Why Shared Confounding Pathways Matter
A proposed negative control exposure is useful only insofar as it can speak to the bias of concern. In particular, it should have a credible connection to causes that make treatment assignment non-comparable and also affect the outcome. A control can be strongly associated with treatment or outcome for other reasons and still provide little information about that confounding.
Causal diagrams make the question explicit: which variables cause treatment, which cause the outcome, and which causes also influence the proposed control? Covariate histories add a time dimension. They help establish whether the control occurred before treatment, whether treatment could have changed it, and whether the control plausibly reflects the same underlying processes as treatment assignment.
The basic confounding pathway is \(T\leftarrow U\to Y\). Because \(U\) affects both treatment and outcome, a comparison of treated and untreated units can partly reflect differences in \(U\), rather than the causal effect of treatment. If \(U\) also affects \(C\), then the control may reflect some of the same latent differences. This is a claim about a causal structure, not merely about a pairwise association in observed data.
For a proposed negative control exposure, the causal interpretation also depends on its not causing the outcome under study. In potential-outcome notation, this means that changing \(C\), while holding treatment fixed, would not change the outcome: \(Y(t,c)=Y(t,c')\) for the relevant values \(c,c'\). The control should also be measured before treatment, or otherwise be protected from being changed by treatment. These conditions make it less likely that a control–outcome relationship is itself a treatment effect or a consequence of treatment.
Reading Pathways Alongside Histories
A diagram is a compact hypothesis about how variables are generated. It does not prove that a latent cause exists, nor does it show that the data contain enough information to measure it. Covariate histories help assess whether the diagram is credible. For each candidate control, ask what happened before treatment, what happened after it, and what recurring processes might have generated both the control and treatment.
- Timing: Was \(C\) recorded before treatment began? Could earlier treatment or a treatment-related decision have changed it?
- Mechanism: Could the same latent factor, such as disease severity, access to care, or health-seeking behavior, affect \(C\), \(T\), and \(Y\)?
- Specificity: Is the proposed common cause tied to the actual confounding concern, or is the connection only a broad claim that “everything is related”?
- History quality: Are dates, prior diagnoses, prior treatments, and eligibility records detailed enough to support the proposed ordering and mechanism?
A history is particularly informative when it shows the same process recurring before treatment: for example, prior symptoms predicting both medication initiation and a later outcome, while also influencing use of a candidate control exposure. Conversely, a control that occurs after treatment, or whose timing is unclear, may be affected by treatment or by treatment-related care. In that case, its observed relationship to the outcome may not reflect the pre-treatment confounding pathway of interest.
Measured covariates can help refine this assessment. If a documented pre-treatment factor explains why \(C\) and \(T\) co-occur, the analyst should ask whether it is itself part of the confounding process or whether the apparent connection disappears once the history is understood. Observed covariates cannot establish that no unmeasured common cause remains. Their role here is to make the causal account more or less plausible, not to certify it.
Worked Example: A Shared Latent Cause in a Simple Model
Suppose a hypothetical study concerns a continuous treatment intensity \(T\). A latent health-seeking factor \(U\) affects treatment, outcome \(Y\), and a proposed control exposure \(C\). Consider the illustrative model \(T=0.7U+\varepsilon_T\) and \(C=0.6U+\varepsilon_C\), where \(U\), \(\varepsilon_T\), and \(\varepsilon_C\) are independent, each with mean zero, and each has variance 1. Assume the outcome model includes an effect of \(T\) and an effect of \(U\), but no causal effect of \(C\).
The diagram implied by the model includes \(U\to T\), \(U\to C\), and \(U\to Y\). Thus \(C\) and \(T\) share a cause that also contributes to outcome confounding. The covariance is
The variances are \(\operatorname{Var}(C)=0.6^2+1=1.36\) and \(\operatorname{Var}(T)=0.7^2+1=1.49\). Therefore,
In this hypothetical model, the positive correlation is consistent with the shared \(U\) pathway. It does not show that \(C\) causes \(Y\), and it does not establish that the same pathway operates in an actual study. The causal relevance comes from the stated structure and assumptions; the correlation is a model consequence, not proof of that structure.
A Short Derivation: How an Unmeasured Cause Biases a Slope
The connection between shared causes and treatment-effect bias can be seen in a linear model. Suppose \(Y=\tau T+\beta U+\varepsilon_Y\), where \(\tau\) is the causal effect per unit of treatment, \(U\) affects the outcome, and \(\varepsilon_Y\) has mean zero and is uncorrelated with \(T\). If a regression uses \(T\) alone, its population slope is \(\operatorname{Cov}(T,Y)/\operatorname{Var}(T)\). Substituting the outcome model gives
The second term is the omitted-variable bias in this model. It vanishes if \(\beta=0\) (the latent variable does not affect the outcome) or if \(\operatorname{Cov}(T,U)=0\) (it does not affect treatment in this linear setting). A proposed control is relevant to this bias only if it connects to the cause that makes this term nonzero. Merely predicting \(Y\) is not enough.
Worked Example: Checking a Candidate Against Covariate Histories
Consider a hypothetical medication study. Researchers propose prior attendance at a preventive appointment as control exposure \(C\). They suspect that unmeasured health-seeking behavior may influence both the appointment and medication initiation \(T\), and may also affect hospitalization \(Y\). Before examining any associations, they review dated records and classify patients by documented access to care \(H\).
| Documented access | Control exposure \(C\) | Number of patients | Medication initiations |
|---|---|---|---|
| High | Yes | 36 | 24 |
| High | No | 24 | 6 |
| Low | Yes | 8 | 2 |
| Low | No | 32 | 8 |
All recorded appointments occurred before medication initiation. Among patients with high documented access, the initiation rates are \(24/36=0.667\) for \(C=1\) and \(6/24=0.250\) for \(C=0\), a difference of \(0.667-0.250=0.417\). In the low-access stratum, the rates are \(2/8=0.250\) and \(8/32=0.250\), a difference of \(0.000\).
