Written and maintained by CASRAI Editorial Board
Last updated
Quantitative bias analysis (QBA) treats a suspected bias as something to model, not just name. Instead of writing “residual confounding cannot be ruled out” or “exposure was self-reported and may be misclassified” as a sentence in the limitations section and stopping there, QBA assigns the bias plausible parameter values — a sensitivity and specificity for a misclassified exposure, a plausible confounder–exposure and confounder–outcome association for an unmeasured confounder, a selection probability that differs by exposure and outcome status — and uses them to compute what the estimate would look like corrected for that specific bias. The output is a bias-adjusted estimate, or a distribution of them, that sits alongside the conventional result, not a paragraph asserting the bias “could” matter.
What QBA adds that a limitations paragraph doesn’t
A qualitative limitations statement answers one question: does this bias plausibly exist? QBA answers three more specific ones: how large would it have to be to matter, in which direction does it push the estimate, and — if you’re willing to state assumptions about it — what does the corrected estimate look like. Those are different questions with different evidentiary value. A reviewer or reader can act on “correcting for plausible non-differential exposure misclassification moves this odds ratio from 2.27 to somewhere in the 2.6–3.8 range” in a way they cannot act on “misclassification is a limitation of this study.”
This isn’t free: QBA trades an unstated, unexamined bias for a stated, examined one whose correctness depends entirely on the plausibility of the bias parameters you feed it. Garbage parameters produce a garbage bias-adjusted estimate with the same false confidence a garbage conventional estimate would carry. The parameters need their own justification — from a validation substudy, a comparable published study, or an honestly wide plausible range — not just plugged in because they make the arithmetic run.
Three levels of QBA
The methodology, most fully worked out in Lash, Fox and colleagues’ widely used text on applying quantitative bias analysis to epidemiologic data, is usually presented in three levels of increasing completeness.
Simple bias analysis
A single bias, at a single fixed set of assumed parameter values, corrected for algebraically. One sensitivity and specificity value for exposure misclassification; one pair of confounder–exposure/confounder–outcome associations for unmeasured confounding; one differential selection probability by exposure and outcome status. The output is one bias-adjusted point estimate. It’s the fastest way to ask “does this bias plausibly matter at all,” and the worked example below is a simple bias analysis.
Probabilistic bias analysis
Instead of one fixed value per bias parameter, each parameter is assigned a probability distribution reflecting genuine uncertainty about its true value — commonly a trapezoidal distribution (a plausible range with a narrower, more-likely-still range inside it), sometimes triangular, normal or uniform where that shape fits the evidence better. Repeated Monte Carlo draws from those distributions, each corrected algebraically the same way as the simple case, build up a distribution of bias-adjusted estimates. The result is a simulation interval: it reflects uncertainty about the bias parameters themselves, which is conceptually distinct from a conventional confidence interval, which reflects only random sampling error given the data as observed. A full accounting reports both, and they are not interchangeable with each other.
Multidimensional bias analysis
More than one bias source corrected jointly — for example, exposure misclassification and unmeasured confounding together, rather than one guide’s worth of arithmetic per bias treated as if the others weren’t happening at the same time. This is the most complete and the most assumption-heavy level: each additional bias source adds its own parameters, each needing its own justification, and the combined correction is only as trustworthy as the weakest-justified parameter in the set.
Worked example: correcting an odds ratio for exposure misclassification
The following is an illustrative composite dataset built for this walkthrough, not drawn from a real published study. All arithmetic was computed directly with a small reproducible script (a mulberry32-seeded Monte Carlo run for the probabilistic section, seed 20260829) rather than estimated by eye; the exact figures are reported below as computed.
Start with an observed 2×2 table relating a binary exposure, measured with error, to a binary outcome:
| Outcome + | Outcome − | Row total | |
|---|---|---|---|
| Exposed (observed) | 220 | 560 | 780 |
| Unexposed (observed) | 180 | 1,040 | 1,220 |
| Column total | 400 | 1,600 | 2,000 |
The crude, uncorrected odds ratio from this table is (220 × 1,040) / (560 × 180) = 2.2698.
