Joint Latent Process Models

The R package JLPM implements in the jointLPM function the estimation of a joint shared random effects model.

 

The longitudinal data ($y_1, y_2, \dots, y_K$) can be continuous or ordinal and are modelled using a mixed model.

 

For the continuous case, a transformation $H_k$ is estimated for outcome $y_k$ ($k \in 1,\dots,K$) :

$$H_k(y_k(t_{ijk}), \eta_k) = X_i(t_{ijk})\beta + Z_i(t_{ijk})u_i + \varepsilon_{ijk}$$

For an ordinal outcome, mixed models are combined to the Item Response Theory:

$$\mathbb{P}( y_k(t_{ijk}) = m) \Leftrightarrow \mathbb{P}(\eta_{k, m-1} < X_i(t_{ijk})\beta + Z_i(t_{ijk})u_i + \varepsilon_{ijk} < \eta_{k, m})$$

where $X_i(t_{ijk})$ and $Z_i(t_{ijk})$ are vectors of covariates measured at time $t_{ijk}$ for subject $i$, $\beta$ is the vector of fixed effects, $u_i \sim \mathcal{N}(0,B)$ are the random effects, $\varepsilon_{ijk} \sim \mathcal{N}(0,\sigma_k^2)$ the measurement errors, $\eta_k$ are the parameters of the link function $H_k$ or the thresholds associated to the outcome $y_k$.

Note that even with multiple longitudinal outcomes, a univariate mixed model is estimated.

 

The time-to-event data are modelled in a proportional hazard model where different associations between the longitudinal outcome and the event can be included.

 

With an association through the random effects of the longitudinal model:

$$\alpha_i(t) = \alpha_0(t, \omega) \exp(\tilde{X}_i \gamma + \delta u_i)$$

With an association through the current level of the longitudinal model:

$$\alpha_i(t) = \alpha_0(t, \omega) \exp(\tilde{X}_i \gamma + \delta (X_i(t)\beta + Z_i(t)u_i))$$

With an association through the current slope of the longitudinal model:

$$\alpha_i(t) = \alpha_0(t, \omega) \exp(\tilde{X}_i \gamma + \delta ( \frac{d}{dt}X_i(t)\beta + \frac{d}{dt}Z_i(t)u_i))$$

with $\alpha_0(t, \omega)$ the baseline hazard function at time $t$, parameterized with $\omega$, $\tilde{X}_i$ a vector of time-fixed covariates, $\gamma$ the fixed effects, and $\delta$ the association parameter.

 

The parameters $\beta, B, \sigma_k, \eta_k, \omega, \gamma, \delta$ are estimated using a Marquardt-Levenberg algorithm.