As AI advances, large language models (LLMs) are shifting from passive tools to active agents that collaborate with experts to cocreate knowledge and artifacts. In this talk, I will explore the role of LLMs as intelligent agents in data science workflows—partners that not only automate tasks but also enhance decision-making by understanding core data science principles, identifying cognitive biases, and nudging experts toward more robust conclusions.

I will discuss how an LLM, equipped with statistical reasoning, ethical AI considerations, and an awareness of human cognitive pitfalls, can challenge assumptions, suggest alternative methodologies, and improve model interpretability. From guiding feature selection to questioning spurious correlations, these AI agents act as reflective collaborators rather than mere calculators. I will examine case studies where LLMs have meaningfully influenced analytical processes, highlight challenges in aligning AI nudges with human intent, and explore the future of AI-augmented data science, generally and while using Stata. This talk is primarily conceptual and designed to inspire but also to rethink our relationship with AI—not as a tool but as a cocreator in the pursuit of knowledge.

Multiverse analysis is becoming an important tool in the methodological repertoire of social scientists. The idea behind the method—variously referred to as “multiverse analysis”, “multimodel analysis”, “specification curve analysis”, or “vibration of effects”—is straightforward: because there are many credible ways of formulating an analysis, and any single statistical estimate may suffer from selective reporting, multiverse analysis explores all reasonable specifications, contrasting authors’ preferred estimates with a range of possible estimates.

Instead of luring readers into a dark corner of the “garden of forking paths”, multiverse analysis provides a bird’s-eye view of the maze of researcher decisions and the resulting range of defensible findings. While multiverse analysis holds significant promise for quantitative empirical research, it poses conceptual, computational, and practical challenges. This talk provides a primer on implementing multiverse analysis in Stata. It highlights the strengths and limitations of existing multiverse tools (for example, mrobust and multivrs) and introduces a new plot type designed to visualize multiverse results effectively. By addressing key challenges in conducting and visualizing multiverse results, the talk seeks to encourage Stata users to adopt multiverse analysis and unlock its potential for robust and transparent research.

Researchers often want to mitigate the increased risk of type I errors that arises from multiple pairwise comparisons of means. Stata provides seven methods to adjust the corresponding confidence intervals and p-values. However, four of these methods assume equal sample sizes, variances, or both, and none explicitly addresses unequal variances, which might pose limitations on applied research. In this presentation, we briefy review how the implemented methods modify the significance level or obtain critical values from alternative distributions to adjust for multiple comparisons. We then discuss three methods that explicitly account for unequal variances by making additional adjustments to standard errors and degrees of freedom. Finally, we (re)introduce the pwmc command in Stata, which implements these three methods, and compare their performance using a Monte Carlo simulation.

This talk presents an egen function to convert units of measurements for length, areas, volumes, angles, masses, temperatures, and currency. The function allows one to convert many non-SI units (for example, inch, furlong, sunradius) to SI units (from pico to peta) or directly from a non-SI unit to another non-SI unit. Currencies are converted by calling the European Central Bank through an API. The conversion rate can be selected on a daily base or by an average of a specified period. German users may be relieved to realize that the function allows converting areas also into units of “Saarland”.

Stata 18 introduced two commands (each with four estimators) to fit heterogeneous DID models: hdidregress for repeated cross-sectional data and xthdidregress for panel/longitudinal data. In this talk, we briefly introduce the theory behind both estimators and then show how to fit heterogeneous DID models using the new commands. We also demonstrate postestimation tools to aggregate and visualize heterogeneous treatment effects and perform diagnostic tests.

The Stata license model is gradually changing from perpetual licenses to a pay-as-you-go model called StataNow™. While this gives researchers and users of Stata the advantage of always having access to the latest features of the software, the pay-as-you-go model requires different planning and budgeting for software. Many different options to license Stata exist, depending on edition, usage, organization, and many other factors. This talk makes suggestions for finding a license option that meets the functional requirements, including multiyear models and covering of EUR/USD fluctuations.

In 2008, I published the Stata command oaxaca, which implements the popular Oaxaca–Blinder (OB) decomposition technique. This technique is used to analyze differences in outcomes between groups, such as the wage gap by gender or race. Over the years, both the functionality of Stata and the literature on decomposition methods have evolved, so that an update of the oaxaca command is now long overdue. In this talk, I will present a revised version of oaxaca that uses modern Stata features such as factor-variable notation and supports additional decomposition variants that have been proposed in the literature (for example, reweighted decompositions or decompositions based on recentered infuence functions).

Multistate life tables (MSLTs), or multistate survival models, have become a widely used analytical framework in the social and health sciences. These models can be cast in continuous or discrete time. The dtms Stata module (dtms stands for "discretetime multistate"), which was presented at the German Stata Conference 2023 (Schneider 2023), implements the estimation of the discrete-time flavor of these models. This presentation first outlines discrete-time multistate estimation and then gives an overview of recent package enhancements including the following: external multinomial logistic estimation results, for example, from the interpolated Markov chain (IMaCh) executable (Brouard 2021), can be imported for further processing; difficulties with reloading saved dtms files across package versions have been resolved; the initial state distribution has been incorporated into the asymptotic analysis; new result type “evol” calculates the evolution of population fractions, along with the corresponding covariance matrix; estimation based on restricted transitions has been improved; transition probabilities can be based on time-varying covariate values; and several dtms trees can now be held in memory.

References:

Brouard, N. 2021. Computing Health Expectancies Using IMaCh: A maximum likelihood computer program using interpolation of Markov chains. Paris: Institut National d'Etudes Demographiques (INED) and EUROREVES. Available at: euroreves.ined.fr/imach/.

Schneider, D. C. 2023. Discrete-time multistate regression models in Stata: The dtms module. German Stata Conference, Berlin. Available at: stata.com/meeting/germany23/slides/Germany23_Schneider.pdf.