In the Potential Outcomes literature, terms that start with 'G' can be hard to differentiate at first sight. These terms include the G-methods, the G-formula and G-estimation (and potentially more). Hernan and Robin's textbook on causal inference via the Potential Outcomes framework has an extensive discussion, throughout the book, of what these terms mean. Hence, we will borrow from there and repeatedly refer to it for further reading.
The confusion is mainly generated by the repeated use of the letter G, and is best resolved by understanding time-varying treatments for which the distinction is necessary. That is, understanding the motivation for time-varying treatment is a good path to understanding the difference between the different G-terms.
It is easiest to start with a quote from Hernan and Robin's book and explain it word by word:
IP weighting, standardization, and g-estimation are often collectively referred to as g-methods because they are designed for application to generalized treatment contrasts involving treatments that vary over time. The application of g-methods to treatments that do not vary over time in Part II of this book may then be overkill since there are alternative, simpler approaches. However, by presenting g-methods in a relatively simple setting, we can focus on their main features while avoiding the more complex issues described in Part III. (Introduction paragraph, Chapter 14)
We can see that G-methods is a group of methods that include IP weighting, standardization and g-estimation. They are referred to as G-methods, as they are designed for generalized treatment constrasts involving treatments varying over time. This refers to two different types of causal inference studies:
- A one-off/one-shot study where we analyse one snapshot of time
- A time-varying study, where we look at data collected at different points in time, for example an AIDS drug applied every month for a year, i.e. 12 time slices.
The G-methods are specialised for studies with multiple time steps. They can be applied to studies looking at a single point in time, but there are simpled methods to do that.
A further defintion can be found in (Hernan, 2020, Section 7.6):
Here, methods are separated into
- model free/ non-parametric:
- standardization
- IP weighting
- g-estimation.
- model based/ parametric:
- parametric g-formula
- g-estimation of nested structural models
In that section, G-methods are contrasted with 'Stratification-based methods':
G-methods simulate the A-Y association in the population if backdoor paths involving the measured variables L did not exist. For example, IP weighting achieves this by creating a pseudo-population in which treatment A is independent of the measured confounders L that is, by “deleting” the arrow from L to A. In contrast, stratification-based methods do not delete the arrow from L to A but rather compute the conditional effect in a subset of the observed population, which is represented by adding a selection box. The advantage of “deleting” the arrow from confounders L to treatment A will become apparent when we discuss time-varying treatments in Part III
Here, A is the treatment (more often denoted by X), L are the confounders (more often denoted by U or C). Importantly, 'deleting the arrow from L to A' is seen better for dealing with time-varying treatments, i.e. G-methods are generally better for time-varying studies compared to startification-based methods. See the entry on Confounding for an explanation on 'deleting the arrow from A to L'.
Hernan and Robin's, 2020: "What if?" Section 7.6 and beginning of Chapter 14, here