationship is mathematically identical to the Pythagorean theorem used to find the hypotenuse of a right triangle.[4][5]

If Poll X and Poll Y both carry a 3.1 percent margin of error, the error of the difference between them is not 3.1 percent, nor is it 6.2 percent. It is the square root of 3.1 squared plus 3.1 squared, which equals approximately 4.38 percent. The penalty for comparing two equally sized independent samples is always a multiplier of the square root of two, or 1.414.[5]

This 41 percent expansion in uncertainty fundamentally alters how survey data should be interpreted. A two-point lead in a poll with a 3.1 percent baseline margin of error is frequently described by commentators as a statistical tie. But the reality is far more fragile: the gap would need to exceed 4.38 points for the lead to be statistically distinguishable from zero.[1]

Writing in January 2026, data analyst G. Elliott Morris highlighted how this mathematical reality is routinely ignored in political coverage, leading to false narratives about shifting momentum. When a candidate's lead shrinks from four points in October to two points in November across two different polls, the shift is almost entirely consumed by the expanded margin of error.[1]

The mathematics become even more complex when comparing two proportions within the exact same sample, rather than across two independent polls. When a pollster asks a single group of 1,000 respondents to choose between two candidates, the two proportions are negatively correlated. Every respondent who shifts their support to Candidate A is simultaneously subtracting a vote from Candidate B.[6]

This negative covariance means the errors are not perfectly independent. As detailed in statistical analyses of intra-sample proportions, the negative correlation slightly reduces the combined variance compared to two entirely separate polls. However, the required margin of error for the gap still remains significantly larger than the baseline error reported for a single candidate's vote share.[6]

The same mathematical penalty applies far beyond political polling. In corporate market research, the Net Promoter Score (NPS) is calculated by subtracting the percentage of "detractors" from the percentage of "promoters." Because the final metric is a difference of two proportions, the margin of error for an NPS result is substantially wider than the error for the promoter metric alone.[2]

Despite this, corporate earnings reports and marketing dashboards routinely display NPS shifts of one or two points as actionable business trends. Without applying the square root of the sum of squares to the underlying variances, executives are frequently reacting to statistical noise rather than genuine changes in customer sentiment.[2]

The textbook definition of inference for two independent population proportions requires calculating a pooled standard error when testing the null hypothesis that the two proportions are equal. If the null hypothesis is true—meaning there is no actual difference between the two populations—the combined sampling noise will frequently produce artificial gaps of three or four points in samples of 1,000 respondents.[4][5]

The Stanford researchers who warned about understated margins of error also pointed out that sampling variance is only one source of polling uncertainty. Non-sampling errors—such as non-response bias, framing effects, and demographic weighting models—often introduce more distortion than the random draw of the sample itself.[3]

When these design effects are incorporated into the variance calculations, the true margin of error for a standard poll often doubles. If a single poll's real-world margin of error is closer to 6 percent, the margin of error for the difference between two such polls approaches 8.5 percent.[3]

The evidence pack for polling accuracy remains constrained by the public's appetite for certainty. News organizations face structural incentives to report a one-point shift as a breaking development rather than a mathematical artifact. Until the geometric addition of variances becomes a standard feature of data journalism, the gap between statistical reality and public perception will persist.[1]