Quantitative Bias Analysis

Quantitative bias analysis (abbreviated QBA) is a methodological framework in epidemiology, biostatistics, and public health that quantifies the direction, magnitude, and uncertainty of systematic errors—specifically unmeasured confounding, selection bias (collider bias), and information bias (exposure misclassification or outcome misclassification)—on observational effect estimates. Developed from foundational sensitivity analyses by Jerome Cornfield (1959) and Sander Greenland (1996, 2005) and systematized by Timothy L. Lash, Matthew P. Fox, and Aliza K. Fink (2009, 2021), quantitative bias analysis addresses a central flaw of conventional observational reporting: standard 95% confidence intervals quantify only random sampling error under the heroic assumption that zero unmeasured confounding, 100% participation, and 100% diagnostic sensitivity and specificity hold. By specifying algebraic bias models governed by explicit bias parameters (such as sensitivity, specificity, selection odds ratios, or confounder prevalence) informed by validation substudies or external literature, QBA produces bias-adjusted point estimates and Monte Carlo total-error simulation intervals that combine systematic error with random sampling error.

In any observational study of sample size N, total estimation error equals the sum of random error (which shrinks at rate 1 / N^1/2 as sample size increases) and systematic error (which does not shrink as N → ∞). In modern administrative healthcare databases and national biobanks containing N = 10,000,000 records, random sampling error approaches zero, yielding narrow 95% confidence intervals (for example, OR = 1.18; 95% CI: 1.16–1.20). However, qualitative discussion sections typically dismiss systematic errors with speculative heuristics—such as claiming that exposure misclassification is "nondifferential and therefore only biases toward the null." Quantitative bias analysis replaces these qualitative claims with exact mathematical equations across four progressive tiers of analytical sophistication:

1. Simple Sensitivity Analysis (Fixed Single-Parameter Correction): Applies an algebraic bias-inversion formula using a single fixed set of bias parameters to compute a single bias-adjusted effect estimate (RR_adj or OR_adj) or summary threshold such as the E-value.
2. Multidimensional Sensitivity Analysis: Evaluates the bias-inversion equation across a deterministic grid of plausible bias-parameter combinations to map contour plots of RR_adj.
3. Probabilistic Bias Analysis (PBA / Monte Carlo Simulation): Assigns prior probability distributions (such as Beta, Triangular, Trapezoidal, or Log-normal distributions) to the bias parameters and draws 10,000 to 100,000 Monte Carlo iterations, combining systematic parameter uncertainty with random binomial/multinomial sampling error.
4. Multiple Bias Modeling (MBM) and Full Bayesian Bias Analysis: Simultaneously corrects all three bias domains in the reverse order in which they arose during data generation (first correcting misclassification, then selection bias, and finally unmeasured confounding) or estimates the joint posterior via Markov chain Monte Carlo (MCMC).

Consider an observed 2 × 2 contingency table comparing binary exposure E ∈ {1, 0} and binary disease D ∈ {1, 0}, with observed cell counts a (E = 1, D = 1), b (E = 1, D = 0), c (E = 0, D = 1), and d (E = 0, D = 0), yielding the observed odds ratio OR_obs = (a × d) / (b × c) and observed risk ratio RR_obs = [a / (a + b)] / [c / (c + d)].

1. Correcting Exposure Misclassification (Matrix Inversion Formula)

Suppose true exposure status is misclassified with Sensitivity Se₁ = Pr(E_obs = 1 | E_true = 1, D = 1) and Specificity Sp₁ = Pr(E_obs = 0 | E_true = 0, D = 1) among cases (D = 1) of total count N₁ = a + c, and with Se₀ and Sp₀ among non-cases (D = 0) of total count N₀ = b + d (when Se₁ = Se₀ and Sp₁ = Sp₀, misclassification is nondifferential; otherwise it is differential). Because a = Se₁ A_true + (1 − Sp₁)(N₁ − A_true), solving for the true exposed cases A_true and true exposed controls B_true yields the Greenland–Kleinbaum misclassification matrix inversion equations:

A_true = [ a − (1 − Sp₁) N₁ ] / ( Se₁ + Sp₁ − 1 ), and C_true = N₁ − A_true

B_true = [ b − (1 − Sp₀) N₀ ] / ( Se₀ + Sp₀ − 1 ), and D_true = N₀ − B_true

The misclassification-adjusted odds ratio is then OR_adj = (A_true × D_true) / (B_true × C_true).

