Why Read Negative Controls From a Graph?
A negative control can help reveal a problem with a causal analysis, but its observed association is not self-interpreting. A treatment associated with a negative control outcome might reflect a shared cause, a path opened by adjustment, or a mistaken belief that the outcome is unaffected by treatment. A negative control exposure associated with the outcome may likewise share causes with the outcome, even when it cannot cause that outcome. Directed acyclic graphs (DAGs) make these possibilities explicit by showing which variables cause which others and which paths are available for association.
The task here is to trace paths, not to decide from a graph alone whether a negative-control test will be statistically significant. A graph describes causal assumptions. It can show whether an association has a possible causal route or an open noncausal route, but it does not establish that the assumptions are true or quantify the resulting bias.
Paths, Causal Routes, and Open Routes
For example, \(T \rightarrow M \rightarrow Y\) is a causal path: treatment affects \(M\), which in turn affects the outcome. The path \(T \leftarrow U \rightarrow Y\) is noncausal. It connects treatment and outcome through a shared cause \(U\), so it can create an association even if treatment itself has no effect on the outcome.
A particularly important noncausal route is a backdoor path, which begins with an arrow pointing into \(T\). The shared-cause path \(T \leftarrow U \rightarrow Y\) is a backdoor path. If it is open, it is a confounding path. More generally, “open bias path” is useful because adjustment can open some noncausal paths that do not begin with an arrow into treatment. An open path can permit association; it does not guarantee a nonzero association, since effects can be weak or cancel.
The path rule gives a quick reading procedure. On a path such as \(T \leftarrow U \rightarrow Y\), \(U\) is a noncollider, so conditioning on \(U\) blocks the path. On \(T \rightarrow C \leftarrow U \rightarrow Y\), \(C\) is a collider, so the path is initially blocked; conditioning on \(C\), or on a descendant of \(C\), can open it. Conditioning on a mediator on a causal path can instead block part of that causal route. Whether to adjust for a variable therefore depends both on the target effect and on its place in the graph.
Reading the Two Negative-Control Roles
Use \(Z\) for a negative control outcome and \(N\) for a negative control exposure. A negative control outcome is not causally affected by \(T\), so a correctly specified graph has no directed causal path from \(T\) to \(Z\). A negative control exposure is not a cause of \(Y\), so there is no directed causal path from \(N\) to \(Y\). These absences rule out particular causal explanations for associations; they do not rule out noncausal paths.
For a negative control outcome, trace paths from \(T\) to \(Z\). If \(T \leftarrow U \rightarrow Z\) is open, an association between \(T\) and \(Z\) can arise through the shared cause \(U\), despite the absence of a treatment effect on \(Z\). For a negative control exposure, trace paths from \(N\) to \(Y\). A path such as \(N \leftarrow U \rightarrow Y\) can produce an association without any causal effect of \(N\) on \(Y\). These graphs can make a negative-control association compatible with residual confounding; they do not prove that the same bias affects the treatment–outcome estimate.
An open path must be assessed under the actual adjustment set. A variable that is a noncollider on one path may block that path when conditioned on, while conditioning on a collider may create an association that was absent before. Graphs also help check the negative-control premise: a directed path from \(T\) to \(Z\), or from \(N\) to \(Y\), contradicts the corresponding no-causal-effect assumption.
Worked Example: A Shared Cause of Treatment and a Negative Control Outcome
Suppose, in a hypothetical cohort of 200 people, \(U\) denotes a high-risk condition, \(T\) denotes treatment, \(Y\) is the outcome of interest, and \(Z\) is a negative control outcome. The assumed graph includes \(U \rightarrow T\), \(U \rightarrow Y\), and \(U \rightarrow Z\), but no directed path from \(T\) to \(Z\). The following counts summarize treatment and \(Z\) within levels of \(U\).
