Okay, so I’m not a pro in this field yet but it looks like I have to become one real quick like as I’m going to be writing two papers about them. But the more I read about, the more I, you guessed it, can relate it to my book(s). So what are propensity scores and what do they do? Well, they match subjects in two groups for certain factors so that only the difference is one that we can attribute to the factor we are most interested in. So like if we want to see if two treatments are different in treating a disease, we use propensity score matching to make sure that the two treatment groups otherwise match on age, race, and so on. Sure, there may be some people that can’t be matched between groups, so we don’t use them. Of course, we could impute data so that unmatched data can have something to match with or someone nice can play matchmaker between me and Joe Mange … anyway, I’m digressing again. But yeah, that’s propensity score matching in a nutshell. And another example from my trilogy (yeah, you know I’d be getting to this) would be if, say, our hero, FBI agent, Randy Lipinski, wanted to know if he was in a good or bad dimension or he was getting close to capturing bad Anton. He could figure the probability of catching bad Anton by look at Anton’s marital history, how many children he has, or where he lives. Although that could be tricky too because sometimes bad Anton is in a good dimension. Or good Anton is in neither a good or bad dimension but in a null space or … well, anyway, you’ll see if you read Book 2, Revised Orders. But anywhoo, in one of the good dimensions, good Anton lives with his wife and two kids in Indianapolis and what a coincidence that I came across a travel guide (in our break room when I needed a break from propensity score reading of all things) with this on the front cover.
Now, don’t they look relaxed and happy and full of bliss? Which is how I hope to look like once I get this propensity score thing downpat.
So what are mixed effects models and how do they relate to GEE models? Well, they are called mixed effects models because they contain both fixed effects and random effects. And we pretty much know what’s going on with fixed effects but not so much with random effects, as discussed before. Like if you are giving a treatment to patients over time, you most likely know what the treatment is (I would hope!) so that’s your fixed effect, while you random effects are something you may not be able to know or predict exactly, as we talked about before. Like we may not know how each patient reacts to treatment so patient would be a random effect. Or like the time I went to see my husband Joe Manganiello at the Old Orchard Barnes & Nobel and attended his Q&A session and was planning to get his book and go to his book signing and maybe slip him my book in the process if I had not a prior commitment. Or that and I was chicken sh … so in this instance, since I knew the time and place where he would appear, time and place would be a fixed effect. On the other hand, just the other day, I was rushing between my two offices and ran into Bill and Guiliana Rancic with their little son. Now, in that case, time and place would still be fixed effect since I still knew where I was (Streeterville) and when I was there (early Thursday afternoon) so those would still be fixed effects but I didn’t know that I would run into them, so that would be a random effect. Actually, I liked the latter encounter better (apologies to my future husband, Joe), maybe because the element of randomness made it feel more natural. Like I just talked to them and interacted with their little boy as I would with any other couple with a small child. Was nice. But anywhoo, I don’t have a picture or anything of my meeting with them as I was rushing to my office and I did not have my book(s) on hand to give them (drats!) but it was still pretty cool to see them. I do have a picture of Joe Manganiello and a picture of his book under my Christmas tree though, for what it’s worth.
Just no autograph in the book from him inside since I was chicken sh …oot, now why I did leave early? Ah well, next time like at the Order of The Dimensions movie premiere … or like on our wedding night, of course. Anyway, until next time … still working on a topic but hope to make it good … and Order of The Dimensions-related, of course!
So why am I so excited about GEE models, also known as generalized estimating equations? Well, because we can do a lot of cool stuff with them! Like in my line of work, we can use the GEE model to study the effects of a treatment on a disease over time. Sure, we also use another model known as the linear mixed effects model if the outcome of our variable is continuous, having infinite values over a particular range, but GEE offers us more flexibility by allowing us to look at other responses, like binary responses. Like if we just want to know if the treatment is helpful or not. That’s an example of a binary response. And responses are observed over several times, at which time point does the treatment become helpful. OR, for example, …. and you know I’m getting to this … in my trilogy, whether Anton Zelov becomes a good guy or a bad guy in a certain dimension. And what could have happened that made him into a bad guy. So in that model, Good Anton vs. Bad Anton is the response and we would like to know if and when a particular event happens that changes his fate. So what is this event? Well, you’ll just have to check out the trilogy to find out, won’t you? Now, GEE also allows us to account for both fixed effects and random effects, where fixed effects are like something we know that might effect the outcome, like the event from Anton’s youth that affects his future, and random effects are like something that could also influence our outcome but we do not know exactly why. Like when … never mind … again, you’ll just have to check out the trilogy. Anyway, I’ve decided to delve into fixed effects and random effects next time. For now, I leave you this image that I got when I was trying to google “Angel vs. Devil Bluto” as I can picture Anton looking a lot like Bluto from Popeye the Sailor Man. But of course, this isn’t Bluto from Popeye the Sailor Man and it isn’t even Bluto from Animal House, but it has the whole Angel vs. Devil theme I wanted to convey as binary outcomes for Anton’s fate that we could apply to a GEE model so …
Not to be confused with simulating itineraries … and take your head out of the gutter! I don’t mean that! You know this is not that type blog! I mean … what? What you do mean what was I thinking that you thinking? I thought that … um, never mind then. But anyway, in terms of iterations, let’s first refresh our memory about what it means for an algorithm to converge. But unfortunately an algorithm doesn’t usually converge over the first try so we often run the algorithm over and over … and over to see if it will converge at any point. Usually, we set a finite number of iterations though, like 100 or 200 before we give up and determine that the algorithm does not converge. Like with FBI agent Randy Lipinski in my trilogy (by the way, I’ve given up on disguising my pimpin’ since you always know where I’m going with this). But anyway, Agent Lipinski goes from one dimension to another dimension to another dimension … to another dimension in hopes of getting the bad guy. When he misses him in one dimension, he goes to another dimension. And when he misses the bad guy there, he goes to another dimension. And when he misses the villain there, he goes to — can you guess? — can you tell where I’m going with this? — he goes to, yup, yet another dimension. And he does this until he finally catches up with the bad guy. So does he finally catch up with the bad guy? Well, decided not to give away all my secrets just yet. But I did have one of these fine fellows in mind to play him if this trilogy thing ever comes to fruition.
Speaking of the first one, John Krasinski, I keep tweeting him again and again … and again … about my book/movie pitch but so far he hasn’t responded. Now, his response would indicate that my algorithm of tweeting him endlessly would converge. But I’m getting a gnawing suspicion that he has blocked me now which would indicate that that particular algorithm might never converge. Oh , well. I think Jared, Ryan, and James also have twitter accounts though so off to check on that. Happy tweeting!