PhD Candidate at Harvard University.
What I've been working on lately.
Overall survival (OS), the length of time between randomization to a treatment and death, is the outcome most often used to determine the effectiveness of new drugs for cancer. However, overall survival may be difficult to measure because of how long it takes for death to occur, or other confounders such as issues with follow-up, treatment crossover, and successive lines of therapy. The use of surrogate endpoints, another indicator used in place of overall survival to judge efficacy, is appealing to investigators as it permits shorter trial duration, smaller patient cohorts, and may even allow for single-arm trial designs. Two common surrogate endpoints include progression-free survival (PFS), defined as the length of time between randomization and progression of the disease (detailed in RECIST criteria), and objective response rate (ORR), defined as the proportion of subjects who partially or completely respond to their treatment.
However, it is difficult to determine if PFS and ORR are valid surrogates for OS. Furthermore, does this relationship change depending on the type of cancer and type of drug (chemotherapy vs. immunotherapy. vs. a combination of both)? In order to quantify the validity of these surrogates, we need to establish a clear relationship between all three endpoints. One common criteria used to do this is the Prentice criteria, which states that a good surrogate must be tightly correlated with the actual ednpoint (correlation at the patient level) and the treatment effect on the surrogate must be tightly correlated with the treatment effect on the actual end point (correlation at the trial level). We explore several correlation coefficients (Pearson’s, Spearman’s Rho, Kendall’s Tau, Harrell’s C-index) and apply a variety of models to explore the relationship between PFS, ORR, and OS for several indications of cancer.
During the COVID-19 pandemic, nursing homes were at the forefront of the response due to the frail nature of their population. Two key markers for quality of care in nursing homes discussed were costs and healthcare utilization. However, this type of data often follows a semi-continuous distribution consisting of a mass of zero values and a long right tail. This distribution introduces challenges in both model fitting and interpretation. Furthermore, residents of nursing homes are at high risk of mortality which acts as a competing risk. We developed a novel Bayesian hierarchical discrete time model for analyzing clustered and/or longitudinal semi-continuous data truncated by death. The model jointly captures the semi-continuous nature by modeling the probability mass at zero and skewed cost separately and combines it with a discrete time framework for semi-competing risks.
We demonstrate the applicability of this novel model first through simulation and then by examining factors longitudinally associated with cost and health care utilization in and across nursing homes using nationally collected data from Centers for Medicare & Medicaid Services. This project is the center of my dissertation with Sebastien Haneuse. Check out my simulation and model code on GitHub.
The hierarchical model for clustered longitudinal semi-continuous data subject to mortality proposed in my first paper provided a framework to analyze complex nursing home data. Nursing home hospitalization and mortality rates are often compared using adjusted and standardized overall outcome rates. However, these rates do not allow for multiple observations or model the interaction between cost accrual and mortality. To account for this interaction, we created joint metrics that utilize our proposed hierarchical model. Our proposed metrics, the cumulative adjusted and standardized overall outcome rates, allow for flexible time scales and jointly consider the dependence between outcomes. Check out my simulation and model code for our metrics on GitHub.
Isolation and distancing practices are fundamental elements of COVID19 epidemic control. However, the CDC currently recommends only 5 days of quarantine after a positive test. We were interested in whether this recommendation should be re-extended to 10 days.
For this study, we collected longitudinal viral load, viral culture samples, and CT values from Massachusetts General Hospital employees who tested positive for SARS-CoV-2. For our analysis, we adjusted for covariates including variants of the virus (original, delta, omicron) and vaccine status. We determined that the previous CDC recommendation of 10 days was needed. Check out our paper and analysis code.
Prior to the COVID-19 pandemic, the WHO declared vaccine hesitancy, defined as a delay in the acceptance or refusal of vaccination despite availability of vaccination services, as one of the top 10 threats to global health. Today, despite massive recent and ongoing global efforts, vaccine hesitancy remains a major threat, particularly in populations that experience health disparities rooted in structural racism. One such population is Hispanic/Latinx adults.
There is an urgent need for innovative, practical, and sustainable strategies to address vaccine hesitancy among the priority population of Latinx adults. This project proposes to address this gap with a novel intervention that integrates evidence-based motivational interviewing (MI) into behavioral health (BH) services, coupled with electronic prompting, and vaccination access at the point of care. As a collaboration between Boston College, Harvard University, and East Boston Neighborhood Community Health Center, we will conduct a multiple-period cluster-randomized crossover trial within four programs at EBNHC. This trial will test the effectiveness of MI and BH interventions on COVID-19 vaccine uptake among Latinx adults and examine theory-based elements of vaccine hesitancy on the causal pathway between the intervention and vaccine uptake.