Now suppose the exposure was assessed with a sensitivity of 0.85 and a specificity of 0.90 against a gold-standard measure — illustrative, assumed values for this walkthrough, applied here as non-differential (the same in both outcome groups). The standard matrix-correction formula for a binary exposure misclassified with known, non-differential sensitivity (Se) and specificity (Sp) recovers the true exposed count within each outcome stratum:
True exposed = (observed exposed − stratum total × (1 − Sp)) / (Se + Sp − 1)
Applied to each column of the table above (stratum totals 400 and 1,600):
- True exposed among Outcome+: (220 − 400 × 0.10) / 0.75 = 240.00
- True exposed among Outcome−: (560 − 1,600 × 0.10) / 0.75 = 533.33
Which gives a corrected table of 240 / 160 (Outcome+) and 533.33 / 1,066.67 (Outcome−), and a bias-adjusted odds ratio of (240 × 1,066.67) / (533.33 × 160) = 3.0000 — a 32.2% upward shift from the crude 2.2698. This is the textbook direction for non-differential misclassification of a binary exposure: it typically (not universally — see below) biases the estimate toward the null, so correcting for it moves the estimate away from the null.
Adding parameter uncertainty: a probabilistic bias analysis on the same table
The simple correction above assumed Se = 0.85 and Sp = 0.90 as fixed, known values — rarely realistic. Assigning each a trapezoidal distribution instead — Se over the range 0.75–0.95 with a most-likely range of 0.82–0.88, Sp over 0.80–0.98 with a most-likely range of 0.87–0.93 — and drawing 20,000 Monte Carlo samples (discarding any draw where Se + Sp ≤ 1, which is not identifiable, or where the corrected counts fall outside the possible range) produces this simulation interval for the bias-adjusted odds ratio:
- Median: 3.0591
- 2.5th–97.5th percentile simulation interval: 2.6118 – 3.7786
The simulation interval is wider than a conventional 95% CI computed on the crude table would be, because it carries an additional source of uncertainty — not knowing the exposure-classification error precisely — that a conventional CI is structurally blind to. Reporting the crude estimate with its conventional CI alone, when a validation study or literature-based Se/Sp range exists, understates true uncertainty.
The same data, a different assumed error structure: an honest mixed result
The correction above assumed the misclassification was non-differential — the same error rate regardless of outcome status. That assumption doesn’t always hold. Suppose instead the exposure was self-reported after the outcome was already known (a recall-bias-style setup common in case-control designs), so cases recall past exposure more completely than controls: sensitivity 0.90 among Outcome+ but only 0.75 among Outcome−, specificity held at 0.90 in both groups. Run through the same correction formula, applied separately per stratum:
- True exposed among Outcome+ (Se=0.90): 225.00
- True exposed among Outcome− (Se=0.75): 615.38
Bias-adjusted odds ratio: (225 × 984.62) / (175 × 615.38) = 2.0571 — a 9.4% downward shift from the same crude 2.2698, not an upward one. Same starting table, same crude estimate, a different but equally plausible assumed error structure, and the correction moves in the opposite direction from the non-differential case above. That is the honest lesson of this comparison, not a caveat to smooth over: differential misclassification does not reliably bias toward the null the way the non-differential case usually does, and the direction of a QBA correction is a direct function of which error structure you assumed — which is exactly why the parameter justification, not the arithmetic, is the part of QBA that carries the evidentiary weight. Two analysts with equally defensible but different assumptions about differential recall would report different-direction conclusions from the identical raw data.
