Combining Multiple Unmeasured Confounders in Bias Analysis

Why Multiple Hidden Confounders Need a Joint Analysis

Sensitivity analysis often begins by asking how much one unmeasured confounder could shift an estimated treatment effect. With several omitted variables, the natural next question is whether their separate shifts can simply be added. In general, that is not justified: the answer depends on how the confounders relate to one another, to treatment, and to the outcome.

A useful way to make these dependencies explicit is a linear bias decomposition. It represents the omitted variables together through their joint treatment associations and their outcome associations conditional on one another. The resulting formula is simple, but its parameters must be defined carefully. In particular, marginal associations for each confounder are not automatically the quantities that belong in the joint formula.

Definition: Let $D$ be a scalar treatment, $Y$ an outcome, $X$ measured covariates included in the analysis, and $U=(U_1,\ldots,U_K)^\top$ a vector of omitted confounders. In a linear outcome model, the joint omitted-variable bias is the difference between the treatment coefficient from a regression adjusting for $X$ but omitting $U$, and the treatment coefficient $\tau$ from the corresponding model that includes $U$. The decomposition below expresses this difference using the joint treatment associations of the omitted variables and their outcome coefficients conditional on one another.

The Joint Linear Bias Decomposition

Consider the full linear model

Y=\tau D+\gamma^\top X+\beta^\top U+\varepsilon.

Here $\beta_j$ is the outcome coefficient for $U_j$, holding treatment, measured covariates, and the other omitted variables fixed. For the decomposition to describe the omitted-variable bias, assume the model is correctly specified in its linear terms and that the error is orthogonal to the regressors: $E(\varepsilon\mid D,X,U)=0$. This is a statistical model for the outcome; a causal interpretation of $\tau$ additionally requires appropriate causal identification assumptions.

To define the treatment associations, first remove the linear contributions of $X$ from $D$ and each component of $U$. Equivalently, work with variables residualized on $X$. Regress each residualized $U_j$ on the residualized $D$, and collect the resulting slopes in $a=(a_1,\ldots,a_K)^\top$. Thus $a_j$ describes the linear association of $D$ with $U_j$ after adjustment for $X$, without adjusting that regression for the other components of $U$.

Theorem: Under the stated linear model and orthogonality condition, the treatment coefficient from the regression of $Y$ on $D$ and $X$, omitting $U$, equals $$ \tau+\beta^\top a. $$ Therefore the joint omitted-variable bias is $\beta^\top a=\sum_{j=1}^{K}\beta_j a_j$. The $\beta_j$ are outcome coefficients conditional on the other omitted variables; the $a_j$ are treatment-association slopes from the joint set of regressions of the omitted variables on $D$ and $X$.

A short derivation shows why the definitions matter. After residualizing on $X$, write the linear projections of the omitted variables on treatment as $U=aD+V$, where each component of $V$ is orthogonal to the residualized treatment. Substituting this projection into the full outcome model gives

Y=\tau D+\beta^\top(aD+V)+\varepsilon =(\tau+\beta^\top a)D+\beta^\top V+\varepsilon.

Both $V$ and $\varepsilon$ are orthogonal to the residualized treatment under the projection and model assumptions. Consequently, the coefficient of treatment in the regression that omits $U$ is $\tau+\beta^\top a$. This is an exact decomposition for the specified linear model, not a claim that arbitrary associations or nonlinear models obey the same formula.

Worked Example: Calculating Bias From Two Omitted Confounders

Suppose all variables have already been adjusted for the measured covariates $X$. For hypothetical data, let the residualized treatment have variance 1, let the omitted confounders have covariance matrix

\Sigma_U= \begin{pmatrix} 1&0.6\\ 0.6&1 \end{pmatrix}, \qquad \operatorname{Cov}(U,D)= \begin{pmatrix} 0.3\\ 0.2 \end{pmatrix},

and suppose their conditional outcome coefficients are $\beta_1=2$ and $\beta_2=1$. Find the joint bias.

