Interpreting Current-Outcome Selection Coefficients

Why the Coefficient Needs Careful Interpretation

A current-outcome selection model connects a person’s outcome level with the chance that the outcome is missing. Its logistic coefficient is compact, but it is easy to overstate what it means. The coefficient describes a change in conditional dropout odds for a specified change in the current outcome, with the other modeled predictors held fixed. It is not, by itself, a probability difference, a population-average contrast, or a causal effect.

The interpretation also depends on which event the model treats as a “success.” Here the event is dropout at visit $t$, among individuals still observed and at risk immediately before that visit’s observation. This convention fixes the direction of the coefficient: a positive coefficient means higher current outcome values correspond to higher conditional dropout odds.

Definition: Let $R_{it}=1$ indicate that individual $i$'s outcome is observed at visit $t$, and $R_{it}=0$ indicate that it is not observed. Under monotone dropout, an individual is at risk at visit $t$ if $R_{i,t-1}=1$. Define the dropout hazard $h_{it}$ as the conditional probability of dropout at $t$, given being at risk and the predictors in the model. A current-outcome logistic selection model is $$ \operatorname{logit}(h_{it}) =\log\!\left(\frac{h_{it}}{1-h_{it}}\right) =\alpha_t+\beta_Y Y_{it}+\boldsymbol{\gamma}^{\mathsf T}\mathbf{Z}_{it}. $$ Here $Y_{it}$ is the current outcome, $\mathbf{Z}_{it}$ collects the other modeled predictors, and $\alpha_t$ is a visit-specific intercept. The outcome $Y_{it}$ is understood as the value relevant to the dropout decision even when it is unobserved for someone who drops out at that visit.

Assumptions Behind the Interpretation

The coefficient interpretation is conditional on the model and its timing. The dropout event must be defined consistently, and $Y_{it}$ must refer to the outcome level relevant to the visit-$t$ decision. In the selection-model setup, a person may drop out before that outcome is recorded; the model can nevertheless relate the dropout probability to the current outcome value, which is then latent for that person. Confusing the value that drives the decision with the last outcome actually observed changes the model being interpreted.

The logistic form assumes that the log odds of dropout are additive in the specified predictors, with a linear term $\beta_Y Y_{it}$. Unless interactions or nonlinear terms are included, $\beta_Y$ is the same across the modeled values of the other predictors and across the outcome range. The interpretation also assumes the stated model is the one of interest: changing the set or form of the other predictors can change the conditional coefficient.

These are model-based interpretations, not assumptions that can all be checked from observed data. In particular, the current outcome is unobserved for people who drop out at that visit. How dropout depends on those unobserved values is therefore not generally determined by the observed data alone. The earlier material on identifiability and sensitivity parameters is relevant when assessing what additional assumptions are needed to estimate or interpret this dependence.

From a Log-Odds Coefficient to an Odds Ratio

The logistic coefficient is a change in log odds per one-unit increase in $Y_{it}$. To express it as a multiplicative change in odds, compare two outcome values, $y$ and $y+\Delta$, with $\mathbf{Z}_{it}$ and the visit held fixed. Subtracting the two linear predictors gives the log odds contrast. Exponentiating that contrast gives the odds ratio:

\begin{aligned} \operatorname{logit}\{h(y+\Delta,\mathbf{Z})\} -\operatorname{logit}\{h(y,\mathbf{Z})\} &=\beta_Y\Delta,\\ \frac{\operatorname{odds}\{h(y+\Delta,\mathbf{Z})\}} {\operatorname{odds}\{h(y,\mathbf{Z})\}} &=\exp(\beta_Y\Delta). \end{aligned}

For a one-unit increase, the conditional dropout-odds ratio is $\exp(\beta_Y)$. If $\beta_Y>0$, the odds increase as the current outcome increases; if $\beta_Y<0$, they decrease. If $\beta_Y=0$, this model contains no current-outcome term in the dropout log odds, conditional on the other modeled predictors. That alone does not establish that the full missingness mechanism is ignorable: other unobserved dependencies could still matter.