The history supports temporal ordering: medication could not have caused the earlier appointment. The pattern in the high-access stratum is compatible with a connection between the control and treatment assignment, while the equal rates in the low-access stratum caution against treating that connection as uniform. Health-seeking behavior could still be a common cause of appointment attendance, medication initiation, and hospitalization, but these counts do not demonstrate that it is. The analyst would need a defensible mechanism linking the same latent process to all three variables and would need to assess whether the appointment itself can affect hospitalization. The table is a diagnostic clue, not a causal test.
When an Association Points to the Wrong Cause
A control can be associated with the outcome because it shares a cause with the outcome that does not affect treatment. Such a control may look outcome-relevant but fail to address the treatment–outcome confounding pathway. The distinction is easiest to see by separating latent causes rather than grouping all unmeasured factors into one label.
Worked Example: An Outcome-Related Control That Misses Confounding
Suppose \(U_1\) and \(U_2\) are independent latent variables with variance 1. Let the treatment, control, and outcome follow the hypothetical model
where the three errors are mutually independent, independent of \(U_1,U_2\), and each has variance 1. Here \(U_1\) confounds treatment and outcome, while \(U_2\) affects the control and outcome but not treatment. The control has no causal effect on \(Y\).
Because \(C\) depends on \(U_2\) and independent error, while \(T\) depends on \(U_1\) and independent error, \(\operatorname{Cov}(C,T)=0\). But substituting the treatment equation into the outcome equation gives \(Y=1.05U_1+0.5\varepsilon_T+0.6U_2+\varepsilon_Y\). Thus
So \(C\) is associated with \(Y\) through \(U_2\), yet it does not share the \(U_1\) pathway that confounds \(T\) and \(Y\). To verify the treatment bias in the treatment-only linear slope, \(\operatorname{Var}(T)=0.5^2+1=1.25\) and \(\operatorname{Cov}(T,U_1)=0.5\). The omitted-variable bias is
so the population treatment-only slope is \(0.5+0.32=0.82\), rather than the causal coefficient \(0.5\). The association of \(C\) with the outcome does not reveal this particular bias because \(C\) reflects \(U_2\), not \(U_1\). In practice, the latent causes are not observed, so the example illustrates why the proposed causal pathways—not association alone—must guide control selection.
Practical Checks and Limits
A disciplined assessment can be organized in a few steps. First, name the specific unmeasured cause or process suspected of creating confounding. Next, draw its links to treatment and outcome, then ask whether the candidate control plausibly shares that cause. Finally, check the dated history for treatment effects on the control, alternative mechanisms, and gaps in measurement.
Describe how an unmeasured factor could affect both treatment assignment and the outcome, rather than referring vaguely to “residual confounding.”
Ask whether the same factor, or a justified connected process, could also generate the proposed control exposure.
Use covariate histories to assess whether the control precedes treatment and whether another cause could explain its relationship to treatment or outcome.
Separate what the observed history supports from what remains an assumption about latent causes and the control’s lack of causal effect on the outcome.
This pathway-based assessment complements, but does not replace, other approaches. Measured-confounder adjustment targets measured common causes directly; a negative control can provide a diagnostic probe of whether some bias remains, but it does not automatically remove that bias. A negative control outcome asks a different question about the outcome side of the causal structure. Sensitivity analyses quantify how conclusions could change under specified unmeasured-confounding scenarios, but they do not establish which latent pathway is true.
The main limitation is fundamental: an unmeasured cause cannot usually be verified from the observed data alone. A plausible history may support a shared-pathway story, but cannot prove it; an association between the control and treatment may arise for another reason, and no observed association does not by itself rule out a shared cause. Causal diagrams are therefore best treated as explicit, criticizable assumptions. Record the proposed mechanism and the alternative explanations, rather than presenting pathway relevance as a discovered fact.
Check Your Understanding
Answer each question using the pathway and timing logic developed above. Solutions follow the questions.
- A latent factor \(U\) affects treatment and the outcome. A proposed control \(C\) is affected by a separate factor \(V\), which affects the outcome but not treatment. Does \(C\)'s outcome association establish relevance to the treatment confounding?
- In a covariate history, a proposed control exposure is recorded only after treatment starts. What key pathway concern should be raised?
- In the first worked example, why is the correlation \(0.2950\) not proof that the proposed control shares an unmeasured confounder in an actual study?
- In the final worked example, what is the omitted-variable bias in the treatment-only slope, and which latent variable produces it?
Solutions. 1. No. \(C\) may be associated with \(Y\) through \(V\), but if \(V\) does not affect treatment, that association does not establish a shared treatment–outcome confounding pathway. 2. Treatment may have changed \(C\), or treatment-related events may have changed it; the observed control may therefore be post-treatment rather than a marker of pre-treatment causes. 3. The value follows from the assumed structural model. Observed correlation alone cannot verify that \(U\) causes both variables or exclude other explanations. 4. The bias is \(0.8\{0.5/1.25\}=0.32\). It is produced by \(U_1\), which affects both treatment and outcome; \(U_2\), despite its relationship to the control and outcome, does not generate this treatment bias.