Why Model Efficacy and Safety Together?
Exposure can be associated with both clinical benefit and harm. An analysis that estimates each endpoint separately may describe its exposure relationship, but it does not directly represent how efficacy and safety outcomes co-occur within the same patient. That dependence can matter when estimating joint outcome probabilities and when quantifying uncertainty in predictions involving both endpoints.
A joint model does not mean that efficacy and safety must have the same exposure-response curve. It means that the two outcomes are modeled together, with an explicit account of their possible patient-level dependence. The exposure relationship can differ by endpoint, while patient-level factors that raise or lower risk can be correlated.
A Bivariate Random-Effects Model
For a concrete formulation, suppose each patient contributes a binary efficacy outcome \(Y_{Ei}\) and a binary safety outcome \(Y_{Si}\). Let \(Y_{Ei}=1\) denote an efficacy response and \(Y_{Si}=1\) denote occurrence of a specified adverse event in the defined assessment period. Let \(x_i\) be the patient's exposure, measured on a specified scale and centered at a meaningful reference value.
A bivariate logistic random-intercept model is:
Here \(\operatorname{logit}(p)=\log\{p/(1-p)\}\). The endpoint-specific coefficients \(\beta_{E0},\beta_{E1}\) and \(\beta_{S0},\beta_{S1}\) describe the conditional log odds at a given pair of random effects. The random intercepts represent patient-level deviations from the population-average intercepts; their correlation \(\rho\) allows patients with higher-than-typical efficacy-response propensity also to have higher or lower-than-typical safety-event propensity.
The model assumes that, conditional on exposure and both random effects, the two binary outcomes are independent. Their marginal association arises because patients share correlated random effects. It also assumes the stated logistic mean relationships and a bivariate normal distribution for those effects. These assumptions are modeling choices, not facts guaranteed by collecting both outcomes.
The coefficients are endpoint-specific even though \(x_i\) appears in both equations. A positive \(\beta_{E1}\) means higher exposure is associated with higher conditional odds of efficacy response per unit of exposure; a positive \(\beta_{S1}\) means higher exposure is associated with higher conditional odds of the adverse event. Neither slope alone establishes a causal exposure effect. Interpretation depends on the exposure definition, study design, covariates, and potential confounding.
Worked Example: Interpreting the Two Exposure Slopes
Suppose, hypothetically, exposure \(x\) is measured in mg/L relative to a reference exposure. A fitted model has efficacy coefficients \(\beta_{E0}=-0.20\), \(\beta_{E1}=0.60\) per mg/L, and safety coefficients \(\beta_{S0}=-2.00\), \(\beta_{S1}=0.50\) per mg/L. First consider a patient with random effects set to zero, so these are conditional predictions at \(b_{Ei}=b_{Si}=0\).
At \(x=2\) mg/L, the efficacy linear predictor is \(-0.20+0.60(2)=1.00\). Thus the conditional response probability is \(1/(1+\exp(-1.00))=0.7311\), or about 73.1%.
At the same exposure, the safety linear predictor is \(-2.00+0.50(2)=-1.00\). The conditional adverse-event probability is \(1/(1+\exp(1.00))=0.2689\), or about 26.9%.
Per 1 mg/L increase in exposure, the conditional odds ratio is \(\exp(0.60)=1.822\) for efficacy response and \(\exp(0.50)=1.649\) for the adverse event. These are distinct exposure relationships. The calculation does not give population-average probabilities, because it fixes the random effects at zero rather than averaging over their distribution.
Dependence and the Joint Likelihood
For a patient with outcomes \(y_{Ei},y_{Si}\), the conditional likelihood contribution is the product of the two Bernoulli probabilities given the random effects. The marginal contribution integrates that product over the bivariate random-effects distribution:
The product inside the integral encodes conditional independence; integration over shared, correlated patient effects generally produces marginal dependence. A useful justification follows from the law of total covariance. Conditional covariance is zero under the model, so:
Thus dependence in the outcomes reflects dependence in the patient-specific probabilities in this model. A positive random-effect correlation often induces positive outcome association, but the magnitude of the outcome association is not equal to \(\rho\). It also depends on the endpoint intercepts, slopes, random-effect variances, and exposure.
