Interpreting Association Parameters in Shared-Parameter Models

What a Shared-Effect Coefficient Says

A shared-parameter model connects an outcome process and dropout through a latent subject-specific quantity. The coefficient on that shared quantity describes how the model’s dropout process changes as the latent effect changes, with the other terms held fixed. Correct interpretation depends on three choices: the dropout link, the units and coding of the latent effect, and the contrast in that effect being compared.

The coefficient is not automatically a correlation between outcome and dropout. It is a conditional regression parameter in a specified model. For a hazard link it changes the log hazard; for a logistic link it changes the log odds of dropout by a specified time. Those scales are different, so the numerical values of coefficients from the two links are not directly comparable.

Definition: Let \(b_i\) denote the shared latent effect for subject \(i\), on the scale used in the outcome model. A shared-effect association coefficient, denoted \(\alpha\), is the coefficient multiplying \(b_i\) in the conditional dropout model. Its sign describes the direction of the modelled change as \(b_i\) increases; its magnitude is interpreted through the link function and a specified change in \(b_i\).

This interpretation presumes that the stated conditional model is appropriate: the same \(b_i\) is the quantity linked to both processes, the other model terms are held fixed in the comparison, and the chosen hazard or logistic link describes dropout on its stated scale. A coefficient has no unique practical interpretation without stating what a one-unit change in \(b_i\) means.

Hazard-Link Interpretation

For a continuous dropout time \(T_i\), consider a conditional proportional-hazards model of the form

h_i(t\mid b_i,z_i)=h_0(t)\exp\{\gamma^\mathsf{T}z_i+\alpha_H b_i\}.

Here \(h_i(t\mid b_i,z_i)\) is the instantaneous dropout hazard at time \(t\), conditional on the latent effect \(b_i\) and observed covariates \(z_i\); \(h_0(t)\) is the baseline hazard; and \(\alpha_H\) is the shared-effect coefficient. The model asserts that the hazard ratio for two latent-effect values is constant over time when their covariates are the same. It does not assert that the dropout probability or cumulative risk has that same ratio.

Interpretation: For two subjects with the same \(z_i\) and latent effects differing by \(\Delta b\), the conditional hazard ratio is \(\exp(\alpha_H\Delta b)\). The coefficient \(\alpha_H\) is the change in log hazard per one unit of \(b_i\). If \(b_i\) has units \(U\), then \(\alpha_H\) has units \(U^{-1}\); the exponent and hazard ratio are dimensionless.

If \(\alpha_H>0\), larger values of \(b_i\) correspond to a higher conditional dropout hazard, provided that increasing \(b_i\) means the same thing as assumed by the model’s coding. If \(\alpha_H<0\), larger values correspond to a lower hazard. If \(\alpha_H=0\), this link contains no direct dependence of the conditional hazard on \(b_i\). None of these statements by itself establishes that the latent effect causes dropout.

Worked Example: A Hazard Ratio for a Latent-Effect Contrast

Suppose a hypothetical model uses \(b_i\) measured in outcome-scale units and estimates \(\alpha_H=-0.40\) per unit. Consider two otherwise identical subjects whose latent effects differ by 2.5 units, with the first subject’s \(b_i\) higher. Calculate and interpret their conditional hazard ratio.

The higher-effect subject has \(\Delta b=2.5\). The model gives

\frac{h(t\mid b_i+2.5,z_i)}{h(t\mid b_i,z_i)} =\exp(\alpha_H\Delta b) =\exp((-0.40)(2.5)) =\exp(-1) =0.3679.

Thus, at any time \(t\), the subject with a latent effect 2.5 units higher has an estimated conditional dropout hazard about 0.368 times that of the comparison subject, or about 63.2% lower, under this model and at the same covariate values. The coefficient’s negative sign indicates decreasing hazard as \(b_i\) increases. The result concerns the instantaneous conditional hazard, not the ratio of dropout probabilities by a given time.

Logistic-Link Interpretation

If dropout is represented as an indicator \(D_i\) for dropout by a specified time \(\tau\), a conditional logistic model can be written as

\operatorname{logit}\{P(D_i=1\mid b_i,z_i)\} =\beta_0+\eta^\mathsf{T}z_i+\alpha_L b_i, \qquad \operatorname{logit}(p)=\log\left(\frac{p}{1-p}\right).

The model’s event is specifically dropout by \(\tau\); a different time horizon defines a different indicator and can yield different coefficients. Holding covariates fixed, a change \(\Delta b\) multiplies the conditional odds of dropout by \(\exp(\alpha_L\Delta b)\). It does not multiply the dropout probability by that factor.

