Skip to contents

Let YY be the outcome of interest, AA be the (binary, coded as 0 for control and 1 for treatment) treatment, SS be the indicator for study participation (so that S=1S=1 means that the subject is in the population of the original study, while S=0S = 0 means that the subject is in the target population) and ๐„\mathbf{E} be effect modifiers. Let Y0Y^0 and Y1Y^1 be counterfactual outcomes associated with control and treatment, respectively. The primary objective of transportability analysis is to estimate the ATE in the target population: ATE=E[Y1โˆ’Y0|S=0].ATE = \mathrm{E}[Y^1 - Y^0 \,|\,S = 0].

Simply taking the difference in sample means using the original study data will only unbiasedly estimate the quantity E[Y|A=1,S=1]โˆ’E[Y|A=0,S=1],\mathrm{E}[Y \,|\,A = 1, S = 1] - \mathrm{E}[Y \,|\,A = 0, S = 1], which is different from the target ATE due to the different distributions of effect modifiers.

Interpolated g-computation is an adaptation of network meta-interpolation (Harari et al. 2022) in the transportability analysis context. To this end, we assume that we have aggregate-level data from a randomized clinical trial (RCT), so that no confounding adjustment is necessary. More specifically, suppose that the estimated treatment effect in the original study is ฮ”ฬ‚\hat{\Delta}, and that its estimated standard error is ฯƒฬ‚\hat{\sigma}. Let ฮ”ฬ‚ij\hat{\Delta}_{ij} and ฯƒฬ‚ij\hat{\sigma}_{ij} be the estimated treatment effect and its estimated standard error within level j=0,1j = 0,1 of the iith dichotomized effect modifier, for i=1,...,Mi = 1,...,M. Collect these estimates into vectors ๐šซ=(ฮ”ฬ‚ฮ”ฬ‚11ฮ”ฬ‚10โ‹ฎฮ”ฬ‚M0)\mathbf{\Delta} =\begin{pmatrix} \hat{\Delta} \\ \hat{\Delta}_{11} \\ \hat{\Delta}_{10} \\ \vdots \\ \hat{\Delta}_{M0} \end{pmatrix} and ๐›”=(ฯƒฬ‚ฯƒฬ‚11ฯƒฬ‚10โ‹ฎฯƒฬ‚M0).\mathbf{\sigma} =\begin{pmatrix} \hat{\sigma} \\ \hat{\sigma}_{11} \\ \hat{\sigma}_{10} \\ \vdots \\ \hat{\sigma}_{M0} \end{pmatrix}. Let pip_i be the proportion of observations with level 1 of the iith effect modifier in the original study sample, and let pโ€ฒip'_i be that in the target sample. Denote ๐ฉโ€ฒ=(1pโ€ฒ1โ‹ฎpโ€ฒM)\mathbf{p}' = \begin{pmatrix}1 \\ p'_1 \\ \vdots \\ p'_M \end{pmatrix} and ๐ฉโ€ณ=(1pโ€ฒ12โ‹ฎpโ€ฒM22pโ€ฒ1โ‹ฎ2pโ€ฒM2pโ€ฒ1pโ€ฒ22pโ€ฒ1pโ€ฒ3โ‹ฎ2pโ€ฒMโˆ’1pโ€ฒM).\mathbf{p}'' = \begin{pmatrix}1 \\ p'^2_1 \\ \vdots \\ p'^2_M \\ 2p'_1 \\ \vdots \\ 2p'_M \\ 2p'_1p'_2 \\ 2p'_1p'_3 \\ \vdots \\ 2p'_{M-1}p'_M\end{pmatrix}. Finally, let nn be the sample size of the original study, and let ฯik\rho_{ik} be the correlation between the dichotomized effect modifiers ii and kk, estimated from the target data or otherwise specified.

Let ๐Ÿl\mathbf{1}_l be a vector in โ„l\mathbb{R}^l whose entries are all 11. To transport the treatment effect estimate, we first calculate ๐ฑi=(xi0โ‹ฎxi,2M)\mathbf{x}_{i} = \begin{pmatrix} x_{i0} \\ \vdots \\ x_{i,2M} \end{pmatrix} for i=1,...,Mi = 1,...,M. We let xi0=pix_{i0} = p_i for i=1,...Mi = 1,...M. For the other entries, we calculate xik={1if i=2k0if i=2k+1ฯikpk(1โˆ’pk)npโŒˆi2โŒ‰(1โˆ’pโŒˆi2โŒ‰)n(xi,โŒˆi2โŒ‰โˆ’pโŒˆi2โŒ‰)+pkotherwise.x_{ik} = \begin{cases} 1 & \textrm{if } i = 2k \\ 0 & \textrm{if } i = 2k + 1 \\ \rho_{ik}\frac{\sqrt{\frac{p_k(1-p_k)}{n}}}{\sqrt{\frac{p_{\lceil\frac{i}{2}\rceil}(1-p_{\lceil\frac{i}{2}\rceil})}{n}}}(x_{i,\lceil\frac{i}{2}\rceil} - p_{\lceil\frac{i}{2}\rceil}) + p_k & \textrm{otherwise} \end{cases}. Essentially, within each marginal subgroup of an effect modifier, proportions of 1s of other effect modifiers are imputed using the best linear unbiased predictor (BLUP) in terms of the level of the effect modifier in the subgroup. Then let M1=[๐Ÿ2M+1โ‹ฏ๐ฑM]M_1 = \begin{bmatrix}\mathbf{1}_{2M+1} & \cdots & \mathbf{x}_M\end{bmatrix} and M2=[๐Ÿ2M+1๐ฑ12โ‹ฏ๐ฑM22๐ฑ1โ‹ฏ2๐ฑM2๐ฑ1๐ฑ22๐ฑ1๐ฑ3โ‹ฏ2๐ฑMโˆ’1๐ฑM].M_2 = \begin{bmatrix}\mathbf{1}_{2M+1} & \mathbf{x}_1^2 & \cdots & \mathbf{x}_M^2 & 2\mathbf{x}_1 & \cdots & 2\mathbf{x}_M & 2\mathbf{x}_1\mathbf{x}_2 & 2\mathbf{x}_1\mathbf{x}_3 & \cdots & 2\mathbf{x}_{M-1}\mathbf{x}_M\end{bmatrix}.

The transported effect estimate is ฮ”ฬ‚โ€ฒ=(๐ฉโ€ฒ)T(M1TM1)โˆ’1M1T๐šซ,\hat{\Delta}' = (\mathbf{p}')^T(M_1^TM_1)^{-1}M_1^T\mathbf{\Delta}, and its estimated standard error is ฯƒฬ‚โ€ฒ=(๐ฉโ€ณ)TM2T(M2M2T)โˆ’1๐›”2.\hat{\sigma}' = \sqrt{(\mathbf{p}'')^TM_2^T(M_2M_2^T)^{-1}\mathbf{\sigma}^2}. In other words, regression models of the treatment effects and their standard errors against effect modifiers are fit using the original study data, and the fitted models are used to calculate estimated treatment effects and standard errors at the proportions of 1s of effect modifiers in the target data.

References

Harari, O, M Soltanifar, JC Cappelleri, A Verhoek, M Ouwens, C Daly, and B Heeg. 2022. โ€œNetwork Meta-Interpolation: Effect Modification Adjustment in Network Meta-Analysis Using Subgroup Analyses.โ€ Research Synthesis Methods 14: 211โ€“33. https://doi.org/10.1002/jrsm.1608.