Why a Selection Coefficient Is Not Automatically Causal
In a current-outcome selection model, the probability that an outcome is observed is specified conditionally on the outcome itself. This is useful for representing dependence between observation and an outcome that may be missing. It also creates a tempting but unwarranted interpretation: if the outcome coefficient is negative, one might say that increasing the outcome causes dropout. The coefficient alone does not support that conclusion.
A selection equation is first a statistical description of a conditional distribution. A causal effect, by contrast, compares outcomes under defined interventions. The distinction matters especially when the outcome follows treatment, when other variables affect both the outcome and response, or when the outcome is not something that can meaningfully be set by intervention.
The model conditions on \(Y\), including values of \(Y\) that are unobserved when \(R=0\). Thus its interpretation depends on the specified joint model and its assumptions about the unobserved outcomes. Even if the selection equation is correctly specified as a conditional model, its coefficient describes a conditional relationship, not how response would change if an investigator intervened to change a person's outcome.
What the Coefficient Says—and What a Causal Claim Would Say
A regression coefficient can be read on its own scale. In the equation above, \(\gamma\) is a log odds contrast per unit of \(Y\), at fixed \(A\) and \(H\). Exponentiating gives the corresponding conditional odds ratio. For example, \(\gamma=-0.6\) means the modeled odds of observation are multiplied by \(\exp(-0.6)\) for each one-unit increase in \(Y\), holding the other terms fixed.
A causal statement requires a different object. It would compare response under two specified interventions on the outcome. In potential-outcome notation, one might define \(R(a,y)\) as the response indicator that would occur under treatment \(a\) and an intervention setting the current outcome to \(y\). A causal contrast could then compare \(P\{R(a,y+1)=1\}\) with \(P\{R(a,y)=1\}\), or compare their odds. Such a contrast is meaningful only if the interventions on \(Y\) are well-defined. For a biomarker, for example, “set the biomarker one unit higher” may not specify a unique intervention or its consequences.
Even when an intervention is meaningful, the observed conditional coefficient equals a causal contrast only under additional conditions. These may include consistency between observed and intervened values, positivity for the relevant treatment and outcome levels, adequate control of common causes of \(Y\) and \(R\), a correctly specified causal model, and a defensible temporal ordering. The needed assumptions depend on the causal question. They do not follow merely from fitting a selection equation.
Treatment requires particular care. If \(A\) is randomized, randomization supports causal comparisons of treatment assignment, but it does not randomize the post-treatment outcome \(Y\). A coefficient for \(Y\) in the response equation is therefore not made causal by treatment randomization. Further, a treatment coefficient conditional on \(Y\) is not generally the total causal effect of treatment on response: it compares treatment groups at a fixed value of a post-treatment variable, rather than allowing the outcome to take the values it would naturally have under each treatment.
Worked Example: Reading an Outcome Coefficient
Suppose a hypothetical study uses the selection equation $$ \operatorname{logit}P(R=1\mid Y)=1.2-0.6\left(\frac{Y-50}{10}\right), $$ where \(Y\) is a biomarker measured in units and \(R=1\) means the visit outcome is observed. Interpret the coefficient and calculate the modeled response probabilities at \(Y=50\) and \(Y=60\).
Conditional association. The scaled predictor increases by one when \(Y\) increases by 10 units. Thus the model-implied observation odds ratio for a 10-unit higher biomarker is $$ \exp(-0.6)=0.5488. $$ At equal values of other modeled predictors, the odds of observation are about 0.549 times as large per 10-unit increase in \(Y\).
Probability at \(Y=50\). The scaled predictor is zero, so the linear predictor is \(1.2\). Therefore $$ P(R=1\mid Y=50)=\frac{\exp(1.2)}{1+\exp(1.2)}=\frac{3.320}{4.320}=0.7685. $$
Probability at \(Y=60\). The scaled predictor is one, so the linear predictor is \(1.2-0.6=0.6\). Hence $$ P(R=1\mid Y=60)=\frac{\exp(0.6)}{1+\exp(0.6)}=\frac{1.822}{2.822}=0.6457. $$
These are model-implied conditional probabilities, and the odds ratio is a conditional association. The calculations do not show that changing a person's biomarker from 50 to 60 would cause that person's probability of being observed to fall. That conclusion would require a causal interpretation and its supporting assumptions.