The odds ratio is constant across $y$ under this model, but the probability difference is not. If the dropout probability at one outcome value is $h$, its odds are $h/(1-h)$. Multiplying those odds by $\exp(\beta_Y\Delta)$ gives the odds at the other value; converting back to a probability requires $p=\text{odds}/(1+\text{odds})$. Thus, the same odds ratio can correspond to different probability changes at different starting probabilities.

Worked Example: A Positive Coefficient and a Ten-Point Contrast

Suppose a hypothetical study uses the dropout model $\operatorname{logit}(h_{it})=-5.2+0.08Y_{it}$, with the visit and any other modeled predictors fixed. The outcome is measured in points. Interpret the coefficient and compare dropout hazards at $Y_{it}=40$ and $Y_{it}=50$.

The coefficient is $0.08$ log-odds units per outcome point. For a ten-point increase, the conditional odds ratio is $\exp(0.08\times10)=\exp(0.8)=2.226$. Thus the model assigns 2.226 times the dropout odds at 50 points as at 40 points, conditional on the modeled predictors. This is a 122.6% increase in odds, since $(2.226-1)\times100\%=122.6\%$; it is not a 122.6 percentage-point increase in dropout probability.

At 50 points, the linear predictor is $-5.2+0.08(50)=-1.20$. The dropout odds are $\exp(-1.20)=0.3012$, and the model-implied probability is $1/(1+\exp(1.20))=0.2315$.

At 40 points, the linear predictor is $-5.2+0.08(40)=-2.00$. The odds are $\exp(-2.00)=0.1353$, and the probability is $1/(1+\exp(2.00))=0.1192$. The odds ratio from these calculations is $0.3012/0.1353$, approximately $2.226$, agreeing with $\exp(0.8)$. The probability difference is about $0.1123$, or 11.23 percentage points. It is specific to these two values and this model’s starting hazard; it is not the interpretation of the coefficient itself.

Worked Example: A Negative Coefficient

Consider a hypothetical symptom-severity outcome measured in points and a fitted conditional dropout model with current-outcome coefficient $\beta_Y=-0.30$. Interpret a four-point increase in symptom severity.

The model-implied conditional odds ratio is $\exp\{-0.30(4)\}=\exp(-1.20)=0.3012$. Holding the other modeled predictors fixed, the dropout odds at a score four points higher are about 0.3012 times the odds at the lower score. Equivalently, they are about 69.88% lower, because $(1-0.3012)\times100\%=69.88\%$.

This says nothing by itself about the absolute dropout probability. For example, an odds ratio of 0.3012 does not mean that dropout probability falls by 69.88 percentage points. The probability change depends on the hazard at the lower outcome value as well as on the odds ratio.

Worked Example: Translating a Coefficient Into Probabilities

In another hypothetical model, let the current-outcome coefficient be $-0.045$ per point and let the contribution of all other predictors to the log odds be $-1.00$. Compare the model-implied dropout probabilities at outcome values 20 and 30.

At 20 points, the linear predictor is $-1.00-0.045(20)=-1.90$. The odds are $\exp(-1.90)=0.1496$, so the probability is $0.1496/(1+0.1496)=0.1301$.

At 30 points, the linear predictor is $-1.00-0.045(30)=-2.35$. The odds are $\exp(-2.35)=0.09537$, so the probability is $0.09537/(1+0.09537)=0.08706$. The ten-point odds ratio is $\exp(-0.045\times10)=\exp(-0.45)=0.6376$. The hazards differ by about 4.30 percentage points in this example, but that difference is not constant across outcome values: it depends on the intercept and the other predictors as well as on $\beta_Y$.

Why the Probability Change Varies

For a fixed value of the other predictors, the logistic hazard at outcome $y$ can be written as $h(y)=1/\{1+\exp[-(\eta_0+\beta_Yy)]\}$, where $\eta_0$ is the contribution not involving the current outcome. Differentiating with respect to $y$ gives $\mathrm{d}h(y)/\mathrm{d}y=\beta_Yh(y)\{1-h(y)\}$. Because $h(y)\{1-h(y)\}$ varies with $y$, the probability slope varies with $y$, even though the log-odds slope $\beta_Y$ is constant. The derivative also shows that the probability changes in the direction of the coefficient, but its magnitude depends on the current hazard.