Worked Example: Quantifying Within-Patient Outcome Association
Suppose, hypothetically, that at a comparable exposure level 100 patients have the following paired outcomes. The table counts patients, not independent endpoint records.
| Safety event: Yes | Safety event: No | Total | |
|---|---|---|---|
| Efficacy response: Yes | 20 | 20 | 40 |
| Efficacy response: No | 10 | 50 | 60 |
| Total | 30 | 70 | 100 |
The observed efficacy-response proportion is \(40/100=0.40\), and the adverse-event proportion is \(30/100=0.30\). The proportion with both outcomes is \(20/100=0.20\). If the outcomes were independent at this exposure, the expected joint proportion based on these marginal proportions would be \(0.40(0.30)=0.12\). The observed excess over that product is \(0.20-0.12=0.08\).
The sample odds ratio for association is \((20\times50)/(20\times10)=5.00\). The sample correlation of the two binary indicators is \((0.20-0.40(0.30))/\sqrt{0.40(0.60)0.30(0.70)}=0.08/\sqrt{0.0504}=0.356\). These descriptive quantities show positive within-patient association in this hypothetical table. They do not identify its cause, prove that a random-effects distribution is correct, or replace adjustment for exposure and relevant covariates.
Uncertainty in Both Endpoint Relationships
A joint fit estimates the endpoint coefficients and the random-effects parameters together. Its estimated covariance matrix contains uncertainty for each endpoint relationship and, when applicable, covariance between estimates from the two endpoint models. That cross-endpoint covariance matters when calculating uncertainty for a quantity involving both endpoints. Treating the two coefficient estimates as independent without checking their covariance can give an incorrect standard error.
For a conditional probability \(p=\operatorname{logit}^{-1}(\eta)\), a confidence interval for the linear predictor \(\eta\) can be transformed through the logistic function because it is monotone increasing. Such a transformed interval is an approximate interval for the conditional probability, provided the original interval is appropriate for the fitted model. It is not automatically an interval for a population-average probability, which requires integrating over the random effects.
Worked Example: Transforming Endpoint-Specific Uncertainty
Suppose a hypothetical joint fit gives, at a stated exposure and with random effects set to zero, an efficacy logit estimate of \(0.600\) with standard error \(0.200\), and a safety logit estimate of \(-0.600\) with standard error \(0.3163\). Using a normal approximation, the efficacy logit interval is \(0.600\pm1.96(0.200)=[0.208,0.992]\). The safety logit interval is \(-0.600\pm1.96(0.3163)\), approximately \([-1.220,0.020]\).
Transforming the efficacy endpoints gives \(1/(1+\exp(-0.208))=0.5518\) and \(1/(1+\exp(-0.992))=0.7295\). The approximate conditional efficacy-probability interval is therefore \([0.5518,0.7295]\).
For safety, the corresponding transformations are \(1/(1+\exp(1.220))=0.2279\) and \(1/(1+\exp(-0.020))=0.5050\). The approximate conditional safety-probability interval is \([0.2279,0.5050]\). These are separate endpoint intervals; they do not state that there is a 95% probability that both true probabilities lie in their respective intervals simultaneously.
For a prediction involving both endpoints, uncertainty should be calculated from the joint parameter estimates, retaining their covariance, or from joint parameter draws or a suitable bootstrap. For instance, each joint draw can be used to calculate both endpoint probabilities at the same exposure; the resulting paired predictions preserve their estimated dependence. A pointwise interval for each endpoint does not itself provide a joint uncertainty statement.