| Risk condition | Treatment | \(Z\) positive | \(Z\) negative | Total |
|---|---|---|---|---|
| High \(U\) | Yes | 48 | 32 | 80 |
| High \(U\) | No | 12 | 8 | 20 |
| Low \(U\) | Yes | 2 | 18 | 20 |
| Low \(U\) | No | 8 | 72 | 80 |
Marginally, the positive-outcome risk is \(50/100=0.500\) among treated people and \(20/100=0.200\) among untreated people. The crude risk difference is
Within the high-risk group, the risks are \(48/80=0.600\) and \(12/20=0.600\); within the low-risk group they are \(2/20=0.100\) and \(8/80=0.100\). Thus, in these hypothetical data there is no within-group difference, while the marginal difference is 0.300, or 30 percentage points. The graph explains how this can happen: the open path \(T \leftarrow U \rightarrow Z\) connects treatment and the negative control outcome. The contrast is consistent with a shared-cause association, not a causal effect of treatment on \(Z\). Because \(U\) also causes \(Y\) in the assumed graph, it is also a potential confounder of the treatment–outcome relation.
This arithmetic illustrates a possible pattern; it does not validate the graph or establish that the same amount of bias affects an estimate for \(Y\). The graph’s arrows are assumptions, and the control-outcome association alone does not identify the treatment effect.
Conditioning Can Create an Open Bias Path
An open confounding path is not the only way a noncausal association can arise. A graph can show that adjustment itself opens a path. In particular, conditioning on a common effect (a collider) can make its causes statistically associated within the conditioned group. If one cause is treatment and the other is related to the outcome, this induced association can bias a treatment–outcome comparison.
Worked Example: Conditioning on a Collider
Consider a hypothetical population of 400 people. Let \(T\) and an unmeasured factor \(U\) be independent binary variables, each taking values 0 and 1 for exactly 200 people. Let \(C=1\) whenever \(T=1\) or \(U=1\), and let \(Y=U\). The graph is \(T \rightarrow C \leftarrow U \rightarrow Y\); there is no causal path from \(T\) to \(Y\).
Before conditioning, half of both treatment groups have \(U=1\), so the outcome risks are \(100/200=0.500\) in each group and the risk difference is \(0.500-0.500=0\). Now restrict the analysis to \(C=1\). Among treated people in this subset, both \(U=0\) and \(U=1\) occur: there are 100 of each, so the outcome risk is \(100/200=0.500\). Among untreated people with \(C=1\), \(U\) must equal 1, so the outcome risk is \(100/100=1.000\). The selected-subset risk difference is
The negative difference is not a treatment effect: by construction \(Y=U\), and \(T\) does not cause \(Y\). Conditioning on \(C\), the collider on \(T \rightarrow C \leftarrow U\), opens the path from \(T\) through \(C\) and \(U\) to \(Y\). This example shows why a graph audit must ask not only which variables are adjusted for, but also whether the adjustment set contains a collider or its descendant.
Negative Control Exposure: A Noncausal Route to the Outcome
Worked Example: A Shared Cause of a Negative Control Exposure and Outcome
In a second hypothetical setting, \(N\) is a negative control exposure and \(Y\) is the outcome. Let \(U\) be a common cause of \(N\), treatment \(T\), and \(Y\), with arrows \(U \rightarrow N\), \(U \rightarrow T\), and \(U \rightarrow Y\). Assume there is no directed path from \(N\) to \(Y\). The table gives counts for \(N\) and \(Y\) within levels of \(U\).
| Risk condition | \(N\) | \(Y\) positive | \(Y\) negative | Total |
|---|---|---|---|---|
| High \(U\) | Yes | 48 | 32 | 80 |
| High \(U\) | No | 12 | 8 | 20 |
| Low \(U\) | Yes | 2 | 18 | 20 |
| Low \(U\) | No | 8 | 72 | 80 |
Overall, the \(Y\)-positive risk is \(50/100=0.500\) for \(N=1\) and \(20/100=0.200\) for \(N=0\). The crude risk difference is
Within high \(U\), the risks are \(48/80=0.600\) and \(12/20=0.600\); within low \(U\), they are \(2/20=0.100\) and \(8/80=0.100\). The marginal association is therefore compatible with the open noncausal path \(N \leftarrow U \rightarrow Y\), even though the graph rules out a causal effect of \(N\) on \(Y\). Since \(U\) also causes \(T\), the same variable could create a backdoor path between treatment and outcome. A negative-control exposure association can thus flag a possible shared-cause problem, but it does not by itself quantify or prove bias in the treatment effect.
A Practical Graph-Reading Sequence
For each proposed negative control, inspect the graph in a fixed order. This keeps causal routes, confounding routes, and adjustment-induced bias routes distinct.
