Deriving Selection Bias In A Gaussian Marker Test

Why Reuse Can Create an Apparent Marker

A marker contrast is often a difference in mean expression between two groups of cells. If the groups are fixed independently of the gene being tested, sampling noise can make the observed contrast positive or negative, but under a null hypothesis its expectation is zero. A different situation arises when the gene helps define the groups: cells with larger values of that gene are more likely to enter one group, so the resulting difference is partly built into the assignment rule.

Selection can add a second source of inflation. If several genes or contrasts are examined and only the largest observed difference is reported, the selected estimate is no longer an ordinary draw from the null distribution. This tutorial derives both effects in Gaussian models. The models are deliberately simple: they isolate the mechanism, rather than reproduce every detail of a clustering pipeline.

A Fixed-Label Gaussian Benchmark

First consider a benchmark where the labels are fixed independently of the tested gene. Under the null, suppose the expression values in both groups are independent observations from the same distribution \(N(\mu,\sigma^2)\). Conditional on group sizes \(n_A\) and \(n_B\), the contrast has expectation

and variance \(\sigma^2/n_A+\sigma^2/n_B\). Thus, if \(\sigma\) is known, \(D\) has a centered Gaussian sampling distribution with standard error \(\sigma\sqrt{1/n_A+1/n_B}\). The groups may have different observed means in a particular sample, but the procedure has not preferentially selected a direction for this gene. The independence of the labels and the gene values is the key condition; having a Gaussian distribution alone does not ensure an unbiased selected contrast.

Deriving the Bias From a Marker-Based Split

Now suppose the gene contributes to the score used to assign cells. Let \(Y\) denote a cell’s expression, with \(Y\sim N(0,\sigma_Y^2)\) under a null model with no biological population difference. Let \(S\) be the standardized, centered clustering score, with \(S\sim N(0,1)\), and suppose the pair \((Y,S)\) is jointly Gaussian with correlation \(\rho\). Assign a cell to group \(A\) when \(S>0\), and to group \(B\) when \(S\leq 0\). A positive \(\rho\) means larger expression tends to accompany a larger assignment score.

For jointly Gaussian variables, the conditional mean is \(E(Y\mid S=s)=\rho\sigma_Ys\). Also, for a standard normal variable, \(E(S\mid S>0)=\sqrt{2/\pi}\), while symmetry gives \(E(S\mid S\leq0)=-\sqrt{2/\pi}\). Averaging the conditional mean over each assigned group therefore gives the expected group difference:

This is a derivation of the group difference induced by the assignment rule, not a biological cluster effect. In a sample, conditional on the number of cells assigned to each group, each group mean averages observations from its corresponding truncated Gaussian distribution. Its expectation is the conditional mean above, provided both groups are nonempty. Consequently, the expected sample contrast has the same value under this idealized model.

If the clustering score is the gene itself, standardized to \(S=Y/\sigma_Y\), then \(\rho=1\). The expected difference is \(2\sigma_Y\sqrt{2/\pi}\), or approximately \(1.596\sigma_Y\), despite the null model having no pre-existing groups with different gene means. If the gene contributes only partly to the score, the induced contrast is scaled by \(\rho\). This simple split makes the mechanical link transparent: the marker values help determine which side of the split cells enter.

Deriving Inflation From Reporting the Largest Contrast

Even with labels independent of the tested genes, screening several noisy contrasts and reporting the largest can inflate the reported magnitude. To isolate this effect, suppose there are \(m\) candidate contrasts, all with true value zero, and their estimates \(D_1,\ldots,D_m\) are independent \(N(0,\tau^2)\) variables. Here \(\tau\) is the standard error of each contrast. If the largest positive estimate is reported, its expected value is positive; if the largest absolute estimate is reported, its expected magnitude is positive as well.

For the absolute-value rule, let \(Z_j=D_j/\tau\), so each \(Z_j\sim N(0,1)\), and define \(W=\max_j|Z_j|\). For any nonnegative random variable, \(E(W)=\int_0^\infty P(W>t)\,dt\). Independence gives \(P(W\leq t)=[P(|Z_1|\leq t)]^m=[2\Phi(t)-1]^m\), where \(\Phi\) is the standard normal cumulative distribution function. Therefore,

This expectation is greater than zero for \(m\geq1\), and it increases with the number of candidates: adding a candidate cannot decrease the maximum absolute value. With a two-sided selection rule, symmetry means the signed contrast associated with the selected absolute maximum still has unconditional expectation zero. The upward bias is in the reported magnitude, and the selected signed estimate is systematically extreme conditional on its selected direction. With a rule that selects the largest positive contrast, the signed reported estimate itself has positive expectation.

Assumptions, Diagnostics, and Alternatives

These derivations rely on explicit idealizations. The marker-based split calculation assumes a centered bivariate Gaussian pair, a threshold at zero on a standardized score, and a stable correlation \(\rho\) between the score and expression. The maximum calculation assumes independent Gaussian null contrasts with a common standard error and a known candidate count \(m\). In real single-cell analyses, genes may be correlated, expression distributions may be skewed or zero-inflated, cluster sizes can vary, and the clustering rule may be much more complex than a threshold. The numerical bias from either formula need not apply directly to such a pipeline.

The assumptions clarify the mechanism even when they are not realistic enough for direct calibration. A marker used strongly in the assignment score has a direct route to an inflated contrast. A larger search over candidate genes, thresholds, or clusterings creates more opportunities to select an extreme. Dependence among candidates changes the distribution of the maximum, so the independent-candidate formula is not an exact correction for correlated genes.

Useful diagnostics begin with the analysis record: identify whether the tested gene, a correlated feature, or a score containing that gene contributed to the assignments; identify how many candidate contrasts were considered; and distinguish prespecified reporting from selection after examining results. Comparing the reported contrast with the range of contrasts considered can reveal a winner’s-curse pattern, but such inspection is not a formal correction. A nominal test based on a fixed-label null distribution does not automatically account for label selection or for choosing the largest result.

The main alternative is a contrast in which labels are fixed independently of the tested gene, as in the benchmark above. Independent data for marker assessment can also separate cluster discovery from marker evaluation. These approaches target the dependence that creates the bias; merely choosing a different test statistic on the same adaptively formed groups does not, by itself, remove that dependence. The Gaussian calculations explain direction and scale of the effect under simple assumptions, not a universal adjustment that validates every marker claim.

Common Mistake

Wrong reasoning: “The marker difference is positive and large, so the null hypothesis of no real group difference is implausible.” This overlooks that the gene may have helped determine which cells were called group \(A\) and which were called group \(B\). Under the centered Gaussian model, that rule alone produces an expected positive contrast of \(2\rho\sigma_Y\sqrt{2/\pi}\).

Correction: First ask whether the gene and candidate set were selected independently of the labels and reported result. A contrast calculated after using the same expression data to form groups is not interpreted like a prespecified fixed-label contrast. The observed difference remains descriptive, but its size alone does not establish a population difference.