|
Abstract:
|
In many data-driven applications, different portions of the observed data are not always generated under equal conditions. Often experimental conditions change over time or across repetitions of the same experiment. In such settings, statistical findings are only reliable if the entire statistical inference pipeline is stable, that is, it produces the same conclusions across all experimental conditions. In this talk, we formalize this type of stability using causal models, and describe how it can be used as an inference principle which increases the reliability of statistical findings in heterogeneous data.
|
