Speaker 1: Well, welcome back if you've seen any of the videos before, if not, welcome. This is going to be all about propensity scores, everything you need to know in five minutes. I'm also going to test out this hypothesis about whether or not saying like and share my video makes a difference. So if you want to spread the word, feel free to like and share. All right, let's jump in. A very quick overview. A propensity score is simply a probability, that's it. Propensity scores do not mimic randomized trials. I always cringe when I hear people say this because it's totally wrong. Instead, propensity score methods are just a tool to use when you want to adjust for confounders in your non-randomized study. If you don't know what a confounder is, watch one of my other videos. And there's other tools that you can use to adjust for confounders. The mnemonic that I use is Mrs. Robinson. So M stands for matching, R stands for restriction, S stands for stratification, and R stands for regression. And you know, an important limitation of regression is this 10 to 1 rule. For every 10 patients in your study that have experienced the outcome, you can only adjust for one variable. And that might be problematic if you want to adjust for many more variables, but not many people had the outcome. All right, so what is a propensity score? It's just the probability. So the probability of receiving, let's say, some drug of interest conditional on a number of observed characteristics. In a randomized clinical trial, the only thing that determines who gets a drug and who gets placebo is a coin toss. So the probability of getting the drug is 50%. That is easy and simple. In the non-randomized setting, aka in the real world, we don't flip a coin to determine who gets drug A versus drug B. We make those decisions based on their age, sex, socioeconomic status, medical conditions, etc. But what you can do is start to calculate, all right, so what was the probability that the person was going to get that drug? Hopefully a real life example can make this a little more clear. So here's a study that I was a part of that looked at whether or not SGLT2 inhibitors, that's a class of diabetes medications, can reduce a person's risk of gout. We had 290,000 patients in the study. There were 100 variables for each patient, and 100 people experienced gout. So using the 10 to 1 rule, we can only adjust for 10 of those variables. But with propensity score methods, fortunately, you can get around that. So what are the steps? Number one, you got to calculate a propensity score for each person. Number two, you got to decide how you want to use those scores. And number three, you run a regression model to see if indeed SGLT2 inhibitors are associated with a decreased risk of gout. So let's go through step one. So here's our data set. Here's the sort of first 10 people. You can see age, sex, one is male, diabetes, one equals yes, and whether or not they got an SGLT2. And sort of looking through this, you might see some associations. You might realize, hey, you know, people who are 50 and up seem more likely to get an SGLT2. And if you have a history of diabetes, you seem more likely to get an SGLT2. All right, so let's generate a score. You know, can we just assign a couple points based on age and a couple points on if they had diabetes? Well, no, that wouldn't be the most scientific way to calculate a propensity score. Instead, what you do is you run a multivariable logistic regression model. Yes, this is ugly looking, but this is a formula for multivariable logistic regression. So what it gives you is a probability of whether somebody got an SGLT2 conditional on their characteristics, so their age, their sex, and whether or not they had diabetes. So maybe we find after we run our logistic regression model, that here are the weights, you know, it's the probability is 0.6 times their age. So if they're older, they have a higher probability, zero times their sex, which means sex was not associated with a not that got an SGLT2 inhibitor, and 0.9 times whether or not they had diabetes. So what we can do is then run all of these patients through that model, and then calculate their propensity score. And that's exactly what we now have. So now you got the propensity score, and you need to decide, all right, what am I going to do with them? The most common three methods or the most common two methods are matching and weighting. And sometimes you also see stratification. There are some considerations to think about within each approach. I've just listed them here. The details are information overload for this introductory talk. All right, so let's bring this to life. Maybe let's use matching. So everyone in green got an SGLT2, everyone in red did not. And now we're trying to find is a match people with the same propensity score, whether they got the drug or they didn't. Boom, we got one match here. Is there another one? Oh, yeah, 0.5. Okay, so good, we got another match. So now we can see, all right, what did the patient population look like before matching? And what do they look like after matching? You can take your time and pause the slide, but quickly you'll realize that matching allows for very good balance of these observed confounders. And the last step is we then run our regression model to see if indeed SGLT2 inhibitors are associated with a decreased risk of gout. And when you're running your regression model, you know, you can analyze the risk of gout in your unmatched population, but in the unmatched population, you didn't adjust for the confounders. So really what you also want to do is run it within your matched population because within that propensity score matched population, confounders have by design been adjusted for. And here are the different types of regression, I'll have to post a five minute talk on a crash course in regression. So what did we find? So before matching, we saw a decreased risk of gout, hazard ratio of 0.6. And then after matching, so when we ran the model within our matched population, and didn't include anyone who didn't have a match, we again see a reduction in the risk of gout. So here's a summary slide and I'm sorry, I know I'm already over time, a propensity score is simply a probability. Propensity scores do not mimic randomized trials. And here are the steps. Number one, you got to generate a propensity score, and that's done using logistic regression, you need to decide how you want to use the scores. My favorite is matching. And then you run your regression model within your propensity score matched population to see if your exposure, SGLT2s, decreased risk of the outcome, which in this case was gout. Thanks so much. I hope you enjoyed the talk. And that's it for now. Bye.