Learning the Covariance of Treatment Effects Across Many Weak Experiments

When primary objectives are insensitive or delayed, experimenters may instead focus on proxy metrics derived from secondary outcomes. For example, technology companies often infer long-term impacts of product interventions from their effects on weighted indices of short-term user engagement signals. We consider meta-analysis of many historical experiments to learn the covariance of treatment effects on different outcomes, which can support the construction of such proxies. Even when experiments are plentiful and large, if treatment effects are weak, the sample covariance of estimated treatment effects across experiments can be highly biased and remains inconsistent even as more experiments are considered. We overcome this by using techniques inspired by weak instrumental variable analysis, which we show can reliably estimate parameters of interest, even without a structural model. We show the Limited Information Maximum Likelihood (LIML) estimator learns a parameter that is equivalent to fitting total least squares to a transformation of the scatterplot of estimated treatment effects, and that Jackknife Instrumental Variables Estimation (JIVE) learns another parameter that can be computed from the average of Jackknifed covariance matrices across experiments. We also present a total-covariance-based estimator for the latter estimand under homoskedasticity, which we show is equivalent to a k-class estimator. We show how these parameters relate to causal quantities and can be used to construct unbiased proxy metrics under a structural model with both direct and indirect effects subject to the INstrument Strength Independent of Direct Effect (INSIDE) assumption of Mendelian randomization. Lastly, we discuss the application of our methods at Netflix.