Disruption Sensitivity Modeling With Propensity-Score Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current predictive modeling approaches fail to accurately quantify the sensitivity of outcomes to potential future disruptions due to their reliance on large, randomizable data sets and inability to account for specific current conditions without generalizing assumptions, leading to speculative or computationally unwieldy results.
Innovation Solution
A supervised machine learning-based method that divides historical data into matched sub-populations using propensity scores to model sensitivity indices, accounting for non-randomized data distributions and enabling robust predictions of entity performance under disrupted and less disrupted conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large, randomizable data sets are used for predictive modeling, then statistical power is improved, but applicability to specific current conditions deteriorates due to necessary generalizing assumptions
Solution Approach 1:
The patent segments the population into treated and control sub-populations based on actual treatment assignment rather than randomization. This allows the model to maintain statistical power while being applicable to specific current conditions by using observed data structures rather than requiring randomized controlled trial data.
Solution Approach 2:
The patent changes the fundamental parameter of data structure from randomized to non-randomized. By developing methods that work with observed treatment assignments and propensity scores rather than requiring randomization, the model maintains reliability while improving adaptability to real-world conditions where randomization is not feasible.
2Productivity
If current predictive modeling approaches are used to quantify sensitivity to disruptions, then computational simplicity is maintained, but accuracy deteriorates due to speculative results
Solution Approach 1:
The patent introduces propensity scores as an intermediary variable that bridges the gap between observed historical data and causal inference. This mediator allows the model to accurately quantify sensitivity to disruptions by accounting for selection bias and confounding factors, while maintaining computational feasibility through efficient matching and weighting procedures.
Solution Approach 2:
The patent performs preliminary actions by calculating propensity scores and creating matched samples before conducting the sensitivity analysis. This preprocessing step eliminates selection bias and confounding factors in advance, allowing the main analysis to proceed with computationally simple operations on already-balanced groups, thereby maintaining both accuracy and computational efficiency.
3Quantity of substance
If non-randomized historical data is used for modeling, then data availability is improved, but measurement bias increases due to non-random data distributions
Solution Approach 1:
The patent extracts the selection bias and confounding factors from the non-randomized data by calculating propensity scores. This extraction process separates the systematic differences between treated and control groups, allowing the model to use abundant historical data while correcting for measurement bias through the propensity score weighting and matching procedures.
Data Source
AI summary
Computer-implemented systems, methods and products for modeling sensitivities to potential disruptions by observing performances of entities in a first sub-population and a second sub-population using a machine learning model comprising a set of predictors and a binary indicator variable associated with a first entity subjected to a first event associated with the first sub-population, the machine learning model trained to predict an expected performance for the first entity based on at least one of a known attribute associated with the first entity in relation to the first event and a value of the binary indicator variable associated with the first event.