This was probably the most hilarious group I worked with. On day one, I got a badge, that had a uterus on it. I was trying to be politically correct, but the coordinator kept pushing it. I was following her through the hospital, and she said, you walk up the stairs first, so you’re not staring at my butt. Next, I was in the office with Dr. Schiff, and they asked me if I wanted to put a lab coat on and shadow… but anyway.
Getting the Call
I got a call from UNC. This was Dr. Michelle Louie, who had a question. She said I have a dataset with slightly small sample size but multiple groups I’m trying to do an analysis for. She explained the problem. She was taking morbidly obese (meaning really, really fat), cutting up their endometrium, the lining of the uterus, to stop bleeding. I’m like, yeah, cool, cutting up fat girls. But they were trying to see if the fat women were gonna bleed during the surgery, so they were using a scalpel, and then an electromechanic morcellator. So they’re cutting up these fat girls, I’m like “word.” But the problem was sample size. She asked if I could do inference with small sample size. Since I’m Bayesian, I was thinking, there might be cells in which we can pool information together.
Consider a situation in which we’d like to share information between similar subgroups. For example, political views associated with what kind of car someone would own. Dietary preferences by demographic information and country. We’d expect it to change, correct? This is called conditional probability:
Could be:
Where the first 
This is just a general form of writing out Bayesian model. 
Louie was concerned with electromechanic morcellation, and, again, cutting open morbidly obese women’s vaginas. I was like, sweet.
A Little Modeling
This wasn’t a problem that couldn’t be solved by using hypothesis testing. So I suggested that we use a multilevel model. We could, estimate, given blood loss as an outcome, conditional on whether the surgeon was using a scalpel (hand and knife) or an electromechanic morcellator, whether there was more blood loss during a hysterectomy. Using BRMS, a linear modeling interface to Stan, a package for fitting Bayesian models, priors aside, we can write this out like this, using RStan. This is psuedocode, and I recommend checking out the documentation for BRMS, which can be found in the Stan documentation. Here’s the initial pub: chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://cran.r-project.org/web/packages/brms/vignettes/brms_overview.pdf. This is pseudocode, and I would consult the official documentation. We also need to include priors in a Bayesian model, so it’s recommended to standardize (meaning subtract by the mean and divide by the standard deviation) and un-standardize after the model is fit. Normal(0, 1) is usually fine, for quantitative covariates, unless there’s strong prior knowledge you’d like to include in the model.
mod1 = brm(formula = bloodloss ~ (1 | knife)).
Knife, being a binary categorical covariate as to whether the surgeon was using a knife or not. Bloodloss a continuous measurement, in mL, I believe, of blood collected during surgery.
We might want to adjust for conditional covariates. For example, there could be more blood loss if a woman is even fatter. We can then adjust for fatness, this being (ignoring priors):
mod2 = brm(formula = bloodloss ~ weight+ (1 | knife)).
And you can adjust for different categorical variables. As it turns out, Louie was correct, and using and electromechanic morcellator was associated with more blood loss. It was not “statistically significant,” but there was a noticeable association. What happened, was that using EPIC the data was so hard to collect, that I gave up and found another job. I think they collected more data, and then drew from interesting other inferences with the model.
Among other things I worked on there, were survey analysis. It was a “before-after” survey, but the same subjects didn’t answer so I couldn’t fit a simple linear model, but to draw inferences, I boot strapped and computed and expectation via Monte Carlo simulation. I should have saved it. I could provide an example but it’s like re-doing work you’re not getting paid for.
Here’s a funny one, for the statisticians. Most data I’ve worked with is observational. This was my first clinical trial data analysis. It was a topical ointment to reduce pain after a vaginal surgery, I think. There were multiple collection sites. Since I had been reading too much Gelman at the time, I tried to adjust for heterogeneity between collection sites using a multilevel model, but the results I was getting were erroneous. I did more exploratory data analysis, and the study was designed well, and the samples I was working with were approximately normally distributed, I just needed a t-test. You live and you learn.
Finding a Good Doctor
I heard on LinkedIn, a woman had been dealing with endometriosis for 20 years, and it hadn’t been treated properly. Consider seeing a doctor, which might cost $100 a pop. 10 visits in 4 months, you’re at $4000. Sometimes it’s good to get a second opinion. Dr. Louie is now at Mayoclinic. Sometimes it might be worth it to get checked out by some of the best. Think about it: within USA, a plane flight might be $300 and the doctors appointment, might be a few hundred out of pocket with no insurance. You’re saving money and getting better healthcare. My girlfriend in high school had endometriosis and she was constantly in pain. We call this medical tourism. Sometimes it’s worth it to see a good doc. More cost effective, and time effective.
Other vetted Gyno’s I had the pleasure of hanging out with, were Dr. Lauren Schiff and Dr. Erin Carey. Thank you.
The job didn’t work out, because they just needed someone to collect data, and I’m a modeler/developer, not a data collector! and I was getting paid like restaurant wages, so I left to China to get my masters. Thanks for the opportunity, anyway. That was one of my favorite summers. I love North Carolina.
Options
To pay for the flight to China, I opened a put-credit spread on apple. The trade went the wrong way, black Friday style. My account was credited (meaning negative) over $110,000, after the bottom leg of the trade was exercised (I didn’t anticipate that, I was new to the options game). But the trade went so bad, the value of the upper leg went about $10,000 above what my account was credited. So I was walking from Carborough, North Carolina, to Chapel hill on the phone with the guy interfacing with the broker. I was ready to jump out of a fucking window. I had to estimate the latent value of the limit to exit the trade so I didn’t blow my brains out. I was screaming at him, “… value X, value X-10, …” something like that. I ended up making out with $10,000. That’s what’s cool about options, even if you mess up, you can leg out of a trade profitably.
But anyway, sex, pussy, morbidly obese women and options.
Happy women’s history month.
likely llc 
Leave a comment