Research & Collaborations

We develop statistical methods for studying how treatments, policies, and exposures affect outcomes, especially when the data do not come from a simple randomized experiment. Our work combines questions about what can be learned from data with methods for estimating effects and describing uncertainty. Collaborations in medicine and public health help shape these questions.

Methodological research

Search all publications

01

Causal inference in networked populations

A treatment can affect people beyond those who receive it. Vaccinating one person, for example, may also protect their contacts. These spillovers make the pattern of connections between people part of the research problem.

Our work on causal inference under network interference asks how to estimate these effects when we only partly know who can affect whom. We develop methods that account for missing or mismeasured connections, use imperfect network measurements, and assess how conclusions change when the assumed network is wrong.

Heatmaps comparing the true network, an observed proxy network, and posterior edge probabilities.
Comparing the true network with an imperfect measurement and the connections estimated by the model. Figure 3 from Weinstein and Nevo (2025), CC BY 4.0.
Main collaborators
Selected papers
02

Causal survival analysis & truncation by death

An intervention may affect both how long people live and whether they develop another condition. That makes questions such as whether it delays disease difficult: people who die first cannot go on to receive the diagnosis, and the people who survive may differ between treatment groups.

We study how to define and estimate treatment effects in these settings, including when an outcome is only meaningful for people who remain alive, a problem known as truncation by death. Our work considers both the timing of events and comparisons among people who would survive under either treatment. We also examine how conclusions depend on assumptions that the data cannot fully check.

Seven horizontal time axes showing possible orderings of potential non-terminal and terminal event times under two exposure levels.
Seven possible patterns of disease and death timing for hypothetical individuals under two exposure levels. Figure 1 from Nevo and Gorfine (2022).
Main collaborators
Selected papers
03

Survival analysis

Many studies ask when something happens: a disease diagnosis, a complication, or death. Follow-up often ends before the event occurs, measurements may be taken only occasionally, and one event can change the chance of another.

We build statistical models that make use of these incomplete and interconnected histories. Our work includes modeling disease onset and death together, accounting for risk factors measured at only a few visits, and analyzing different disease types when some diagnoses lack information about the type. These methods describe how risks evolve over time and how events are related.

Two hypothetical patients' disease and death histories represented over successive time intervals, comparing nearest-neighbor, conservative, and anti-conservative recording strategies.
Two hypothetical patients’ histories, showing how disease onset and death are recorded together over successive time intervals. Author version of Figure 1 from Nevo et al. (2022).
Main collaborators
Selected papers
04

Econometrics

Policies and treatments are rarely assigned at random outside experiments. Researchers often use variables such as eligibility thresholds or factors that influence treatment choice to learn about effects. The credibility of those comparisons depends on assumptions about how the data were generated.

We study instrumental-variable designs, which use a factor that influences treatment choice to estimate treatment effects. Our work shows how additional measurements, called negative controls, can test the assumptions behind these designs, and when such tests can mislead. We also develop corrections for regression-discontinuity designs, which compare observations just above and below a threshold, using data from multiple time periods to address other differences at that threshold.

Four directed acyclic graphs illustrating negative-control outcome and instrument tests for instrumental-variable designs.
Diagrams showing how additional measurements, called negative controls, can help assess an instrumental-variable study. Figure 1 from Danieli et al. (2026). Copyright American Economic Association; reproduced with permission.
Main collaborators
Selected papers
05

Causal inference & machine learning

Our work with Yonatan Belinkov and colleagues uses causal inference to understand what drives language-model behavior. Finding information inside a model does not necessarily mean the model uses it to make a prediction. We use causal mediation analysis to trace how information passes through particular components, such as individual neurons and attention heads. In our work on gender bias, this lets us examine which components contribute to biased predictions and how strongly.

Other work in this area addresses reliable prediction in new settings, using small sets of plausible answers when a single answer may be unreliable. We also develop methods for finding collections of models that predict similarly well even when they use different variables.

Two plots comparing minimum recall and mean prediction-set size for five domain-generalization methods.
Balancing small sets of predicted answers with reliable coverage of the correct answer in new settings. Figure 1 from Tsibulsky, Nevo, and Shalit (2025), CC BY 4.0.
Main collaborators
Selected papers
06

Learn-As-You-Go (LAGO) trials

Public-health interventions often combine several components. During a trial, researchers may discover that the package needs adjusting. Learn-As-You-Go (LAGO) trials allow its composition or intensity to change at planned stages, using the results collected so far. The aim is to find a package that reaches a chosen outcome goal at the lowest cost.

