In the continuous normal setting, surroagacy can be assessed by studying the association between the individual causal effects on \(S\) and \(T\) (see ICA.ContCont). In that setting, the Pearson correlation is the obvious measure of association.
When \(S\) and \(T\) are binary endpoints, multiple alternatives exist. Alonso et al. (2014) proposed the individual causal association (ICA; \(R_{H}^{2}\)), which captures the association between the individual causal effects of the treatment on \(S\) (\(\Delta_S\)) and \(T\) (\(\Delta_T\)) using information-theoretic principles.
The function ICA.BinBin computes \(R_{H}^{2}\) based on plausible values of the potential outcomes. Denote by \(\bold{Y}'=(T_0,T_1,S_0,S_1)\) the vector of potential outcomes. The vector \(\bold{Y}\) can take 16 values and the set of parameters \(\pi_{ijpq}=P(T_0=i,T_1=j,S_0=p,S_1=q)\) (with \(i,j,p,q=0/1\)) fully characterizes its distribution.
However, the parameters in \(\pi_{ijpq}\) are not all functionally independent, e.g., \(1=\pi_{\cdot\cdot\cdot\cdot}\). When no assumptions regarding monotonicity are made, the data impose a total of \(7\) restrictions, and thus only \(9\) proabilities in \(\pi_{ijpq}\) are allowed to vary freely (for details, see Alonso et al., 2014). Based on the data and assuming SUTVA, the marginal probabilites \(\pi_{1 \cdot 1 \cdot}\), \(\pi_{1 \cdot 0 \cdot}\), \(\pi_{\cdot 1 \cdot 1}\), \(\pi_{\cdot 1 \cdot 0}\), \(\pi_{0 \cdot 1 \cdot}\), and \(\pi_{\cdot 0 \cdot 1}\) can be computed (by hand or using the function MarginalProbs). Define the vector
$$\bold{b}'=(1, \pi_{1 \cdot 1 \cdot}, \pi_{1 \cdot 0 \cdot}, \pi_{\cdot 1 \cdot 1}, \pi_{\cdot 1 \cdot 0}, \pi_{0 \cdot 1 \cdot}, \pi_{\cdot 0 \cdot 1})$$ and \(\bold{A}\) is a contrast matrix such that the identified restrictions can be written as a system of linear equation
$$\bold{A \pi} = \bold{b}.$$
The matrix \(\bold{A}\) has rank \(7\) and can be partitioned as \(\bold{A=(A_r | A_f)}\), and similarly the vector \(\bold{\pi}\) can be partitioned as \(\bold{\pi^{'}=(\pi_r^{'} | \pi_f^{'})}\) (where \(f\) refers to the submatrix/vector given by the \(9\) last columns/components of \(\bold{A/\pi}\)). Using these partitions the previous system of linear equations can be rewritten as
$$\bold{A_r \pi_r + A_f \pi_f = b}.$$
The following algorithm is used to generate plausible distributions for \(\bold{Y}\). First, select a value of the specified grid of values (specified using Sum_Pi_f in the function call). For \(k=1\) to \(M\) (specified using M in the function call), generate a vector \(\pi_f\) that contains \(9\) components that are uniformly sampled from hyperplane subject to the restriction that the sum of the generated components equals Sum_Pi_f (the function RandVec, which uses the randfixedsum algorithm written by Roger Stafford, is used to obtain these components). Next, \(\bold{\pi_r=A_r^{-1}(b - A_f \pi_f)}\) is computed and the \(\pi_r\) vectors where all components are in the \([0;\:1]\) range are retained. This procedure is repeated for each of the Sum_Pi_f values. Based on these results, \(R_H^2\) is estimated. The obtained values can be used to conduct a sensitivity analysis during the validation exercise.
The previous developments hold when no monotonicity is assumed. When monotonicity for \(S\), \(T\), or for \(S\) and \(T\) is assumed, some of the probabilities of \(\pi\) are zero. For example, when montonicity is assumed for \(T\), then \(P(T_0 <= T_1)=1\), or equivantly, \(\pi_{1000}=\pi_{1010}=\pi_{1001}=\pi_{1011}=0\). When monotonicity is assumed, the procedure described above is modified accordingly (for details, see Alonso et al., 2014). When a general analysis is requested (using Monotonicity=c("General") in the function call), all settings are considered (no monotonicity, monotonicity for \(S\) alone, for \(T\) alone, and for both for \(S\) and \(T\).)
To account for the uncertainty in the estimation of the marginal probabilities, a vector of values can be specified from which a random draw is made in each run (see Examples below).