Cluster randomized crossover multi-period trials (CRCMPT) combine elements from different trial designs into a single study approach. The first component is cluster randomization, which involves the randomization of groups of participants rather than individual randomization. The second component is treatment crossover. In crossover trials, each participant receives multiple interventions in a specific sequence. Lastly, the third component is having multiple periods of treatment crossover. Multi-period trials involve multiple stages or periods within a single trial, with each stage encompassing different phases of treatment conducted over time. Combining these elements provides a complex trial design that borrows the strengths and statistical benefits of its individual components. However, it also mitigates issues inherent to CRTs by leveraging the advantages of crossover and multi-period trials. Both designs increase the efficiency of the study by requiring fewer participants and clusters to achieve similar statistical power, in addition to minimizing confounding and maximizing time efficiency.
CRCMPTs provide many advantages for logistics and statistical power; however, several challenges and considerations arise when there are very few clusters being randomized that can impact the trial’s validity, precision, and generalizability. To our knowledge, little work has been done on analyzing the effects on statistical power in extreme cases (e.g. 2 clusters). We propose a novel method that incorporates several small-sample adjustments to better analyze small-sample cluster-randomized trials, and apply our model to the MI-VACUNA trial examining vaccine hesitancy among Latinx adults.
HaSET’s mission is to advance the health of mothers, newborns and children by generating high quality evidence, training researchers, and translating results to programs and policies. It is a partnership of the world’s leading medical and research institutions Harvard T.H. Chan School of Public Health, Boston Children’s Hospital, St. Paul’s Hospital Millennium Medical College, and the Ethiopian Public Health Institute.
Although there has been a reduction in stunting (low height/length for age), the prevalence of malnutrition in Ethiopia is still high. Child growth patterns and estimates of stunting are needed to determine vulnerabilities and potential for recovery. We collected longitudinal data to determine the prevalence, incidence and reversal of stunting among children aged 0-24 months. Overall, the evidence from this study highlights a chronically malnourished population with much of the burden driven by growth faltering during the neonatal periods as well as after 6 months of age. To end all forms of malnutrition, growth faltering in populations such as that in young children in Amhara, Ethiopia needs to be addressed. Check out the HaSET project here and our paper.
Social determinants of health (SDH) include factors such as age, race/ethnicity, sex, language, comorbidities, cultural group membership, and stress. In collaboration with Caring Health Center and the Implementation Science Center for Cancer Control Equity, we want to analyze the role that SDH play in health inequity.
The data was collected in electronic health record (EHR) format. However, this introduces the statistical complexity of incomplete / missing data. We want to develop methods that acknowledge the complexity and heterogeneity of the incomplete / missing EHR data, specifically in regard to SDH factors. Check out the analysis code for the project here.
Before a drug can be introduced to market, it needs to undergo extensive testing. The first stage of this testing procedure is the phase I clinical trial, where drugs are tested for toxicity and the correct dosage. Currently, the most common design for phase I clinical trials is the 3+3.
Supervised by Susan Groshen of USC Norris Cancer Center, I wanted to compare the 3+3 to more modern methods at evaluating dose limiting toxicity and maximum tolerated dose. We found that despite the lack of statistical theory backing the 3+3, it performed very well when compared to the continual reassessment method, several Bayesian methods, and the rolling 6. Additionally, its simplistic implementation was more desirable for researchers. Check out my paper and code here.
The SEIR (susceptible, exposed, infectious, recovered) model is commonly used to model infectious disease spread. In a recent paper by Hao et al, they propose expanding this model by appending 3 transition states: presymptomatic infectious, unascertained infectious, and isolation in hospital.
Although the model was effective in predicting how COVID19 would spread through Wuhan, China, we wanted to improve on their implementation. Within their Metropolis-Hastings step, we introduced a relatively new algorithm called DRAM (Delayed Rejection Adaptive Metropolis), allowing for more flexibility and adaptability of the algorithm. This allowed us to increase simulation speed significantly (from 1103 to 593 seconds) while not affecting statistical performance. See our implementation, proposal, and presentation here.
Sample size and power calculations are critical when designing a clinical trial. In the past, clinical trials only performed these calculations at the start of the trial; however, researchers have begun implementing more adaptive procedures recently. For example, trials are often stopped to monitor if the treatment being evaluated is efficacious enough, or if the trial should be stopped for futility. During this stage, a common method of adaptation is performing sample size re-estimation to ensure that the trial has enough power to detect a difference between treatments.