This distinction matters when translating model output for a clinical or scientific audience. Report the coefficient’s units, its conditional odds-ratio interpretation for a stated contrast, and—if probabilities are presented—the predictor values and other conditions used to calculate them. Do not replace these details with a vague statement that “dropout rises by the coefficient.”

Diagnostics, Alternatives, and Limitations

Useful checks include confirming how dropout is coded, verifying that the outcome’s units and visit timing are correct, and checking whether the fitted model has plausible predictions over the range of outcomes under discussion. Sparse data or near separation can produce unstable logistic estimates. A linear current-outcome term may also be inadequate if the log odds bend across the outcome range or if the current-outcome association differs across visits or predictor strata. Such patterns motivate checking the specified functional form and considering whether the model needs modification; they do not license an unreported change in interpretation.

Observed dropout frequencies can be compared with predictions for individuals whose relevant predictors are observed, but this cannot directly validate the model’s dependence on current outcomes that are missing for dropouts. In particular, a good fit to observed cases does not prove that the assumed relationship with unobserved current outcomes is correct. Estimates of that relationship can depend materially on the outcome model and the identifying or sensitivity assumptions used in the analysis.

Approach	What is modeled directly	How it differs here
Current-outcome selection model	Dropout odds conditional on the current outcome and specified predictors	The coefficient describes conditional outcome dependence in the dropout mechanism.
Selection model without a current-outcome term	Dropout odds conditional on the predictors included in that model	Omitting the term removes this particular dependence from the specified hazard; it does not prove that all missingness is ignorable.
Pattern-mixture approach	Outcome distributions conditional on missingness patterns	It organizes the model by missingness pattern rather than expressing the current-outcome dependence through this dropout-hazard coefficient.

The approaches answer different modeling questions and rely on different assumptions; the coefficient is not a test that selects one approach as correct. Most importantly, it is a conditional association parameter in a selection model. It does not show that changing an individual’s outcome would cause that individual to drop out.

Common Mistake

A common error is to read $\beta_Y=0.08$ as an 8-percentage-point increase in dropout probability for each outcome point. That is wrong because the coefficient is on the log-odds scale. The one-point odds ratio is $\exp(0.08)=1.083$, meaning the conditional odds are about 8.33% higher per point under the model. The corresponding probability change depends on the starting hazard. Another error is to omit “conditional on the other modeled predictors,” which can make a model-specific association sound like an unconditional or causal conclusion.

Key takeaway: In a logistic dropout hazard, $\beta_Y$ is the change in conditional log dropout odds per unit increase in the current outcome. For a contrast of $\Delta$ units, the conditional odds ratio is $\exp(\beta_Y\Delta)$. Probability changes vary with the starting hazard, and the coefficient is not a causal effect.

Check Your Understanding

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

A fitted dropout model has $\beta_Y=0.20$. What is the conditional dropout-odds ratio for a two-unit increase in the current outcome?
A model has $\beta_Y=-0.10$ per outcome point. Interpret a five-point increase in terms of conditional dropout odds.
Why is it incorrect to say that an odds ratio of 1.5 means dropout probability increases by 50 percentage points?
Can a good fit among observed cases verify the assumed current-outcome dependence for people who dropped out? Explain.

Solutions. (1) The odds ratio is $\exp(0.20\times2)=\exp(0.40)=1.492$. With other modeled predictors fixed, dropout odds are about 1.492 times as large for the outcome value two units higher. (2) The odds ratio is $\exp(-0.10\times5)=\exp(-0.50)=0.6065$. The conditional odds are about 39.35% lower, since $(1-0.6065)\times100\%=39.35\%$; this is not a probability-point change. (3) An odds ratio multiplies odds, not probabilities. Converting the changed odds to a probability requires the starting hazard, so a 50-percentage-point increase does not follow. (4) No. The relevant current outcomes are unobserved for dropouts, so fit among observed cases cannot directly validate the model’s dependence on those values.