2. Correcting Selection Bias (Selection Probability Ratios)

Suppose individuals in the underlying source population are selected into the analytic study sample (S = 1) with cell-specific selection probabilities S₁₁ = Pr(S = 1 | E = 1, D = 1), S₁₀ = Pr(S = 1 | E = 1, D = 0), S₀₁ = Pr(S = 1 | E = 0, D = 1), and S₀₀ = Pr(S = 1 | E = 0, D = 0). Because each observed cell equals the true source cell multiplied by its selection probability (for example, a_obs = A_source × S₁₁), the Selection Bias Factor (OR_sel) is:

OR_sel = ( S₁₁ × S₀₀ ) / ( S₁₀ × S₀₁ )

Dividing the observed odds ratio by the selection bias factor recovers the true source-population odds ratio (equivalent to inverse-probability of selection weighting):

OR_adj = OR_obs / OR_sel

3. Correcting an Unmeasured Binary Confounder (Schlesselman–Bross Formula)

Suppose a binary unmeasured confounder U ∈ {1, 0} has prevalence p₁ = Pr(U = 1 | E = 1) in the exposed group, prevalence p₀ = Pr(U = 1 | E = 0) in the unexposed group, and confounder–disease risk ratio RR_UD = Pr(D = 1 | U = 1, E) / Pr(D = 1 | U = 0, E). By Irwin Bross's (1966) and James Schlesselman's (1978) external adjustment formula, the multiplicative confounding bias factor (BF_U) and the confounder-adjusted risk ratio (RR_adj) are:

BF_U = [ RR_UD × p₁ + (1 − p₁) ] / [ RR_UD × p₀ + (1 − p₀) ], and RR_adj = RR_obs / BF_U

Why Naive Misclassification Inversion Can Produce Negative Cell Counts

In a deterministic sensitivity analysis, if the user draws a specificity value Sp₁ such that (1 − Sp₁) > a / N₁ (meaning the assumed false-positive rate exceeds the observed prevalence of positive tests), the matrix inversion formula A_true = [a − (1 − Sp₁)N₁] / (Se₁ + Sp₁ − 1) yields an impossible negative cell count (A_true < 0).

To guarantee valid cell counts and integrate multivariable covariate adjustments, Timothy Lash and Matthew Fox recommend converting sensitivity and specificity into Positive Predictive Value (PPV) and Negative Predictive Value (NPV) at each Monte Carlo iteration k, and simulating record-level true exposures via Bernoulli trials:

1. Draw (Se₁^(k), Sp₁^(k), Se₀^(k), Sp₀^(k)) from Beta or Trapezoidal prior distributions (incorporating a high positive correlation ρ ≈ 0.80 between Se₁ and Se₀ when misclassification is believed to be nearly nondifferential).
2. Compute expected true exposed counts A_true^(k) and B_true^(k) (discarding or re-parameterizing any draw outside data-admissible bounds) and compute PPV_d^(k) = Pr(E_true = 1 | E_obs = 1, D = d) and NPV_d^(k) = Pr(E_true = 0 | E_obs = 0, D = d).
3. Simulate the bias-corrected log odds ratio ln(OR_syst^(k)) and add random sampling error by drawing a standard normal deviate Z^(k) ∼ N(0, 1):

ln(OR_total^(k)) = ln(OR_syst^(k)) − Z^(k) × SE_obs

The 2.5th, 50th, and 97.5th percentiles of exp[ln(OR_total^(k))] across 100,000 iterations define the median bias-adjusted estimate and the 95% total-error simulation interval.

In what order should multiple biases be corrected in Multiple Bias Modeling?

In a study affected simultaneously by exposure misclassification, selection bias, and unmeasured confounding, bias adjustments must be applied in the exact reverse order of the physical data-generating process:

1. In reality, nature first generates the confounded association in the source population; next, selection restricts the source population to the study sample; finally, measurement/classification occurs on the enrolled participants.
2. Therefore, during analysis, the investigator must first correct the observed data for misclassification, second correct the reclassified data for selection bias, and third adjust the source-population estimate for unmeasured confounding.

How does Quantitative Bias Analysis differ from the E-value?

The E-value is a bounding sensitivity metric that requires zero external assumptions about confounder prevalence or direction, asking only: "How strong would an unmeasured confounder have to be to reduce our observed risk ratio to 1.0?" In contrast, Quantitative Bias Analysis actively incorporates external knowledge (validation studies, survey prevalences, selection logs) to model the actual expected bias and produce a new bias-adjusted point estimate and 95% simulation interval across misclassification, selection bias, and confounding.

1. Lash, T. L., Fox, M. P., & Fink, A. K. (2009). Applying Quantitative Bias Analysis to Epidemiologic Data. Springer. https://doi.org/10.1007/978-0-387-87959-8
2. Greenland, S. (2005). Multiple-bias modelling for analysis of observational data. Journal of the Royal Statistical Society: Series A (Statistics in Society), 168(2), 267–306. https://doi.org/10.1111/j.1467-985X.2004.00349.x
3. Lash, T. L., Fox, M. P., MacLehose, R. F., Maldonado, G., McCandless, L. C., & Greenland, S. (2014). Good practices for quantitative bias analysis. International Journal of Epidemiology, 43(6), 1969–1985. https://doi.org/10.1093/ije/dyu149
4. Fox, M. P., MacLehose, R. F., & Lash, T. L. (2021). Applying Quantitative Bias Analysis to Epidemiologic Data (2nd ed.). Springer. https://doi.org/10.1007/978-3-030-82673-4