|
| 1 | +--- |
| 2 | +title: 'outstandR: An R Package for Indirect Treatment Comparison with Limited Subject-Level Data' |
| 3 | +tags: |
| 4 | + - R |
| 5 | + - health economics |
| 6 | + - outcome standardisation |
| 7 | + - Bayesian inference |
| 8 | +authors: |
| 9 | + - name: Nathan Green |
| 10 | + orcid: 0000-0003-2745-1736 |
| 11 | + equal-contrib: true |
| 12 | + affiliation: 1 |
| 13 | + - name: Antonio Remiro-Acócar |
| 14 | + orcid: |
| 15 | + equal-contrib: true |
| 16 | + affiliation: 2 |
| 17 | + - name: Gianluca Baio |
| 18 | + corresponding: true |
| 19 | + affiliation: 1 |
| 20 | +affiliations: |
| 21 | + - name: University College London (UCL), UK |
| 22 | + index: 1 |
| 23 | + ror: 02jx3x895 |
| 24 | + - name: Novonodisk |
| 25 | + index: 2 |
| 26 | + ror: |
| 27 | +date: 1 August 2025 |
| 28 | +bibliography: paper.bib |
| 29 | +editor_options: |
| 30 | + markdown: |
| 31 | + wrap: 72 |
| 32 | +--- |
| 33 | + |
| 34 | +# Summary |
| 35 | + |
| 36 | +The goal of `outstandR` is to provide functionality to perform |
| 37 | +population adjustment methods are increasingly used to compare marginal |
| 38 | +treatment effects when there are cross-trial differences in effect |
| 39 | +modifiers and limited patient-level data [@RemiroAzocar2022a]. This |
| 40 | +presents a significant challenge in evidence synthesis, particularly in |
| 41 | +health technology assessment. Existing methods often face limitations |
| 42 | +such as sensitivity to poor covariate overlap or inability to |
| 43 | +extrapolate beyond observed covariate spaces. The `outstandR` package |
| 44 | +addresses these challenges by offering a robust framework for out- come |
| 45 | +regression standardisation, focusing on G-estimation. It enables |
| 46 | +researchers to perform model-based standardization with two additional |
| 47 | +crucial steps: covariate simulation (to over- come limited subject-level |
| 48 | +data for one of the studies) and indirect comparison across studies |
| 49 | +[@RemiroAzocar2022a]. The target audience of `outstandR` is those who |
| 50 | +want to perform model-based standardization in the specific context of |
| 51 | +two-study indirect treatment comparisons with limited subject-level |
| 52 | +data. |
| 53 | + |
| 54 | +# Statement of need |
| 55 | + |
| 56 | +Indirect treatment comparisons (ITCs) are a cornerstone of modern |
| 57 | +evidence synthesis, especially in health technology assessment (HTA) |
| 58 | +where decision-makers must compare novel treatments against a range of |
| 59 | +competitors. A common and challenging scenario arises when individual |
| 60 | +patient data (IPD) is available for one trial, but only aggregate-level |
| 61 | +data (ALD) is available for the comparator trial. Naively comparing |
| 62 | +these studies can introduce significant bias because of differences in |
| 63 | +patient populations. There is a clear need for a unified and robust |
| 64 | +software tool that implements a range of modern population adjustment |
| 65 | +techniques to address these challenges systematically. |
| 66 | + |
| 67 | +# Method |
| 68 | + |
| 69 | +We developed the `outstandR` R package to provide a comprehensive |
| 70 | +framework for performing anchored ITCs using a suite of population |
| 71 | +adjustment methods, with a focus on robust G-computation techniques. |
| 72 | +`outstandR` streamlines the entire analytical workflow: from fitting |
| 73 | +outcome models on IPD and standardizing them to the ALD population, to |
| 74 | +performing the final indirect comparison. By implementing multiple |
| 75 | +methods — including Matching-Adjusted Indirect Comparison (MAIC), |
| 76 | +Simulated Treatment Comparison (STC) [@phillippo2016methods], parametric |
| 77 | +G-computation (both frequentist and Bayesian) [@RemiroAzocar2022a], and |
| 78 | +the Multiple Imputation Marginalization (MIM) method |
| 79 | +[@RemiroAzocar2022b] — within a single interface, our package empowers |
| 80 | +researchers to conduct sensitivity analyses and select the most |
| 81 | +appropriate approach for their data. `outstandR` lowers the technical |
| 82 | +barrier to entry for these complex analyses, promoting more reliable and |
| 83 | +transparent evidence synthesis for healthcare decision-making. |
| 84 | + |
| 85 | +Related R packages we are aware of are more general-purpose tools |
| 86 | +implementing specific methods. The `marginaleffects` [@aari_marginaleffects_2024] package is not |
| 87 | +designed for population adjustment between studies. `stdReg2` focuses on |
| 88 | +standardising outcomes with a single data set [@g_sofer_2023_10022204]. |
| 89 | +For G-formula |
| 90 | +implementations, `gfoRmula` can estimate effects in the presence of |
| 91 | +time-varying treatments and confounders [@sjolander2023gformula]. It is designed for estimating |
| 92 | +causal effects from longitudinal data with one study. `gFormulaMI` |
| 93 | +employs multiple imputation using the `mice` package [@Sterne2023]. Finally, |
| 94 | +`maicplus` is a specialist ITC package but focused only on the MAIC |
| 95 | +approach [@maicplus_2024]. |
| 96 | + |
| 97 | +Our analysis performs an anchored indirect treatment comparison (ITC) to |
| 98 | +estimate the relative effect of two treatments, A and B. This comparison |
| 99 | +uses individual patient data (IPD) from a trial comparing treatments A |
| 100 | +and C (the AC trial) and aggregate level data (ALD) from a trial |
| 101 | +comparing treatments B and C (the BC trial). To account for potential |
| 102 | +differences in the patient populations between the two trials, which can |
| 103 | +bias the ITC, we use population adjustment methods. These methods |
| 104 | +standardize the results from the AC trial to the baseline |
| 105 | +characteristics of the BC trial population. |
| 106 | + |
| 107 | +The general procedure involves two main stages. First, we fit an outcome |
| 108 | +regression model using the IPD from the AC trial. This model describes |
| 109 | +the relationship between the outcome, treatment assignment, and a set of |
| 110 | +baseline covariates, including both prognostic factors and treatment |
| 111 | +effect modifiers. Second, we use this fitted model to predict the |
| 112 | +outcomes for treatment A and C in a target population that reflects the |
| 113 | +aggregate baseline characteristics of the BC trial. For G-computation, |
| 114 | +this step can involve simulating a large synthetic dataset that mirrors |
| 115 | +the covariate distributions of the BC population. The resulting adjusted |
| 116 | +treatment effect, $\Delta$AC(BC), is then indirectly compared against |
| 117 | +the observed effect from the ALD, $\Delta$BC(BC), to yield the final |
| 118 | +adjusted estimate for A versus B, $\Delta$AB(BC). Uncertainty is |
| 119 | +quantified using non-parametric bootstrapping for frequentist methods or |
| 120 | +by propagating parameter uncertainty from the full posterior |
| 121 | +distribution in Bayesian implementations. |
| 122 | + |
| 123 | +# References |
0 commit comments