During the pandemic countries in the African region used data systems they built on top of the WHO Integrated Disease Surveillance and Response (IDSR) framework to track COVID-19 and formulate public health responses. Our IDSR project wrangles these data systems into instances of a common data model and conducts network research on top of them.
0
stars
0
forks
source link
Replicating the PROVE project with the IDSR in ATLAS: More concerns #6
To adjust for differences between the two treatment groups in the OHDSI data analysis work bench, several adjustment strategies can be used, such as stratification, matching, or weighting by the propensity score, OR by adding baseline characteristics to the outcome model. It appears in your analysis, you are adding baseline characteristics to a logistic regression model and you are not using the propensity score in any way, shape or form.
Recall that in a propensity score-adjusted observational study, we estimate the probability of a patient receiving the target treatment based on what we can observe in the data on and before the time of treatment initiation (irrespective of the treatment they actually received).
The PS can be used in several ways including matching target subjects to comparator subjects with similar PS, stratifying the study population based on the PS, or weighting subjects using Inverse Probability of Treatment Weighting (IPTW) derived from the PS.
In one-on-one PS matching, we find one or more matched patients that received the comparator but had the same prior probability of receiving the target. Then we compare the outcome for the target patient with the outcomes for the comparator patients within each of these matched groups
Now consider the alternative which adds baseline characteristics to a logistic regression model in ATLAS
In this approach ATLAS does NOT identify confounders by stratifying a vaccine effectiveness statistic on candidate confounders. Instead ATLAS tries the regression with all the possible characteristics. Indeed, there is no "selection". ATLAS is simply driven by the available data.
Note that "Propensity score analysis is overall a more favorable approach than traditional regression analysis when estimating causal effects using observational data." See here.
This discussion suggests we create two emulated clinical trial templates — a “traditional" one that uses the outcome model only to identify confounders and a template that specifies how to use propensity scores either alone or in conjunction with the outcome model. We might call this template the “advanced” emulated clinical trial template based on research that compares the two approaches.
To adjust for differences between the two treatment groups in the OHDSI data analysis work bench, several adjustment strategies can be used, such as stratification, matching, or weighting by the propensity score, OR by adding baseline characteristics to the outcome model. It appears in your analysis, you are adding baseline characteristics to a logistic regression model and you are not using the propensity score in any way, shape or form.
Recall that in a propensity score-adjusted observational study, we estimate the probability of a patient receiving the target treatment based on what we can observe in the data on and before the time of treatment initiation (irrespective of the treatment they actually received).
The PS can be used in several ways including matching target subjects to comparator subjects with similar PS, stratifying the study population based on the PS, or weighting subjects using Inverse Probability of Treatment Weighting (IPTW) derived from the PS.
In one-on-one PS matching, we find one or more matched patients that received the comparator but had the same prior probability of receiving the target. Then we compare the outcome for the target patient with the outcomes for the comparator patients within each of these matched groups
Now consider the alternative which adds baseline characteristics to a logistic regression model in ATLAS
In this approach ATLAS does NOT identify confounders by stratifying a vaccine effectiveness statistic on candidate confounders. Instead ATLAS tries the regression with all the possible characteristics. Indeed, there is no "selection". ATLAS is simply driven by the available data.
Note that "Propensity score analysis is overall a more favorable approach than traditional regression analysis when estimating causal effects using observational data." See here.
This discussion suggests we create two emulated clinical trial templates — a “traditional" one that uses the outcome model only to identify confounders and a template that specifies how to use propensity scores either alone or in conjunction with the outcome model. We might call this template the “advanced” emulated clinical trial template based on research that compares the two approaches.