How an Association Can Appear Without an Outcome Effect
A common cause can generate an outcome-response association even when the outcome has no causal effect on response. The following hypothetical calculation makes the distinction explicit. Let \(U\) be an unmeasured binary severity indicator. Suppose \(U\) affects both the chance of \(Y=1\) and the chance of observation, while response is generated independently of \(Y\) once \(U\) is known.
Worked Example: A Common Cause Creates an Association
Assume \(P(U=0)=P(U=1)=0.5\). The outcome probabilities are \(P(Y=1\mid U=0)=0.2\) and \(P(Y=1\mid U=1)=0.8\). The response probabilities are \(P(R=1\mid U=0)=0.8\) and \(P(R=1\mid U=1)=0.4\). By construction, within each level of \(U\), \(R\) does not depend on \(Y\).
First calculate the joint probability of \(Y=1\) and response: $$ P(Y=1,R=1)=(0.5)(0.2)(0.8)+(0.5)(0.8)(0.4)=0.08+0.16=0.24. $$ Also \(P(Y=1)=(0.5)(0.2)+(0.5)(0.8)=0.50\), so \(P(R=1\mid Y=1)=0.24/0.50=0.48\).
For \(Y=0\), $$ P(Y=0,R=1)=(0.5)(0.8)(0.8)+(0.5)(0.2)(0.4)=0.32+0.04=0.36. $$ Since \(P(Y=0)=0.50\), \(P(R=1\mid Y=0)=0.36/0.50=0.72\). Thus the observed conditional odds ratio comparing \(Y=1\) with \(Y=0\) is $$ \frac{0.48/(1-0.48)}{0.72/(1-0.72)} =\frac{0.9231}{2.571}=0.3590. $$ There is a substantial outcome-response association, despite there being no within-\(U\) effect of \(Y\) on \(R\).
Under an intervention setting \(Y\) to either value while leaving the distribution of \(U\) unchanged, the response probability is \(0.5(0.8)+0.5(0.4)=0.60\) in either case. The causal effect of setting \(Y\) on response is therefore zero in this constructed example. The association arises because the distribution of \(U\) differs between the observed outcome groups. This illustrates why a selection coefficient cannot by itself rule out confounding.
Treatment and Outcome Terms in the Same Equation
Treatment-by-outcome interactions make it especially important to state exactly what each coefficient represents. Consider a logistic response equation with a centered, scaled outcome \(z=(Y-50)/10\): $$ \operatorname{logit}P(R=1\mid A,Y)=-0.4+0.7A-0.5z+0.3Az, $$ where \(A=0\) is control and \(A=1\) is active treatment. The outcome slope in the control arm is \(-0.5\), while the outcome slope in the active arm is \(-0.5+0.3=-0.2\). These slopes describe arm-specific conditional associations with observation.
Worked Example: A Treatment-by-Outcome Selection Interaction
Using the hypothetical equation above, calculate the conditional treatment odds ratio for observation at \(Y=50\) and \(Y=60\). Then state why neither value is automatically a causal treatment effect.
At \(Y=50\), \(z=0\). The control linear predictor is \(-0.4\), and the active-treatment predictor is \(-0.4+0.7=0.3\). Their conditional log odds contrast is \(0.3-(-0.4)=0.7\), so the odds ratio is $$ \exp(0.7)=2.014. $$
At \(Y=60\), \(z=1\). The control predictor is \(-0.4-0.5=-0.9\). The active-treatment predictor is \(-0.4+0.7-0.5+0.3=0.1\). The conditional log odds contrast is \(0.1-(-0.9)=1.0\), giving $$ \exp(1.0)=2.718. $$
The model therefore says that, among records with the same specified outcome value, the conditional odds of observation are higher in the active arm by these factors, under the fitted equation. If \(Y\) is post-treatment, fixing it compares groups at a value that treatment itself may affect. These conditional contrasts are not the total causal effect of treatment on observation. Nor does random assignment of \(A\) establish that the outcome slope is causal.
A causal interpretation of a treatment contrast at fixed \(Y\) would usually correspond to a controlled direct-effect question: what would response be under each treatment if the outcome were also set to the same specified value? That question requires a meaningful intervention on \(Y\) and stronger assumptions than treatment randomization alone. In particular, conditioning on a post-treatment outcome can also induce associations between treatment and other causes of that outcome. A regression coefficient does not resolve those causal issues.
Interpretation, Diagnostics, and Alternatives
When reporting a selection coefficient, use language that stays on the model's scale. For example: “Conditional on treatment and the specified history, the fitted response odds are multiplied by 0.55 per 10-unit increase in the outcome.” Avoid “a 10-unit increase causes response odds to fall by 45%” unless a causal analysis has explicitly justified that interpretation. Even then, state the intervention and target population.