For a negative control outcome, check that no directed path runs from \(T\) to \(Z\). For a negative control exposure, check that no directed path runs from \(N\) to \(Y\).
List paths connecting \(T\) and \(Y\), \(T\) and \(Z\), or \(N\) and \(Y\), as appropriate. Mark shared causes and other possible common ancestors.
For the variables actually conditioned on, determine whether each noncollider is conditioned on and whether each collider or its descendant is conditioned on. Classify each path as blocked or open.
An open noncausal path makes a bias explanation possible. It does not establish the strength, direction, or presence of bias in a particular data set.
This sequence also helps expose a graph that does not support the intended interpretation. For example, if the proposed negative control outcome has a directed path from treatment, the no-treatment-effect premise is not represented by that graph. If a negative control exposure has a directed path to the outcome through an intermediate variable, it is not a negative control for that outcome under the stated definition.
Diagnostics, Alternatives, and Limits
Several warning signs deserve attention when reading or using a DAG. First, a claimed negative control may have a plausible causal path that the graph omits. Second, an adjustment set may leave a backdoor path open, or may condition on a collider. Third, a graph may omit an important common cause or use the wrong temporal order. Reviewing the variables’ timing and causal roles is essential; the graph itself cannot detect an omitted variable from the data without additional assumptions.
DAG reasoning complements, rather than replaces, other approaches. Covariate balance diagnostics in propensity score analyses describe measured covariate balance, but do not by themselves show that all relevant backdoor paths are blocked. A negative-control association is an empirical probe, whereas a DAG is a qualitative statement about causal structure. Neither alone provides a numerical correction for unmeasured confounding. Instrumental-variable analyses use different identifying assumptions and do not make a negative-control graph valid automatically. The introductory material on causal inference provides broader context for these distinctions.
The central limitation is that a graph encodes assumptions, not data-derived proof. In particular, no edge in a drawn graph may mean “assumed absent,” not “tested and shown absent.” Even when an open bias path is present, the association along it may be small or may cancel against another path. Conversely, a null negative-control association does not prove that all bias paths are closed. Path analysis clarifies what could generate an association and what adjustment might do; it does not settle those empirical questions by itself.
Common Mistakes
- Calling every open path confounding. A confounding path usually means an open backdoor path. Conditioning can also open noncausal paths that do not begin with an arrow into treatment. Classify the path and how it became open.
- Treating a negative-control association as proof of a particular bias. An open shared-cause path makes that explanation possible, but does not establish which cause generated the observed association or its magnitude.
- Assuming a negative control must be unassociated with everything. Its defining restriction rules out a specified causal effect; shared causes can still create associations.
- Adjusting for every measured variable. Conditioning on a collider or its descendant can open a path, and conditioning on a mediator can change the causal effect being estimated.
- Reading an absent arrow as an empirical finding. It is a causal assumption that should be justified, not a fact established by the diagram’s appearance.
Check Your Understanding
Classify the paths before interpreting any association. Solutions follow the questions.
- In \(T \leftarrow U \rightarrow Y\), what kind of path is this, and what happens to it if the analysis conditions on \(U\)?
- In \(T \rightarrow C \leftarrow U \rightarrow Y\), is the path open when nothing is conditioned on? What happens if the analysis conditions on \(C\)?
- A proposed negative control outcome \(Z\) has a directed path \(T \rightarrow M \rightarrow Z\). Does the graph satisfy the no-treatment-effect requirement for \(Z\)?
- A negative control exposure \(N\) has no directed path to \(Y\), but \(N \leftarrow U \rightarrow Y\) is open. Can \(N\) and \(Y\) be associated? Explain.
Solutions. (1) It is an open backdoor, or confounding, path if \(U\) is not conditioned on. Conditioning on the noncollider \(U\) blocks it. (2) The path is initially blocked because \(C\) is a collider. Conditioning on \(C\) opens it, so it becomes an open noncausal bias path. (3) No. The directed path \(T \rightarrow M \rightarrow Z\) means the graph permits a causal effect of treatment on \(Z\), contradicting the negative-control-outcome restriction. (4) Yes. The shared cause \(U\) creates an open noncausal path, so an association is possible without a causal effect of \(N\) on \(Y\).