From a Sensitivity Parameter to a Hidden Covariate
Rosenbaum’s Gamma model describes how much hidden bias in treatment assignment a matched analysis allows. Gamma is useful as a sensitivity parameter, but it can be hard to interpret without a concrete scenario: what kind of unmeasured covariate might produce a given value? A binary covariate provides a clear way to translate between the model and a familiar odds ratio.
The translation concerns treatment assignment, conditional on the measured covariates used for matching. A hidden covariate’s association with the outcome is a separate quantity. Both associations matter for confounding, but only the treatment-assignment association determines the Gamma value in the simple binary-covariate model developed here.
For units \(i\) and \(k\) in the same matched set, a common bound is written as \(1/\Gamma\leq \text{odds}_i/\text{odds}_k\leq\Gamma\). The bound allows either unit to have the larger odds. A value of \(\Gamma=1\) means no hidden variation in treatment odds within a matched set under the model; larger values permit greater variation. Gamma is not a probability, and it does not say that treatment probabilities differ by a factor of Gamma.
The Logistic Translation
Suppose treatment assignment, conditional on measured covariates \(X\) and the binary covariate \(U\), follows a logistic model:
Here \(\alpha(X)\) represents the baseline log odds at the measured covariate profile, and \(\beta\) is the change in log treatment odds when \(U\) changes from 0 to 1. Exponentiating a log-odds difference gives an odds ratio. Thus, at any fixed \(X\), the treatment odds ratio for \(U=1\) versus \(U=0\) is \(\exp(\beta)\). Since Gamma is at least one and the bound is symmetric in which unit has higher odds, the corresponding value is \(\Gamma=\exp(|\beta|)\).
This equality relies on the stated model: a binary \(U\), a common coefficient \(\beta\), and no interaction that changes the \(U\) effect across matched sets. If the effect varies, a single Gamma must instead be interpreted as a bound large enough to cover the largest relevant odds ratio. The model translates a hypothesized hidden association into a sensitivity value; observed matched data alone generally cannot identify that association.
Worked Example: Converting a Logistic Coefficient to Gamma
Suppose a hypothetical treatment-assignment model has \(\beta=\log(2.5)\) for an unmeasured binary risk marker \(U\). The model is assumed to apply within matched covariate profiles, with the same coefficient at every profile.
The treatment odds ratio comparing \(U=1\) with \(U=0\) is \(\exp(\beta)=\exp(\log(2.5))=2.5\). Therefore the hidden-bias value represented by this model is \(\Gamma=\exp(|\beta|)=2.5\). A unit with \(U=1\) has 2.5 times the treatment odds of an otherwise matched unit with \(U=0\).
If instead \(\beta=-\log(2.5)\), the odds ratio in the \(U=1\) versus \(U=0\) direction is \(1/2.5=0.400\). Reversing the comparison gives 2.5, so the symmetric bound is again \(\Gamma=2.5\). The sign tells which covariate group has higher treatment odds; Gamma records the permitted magnitude.
An odds ratio is not a risk ratio or a difference in treatment probabilities. For a baseline treatment probability \(p_0=P(T=1\mid X,U=0)\), a treatment odds ratio \(\Gamma\) implies
when \(U=1\) has the higher treatment odds. This follows by writing the baseline odds as \(p_0/(1-p_0)\), multiplying by \(\Gamma\), and converting the resulting odds back to a probability. Consequently, the probability difference associated with a fixed Gamma depends on the baseline probability.
Worked Example: Calculating Gamma From Treatment Probabilities
In a hypothetical matched population, suppose the treatment probability is 0.60 for \(U=1\) and 0.30 for \(U=0\), at the same measured covariate profile. Calculate the conditional treatment odds ratio.
For \(U=1\), the treatment odds are \(0.60/(1-0.60)=0.60/0.40=1.5\). For \(U=0\), they are \(0.30/(1-0.30)=0.30/0.70\approx0.4286\). Therefore, \[ \text{OR}_{T,U\mid X}=\frac{1.5}{0.4286}\approx3.50. \] Under the constant-effect binary-covariate model, \(\Gamma=3.50\).
