Why the Intercept Is Not a Universal Confounding Estimate
The LD score regression intercept is often described as an estimate of inflation not explained by polygenic signal. That interpretation is useful only under a particular mean model and its assumptions. An intercept is not a direct measurement of every confounding process, nor does it generally report what fraction of observed test-statistic inflation is confounding. In particular, confounding that varies with LD score can affect the fitted slope rather than appear as excess intercept.
The key is to distinguish a statistical decomposition from a causal or complete accounting of bias. The intercept is a fitted quantity in a regression of association test statistics on LD scores. Its interpretation depends on how the expected test statistics relate to those scores, how well the scores represent the GWAS, and how much uncertainty remains in the fit.
The Model Behind the Usual Interpretation
For variant \(j\), let \(\chi_j^2\) be its one-degree-of-freedom association test statistic and \(l_j\) its LD score, calculated as the sum of squared correlations with variants in the chosen reference set. A simplified LD score regression mean model is
Here \(N\) is the GWAS sample size, \(h_g^2\) is SNP heritability on the scale used for the analysis, and \(M\) is the number of variants in the model’s genetic component. The constant \(c\) represents excess expected test-statistic inflation that does not vary with LD score under this simplified model. In a regression of \(\chi_j^2\) on \(l_j\), the expected intercept is \(1+c\), and the expected slope is \(N h_g^2/M\).
This decomposition relies on assumptions. The genetic contribution must have the LD-score dependence represented by the model; the confounding contribution being assigned to the intercept must be approximately constant across variants; the LD scores must be appropriate for the GWAS; and the summary statistics must be sufficiently well calibrated for the model to be meaningful. The fitted relationship also needs enough information across the LD-score range to distinguish its intercept from its slope. If these conditions fail, the fitted coefficients remain regression coefficients, but the standard interpretation may not hold.
Even when the model is adequate, the intercept is an extrapolated regression value at \(l_j=0\), not usually the expected statistic for a typical observed variant. The intercept and slope are estimated with uncertainty. A point estimate close to 1 does not establish that confounding is absent, and a point estimate above 1 does not identify the source of the excess.
When Confounding Varies With LD Score
The central limitation is that LD score regression separates components by their dependence on LD score, not by whether they are genetic or confounding in origin. Suppose the expected statistics instead follow
where \(a l_j\) is the genetic contribution and \(d l_j\) is an LD-dependent confounding contribution. Their sum is \(1+c+(a+d)l_j\). A regression with an intercept and a slope will place both LD-proportional components in the slope. It cannot separate \(a\) from \(d\) using this mean relationship alone.
Worked Example: Confounding Enters the Slope
Consider hypothetical expected statistics at LD scores \(5,10,15,20\). Suppose the genetic contribution is \(0.008l_j\), and an LD-dependent confounding contribution is \(0.004l_j\), with no constant excess. Then
Substitution gives \(1+0.012(5)=1.06\), \(1+0.012(10)=1.12\), \(1+0.012(15)=1.18\), and \(1+0.012(20)=1.24\). These values lie exactly on a line with intercept \(1.00\) and slope \(0.012\). The intercept shows no excess above 1, even though the stated confounding contribution is \(0.004l_j\). The fitted slope combines both contributions; it does not identify how much of that slope is genetic.
This example does not claim that every confounding process is proportional to LD score. It shows why an intercept cannot be a universal estimate: whenever confounding follows an LD-dependent pattern, an intercept-only summary does not capture it as a constant excess.
How Misspecification Can Shift the Intercept
If the true relationship is not linear in LD score, the fitted line is a projection of that relationship onto an intercept and a linear term. Its intercept can therefore reflect the shape of the relationship over the observed LD-score range, not just a constant inflation component. For an ordinary least-squares line with slope \(\hat b\), the intercept is \(\hat a=\bar y-\hat b\bar l\), where \(y\) denotes the statistics being regressed. A nonlinear pattern can change \(\bar y\) and the fitted slope in ways that alter \(\hat a\).
Worked Example: A Nonlinear Pattern Alters the Fitted Intercept
Suppose hypothetical expected statistics follow a genetic linear contribution plus a curved deviation:
At LD scores \(5,10,15,20\), the squared deviations are \(56.25,6.25,6.25,56.25\). The resulting expected statistics are \(1.06125,1.10125,1.15125,1.21125\), respectively. Their mean is \(1.13125\). Because the curved deviations are symmetric around the mean LD score \(12.5\), their covariance with these four LD scores is zero; the fitted slope is therefore \(0.01\). The fitted intercept is
There was no constant \(c\) in the stated data-generating relationship. Yet the fitted intercept exceeds 1 because the curved term has a positive average over these LD scores. This is a projection effect, not evidence that the curvature is specifically a constant confounding process. With other LD-score distributions or a different form of curvature, the fitted coefficients could differ.