We develop the statistical methods that make it possible to learn from these trials. Because later interventions depend on earlier results, ignoring this adaptation can invalidate the analysis. Our work establishes conditions for estimating intervention effects and describing uncertainty using data from all stages, including uncertainty about the best package. The methods cover both yes-or-no outcomes and outcomes measured on a continuous scale.

  1. Start with a packageChoose the components and collect outcomes.
  2. Learn and adaptUse accumulated results to revise the package at each planned stage.
  3. Analyze all stagesEstimate effects and uncertainty while accounting for adaptation.
Outline of the LAGO approach, based on Nevo, Lok, and Spiegelman (2021) and Bing et al. (2025). Schematic, not a figure from either paper.
Main collaborators
Selected papers

Collaborative research

01

Infectious diseases & antimicrobial resistance

Choosing an antibiotic can have consequences beyond the infection being treated. Our collaborations ask how previous antibiotic use relates to later antibiotic resistance, how long that relationship lasts, and how different treatment choices compare.

A central difficulty is that patients receiving different antibiotics may already differ in their health and treatment history. We combine clinical data with methods designed to address these differences. These clinical questions also motivate new statistical methods when standard comparisons are not possible.

Estimated differences in antimicrobial resistance risk over time after exposure to penicillins compared with other antibiotics.
Estimated differences in the risk of antibiotic resistance over time after different antibiotic treatments. Figure 3 from Saciuk et al. (2025), CC BY-NC-ND 4.0.
Main collaborators
Selected papers
02

Environmental epidemiology

Our collaborations examine how outdoor temperature and air pollution relate to pregnancy, birth, and children’s health and development. We are interested in both the amount of exposure and its timing, such as which weeks of pregnancy may matter for preterm birth or fetal growth.

These questions require care because temperature, other exposures, and health risks all vary with the seasons. Alongside analyses of population data, we use simulations to study how choices about timing and seasonal adjustment can change the answer, including when an adjustment intended to reduce bias can instead introduce it.

Estimated associations between prenatal temperature exposure at selected gestational weeks and preterm birth.
How the association between outdoor temperature and preterm birth varies with the week of pregnancy. Figure 2 from Girgin et al. (2025), CC BY-NC 4.0.
Main collaborators
Selected papers
03

Inflammatory bowel disease

Crohn’s disease and ulcerative colitis can follow very different courses from one patient to another. Our collaborations study which patients are more likely to develop complications and how the timing of treatment relates to outcomes in children and adults.

We use the nationwide Israeli epi-IIRN cohort to follow patients’ treatments and disease complications over time. A key challenge is that worsening disease can both prompt treatment and predict later complications. Our analyses account for these changing patient histories when comparing treatment strategies, alongside work on describing and predicting disease course.

Line chart showing higher cumulative probability of a complicated ulcerative-colitis course in pediatric-onset than adult-onset disease at one, three, and five years.
Probability of a complicated course of ulcerative colitis over time, comparing disease diagnosed in childhood and adulthood. Website visualization of the 1-, 3-, and 5-year probabilities reported by Atia et al. (2025).
Main collaborators
Selected papers
04

Cancer epidemiology & heterogeneous disease

Cancers with the same clinical name can differ in their molecular characteristics and in the factors associated with their development. Our collaborations combine long-term follow-up with information about tumors to study colorectal and pancreatic cancer.

We ask, for example, whether the relationship between smoking and colorectal cancer differs across tumor types. The disease types also compete to be the first diagnosis, which complicates comparisons of their causes. Our methods work defines effects for individual disease subtypes and examines the assumptions needed to give those comparisons a causal interpretation.

Forest plot of colorectal cancer incidence by tumor mutational burden and neoantigen load according to cumulative smoking exposure.
Comparing the relationship between smoking and colorectal cancer across tumors with different molecular features. Figure 2 from Hamada et al. (2025), CC BY 4.0.
Main collaborators
Selected papers
05

COVID-19 & vaccine studies

Our COVID-19 collaborations study vaccination in everyday healthcare settings, including during pregnancy and among people with inflammatory bowel disease. We also examine protection within households: whether vaccinating parents reduces the chance that their unvaccinated children become infected.

These studies use large healthcare datasets to compare infection outcomes while accounting for differences between vaccinated and unvaccinated groups. Related work considers vaccination strategies across populations and how biological protection and changes in behavior contribute to observed vaccine effects.

Illustration of direct protection of vaccinated parents, reduced household infectiousness, and indirect protection of children.
How vaccination can protect parents and reduce the chance of passing infection to their children. Figure 2 from Hayek et al. (2022), CC BY 4.0.
Main collaborators
Selected papers
Complete record

See the full publication list, with filters by topic and type.

Publications Full CV

Back to top