Warning signs include describing a conditional coefficient as a treatment effect without specifying the conditioning set, treating randomization of treatment as randomization of the outcome, and using causal verbs when the analysis only specifies a conditional distribution. Also question a causal reading when important common causes of outcome and response are unavailable, the proposed intervention on the outcome is unclear, or response and outcome timing do not support the proposed direction of effect. A model can fit observed data and still fail to identify the claimed causal contrast.
The main alternatives answer different questions. A selection model specifies response conditional on outcomes and covariates; a pattern-mixture model organizes the outcome distribution by response patterns; an MAR-based analysis relies on its own missingness assumptions. These are alternative statistical approaches to missing data, not automatic causal remedies. They can yield different results because they make different assumptions about unobserved outcomes. None turns a conditional association into a causal effect without a causal estimand and defensible causal assumptions.
| Quantity or approach | What it directly describes | What it does not establish by itself |
|---|---|---|
| Selection coefficient \(\gamma\) | Conditional log odds association between outcome and observation in the specified equation | A causal effect of intervening on the outcome |
| Treatment coefficient conditional on \(Y\) | Conditional treatment association at a fixed modeled outcome value | The total causal effect of treatment on response |
| Alternative missing-data model | Outcome or response distributions under that model's assumptions | Causal identification without a defined causal question and supporting assumptions |
The practical limitation is not that selection coefficients are unhelpful. They can express scientifically important dependence and support sensitivity analyses. The limitation is interpretive: a coefficient's statistical meaning is set by the model, while its causal meaning requires additional structure. Keep those claims separate when fitting, explaining, or comparing models.
A Short Derivation of the Conditional Odds Ratio
For a logistic equation with a one-unit outcome increase and all other predictors held fixed, let the linear predictor at \(Y=y\) be \(\eta(y)\). Then \(\eta(y+1)=\eta(y)+\gamma\). The response odds at \(y\) are \(\exp\{\eta(y)\}\), and at \(y+1\) they are \(\exp\{\eta(y)+\gamma\}\). Their ratio is $$ \frac{\exp\{\eta(y)+\gamma\}}{\exp\{\eta(y)\}}=\exp(\gamma). $$ This derivation establishes the model's conditional odds-ratio interpretation. It contains no intervention comparison, no adjustment for unmeasured common causes, and no argument that changing \(Y\) would change \(R\). Those additional ingredients are why the odds ratio is not automatically causal.
Common Mistake
Wrong reasoning: “The fitted outcome coefficient is negative, so raising the outcome causes dropout.” This treats a conditional association in a statistical factorization as an intervention effect. It ignores possible common causes, the post-treatment status of the outcome, and whether changing the outcome is a well-defined intervention.
Correction: First report the coefficient as a conditional association on the model scale. If a causal claim is intended, define the intervention and causal contrast, then justify the required assumptions and analysis. If those steps are unavailable, avoid causal wording.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A selection equation has an outcome coefficient of \(-0.3\) per unit. What conditional odds ratio does it imply, and what causal conclusion is not justified by that calculation?
- Why does randomization of treatment not automatically make the outcome coefficient in a dropout equation causal when the outcome is measured after treatment?
- In the common-cause example, what is the response probability under an intervention setting \(Y\) to either value, and how does it compare with the observed conditional probabilities?
- In the interaction example, what is the conditional outcome slope in the active arm? Interpret it without causal language.
Solutions. (1) The odds ratio is \(\exp(-0.3)=0.7408\), so the model implies about 0.741 times the conditional observation odds per one-unit increase, holding other terms fixed. This does not establish that intervening to raise the outcome causes the odds to fall. (2) Randomization assigns \(A\), not the post-treatment \(Y\); causes of \(Y\) may also affect response, and conditioning on \(Y\) does not preserve the interpretation of an unadjusted randomized treatment comparison. (3) Under either intervention on \(Y\), the response probability is \(0.5(0.8)+0.5(0.4)=0.60\). The observed conditional probabilities differ: \(0.48\) for \(Y=1\) and \(0.72\) for \(Y=0\), because the distribution of \(U\) differs between outcome groups. (4) The active-arm slope is \(-0.5+0.3=-0.2\). Per one-unit increase in the scaled outcome \(z\), the fitted conditional observation odds are multiplied by \(\exp(-0.2)=0.8187\), holding treatment fixed; this is an association in the response equation, not necessarily a causal effect of changing \(Y\).