The evolution of \(y_j\) over time is defined using a marginal model without subject-specific random effects.
$$ T_{0j} = x_{T0j} \gamma_{T0} + \varepsilon_{T0j}$$
$$T_{1j} = x_{T1j} \gamma_{T1} + \varepsilon_{T1j}$$
$$S_{0j} = x_{S0j} \gamma_{S0} + \varepsilon_{S0j}$$
$$S_{1j} = x_{S1j} \gamma_{S1} + \varepsilon_{S1j}$$
It can be written as \(y=x \gamma + \varepsilon\). It is assumed that \(\varepsilon \sim N(0, \Sigma)\),
or, equivalently, \(y \sim N(x \gamma, \Sigma_y)\). Essentially,
Galecki (1994) assumed that the covariance matrix is
separable, i.e. it can be written as \(\Sigma_y = \Sigma = R \otimes V\) , where \(V\) is assumed to be unstructured and
invariant across time, and \(R\) reflects a general correlation matrix.
Under Galecki's model,
the vector of individual causal treatment effects \(\Delta\) is:
$$ \Delta = \mu_{\Delta}+ \varepsilon_{\Delta} $$
and the marginal covariance matrix of the individual causal treatment effects is
\(\Sigma_{\Delta} = \Sigma_u \otimes R\).
Under Galecki's model, the ICA is given by:
$$ R_{\Lambda}^2 = 1 - (1-\rho_{u}^2)^p $$
In this model \(\rho_u=\rho_{\Delta}=corr(\Delta T, \Delta S)\), which is the metric used in the cross-sectional setting.
$$
\rho_u =
\frac{
\sqrt{\sigma_{T0T0} \, \sigma_{S0S0}} \, \rho_{T0S0} +
\sqrt{\sigma_{T1T1} \, \sigma_{S1S1}} \, \rho_{T1S1} -
\sqrt{\sigma_{T1T1} \, \sigma_{S0S0}} \, \rho_{T1S0} -
\sqrt{\sigma_{T0T0} \, \sigma_{S1S1}} \, \rho_{T0S1}
}{
\sqrt{
(\sigma_{T0T0} + \sigma_{T1T1} - 2 \sqrt{\sigma_{T0T0} \, \sigma_{T1T1}} \, \rho_{T0T1})
(\sigma_{S0S0} + \sigma_{S1S1} - 2 \sqrt{\sigma_{S0S0} \, \sigma_{S1S1}} \, \rho_{S0S1})
}
}
$$