For this project, I was tasked to perform sample size re-estimation for the Long-term Odanacatib Fracture Trial (LOFT). Furthermore, I wanted to test through simulation how the type I error would be affected from changes to the sample size, power, accrual rate, and how many years patients would be followed up for. Check out my analysis and simulation code here.
Clustered randomized clinical trials are trials where patients are recruited in “clusters”. These clusters are typically locations where individuals are likely to be correlated with one another, such as a school, nursing home, or village. These types of trials are becoming increasingly popular; however, methodology is still being developed.
When performing survival analysis on clinical trials, the proportional hazards Cox regression model is often used. The goal of this project was to determine if performing a naive analysis using the standard Cox model would be sufficient. I compared results from the naive model to the robust sandwich and randomization-based variants in 27 different scenarios. We discovered that despite not accounting for the correlation inherent in these clusters, the naive Cox model still performed just as well as the other two models. Check out the code and presentation for my project here.
Palliative care focuses on providing relief from the symptoms and stress of an illness. The goal is to improve quality of life for both the patient and the family. However, some patients and families prefer receiving treatment in the comfort of their own home compared to at the hospital. We were interested in whether or not home-based palliative care was as effective as hospital-based palliative care.
This was a 5-year clinical trial enrolling 1155 participants. My main tasks involved managing confidential patient files sent from Blue Shield, creating surveys and scripts in the REDCap database, and recording/monitoring new inpatient referrals. I also created scripts to automatically create summaries of demographic information, ineligibility criteria, and patient concerns so that we could present them to the DSMB and funding agencies. Check out the study and my code here.
Eribulin is a potential new treatment for bladder cancer. This was a combined phase I/II clinical trial aimed at determining whether or not Eribulin could move on to a phase III trial. At USC’s Norris Cancer Center, I worked with Susan Groshen on the statistical analysis of this trial data.
First, we looked at the frequency of adverse events to generate toxicity tables for the DSMC report. Second, I coded up scripts to generate table I statistics, response tables, and Kaplan-Meier curves for progression-free and overall survival. Lastly, I ran a multivariate Cox regression model for progression-free survival. Check out the paper and my code here.
Old school runescape is a classic MMORPG that released back in 2001. Despite its age, the game still has around 140,000 active concurrent players and is growing. Although OSRS offers its own vanilla client, a majority of players choose to use a third-party client called RuneLite due to its open-source nature, quality-of-life improvements, and large developer presence.
By using RuneLite, it allows for a plethora of plugins that improve quality of life that make the game significantly easier to play. Using IntelliJ and the RuneLite API, I wrote several customizeable plugins that do a variety of things including tracking enemies and overlaying a counter over them, tracking projectiles, ground objects, and helpers for various fights. Check out my various plugins here and learn how to install them using the official RuneLite wiki.
Pokemon is a popular franchise where players set out to capture and battle creatures they encounter. I, along with 4 other friends, created this simulator which attempts to recreate the battling component of the Pokemon video games.
We implemented animated sprites, attributes from the player-run wiki, and the various battle formulas present in a typical Pokemon battle. Players are able to choose teams of 6 from the original 151 generation I Pokemon. Battle either locally or over WiFi with chat and emojis. Check out our implementation here.
Neopets was a popular website that allowed users to raise a cartoon pet, explore an interactive point-and-click world, and play hundreds of flash games to earn neopoints. One addicting game from Neopets was Destruct-o-match, a puzzle game with a goal of destroying as many blocks as possible by clicking on bundles of adjacently colored blocks.
To learn the workflow of developing an Android mobile game, I recreated this flash game using assets from the original. I coded up the algorithm that enforces the game’s rules, set up a point scoring system based on how many blocks were destroyed, and included a timer that limits how much time players had to complete each level. Levels were randomly generated and made to get increasingly harder. Check out the game here.
The deep web is often associated with criminal activity and is difficult to access unless you are explicitly searching. However, the deep web holds a plethora of information that can be used for research.
Our goal with this project was to create a search engine that researchers can easily use to scour the deep web for polar-related information. We also developed many visualization tools using D3 that are integrated within the search engine to simplify the results. Check out the website and the codebase here.
Google is the most used search engine in the world. I wanted to better understand how the search engine crawled through websites, which motivated this project.
The project was a mainly a test of understanding various algorithms and data structures such as DFS, graph theory, crawling policies, and filtering. Check out my implementation here.