The probability difference is \(0.60-0.30=0.30\), but that difference is not Gamma. The calculation uses odds, and it is conditional on the same \(X\). If these probabilities were aggregated across different covariate profiles, the resulting marginal odds ratio would not, in general, equal the conditional odds ratio used to interpret Gamma.
Why the Outcome Association Is a Separate Ingredient
A hidden covariate can confound a treatment-outcome comparison only if it is related to treatment assignment and to the outcome, after accounting for measured covariates. For the outcome relationship, the relevant question is whether \(U\) predicts the potential outcomes \(Y(1)\) or \(Y(0)\) within levels of \(X\). An association with treatment alone can produce treatment imbalance without outcome confounding; an outcome predictor unrelated to treatment within \(X\) does not create treatment-group imbalance through that covariate.
The outcome association does not enter the formula \(\Gamma=\exp(|\beta|)\). It indicates how consequential the treatment imbalance might be for an outcome comparison. In particular, Gamma alone does not determine the size or direction of bias in an effect estimate. The outcome scale, baseline outcome risks, prevalence of \(U\), and treatment-outcome structure also matter.
Worked Example: Treatment and Outcome Odds Ratios Are Different Quantities
Consider a hypothetical population at a fixed measured covariate profile. In a sample of 200, treatment counts by \(U\) are as follows.
| Unmeasured covariate | Treated | Untreated |
|---|---|---|
| \(U=1\) | 60 | 40 |
| \(U=0\) | 30 | 70 |
The treatment odds are \(60/40=1.5\) for \(U=1\) and \(30/70\approx0.4286\) for \(U=0\). The odds ratio is \(1.5/0.4286=3.50\), equivalently \((60\times70)/(40\times30)=3.50\). Under the stated constant-effect model, this corresponds to \(\Gamma=3.50\).
Now suppose the outcome risks under a specified treatment condition are 0.40 for \(U=1\) and 0.10 for \(U=0\). The outcome odds ratio comparing these covariate groups is \[ \frac{0.40/(1-0.40)}{0.10/(1-0.10)} =\frac{0.40/0.60}{0.10/0.90} \approx\frac{0.6667}{0.1111}=6.00. \] This outcome odds ratio is 6.00, not 3.50. It describes the covariate-outcome association; the treatment odds ratio describes the assignment association and gives the Gamma calibration. Both numbers are needed to characterize this hypothetical hidden factor, but they answer different questions.
Same Gamma, Different Confounding
A useful way to see the limitation of Gamma alone is to hold treatment odds fixed while changing the outcome association. The following calculation illustrates how even the treatment-group prevalence of \(U\) depends on its prevalence in the population and the baseline treatment probability.
Worked Example: A Gamma of 2 With Different Outcome Associations
Assume \(P(U=1)=0.20\), and at a fixed \(X\), \(P(T=1\mid U=0)=0.20\). The baseline odds are \(0.20/0.80=0.25\). If \(\Gamma=2\), the odds for \(U=1\) are \(2(0.25)=0.50\), so \[ P(T=1\mid U=1)=\frac{0.50}{1+0.50}=\frac{1}{3}\approx0.3333. \] The treatment-odds ratio is 2, as specified.
The treated group has \(P(U=1\mid T=1)=[0.20(1/3)]/[0.20(1/3)+0.80(0.20)]\approx0.2941\). The untreated group has \(P(U=1\mid T=0)=[0.20(2/3)]/[0.20(2/3)+0.80(0.80)]\approx0.1724\). Thus the treatment groups differ in their prevalence of the hidden covariate.
Suppose first that the outcome risk is 0.30 for \(U=1\) and 0.05 for \(U=0\), regardless of treatment. The outcome odds ratio for \(U\) is \((0.30/0.70)/(0.05/0.95)\approx8.143\). With no treatment effect in this hypothetical setup, the treated group’s expected outcome risk is \(0.2941(0.30)+0.7059(0.05)\approx0.1235\). The untreated group’s expected risk is \(0.1724(0.30)+0.8276(0.05)\approx0.0931\). The crude risk difference is about \(0.0304\), arising from different \(U\) composition.
