CS3d: Celebrating Our Research

Glomerular filtration rate (GFR) is measure of how well the kidneys filter the blood by removing waste and extra water to make urine. GFR is used to help diagnose kidney disease at an early stage, to monitor people with chronic kidney disease (CKD), to monitor people with other conditions that cause kidney damage, such as diabetes and high blood pressure. Often treatment interventions may result in an initial acute effect on eGFR shortly after the treatment that differs from the long-term treatment effect. When this occurs, the rate of decline in eGFR over time may be characterized by two slopes: an acute slope and chronic slope. In this study, we present longitudinal data modeling strategies characterizing the rate of decline in eGFR among individuals with AKI and CKD in a large cohort of lung transplant recipients. 

Randomized controlled experiments (or A/B testing) remain the gold standard for measuring causal effects. However, we are sometimes faced with challenges, like limited resources or technical capabilities, that prevent us from A/B testing. When A/B testing is not feasible, data scientists often look to causal inference techniques to help draw conclusions about causal effects. User data from web and mobile applications pose challenges to traditional causal inference methods, like Difference-in-Differences and Regression Discontinuity. For applications with a regularly returning user base, pre v. post-intervention comparisons become difficult as the assumption of independent observations is often violated and the impact of confounders varies over time. Instead, the Data Science team at the Los Angeles Times has adopted causal inference techniques for time series data, such as Causal Impact analysis and Regression Discontinuity in Time, which allow control over autocorrelation in the data. In this session, we'll discuss the specific challenges we faced in designing traditional causal inference studies for product changes, and the alternative approaches we took for working with time series data. We'll share findings from projects that utilized Causal Impact analysis and Regression Discontinuity in Time, and limitations that persist even with these approaches.  

This study investigates long-range dependence, or long memory, focusing on its relevance in spatial and temporal data analysis. Processes with long-range dependence have correlations that decay slower than exponentially, which is important in fields like economics and hydrology. The parameter d quantifies the persistence: d <= 0 indicates short-range dependence, 0 < d < 0.5 indicates long-range dependence, and d >= 0.5 indicates non-stationarity. Accurate estimation of d is crucial for characterizing temporal dynamics in data.

Our research focuses on estimating d within a time series with stationary increments. We use a regression technique proposed by Geweke and Porter-Hudak (1983), enhanced by the spectral density estimation method by Chen et al. (2024). Simulations demonstrate the robustness and precision of the modified GPH method in quantifying d across stationary increments without transforming the data to stationarity.

In practical applications, we apply the modified GPH method to Consumer Price Index (CPI) data, using Cressie's graphical method to determine the order of stationary increments. Exponential Smoothing effectively captures persistent correlations, making it suitable for CPI forecasting, where long-term trends are crucial for accurate predictions. 

Conducting reliable causal inference in observational longitudinal data analysis encounters challenges when unmeasured confounding is present. Inspired by two influential methodologies, instrumental variables and nuclear norm-penalized regression, we propose a novel method called instrumented nuclear norm-penalized regression. This method aims to estimate average causal effects of exposure on outcome for longitudinal data while accounting for potential unmeasured confounding. Our method handles confounding variables with low-rank structures and provides flexibility by relaxing the exclusion restriction linked with instrumental variables. We develop the identification assumptions utilizing the potential outcome framework and provide theoretical results demonstrating the consistency of the proposed estimator. We also conduct simulation studies and apply the method to estimate the number of collisions deterred by the issuance of traffic tickets by police officers in New York City, validating its efficacy and robustness in comparison with competing methods.