Interpretation: In the logistic dropout model, \(\alpha_L\) is the change in conditional log odds of dropout by \(\tau\) per one-unit increase in \(b_i\). For a \(\Delta b\)-unit increase, the conditional odds ratio is \(\exp(\alpha_L\Delta b)\). If \(b_i\) has units \(U\), then \(\alpha_L\) has units \(U^{-1}\).

To recover a probability from an odds value \(o\), use \(p=o/(1+o)\). Because this conversion is nonlinear, the same odds ratio produces different probability differences at different starting probabilities. Consequently, a logistic coefficient alone does not give a constant percentage-point change in dropout probability.

Worked Example: Converting a Logistic Coefficient to Odds and Probability

In a hypothetical dropout-by-\(\tau\) model, suppose \(\alpha_L=0.70\) per standard-deviation unit of \(b_i\). At \(b_i=0\) and fixed covariates, suppose the fitted conditional dropout probability is 0.20. Find the odds ratio for increasing \(b_i\) by one unit and the corresponding probability when \(b_i=1\).

At \(b_i=0\), the odds are

o_0=\frac{0.20}{1-0.20}=0.25.

For a one-unit increase, the conditional odds ratio is

\exp(\alpha_L)=\exp(0.70)=2.014.

The odds at \(b_i=1\) are therefore \(0.25(2.014)=0.5035\). Converting these odds back to probability gives

p_1=\frac{0.5035}{1+0.5035}=0.3349.

The model therefore estimates about 2.014 times the conditional odds of dropout by \(\tau\) for a one-standard-deviation increase in \(b_i\). At the stated reference probability of 0.20, the corresponding probability is about 0.335, an increase of about 0.135, or 13.5 percentage points. That probability difference is specific to the reference probability and covariates; it is not the coefficient’s general probability-scale effect.

Scale, Units, and Coding

The coefficient’s numerical value depends on how \(b_i\) is represented. Suppose a latent effect measured in original units is \(b_i\), and a rescaled version is \(u_i=b_i/c\), where \(c>0\). The same linear predictor can be written either as \(\alpha b_i\) or as \((c\alpha)u_i\), because \(\alpha b_i=(c\alpha)u_i\). Thus, changing the unit changes the coefficient even though the model’s fitted linear predictor and comparisons are unchanged. The meaningful quantity is the coefficient multiplied by the stated latent-effect contrast.

Centering and reversing the coding have different consequences. Replacing \(b_i\) by \(b_i-a\) shifts the intercept but leaves the coefficient on the centered effect unchanged. Reversing its direction, for example using \(u_i=-b_i\), changes the coefficient’s sign. Therefore, the verbal meaning of “positive association” cannot be stated without saying what a larger latent effect represents.

Worked Example: Rescaling a Hazard-Link Effect

Suppose a hypothetical hazard model uses an original latent-effect scale measured in units \(U\), with \(\alpha_H=0.12\) per \(U\). A change of 4 \(U\) is being considered. Re-express the effect in units of 4 \(U\), and verify that the hazard ratio is unchanged.

On the original scale, \(\Delta b=4\,U\), so the log hazard ratio is \(0.12(4)=0.48\), and the hazard ratio is

\exp(0.48)=1.616.

Define \(u_i=b_i/4\), so one unit of \(u_i\) equals 4 \(U\). The coefficient on \(u_i\) is \(4(0.12)=0.48\) per unit of \(u_i\). A one-unit increase in \(u_i\) gives hazard ratio \(\exp(0.48)=1.616\), exactly the same comparison. The numerical coefficient changed from 0.12 to 0.48 because the unit changed; the model-implied hazard comparison did not.

Derivation and Practical Interpretation Checks

The hazard-ratio formula follows directly by taking the ratio of the two conditional hazards. For latent values \(b\) and \(b+\Delta b\), with the same covariates and baseline hazard,

\frac{h(t\mid b+\Delta b,z)}{h(t\mid b,z)} = \frac{h_0(t)\exp\{\gamma^\mathsf{T}z+\alpha_H(b+\Delta b)\}} {h_0(t)\exp\{\gamma^\mathsf{T}z+\alpha_H b\}} =\exp(\alpha_H\Delta b).

The baseline hazard and covariate contribution cancel, leaving a time-constant ratio under this model. For the logistic model, subtracting the two linear predictors gives the log odds ratio:

\operatorname{logit}\{P(D=1\mid b+\Delta b,z)\} -\operatorname{logit}\{P(D=1\mid b,z)\} =\alpha_L\Delta b.

