What does the independent calculation test?
G-computation uses fitted models to predict outcomes under the intervention settings specified for the analysis and then averages those predictions over the observed data. The comparison checks the resulting direct effect, joint indirect effect and total contrast using an independent calculation.
The models and the definition of each effect must match before the numbers can be compared. Agreement then supports the calculation for that observed-data design. Whether the contrast can be interpreted causally also depends on the study assumptions, including whether relevant confounding has been addressed and the data support the settings being compared.
Which version produced the checked effects?
Both records refer to the same development version listed below. The first contains the point-estimate comparison; the second is a summary of that work. Listing both helps you locate the evidence, but does not mean that the calculation was independently repeated twice.
On a small screen, swipe the table sideways to see every column.
| Record | QuickPLS build or result checked | Reference | Comparison outcome |
|---|---|---|---|
| Interventional mediation point reference | 1.01.20 correction source | See which mathematical and app-workflow checks were performed | Passed the comparisons documented for this record |
| Interventional mediation evidence | 1.01.20 correction source | See which mathematical and app-workflow checks were performed | Mathematical reference plus checks of the specified calculations and app workflows |
The three effects in the observed-data example
The example uses a fixed synthetic dataset with 140 rows, analysed as observed data. An independent ordinary least squares (OLS) calculation fits the linear equations and calculates the three g-computation effects. The table reports the allowed error, number of point estimates and largest observed difference.
A point estimate is the single effect value calculated from the observed dataset. Confidence intervals describe uncertainty around an estimate. This test did not independently refit the bootstrap samples used to calculate those intervals.
- The result belongs to the recorded source and software version. A later version label does not mean that this calculation was repeated.
- The comparison counts refer to checked effect values from this example, not to independent studies.
| What was checked | Value / scope |
|---|---|
| Absolute tolerance | 1e-10 |
| Estimates compared | 3 |
| Maximum absolute error | 8.881784197001252e-16 |
| Result | passed |
The direct, joint indirect and total effect values
The direct and joint indirect effects are parts of the intervention contrast defined by the model; the total row reports the overall contrast. For example, the displayed joint indirect effect is 0.507079315206138 in QuickPLS and 0.507079315206139 in the reference. The difference column lets you inspect that small discrepancy directly.
All displayed values use 15 significant digits. The differences were calculated before display rounding. Use the download for the original recorded precision and effect definitions, and check those definitions before applying the comparison to another model.
On a small screen, swipe the table sideways to see every column.
| Quantity compared | QuickPLS | Reference | Absolute difference | Comparison result |
|---|---|---|---|---|
Interventional direct effectResult identifierinterventional_direct_effect | 0.24651407663568 | 0.246514076635681 | 3.60822483003176e-16 | See the detailed comparison record |
Joint interventional indirect effectResult identifierjoint_interventional_indirect_effect | 0.507079315206138 | 0.507079315206139 | 4.44089209850063e-16 | See the detailed comparison record |
Total interventional contrastResult identifiertotal_interventional_contrast | 0.753593391841819 | 0.753593391841819 | 8.88178419700125e-16 | See the detailed comparison record |
The later summary of the same calculation
This record summarises the independent check of effect estimates from parametric g-computation on observed data. Parametric means that the calculation uses the specified form of statistical model. It is supporting documentation for the point comparison above, not an additional independent set of estimates.
- The summary does not establish causal identification, meaning that the effect can be recovered from the observed data under justified assumptions. It also does not independently refit the models used for confidence intervals.
What the numerical agreement supports
The three effects agree within the stated absolute-error limit. That supports the model calculations for this saved example. It does not establish that all possible study designs satisfy the assumptions or that the bootstrap confidence intervals have been independently reproduced.
Read the point-estimate record and its later summary as two descriptions of the same comparison. Their counts should not be added. When inspecting the table, use the reported differences and error limit rather than judging agreement only from rounded values.
- The maximum allowed absolute error for the interventional point comparison is 1e-10. The test description identifies the effects that were checked.
How to inspect the mediation comparison
The JSON summary covers the software version, model type, three checked effects and numerical limits. Use the CSV to inspect the available effect values and differences. Together they document the comparison; they are not a complete recipe for repeating the fitted model.
Repeating the calculation requires the original dataset and the same model, settings and software versions. These downloads report completed work; the original dataset and app source code are not included.
Download the comparison results
See the values we checked, the differences we found and the software versions used. The download also identifies the original test files.
About these tests
These are QuickPLS development comparisons. Each record identifies the reference software or mathematical calculation, the tested build and the settings used.
How to understand these comparisons · Find checks for another method