Rosenbaum’s Gamma Model for Hidden Bias

Why Specify a Hidden-Bias Model?

Matching can make treated and comparison units similar on measured covariates, but it cannot establish that they are also similar on every unmeasured factor related to treatment assignment. The earlier material on matching motivates sensitivity analysis for this reason: an analysis should make clear how much residual selection into treatment it allows, rather than quietly assuming that matching removed it all.

Rosenbaum’s Gamma model expresses that sensitivity as a bound on how different two units’ treatment-assignment odds may be within the same matched set. It does not specify a particular unmeasured variable, estimate hidden bias from the matched data, or by itself calculate an outcome effect. Its role here is to state a transparent assumption about treatment assignment.

Definition: Let $i$ index matched sets and $j$ index units within a set. Let $Z_{ij}=1$ indicate that unit $j$ in set $i$ receives treatment, and let $\pi_{ij}=P(Z_{ij}=1\mid\text{the matched set and measured covariates})$. Rosenbaum’s Gamma model assumes that, for every pair $j,k$ in the same set, the treatment-assignment odds satisfy $$ \frac{1}{\Gamma}\ \leq\ \frac{\pi_{ij}/(1-\pi_{ij})}{\pi_{ik}/(1-\pi_{ik})} \ \leq\ \Gamma, \qquad \Gamma\geq 1. $$

The bound applies to odds, not directly to probabilities. A value such as $\Gamma=2$ says that, within a matched set, one unit’s treatment-assignment odds can be at most twice another unit’s odds, or as low as half those odds. It does not say that treatment probability differs by two percentage points or that treatment is twice as likely in probability terms.

The comparison is within a matched set, conditional on the measured covariates represented in the design. The model allows the baseline level of treatment assignment to differ between sets. Its bound concerns the remaining relative differences among units in the same set.

A Latent-Score Parameterization

A convenient way to express the bound is to introduce an unobserved assignment score $u_{ij}$ between zero and one. In a logistic parameterization, write

\operatorname{logit}(\pi_{ij}) =\log\!\left(\frac{\pi_{ij}}{1-\pi_{ij}}\right) =\kappa_i+\gamma u_{ij}, \qquad 0\leq u_{ij}\leq 1, \qquad \gamma=\log(\Gamma).

Here $\kappa_i$ is a set-specific baseline log-odds and $\gamma$ controls how strongly the unobserved score can shift log-odds. The score is a sensitivity device: it can summarize residual differences in treatment assignment, and need not be an actually measured covariate. Because $u_{ij}$ lies between zero and one, the largest possible difference between two units’ scores is one.

Why the Parameterization Gives the Odds Bound

Subtract the two log-odds equations for units $j$ and $k$ in the same set. Their common intercept cancels, leaving the difference in log-odds as $\gamma(u_{ij}-u_{ik})$. Since both scores lie in $[0,1]$, their difference lies between $-1$ and $1$. Exponentiating therefore bounds the ratio of odds between $\exp(-\gamma)$ and $\exp(\gamma)$. With $\Gamma=\exp(\gamma)$, this is exactly the stated Gamma bound.

-\gamma\leq \gamma(u_{ij}-u_{ik})\leq\gamma \quad\Longrightarrow\quad \exp(-\gamma)\leq \frac{\pi_{ij}/(1-\pi_{ij})}{\pi_{ik}/(1-\pi_{ik})} \leq\exp(\gamma), \qquad \exp(\gamma)=\Gamma.

This derivation explains the interpretation of Gamma: it is the maximum multiplicative difference in assignment odds permitted by the model. It is not a parameter that the matched data identify on their own. The analyst selects values of Gamma to describe different sensitivity assumptions.

Worked Example: Checking Whether a Pair Fits Gamma Equal to Three

Suppose two units in a hypothetical matched set have treatment-assignment probabilities $0.25$ and $0.50$, conditional on the design information. Calculate their odds ratio and determine the smallest Gamma that permits this pair.

The first unit’s odds are $0.25/(1-0.25)=0.25/0.75=1/3$. The second unit’s odds are $0.50/(1-0.50)=1$. The larger odds divided by the smaller are

\frac{1}{1/3}=3.

Thus the odds ratio is 3. The pair satisfies the Gamma bound whenever $\Gamma\geq 3$, and the smallest permitted value is $\Gamma=3$. For example, $\Gamma=2$ would not allow this pair, because the odds ratio exceeds the specified upper bound.

This calculation describes compatibility with the assignment model only. It does not show that the probabilities are known from a real matched study, nor does it say anything by itself about the treatment’s effect on an outcome.

Worked Example: Calculating Probabilities From a Latent Score

Consider a hypothetical matched pair under the logistic parameterization with $\Gamma=3$, $\kappa_i=-1$, and scores $u_{i1}=0.2$ and $u_{i2}=0.8$. Calculate the two assignment probabilities and check the odds ratio.

First, $\gamma=\log(3)\approx1.0986$. For the first unit, the log-odds are $-1+\log(3)(0.2)\approx-0.7803$. Its odds and probability are

\exp(-0.7803)\approx0.4583, \qquad \pi_{i1}=\frac{0.4583}{1+0.4583}\approx0.3143.

For the second unit, the log-odds are $-1+\log(3)(0.8)\approx-0.1211$. Thus

\exp(-0.1211)\approx0.8859, \qquad \pi_{i2}=\frac{0.8859}{1+0.8859}\approx0.4698.

The ratio of the second unit’s odds to the first unit’s odds is $\exp\{\log(3)(0.8-0.2)\}=3^{0.6}\approx1.933$, which is between $1/3$ and $3$. The probabilities differ by about $0.1555$, but that difference is not what Gamma bounds. Gamma bounds their odds ratio; here that ratio is approximately $1.933$.