The E-value: the accessible entry point for unmeasured confounding
The worked examples above correct for exposure misclassification. Unmeasured confounding is a distinct bias source with its own accessible QBA-adjacent tool: the E-value (VanderWeele TJ, Ding P, “Sensitivity Analysis in Observational Research: Introducing the E-Value,” Annals of Internal Medicine, 2017), the minimum risk-ratio-scale strength an unmeasured confounder would need with both exposure and outcome, beyond everything already adjusted for, to fully explain away the observed association. It answers a narrower question than a full bias analysis — how strong would an unmeasured confounder have to be, not what is the confounder-corrected estimate — which is exactly what makes it the accessible entry point: no assumed prevalence or distribution for the confounder is needed, only the formula E-value = RR + √(RR × (RR−1)) applied to the point estimate and to whichever confidence bound sits closer to the null. CASRAI’s ROBINS-I guide works a full E-value calculation end to end (RR = 1.80, giving an E-value of 3.00 for the point estimate and 1.92 for the CI bound) in the context of scoring Domain 1 (confounding) — the same formula and logic apply here as a stand-alone sensitivity check, independent of ROBINS-I’s seven-domain framework.
Reporting a quantitative bias analysis
- State every bias parameter and where it came from — a validation substudy against a gold standard, a comparable published study, or an explicitly reasoned plausible range. “We assumed sensitivity of 0.85” with no source is not meaningfully more rigorous than the qualitative limitations sentence QBA is supposed to improve on.
- Report the corrected estimate alongside the conventional one, not instead of it — a reader needs both to judge how much the correction changed the picture.
- Report a probabilistic simulation interval, not just a simple bias analysis point estimate, whenever real uncertainty exists about the parameters — which is nearly always. A single-point correction implies a false precision about values that are themselves estimated.
- State the direction of bias the correction implies, and whether that direction was assumed or derived — as the two misclassification variants above show, it is not safe to assume non-differential-style toward-the-null behavior without checking which error structure actually applies.
- Don’t run QBA on every conceivable bias as a formality. It’s worth the analytical and reporting overhead for a bias source that’s genuinely plausible and consequential for the study’s own exposure-outcome relationship — not a checklist item added to every discussion section.
Where this fits alongside other bias tools on CASRAI
QBA is a distinct move from most of the other bias-related content in this cluster: selection bias and non-response bias guides describe how those specific biases arise and how to detect or design against them, largely qualitatively; ROBINS-I is a structured qualitative risk-of-bias judgement tool (with the E-value as one quantitative add-on, as above); DAGs for confounder selection address which variables to adjust for in the first place, not how wrong the estimate is if an unmeasured one was missed; and endogeneity and its remedies covers the econometric-tradition version of the confounding problem with design-based and instrumental-variable fixes rather than a bias-parameter correction. QBA is the tool for the specific, narrower question this page covers: given a bias you can’t design away or adjust away, how large is it and which direction does it push your result — stated as a number, not a sentence.
Frequently asked questions
Is quantitative bias analysis the same as a sensitivity analysis?
QBA is one specific family of sensitivity analysis — the family that models a named, structural bias source (misclassification, unmeasured confounding, selection) with explicit parameters. “Sensitivity analysis” more broadly also covers things like re-running a model under alternative specifications or exclusion criteria, which isn’t QBA unless it’s specifically parameterizing a bias mechanism.
Do I need individual-level data to run a simple bias analysis?
No — the matrix-correction approach in the worked example above only needs the observed 2×2 (or larger) table and the assumed bias parameters; it doesn’t require re-analyzing individual records. Probabilistic bias analysis is the same, run many times over sampled parameter draws.
Where do plausible sensitivity and specificity values come from if there’s no internal validation substudy?
The next-best source is a comparable published validation study of the same or a similar exposure measure in a similar population — self-report versus biomarker or medical-record validation studies exist for many common exposures. Absent either, an honestly wide plausible range run through a probabilistic bias analysis is preferable to a single guessed value, precisely because it doesn’t overstate precision the analyst doesn’t actually have.
Can quantitative bias analysis replace a limitations discussion?
No — it extends it. A QBA-adjusted estimate for one named, parameterized bias doesn’t address every limitation a study has; the corrected estimate is still conditional on every other assumption in the analysis holding. Report it as an addition to, not a substitute for, an honest limitations discussion of what wasn’t or couldn’t be quantitatively modeled.