The vector $a$ contains the slopes from regressing each confounder on treatment. In matrix form, $\operatorname{Cov}(U,D)=\operatorname{Var}(D)a$, so

a=\frac{\operatorname{Cov}(U,D)}{\operatorname{Var}(D)} = \begin{pmatrix} 0.3\\ 0.2 \end{pmatrix}.

Thus the joint bias is

\beta^\top a =(2)(0.3)+(1)(0.2) =0.6+0.2 =0.8.

If the full-model treatment coefficient were $\tau=1.5$ outcome units per treatment unit, the coefficient from omitting both confounders would be $1.5+0.8=2.3$ outcome units per treatment unit. This is a model-based bias calculation, not evidence that the hypothetical coefficient is causal. The values are mutually compatible: $\Sigma_U$ is positive definite, and the residual variance of $D$ after linear projection on $U$ is $1-\operatorname{Cov}(D,U)\Sigma_U^{-1}\operatorname{Cov}(U,D)=1-0.090625=0.909375$, which is positive.

Why One-at-a-Time Biases May Not Add

The joint formula is a sum of products, but that does not mean that any set of separately calculated single-confounder biases can be summed to recover the joint bias. To add exactly, each product must use the corresponding $\beta_j$ from the model including all omitted confounders and the corresponding $a_j$ from the joint treatment-association definition above.

A common one-at-a-time calculation instead omits one confounder while adjusting for the others. That calculation changes the treatment-association slope: the confounder being omitted is now related to treatment conditional on the other included confounder. Such slopes are not generally the $a_j$ in the joint decomposition. Also, outcome coefficients estimated without the other confounders need not equal the conditional $\beta_j$. The parameters have changed, so the separate shifts need not add to the shift from omitting all confounders together.

Worked Example: Comparing One-at-a-Time and Joint Omissions

Continue with the hypothetical covariance structure above and the full outcome coefficients $\beta_1=2$, $\beta_2=1$. The joint bias from omitting both variables was $0.8$. Now calculate the bias from omitting only $U_1$ while retaining $U_2$, and then from omitting only $U_2$ while retaining $U_1$.

When $U_2$ is included, the slope for $D$ in the regression of $U_1$ on $D$ and $U_2$ is

\frac{\operatorname{Cov}(D,U_1)-\operatorname{Cov}(D,U_2)\operatorname{Cov}(U_1,U_2)} {\operatorname{Var}(D)-\operatorname{Cov}(D,U_2)^2} = \frac{0.3-(0.2)(0.6)}{1-0.2^2} =\frac{0.18}{0.96} =0.1875.

The resulting bias from omitting $U_1$ alone is $(2)(0.1875)=0.375$. In the other one-at-a-time analysis, $U_1$ is included and the slope for $D$ in the regression of $U_2$ on $D$ and $U_1$ is

\frac{0.2-(0.3)(0.6)}{1-0.3^2} =\frac{0.02}{0.91} \approx0.02198.

The bias from omitting $U_2$ alone is therefore $(1)(0.02198)\approx0.02198$. These one-at-a-time biases sum to about $0.39698$, not the joint bias $0.8$. There is no contradiction: the first pair of calculations conditions on the other confounder, whereas the joint bias calculation omits both. They use different treatment-association slopes. The exact joint decomposition still holds when its own consistently defined parameters are used.

Joint Associations Must Be Mutually Compatible

Sensitivity parameters do not describe independent knobs that can always be assigned arbitrary values. The treatment associations and the correlations among omitted variables must be capable of arising from one joint distribution. For example, if the confounders are standardized, their proposed correlation matrix must be positive semidefinite. The joint covariance matrix including treatment must also be valid. Checking compatibility helps prevent a sensitivity analysis from relying on impossible combinations of associations.