Exponentiating this difference yields the conditional odds ratio \(\exp(\alpha_L\Delta b)\). In both links, the interpretation is conditional on the model terms, not an unadjusted comparison between groups of subjects with different latent effects.

Before reporting either parameter, check the latent-effect scale, its direction, the dropout event or time scale, and the contrast used in the exponent. Then translate the result into the link’s native measure: a hazard ratio for a hazard link or an odds ratio for a logistic link. For a logistic link, also calculate probabilities at relevant reference values if probability-scale interpretation is needed. Very large implied probabilities or hazards across plausible latent-effect values may indicate that the assumed effect scale or model form deserves scrutiny; they do not alone prove which part is wrong.

How These Parameters Differ From Other Summaries

Quantity	What it describes	Why it is not interchangeable with \(\alpha\)
\(\alpha_H\) in a hazard link	Change in log conditional hazard per unit of \(b_i\)	Its exponent for a stated contrast is a conditional hazard ratio
\(\alpha_L\) in a logistic link	Change in conditional log odds of dropout by \(\tau\) per unit of \(b_i\)	Its exponent for a stated contrast is a conditional odds ratio
Correlation between outcome and dropout	A standardized summary of co-variation, when such a correlation is defined	It is not the coefficient in either conditional link and does not encode the link’s scale
Marginal dropout probability	Dropout probability averaged over the distribution of latent effects	It is not the probability conditional on a particular \(b_i\)

Hazard and logistic links are alternatives for representing dropout, not two units for the same coefficient. One describes instantaneous rates over time; the other describes odds of a binary dropout event by a chosen horizon. Neither coefficient alone supplies a marginal dropout probability, and neither establishes a causal effect of the latent quantity on dropout. Their interpretation is conditional on the joint model and on the chosen latent-effect scale.

Common Pitfalls and Limitations

Calling a positive coefficient “positive association” without defining direction. A positive coefficient means higher dropout hazard or odds as \(b_i\) increases. If a larger \(b_i\) represents a lower outcome trajectory, the substantive description must reflect that coding.
Reporting \(\exp(\alpha)\) without a contrast. This is the comparison for a one-unit increase, not necessarily a useful or interpretable change. State what that unit represents, or calculate \(\exp(\alpha\Delta b)\) for a meaningful contrast.
Treating an odds ratio as a risk ratio. The logistic coefficient multiplies odds. Convert using \(p=o/(1+o)\) to obtain a probability, and specify the reference probability.
Comparing coefficient magnitudes across different scales or links. A coefficient can change under rescaling, and hazard and odds links have different meanings. Compare model-implied contrasts on a clearly stated scale instead.
Interpreting the parameter as causal or as an observed correlation. It is a conditional association parameter in a shared-parameter model. The latent effect is not generally observed, and the model’s assumptions do not by themselves establish causation.

A fitted coefficient is also not proof that the shared-effect structure is correct. Its meaning relies on the chosen latent-effect definition, dropout link, and conditional model. The earlier material on model fit and missingness assumptions is relevant to the broader caution: fit to observed data does not, by itself, validate assumptions about unobserved outcomes or dropout.

Key takeaway: Interpret a shared-effect coefficient by naming its link, the direction and units of the latent effect, and the size of the contrast. Exponentiate the coefficient times that contrast to obtain the appropriate conditional hazard ratio or odds ratio.

Check Your Understanding

Answer each question, then compare with the solutions below.

A hazard-link coefficient is \(-0.25\) per unit of \(b_i\). What is the conditional hazard ratio for a 2-unit increase?
A logistic-link coefficient is \(0.40\) per unit of \(b_i\). What is the conditional odds ratio for a 3-unit increase?
Why can a logistic odds ratio not be reported as the same percentage increase in dropout probability?
If \(b_i\) is replaced by \(u_i=b_i/5\), how must the coefficient change to preserve the same linear predictor?

Solutions. (1) \(\exp((-0.25)(2))=\exp(-0.50)=0.6065\). The conditional hazard is about 0.607 times as high for a 2-unit increase, with other terms fixed. (2) \(\exp((0.40)(3))=\exp(1.20)=3.320\), so the conditional dropout odds are 3.320 times as high. (3) Probability is \(o/(1+o)\), a nonlinear function of odds, so the same odds multiplier gives different probability changes at different starting probabilities. (4) Since \(b_i=5u_i\), \(\alpha b_i=(5\alpha)u_i\); the new coefficient is five times the old coefficient per unit of \(u_i\).