A related concern is reference-panel mismatch. If the reference LD scores do not adequately represent the GWAS LD structure, the predictor can be measured inaccurately or have a different relationship with the expected statistics. The resulting regression can be distorted; the intercept is not guaranteed to isolate a constant confounding component. The earlier material on reference-panel mismatch treats that issue directly. Here the important implication is that an intercept’s interpretation depends on the adequacy of the LD-score predictor, not just on its numerical value.
Worked Interpretation With a Nonzero Intercept
Worked Example: What an Intercept of 1.04 Does and Does Not Say
Suppose a hypothetical LD score regression fit is \(\widehat{E[\chi_j^2\mid l_j]}=1.04+0.012l_j\). At \(l_j=10\), the fitted mean is
At \(l_j=20\), it is \(1.04+0.012(20)=1.28\). The fitted intercept is \(1.04\), or \(0.04\) above the null reference value of 1. Under the standard mean model, adequate LD scores, and a confounding component that is constant across variants, that excess is consistent with LD-independent inflation.
The calculation does not show that exactly 0.04 units of inflation in every statistic were caused by a particular confounder. It does not rule out additional confounding that contributes to the slope, and it does not show that the slope is entirely genetic. Nor does the point estimate alone show whether the intercept is distinguishable from 1: that requires its standard error or confidence interval. The interpretation is conditional on the model, and should be reported in that conditional form.
Diagnostics, Alternatives, and Limitations
An intercept should be considered alongside evidence about model fit and study design. A useful diagnostic sequence is:
Look for systematic departures from a linear mean pattern across the LD-score range. A single intercept and slope can conceal curvature or influential regions.
Consider whether the reference information represents the GWAS population and variants well enough for the predictor to be informative.
Interpret the intercept estimate with its standard error or interval. An estimate alone does not convey precision.
Use study-design knowledge and other diagnostics to assess plausible sources of inflation. Agreement or disagreement among diagnostics is informative, but none provides a universal decomposition on its own.
The mean chi-square statistic summarizes overall average test-statistic elevation, but it does not separate polygenic signal from confounding. The genomic inflation factor \(\lambda\) summarizes a different feature of the test-statistic distribution and also does not by itself identify the source of inflation. QQ plots show the distributional pattern of observed statistics relative to a null reference, but they too do not label departures as genetic or confounded. LD score regression adds information by relating statistics to LD scores; that added structure is useful, but its interpretation still depends on the mean model and assumptions. Comparing these approaches is most valuable when their limitations are kept explicit.
The intercept also has finite-sample uncertainty and can be sensitive to modeling choices. An estimate near 1 should not be translated into “no confounding”; an estimate above 1 should not be translated into a complete measurement of confounding. The intercept’s value may be especially difficult to interpret when the slope and intercept are poorly distinguished by the available LD-score range, when the mean relationship is misspecified, or when the predictor is mismatched to the association data.
Common Mistake
Correction: The intercept is not designed to capture every confounding pattern. Confounding that varies with LD score can contribute to the slope, as in the worked example where the total slope was \(0.012\) but included a confounding component of \(0.004\). A near-1 intercept is therefore not proof of an unconfounded analysis.
A second version of the same mistake is to call the intercept excess the “fraction of inflation due to confounding” without stating the assumptions and calculation behind that quantity. A difference such as \(1.04-1=0.04\) is an excess in the fitted intercept relative to the null reference, not automatically a proportion of total inflation, and not an attribution to a named cause.
Check Your Understanding
Use the model assumptions as part of each answer; the intercept alone does not settle the interpretation.
- If the expected statistics follow \(1+0.03+0.01l_j\), what intercept and slope does the ideal linear fit have?
- Suppose genetic and confounding contributions are \(0.006l_j\) and \(0.003l_j\), respectively, with no constant excess. What does the fitted intercept reveal about the confounding contribution?
- Why can a fitted intercept exceed 1 when a true mean relationship has no constant inflation component?
- What additional information is needed before deciding whether an estimated intercept of 1.02 is meaningfully above 1?
Solutions. (1) The intercept is \(1.03\) and the slope is \(0.01\); this follows directly from \(1+0.03+0.01l_j\). (2) The expected relationship is \(1+0.009l_j\), so the ideal intercept is 1 and the slope combines genetic and confounding contributions; the intercept does not reveal the LD-dependent confounding. (3) If the relationship is nonlinear or otherwise misspecified, the fitted line projects that pattern onto an intercept and slope; the intercept can then differ from the constant component in the true relationship. (4) Its standard error or confidence interval is needed to assess precision and whether the estimate is statistically distinguishable from 1; model adequacy and LD-score suitability also matter for its substantive interpretation.