Shared-denominator spurious correlation: Definition & Examples

Shared-denominator spurious correlation is association created by dividing two quantities by the same varying denominator, then interpreting the resulting correlation as a relationship between the underlying quantities. In quant research, two ratios such as profit/assets and investment/assets can move together partly because both contain assets in the denominator. Their correlation alone does not establish a connection between profit and investment.

The association between the ratios can be mathematically real. What makes an interpretation spurious is attributing that association to a mechanism the calculation has not established. Kim's exact treatment distinguishes correlation of ratios from correlation of their numerators. This mechanism also differs from spurious regression caused by nonstationary time series: a common denominator can create association in a one-period probability model.

Eight outcomes, no relationship between the numerators

Consider a fictional, dimensionless model. Numerators

U

and

W

each equal 1 or 3 with equal probability. The denominator

V

equals 1 or

1/2

with equal probability. All three variables are mutually independent, so the eight combinations each have probability

1/8

Define $X=U/V$ and $Y=W/V$ . Conditioning on the denominator makes the mechanism visible:

Denominator	Probability of this denominator	Possible values of each ratio	Mean of each ratio within this group	Within-group correlation
$V=1$	$1/2$	1 or 3	2	0
$V=1/2$	$1/2$	2 or 6	4	0

Within either group, the ratios remain independent. Pool the groups, however, and a small denominator raises both ratios together. The pooled means are both 3. Direct probability-weighted calculation gives

\mathbb{E}[XY]=10,\qquad \text{Cov}(X,Y)=10-3\times3=1.

Each ratio has variance

7/2

, so their correlation is

\operatorname{Corr}(X,Y) =\frac{1}{\sqrt{7/2}\sqrt{7/2}} =\frac{2}{7}\approx0.2857.

The numerators have correlation zero. Nothing about their relationship changed; the shared scaling introduced another source of co-movement.

Independent numerators U and W each take 1 or 3, giving four equally likely points and correlation zero. Ratios U/V and W/V give eight equally likely points: four for V equal to 1 and four for V equal to one half. Within each denominator group correlation is zero, while pooled correlation is 2/7.

Exact fictional probability model, not market observations. Circles identify denominator 1; outlined squares identify denominator 1/2. Each right-panel point has probability 1/8.

These are all outcomes of a stipulated distribution, rather than eight independently sampled firms or dates. The coefficient is a population calculation under that model, with no significance test or empirical discovery attached.

The common multiplier is the source

Write $D=1/V$ , so $X=UD$ and $Y=WD$ . Assume $U$ , $W$ and $D$ are mutually independent, all have finite second moments, and the two ratios have positive variance. Here $V$ must be nonzero with probability one; finite moments of $V$ alone would not guarantee finite moments of its reciprocal.

Expanding covariance and factoring the expectations gives

\begin{aligned} \text{Cov}(UD,WD) &=\mathbb{E}[U]\mathbb{E}[W]\mathbb{E}[D^2] -\mathbb{E}[U]\mathbb{E}[W]\mathbb{E}[D]^2\\ &=\mathbb{E}[U]\mathbb{E}[W]\text{Var}(D). \end{aligned}

For the example, the numerator means are both 2 and $\text{Var}(D)=1/4$ , producing covariance 1. The variance calculation is

\text{Var}(UD) =\mathbb{E}[D^2]\text{Var}(U) +\mathbb{E}[U]^2\text{Var}(D) =\frac52\times1+4\times\frac14 =\frac72.

This identity describes an independent-variable benchmark, not a universal formula for financial ratios. In actual company data, profit, investment and assets need not be independent. Their joint economic relationships can strengthen, weaken or reverse the association. Even under the benchmark, opposite-sign numerator means give negative covariance; a common denominator does not imply positive correlation in every setting.

Audit the interpretation before changing the variables

A useful investigation starts by retaining the two numerators and the denominator in the dataset, rather than retaining only the ratios.

Specify the question. A relationship between profit and investment is different from a relationship between their intensities per unit of assets. Ratios can be economically meaningful outcomes in their own right.
Inspect the scaling. Plot each numerator against the denominator and examine the ratios by denominator range. Very small or mismeasured denominators deserve attention because division can magnify their effects.
Write down the model before normalising. If the model in levels is $U=\alpha+\beta W+\varepsilon$ , dividing it by $V$ gives $U/V=\alpha/V+\beta W/V+\varepsilon/V$ . An ordinary constant in a regression of $U/V$ on $W/V$ does not generally replace the term $\alpha/V$ . Scaling also changes the error structure. Bartlett and Partnoy develop these specification and denominator-measurement concerns for ratio outcomes.

There is no automatic cure in “control for size”: a single linear denominator term may miss reciprocal or interaction terms required by the economic model. Likewise, the fact that conditioning removes association in the fictional example does not establish a causal adjustment strategy for real data.

Finally, dividing both variables by the same fixed nonzero constant does preserve correlation. The issue here is a denominator that varies across observations or states. It is not a criticism of changing currency units, or of the constant population standard deviations used to define Pearson correlation.

For the probability-weighted moment calculations behind this example, Expectation, Variance and Covariance develops the same tools and a portfolio application. Its public preview is open; the full lesson requires a free signed-in account.

Sources

Ji-Hyun Kim (1999), “Spurious correlation between ratios with a common divisor”, Statistics & Probability Letters 44(4), 383–386. Publisher abstract, introduction, theorem preview and discussion: the exact independent-variable benchmark and the interpretation of mechanically induced association. The finite model, derivation presentation and figure here are original.
Robert Bartlett and Frank Partnoy (September 2020), The Ratio Problem, introduction, §2.1 and §3.2: ratio outcomes, omitted reciprocal/scaled terms and denominator measurement error. These are broader regression concerns; they do not imply that every ratio is an invalid economic construct.

Eight outcomes, no relationship between the numerators

The common multiplier is the source

Audit the interpretation before changing the variables

Sources

Related terms

Correlation

Covariance

Independence

Variance