3
$\begingroup$

Can anyone explain the process of this derivation, step by step? This derivation is from Joint Models for Longitudinal and Time-to Event Data by Dimitris Rizopoulos.

\begin{equation} \begin{aligned} S(\theta) &= \sum_i \frac{\partial}{\partial \theta^\top}\log\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i \\ &= \sum_i\frac{1}{p(T_i, \delta_i, y_i; \theta)}\frac{\partial}{\partial \theta^\top}\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i \\ & = \sum_i\frac{1}{p(T_i, \delta_i, y_i; \theta)}\int \frac{\partial}{\partial \theta^\top}\{ p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\}db_i \\ & =\sum_i\int\bigg[ \frac{\partial}{\partial \theta^\top}\log\{ p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\}\bigg]\\ & \times \frac{p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)}{p(T_i, \delta_i, y_i; \theta)}db_i \\ &=\sum_i\int A(\theta,b_i)p(b_i \mid T_i, \delta_i, y_i; \theta)db_i, \end{aligned} \end{equation}

where $A(\cdot)$ denotes the complete data score vector, given by $A(\theta,b_i) = \partial\{\log p(T_i, \delta_i \mid b_i; \theta) + \log p(y_i \mid b_i;\theta) + \log p(b_i;\theta)\}/\partial \theta^\top$.

I also do not understand using a posterior distribution for the random effects $b_i$, since this is a frequentist derivation. Bayesians would have no need to integrate over the random effects, since the random effects vector is considered a parameter in the Bayesian regime.

$\endgroup$
2
$\begingroup$

First recall that $(\log f)' = \frac{f'}{f}$.

We have

$$ \frac{\partial}{\partial \theta} \log\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i = \frac{\frac{\partial}{\partial \theta} \int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i}{\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i} $$

Now$^*$, \begin{align*} \int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i &= \int p(T_i,\delta_i,y_i \mid b_i ; \theta)p(b_i; \theta)db_i \\ &= p(T_i,\delta_i,y_i; \theta) \end{align*} since we integrate the conditional probability it gives the marginal probability.

Under some regularity/integrability asumptions we have,

$$ \frac{\partial}{\partial \theta}\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i = \int \frac{\partial}{\partial \theta}\{ p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\}db_i. $$

Since $(\log f)' = \frac{f'}{f} \Longrightarrow f'= f (\log f)'$

Thus we can rewrite, $$ \frac{1}{p(T_i, \delta_i, y_i; \theta)}\int \frac{\partial}{\partial \theta}\{ p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\}db_i \\ $$ as: $$ \int \frac{\partial}{\partial \theta} \log \big( p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\big) \times \frac{p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)}{p(T_i, \delta_i, y_i; \theta)} dbi_i $$

From the Bayes formula we have$^*$: \begin{align*} \frac{p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)}{p(T_i, \delta_i, y_i; \theta)} &= \frac{p(T_i, \delta_i,y_i \mid b_i; \theta)p(b_i;\theta)}{p(T_i, \delta_i, y_i; \theta)} \\ &=p(b_i \mid T_i,\delta_i,y_i ; \theta) \end{align*} Thus finally if we note $$ A(\theta,b_i) = \frac{\partial}{\partial \theta} \log \big( p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)\big) $$

We have, $$ \frac{\partial}{\partial \theta^\top}\log\int p(T_i, \delta_i \mid b_i; \theta)p(y_i \mid b_i;\theta)p(b_i;\theta)db_i \\ = \int A(\theta,b_i) p(b_i \mid T_i,\delta_i,Y_i ; \theta) db_i $$

Thus sums follows by taking this respect to each individual in the sample.

$^*$ here we use $(T, \delta) \perp Y \mid b$ which I think is common assumption in joint modeling

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Perfect! Very easy to follow. And, yes! It is very common to assume that $(T,\delta) \perp Y \mid b$ when using joint models for longitudinal and time-to-event data. $\endgroup$ – JLee Nov 12 '19 at 15:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.