It is a sad truth that the same data can be analyzed different ways to get different results. Also, multiple models might be reasonable choices for a given dataset. Still, it is not a great feeling to wonder whether what seems like a subjective design decision is driving the results. In some ongoing work, we are …

A Comprehensive Guide to Fine-Tuning ChatGPT for Essay Grading Introduction Our “Esssay Grading with ChatGPT” blog post series have unveiled the potential of ChatGPT for essay grading. We started with the fundamentals of the ChatGPT API and gradually explored the art of crafting effective prompts, building a solid …

This second entry in the 'Essay Grading with ChatGPT' series delves deeper into this challenge, comparing the outcomes of essay grading based on different prompts (the instructions we give to ChatGPT) to optimize AI for educational purposes.

How to Grade Essays with ChatGPT Introduction The rise of large language models (LLMs) like OpenAI’s ChatGPT has opened exciting possibilities in essay grading. With its advanced natural language processing capabilities, ChatGPT offers a new dimension in assessing written work, potentially revolutionizing the grading …

Estimating subgroup impacts in an RCT can be hard. An RCT by itself is usually underpowered for this task–we barely have enough data to give us an overall average, and as subgroups are smaller, they are noisier! One idea recently gaining increased traction is to augment an RCT with observational data. We might use …

Do you ever want to visualize the distribution of effects across sites in a multi-site evaluation (or meta analysis)? For example, consider a multisite trial with 30 sites, where each site is effectively a small randomized experiment. A researcher might fit a multilevel model with a random effect for the impact in each …

There are a lot of ways to do power calculations, but in this post we show off a package, `PUMP`, that we recently set out into the world. It makes it easy to explore many combinations of ones design parameters, allowing for rapid exploration of power. It also handles corrections for multiple testings!

A simulation and R tutorial demonstrating how to recover interpretable treatment effects from logistic regression.

When does blocking help? Should we always do it if we can? In this post we talk about how there are actually different kinds of blocking, and different ways of thinking about where blocks come from. The question of whether blocking can ever cause harm turns out to depend on these perspectives.

When estimating average treatment impact for a multisite randomized trial, recent literature (see Bloom et al. (2017)) has suggested using the so-called FIRC (Fixed Intercept, Random Coefficient) model. There are some technical details that make fitting this model slightly tricky in R; this post talks through these …