Not all reductions in tumor size are equal (SITC 2020 poster)

We recently presented our poster at SITC 2020.

Note: the full text of the abstract is here. download the poster.

Main findings: Not all reductions in tumor size are equal in Non-Small Cell Lung Cancer (NSCLC).

  • Subjects with tumor shrinkage on a PD-L1 inhibitor had improved survival vs subjects with similar tumor shrinkage on VEGF inhibitors, MEK inhibitors, or chemotherapy drugs.

  • Among combination therapies, the aPD-L1 + aCTLA-4 combo showed a similar improvement in survival among responders as the aPD-L1 inhibitor alone.

Although we characterized this finding using a Joint Model, it is also evident when plotting Kaplan-Meier survival estimates by category of tumor shrinkage.

The improved survival benefit for the same tumor shrinkage was observed among subjects with high (>25%) PD-L1 expression and low (<= 25%) PD-L1 expression on tumor cells.

Read More

Talks this week: R/Pharma 2020 and PyData Eindhoven 2020

I’ve been invited to give two presentations this week.

  1. Wednesday, October 7, 2020. 1-4 pm Eastern. R/Pharma 2020

    As part of R/Pharma, I am teaching a 3-hour Stan workshop. It will be focused more on practical aspects of working with Stan. Since everything is virtual, it’s a new format.

  2. Thursday, October 8, 2020. 11 - 11:30 am Eastern. PyData Eindhoven

    For PyData Eindhoven 2020, I am giving a 25 minute talk on Stan titled: “Stan: why does it exist? when is it useful? why do I use it?” Since it’s a general data-focused conference, I’ll dig into when it’s time to use Stan and why we use it at Generable.

If you are at either of these events, please reach out. I’m always happy to connect.

Read More

Prototype: Simplified CLI to Stan

I’ve created a prototype that’s meant to be easier to use than CmdStan. If you’re a CmdStan user, I’d appreciate some feedback on what works and what doesn’t. Please comment here, on discourse, on twitter, or however you can find me.

The license for the prototype is 3-clause BSD (just like CmdStan).

Quick example: setting delta and max_depth

Here’s how to set delta to 0.9 and max_depth to 14.


> examples/bernoulli/bernoulli sample --delta 0.9 --max_depth 14 \
  --data_file examples/bernoulli/
Read More

Fitting a Basic SIR Model in Stan

Today, it seems like everyone is an epidemiologist. I am definitely not an epidemiologist but I did want to learn the basics of the popular SIR (Susceptible, Infected, Recovered) models. My guide is “Contemporary statistical inference for infectious disease models using Stan” by Chatzilena et al. [1], which does a great job at describing the model and in the spirit of reproducibility provides both Stan and R code. In this blog post, I will test a slightly modified version of the model on a simulated dataset and implement out-of-sample prediction logic in Stan.

Read More

Introducing the Generable Platform

Over the past 6 months, we have been hard at work building our meta-analytic software platform for making go/no-go decisions from early clinical trials in solid tumor oncology. Our main goal is to estimate the Probability of Technical Success (PTS) in late-stage trials by using joint event and semi-mechanistic models of tumor growth, hierarchical Bayesian models fit with Stan to a large portfolio of completed trials, and proper statistical adjustments (poststratification) to account for the differences between the sample and the target population. We are excited to share what we have built and if you are coming to ACoP in Orlando next week, please grab a demo time slot or reach out by email at

Read More

Objective is not so objective

Model selection is a difficult process particularly in high dimensional settings, dependent observations, and sparse data regime. In this post, I will discuss a common misconception about selecting models based on values of the objective function generated from optimization algorithms in sparse data settings. TL;DR Don’t do it.

Read More

Optimizing, Sampling, and Choosing Priors

Do you really believe your variance parameter can be anywhere from zero to infinity?

In the past, I’ve often not included priors in my models. I often felt daunted by having to pick sensible priors for my parameters, and I usually fell into the common trap of thinking that no priors or uniform priors are somehow the most objective prior because they “let the data do all the talking.” Recent experiences and have completely changed my thinking on this though.

Read More

Meetup/Webinar on June 28: Understanding the Progression of Alzheimer's

We are excited to announce our Meetup next Thursday at 7 PM EDT titled “Understanding the progression of Alzheimer’s.” The Meetup will be hosted by me and we will be streaming it live for those are unable to attend in person (although it’s always more fun to attend in person and Eric is bringing authentic New York pizza).

Read More

2018 PAGE meeting in Montreux

Last week I attended the PAGE conference in Montreux, Switzerland. PAGE conference organizers love beautiful scenery and parties at least as much as they love the scientific program and this year was no different. Montreux is a gem of a Swiss town overlooking Lake Geneva and the Swiss Alps. It is also the final resting place of one my favorite writers, Vladimir Nabokov, as well as his son Dmitri and his wife Vera.

Read More

Deconstructing Stan Manual Part 2: QR Decomposition

On March 15, we held our second meetup of 2018 covering QR decomposition, simulating correlation matrices using the LKJ distribution, and ending with some general advice about priors.

The slides from the meetup are now available on RPubs. If you have any questions or suggestions, please let us know in the comments.

Read More


Last month, I was invited to Melbourne for a pharmacometrics conference: Population Approach Group of Australia & New Zealand (PAGANZ) meeting. It was great getting to know some more pharmacometricians and really digging into the problems they face, specifically in statistical inference as applied to PK/PD models.

It was a busy trip with talks / workshops lined up on all three days:

  • One day Stan course as part of Population Analysis Work Shops (PAWS) taught by myself and Sam Brilleman.
    Slides (pdf)

  • ISoP Lecture at PAGANZ titled “Stan Meets Pharmacology.”
    Slides (pdf)

  • Melbourne Stan Meetup talk titled “Understanding lp__: proportionality constants and (automatic) transforms.”
    Slides (pdf)

I wanted to fill in some of the motivation, especially for the pace of the short course.

Read More

Correlation or no correlation, that is the question

A friend asked me about how he should update his beliefs about correlation after seeing some data. In his words:

If I have two variables and I want to express that my prior is that the correlation could be anything between -1 and +1 how would I update this prior based on the observed correlation?

Read More

Deconstructing Stan Manual Part 1: Linear Regression

On February 15, we held our first meetup of 2018 starting a new series called Deconstructing the Stan Manual. During the meetup we coded a Linear Regression model in Stan and fit it to the Wine Quality dataset from the UCI Machine Learning Repository.

The slides from the meetup are now available on RPubs. If you have any questions or suggestions, please let us know in the comments.

Read More

San Francisco, Jan 6-9. Asilomar, Jan 9-12.

We’re headed west for a week!

If you’re in San Francisco or Monterey and want to meet up, please reach out. Both Eric and Daniel are making the trip. We’ll be in town for the J.P. Morgan Healthcare Conference and StanCon 2018.

Read More

Generable and Stan

Stan is freedom-respecting, open-source software.

Stan is amazing. It’s our tool of choice for building generative models. The project is open-source and we are committed to supporting the open-source community.

Read More

Why Generable

The more important a decision the more “Bayesian” it is apt to be.
—- Irving J. Good

At Generable, formerly Stan Group, we are focused on productizing state-of-the-art generative models for making decisions. We are currently working on a broad class of survival (time to event) and econometric models that encode generative structure that cannot be learned from data alone. These models are special, because they enable us to simulate counterfactual states of the world weighted by their respective probabilities.

Read More