Estimands are the starting point for methodological choices

Research should begin by stating the quantity to be estimated: the estimand. This involves defining the unit of analysis, the outcome quantity defined for each unit, and the target population of units. "What is your estimand?" (2021 ASR, with Rebecca Johnson and Brandon Stewart) lays out this vision of estimands as the starting point of methodological choices.

I develop methods that define new estimands

The gap-closing estimand

To what degree would gaps across race, class, and gender close if we intervened to equalize a treatment variable? In the gap-closing estimand (2024, SMR), I show how to answer this type of question in a causal framework. The gapclosing package implements these methods in R.

Categorical and continuous treatments

Defining the estimand may seem daunting when the causal treatment is not binary. Each unit has many potential outcomes. Recent projects define estimands when the treatment is categorical (with Daniel Molitor and Jennie Brand, 2025 SMR) or continuous (with Jennie Brand, 2026 SM).

Quantile-based summaries

Many estimands involve summaries that are means. But one need not stop there. For instance, quantile-based summaries enable a new visualization of economic mobility (2020, SM).

Sometimes we can discover the estimand from the data

Adaptive randomization

At the beginning of a randomized experiment, we may think of many estimands of interest. Which one should we estimate? Jennah Gosciak, Daniel Molitor, and I focus on survey experiments to explore heterogeneous human preferences. We show how to explore over many possible causal estimands through adaptive randomization in conjoint survey experiments (in press, Political Analysis).

Estimands can be predictive

It is well known that social science models do not predict well. But is this just for lack of trying?

A new way of doing science. We collaborated with hundreds of social scientists and data scientists in a research design optimized for prediction. Teams trained predictive models on a standard social science dataset. We evaluated them on a holdout set locked away until the end.

Predictions were far from the truth (2020, PNAS). Through qualitative interviews, we then developed a framework to undersatnd the origins of unpredictability in life outcome prediction tasks (2024, PNAS).

Estimands can be descriptive

Estimands are often causal. But descriptive estimands involving no counterfactuals can also be powerful.

For example, what proportion of children born into poverty in large U.S. cities in 1998-2000 were ever evicted between birth and age 15? Louis Donnelly and I estimated that more than 1 in 4 are evicted (2019, Demography).

Page updated

Google Sites

Report abuse

Ian Lundberg

I develop quantitative methods to answer questions in stratification and inequality. A theme of my work is that a precise statement of the research question can improve methodological choices. Below is a quick summary of some of my work.