Keep the same treatment model and thus the same \(\Gamma=2\), but let the outcome risks be 0.10 for \(U=1\) and 0.05 for \(U=0\). The outcome odds ratio is now \((0.10/0.90)/(0.05/0.95)\approx2.111\). The crude risk difference attributable to the same prevalence imbalance becomes \([0.2941-0.1724](0.10-0.05)\approx0.0061\). Gamma is unchanged, but the outcome association—and therefore the amount of confounding in this example—is weaker.
Assumptions, Diagnostics, and Limits
To interpret a numerical Gamma as the effect of a single binary covariate, state the assumptions that make the translation meaningful:
- The comparison is conditional on the measured covariates defining the matched sets. A marginal association across different \(X\) profiles is not automatically the relevant quantity.
- The hidden covariate is coded consistently as \(U=0\) or \(U=1\), and treatment odds are modeled with a constant log-odds effect of \(U\), unless a larger bound is explicitly used for heterogeneous effects.
- The odds refer to treatment assignment, not outcome odds, risk ratios, or treatment probabilities.
- To act as a confounder, \(U\) must also be associated with the potential outcome(s) conditional on \(X\). Gamma does not measure that association.
In practice, an unmeasured covariate’s conditional treatment odds ratio cannot usually be checked directly. Diagnostics therefore focus on whether the proposed interpretation is coherent: specify the covariate, the direction of its treatment association, the conditioning set, and a plausible outcome association. Check that an odds ratio has not been substituted for a risk ratio, and that the comparison is not based on pooled groups with different measured-covariate distributions. These checks clarify a sensitivity scenario; they do not verify that the hidden covariate actually has the proposed values.
The main alternative is to measure the covariate and adjust for it in the design or analysis, if valid measurement and adequate overlap are available. A sensitivity analysis instead asks how conclusions behave over assumed degrees of hidden treatment imbalance. A probabilistic bias analysis can additionally specify distributions for uncertain covariate prevalences and associations, but it requires assumptions beyond a single Gamma value. Neither approach makes an unmeasured factor identifiable from the matched data.
The translation is valuable because it makes a sensitivity value interpretable as a hypothetical treatment-odds contrast. Its limit is equally important: the same Gamma can arise from different covariate prevalences and outcome associations, and those scenarios need not produce the same confounding. Gamma is a way to index hidden assignment bias, not a complete description of an omitted variable.
Check Your Understanding
Work each question, then compare your reasoning with the solutions below.
- A binary \(U\) has a treatment log-odds coefficient \(\beta=\log(4)\), constant across matched covariate profiles. What is Gamma?
- At a fixed \(X\), treatment probabilities are 0.50 for \(U=1\) and 0.20 for \(U=0\). Calculate the treatment odds ratio and the corresponding Gamma under the constant-effect model.
- If the treatment odds ratio for \(U\) is 2, does that mean the treatment probability doubles? Explain using a baseline probability of 0.10.
- Two hidden covariates imply the same treatment Gamma, but one has a stronger association with the outcome. Which implies more confounding automatically?
- What two conditional associations are needed for a binary covariate to be a potential confounder in this setting?
Solutions. (1) \(\Gamma=\exp(|\log(4)|)=4\). (2) The odds are \(0.50/0.50=1\) and \(0.20/0.80=0.25\), so the ratio is \(1/0.25=4\), giving \(\Gamma=4\). (3) No. Multiplying baseline odds \(0.10/0.90=0.1111\) by 2 gives odds \(0.2222\), hence probability \(0.2222/(1+0.2222)\approx0.1818\), not 0.20. (4) Gamma alone does not determine which implies more confounding. The outcome association, prevalence, and other features also matter. (5) \(U\) must be associated with treatment assignment and with the potential outcome(s), each conditional on the measured covariates \(X\).