Fitting, Diagnostics, and Alternatives
The model is typically fitted by maximum likelihood using numerical integration or approximation for the random-effects integral. Before interpreting results, verify that the outcome definitions and risk periods are aligned, that exposure is defined consistently for both endpoints, and that the exposure range contains sufficient information to estimate each slope. Include covariates where justified by the design and scientific question, while keeping each endpoint's mean structure appropriate.
Confirm paired outcomes are available for the intended patients and that missingness, unequal observation windows, or censoring are handled consistently with the analysis target.
Inspect endpoint-specific observed rates across exposure and compare them with fitted probabilities. Look for systematic curvature, sparse exposure regions, and influential observations.
Compare observed paired outcomes with model-based joint frequencies, examine estimates and uncertainty for random-effect variances and correlation, and assess convergence and sensitivity to integration settings or reasonable model alternatives.
Several alternatives answer different questions. Separate endpoint regressions can be simpler and may be adequate when only endpoint-specific associations are required, but they do not model paired outcome dependence. A joint model with correlated random effects supports joint predictions under its distributional assumptions. A constrained model that forces equal exposure slopes is more parsimonious, but it imposes a substantive restriction that should be scientifically justified and evaluated rather than assumed because exposure is shared. Marginal approaches such as generalized estimating equations can target population-average relationships and robust standard errors, but do not automatically provide the same patient-level joint probability model.
Warning signs include a near-boundary random-effect correlation, very large standard errors, unstable estimates under modest model changes, poor fit to the paired outcome frequencies, or predicted probabilities inconsistent with observed rates in parts of the exposure range. A complicated dependence structure cannot compensate for little information about rare outcomes or inadequate exposure overlap. If one endpoint is sparse, the joint fit may be driven by modeling assumptions rather than data. The earlier material on diagnostics for exposure-response model fit is relevant to assessing the endpoint-specific mean relationships.
Common Pitfalls and Scope
- Assuming shared exposure means a shared slope. The same exposure can have different relationships with efficacy and safety; equality is an added restriction.
- Equating random-effect correlation with outcome correlation. They are different quantities. Outcome association also depends on the fitted probabilities and random-effect variances.
- Calling a conditional prediction a population average. A prediction at zero random effects describes a typical latent patient under the model; a population-average probability integrates over the random-effects distribution.
- Interpreting exposure association as causation. Differences in patient characteristics, disease severity, treatment changes, or selection into exposure levels can affect observed relationships.
- Combining marginal intervals as if they were a joint interval. Separate 95% intervals do not guarantee 95% coverage for both endpoints together, and cross-endpoint parameter covariance may matter.
Joint modeling is useful when the paired endpoints and their dependence are part of the scientific question, but it adds assumptions and computational demands. The random-effects distribution, conditional independence, exposure functional form, and handling of incomplete outcomes all require scrutiny. The model should support the intended predictions; it does not by itself settle a clinical decision or establish that an exposure change causes a change in either endpoint.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- In a joint model, why can efficacy and safety have different exposure slopes even though both equations use the same exposure variable?
- What does conditional independence mean in the bivariate random-effects model, and how can marginal dependence still arise?
- A conditional efficacy logit interval is \([0,1]\). What probability interval does it imply, to three significant figures?
- Why is the random-effect correlation not generally equal to the correlation between the observed binary outcomes?
Solutions. (1) The exposure measure is shared, but its endpoint-specific coefficients \(\beta_{E1}\) and \(\beta_{S1}\) are separately estimated unless equality is explicitly imposed. (2) Given exposure and both random effects, the outcome probabilities factor into two Bernoulli contributions; integrating over correlated patient effects induces marginal association. (3) Apply the logistic function to each endpoint: \(\operatorname{logit}^{-1}(0)=0.500\) and \(\operatorname{logit}^{-1}(1)=1/(1+\exp(-1))=0.731\), giving \([0.500,0.731]\). (4) The random-effect correlation describes latent intercepts. Observed binary-outcome correlation also depends on the endpoint-specific probabilities and the random-effect variances, so it is not numerically the same quantity.