Interpreting Gamma Equal to One

When $\Gamma=1$, both sides of the odds bound equal one. Therefore every pair of units in a matched set must have equal treatment-assignment odds. Because the odds-to-probability transformation is one-to-one, their assignment probabilities are equal as well. In the latent-score representation, $\gamma=\log(1)=0$, so the unobserved score has no effect on relative assignment odds within the set.

This is the no-hidden-bias benchmark for the model. It does not mean that the study was randomized, that treatment assignment is independent of every possible factor in the population, or that there is no confounding in any broader sense. It states that, conditional on the matched-set design information, there is no within-set difference in assignment odds of the kind represented by this model. Nor must the common probability be the same across different sets: the intercept $\kappa_i$ can vary with $i$.

Worked Example: Gamma Equal to One Across Two Sets

Suppose a hypothetical model has $\Gamma=1$. In matched set A, both units have treatment probability $0.40$; in set B, both have probability $0.70$. Is this consistent with the model?

Within set A, each unit’s odds are $0.40/0.60=2/3$, so their odds ratio is 1. Within set B, each unit’s odds are $0.70/0.30=7/3$, also giving a within-set odds ratio of 1. Both sets satisfy the $\Gamma=1$ restriction. The probabilities differ between sets, which is allowed because the model places the equal-odds restriction within each set, not across sets.

If the two units in set A instead had probabilities $0.40$ and $0.50$, their odds would be $2/3$ and $1$, giving an odds ratio of $1.5$. That pair would violate $\Gamma=1$, even though both probabilities are below one-half.

Assumptions and Practical Checks

The Gamma model is a sensitivity assumption, not a fact that matching diagnostics can verify. Its use is clearest when the analyst states the following points explicitly:

Matched-set comparison: the odds restriction concerns units within the same matched set, conditional on the observed design information.
Bounded residual selection: unmeasured differences may affect treatment assignment, but their combined effect is assumed not to push the within-set odds ratio beyond the chosen Gamma.
Common sensitivity scale: the usual model uses one Gamma value across sets. This is a simplifying restriction; it does not claim that the underlying unmeasured factors are identical in every set.
Assignment-model focus: Gamma describes treatment assignment, not how an unmeasured factor changes the outcome. A sensitivity analysis also relies on its outcome and study-design assumptions.

Measured-covariate balance can reveal mismatches on variables that were observed and checked, but it cannot prove that the Gamma bound holds for variables that were not measured. A well-balanced matched sample is therefore not evidence that Gamma must equal one. Conversely, choosing a larger Gamma does not prove hidden bias of that size exists; it asks how conclusions behave under a more permissive assignment model.

Common Mistakes and Alternatives

Incorrect interpretation	Correction
“Gamma is the ratio of treatment probabilities.”	Gamma bounds a ratio of treatment-assignment odds. Probability ratios need not equal odds ratios.
“Gamma equal to one proves the study was randomized.”	It is the no-differential-hidden-bias benchmark within matched sets under this model; it does not establish actual randomization.
“A good balance table establishes Gamma equals one.”	Balance checks assess measured covariates. They cannot test the absence of unmeasured confounding.
“Gamma is estimated from the outcome data.”	Gamma is a sensitivity parameter selected for analysis. The observed data do not determine the true amount of hidden bias.

The main alternative is to report a matched analysis under the no-hidden-bias assumption alone. That gives a useful reference point, but it hides how strongly conclusions rely on that assumption. Rosenbaum’s model makes the assignment assumption explicit and permits examination across specified Gamma values. This differs from ordinary covariate adjustment, which models or controls observed covariates but does not by itself supply a bound on unmeasured selection. Neither approach makes the unmeasured information observable.

The model is deliberately limited. A single common bound may be too restrictive if hidden selection differs substantially across sets; an overly large bound may be uninformative. The scalar score is a convenient representation of assignment-odds differences, not a guarantee that one real-world hidden variable exists. Most importantly, Gamma alone cannot establish whether a particular bound is plausible. Its value must be argued from substantive knowledge and reported as an assumption.

Key takeaway: Rosenbaum’s Gamma model bounds the ratio of treatment-assignment odds for any two units in the same matched set between $1/\Gamma$ and $\Gamma$. Gamma equal to one rules out differential hidden assignment odds within sets under the model; it does not certify randomization or the absence of all confounding.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

One unit has assignment probability $0.20$, and another in the same set has probability $0.50$. What is their odds ratio, and what is the smallest Gamma that permits the pair?
In the latent-score model, what does $\Gamma=1$ imply about two units’ probabilities in the same matched set? Does it force probabilities to match across different sets?
For $\Gamma=4$, what range of odds ratios is allowed when comparing either unit with the other?
Why does good balance on measured covariates not establish that Gamma equals one?
A researcher says, “Gamma is 2, so treatment probability is twice as high for one unit.” Identify the error.

Solutions. (1) The odds are $0.20/0.80=0.25$ and $0.50/0.50=1$; the larger-to-smaller odds ratio is $1/0.25=4$, so the smallest Gamma is 4. (2) Gamma equal to one forces equal odds, and thus equal probabilities, within each set. It does not force equality across sets because $\kappa_i$ may vary. (3) The allowed ratio is from $1/4=0.25$ to $4$. (4) Balance checks concern observed variables and cannot establish that unmeasured factors have no differential association with treatment assignment. (5) Gamma bounds relative odds, not relative probabilities; the probability ratio depends on the baseline probability.