Worked Example: Checking a Three-Confounder Specification

Consider three hypothetical standardized omitted variables with covariance matrix

\Sigma_U= \begin{pmatrix} 1&0.4&0.4\\ 0.4&1&0.6\\ 0.4&0.6&1 \end{pmatrix}.

Its characteristic polynomial factors as $(\lambda-0.4)(\lambda^2-2.6\lambda+1.28)$. The remaining roots are $(2.6-\sqrt{1.64})/2\approx0.6597$ and $(2.6+\sqrt{1.64})/2\approx1.9403$. All three eigenvalues are positive, so this covariance matrix is positive definite.

For an internally consistent hypothetical specification, choose the treatment-association slopes $a=(0.2,0.1,0.1)^\top$. The implied treatment-confounder covariances are

\operatorname{Cov}(U,D)=a = \begin{pmatrix} 0.2\\ 0.1\\ 0.1 \end{pmatrix}.

Take $\operatorname{Var}(D)=1$. The residual variance of $D$ after projection on $U$ is $1-a^\top\Sigma_U^{-1}a=1-0.040625=0.959375$, which is positive. Thus the specified covariance structure is compatible. If the conditional outcome coefficients are $\beta=(1,-1,2)^\top$, the joint bias is

\beta^\top a=(1)(0.2)+(-1)(0.1)+(2)(0.1)=0.3.

The negative contribution from $U_2$ partially offsets the positive contributions from $U_1$ and $U_3$. Adding absolute magnitudes would give $0.5$, which is not the signed joint bias. Direction matters, and the net shift depends on the signs as well as the sizes of the component products.

Practical Interpretation and Common Pitfalls

A defensible multiple-confounder sensitivity analysis should state what its parameters mean and which model they belong to. In the linear decomposition, report the outcome coefficients conditional on the full set of omitted variables and the treatment-association coefficients for that same joint specification. If those quantities are not directly known, a sensitivity analysis can examine plausible values, but the choices and their compatibility should be made explicit.

Specify the estimand and model.
Identify the treatment coefficient of interest, measured covariates being controlled, omitted confounders, and the assumed linear outcome model.

Define joint sensitivity parameters.
Specify the conditional outcome coefficient vector $\beta$ and the treatment-association vector $a$, rather than mixing marginal and conditional quantities.

Check compatibility.
Verify that the proposed covariance or correlation structure is valid and that the implied joint moments with treatment can exist.

Calculate and interpret the net shift.
Compute $\beta^\top a$ with signs retained, then add it to the coefficient from the full model to obtain the coefficient implied by omission under the assumptions.

Common mistake: “I estimated the bias from each hidden confounder separately, so their sum is the bias from omitting all of them.” This reasoning is generally wrong because the separate calculations may use associations conditional on different covariates, or marginal outcome associations that differ from the full-model coefficients. The correction is to use a coherent joint model and its $\beta^\top a$ decomposition. A sum is valid only when its terms are precisely the jointly defined contributions in that formula.

Finally, the decomposition quantifies model-based omitted-variable bias; it does not by itself establish that the adjusted coefficient is causal or that the assumed sensitivity parameters are realistic. Its value is that it makes the joint assumptions visible, shows which combinations of confounder associations shift the estimate, and clarifies why independent one-confounder scenarios can be misleading.

Check Your Understanding

Use the definitions and assumptions in this tutorial to answer the following.

In the joint decomposition, what does $\beta_j$ represent, and what does $a_j$ represent?
Why is the treatment-association vector calculated from the joint covariance structure of the omitted confounders rather than from each confounder in isolation?
Under what parameter definitions does the sum $\sum_j\beta_j a_j$ equal the joint omitted-variable bias?
In the two-confounder example, why did the one-at-a-time bias calculations fail to sum to the bias from omitting both variables?
What does a negative component $\beta_j a_j$ imply about that confounder’s contribution to the net bias, relative to a positive component?