This lecture's about the Types of Data Science Questions. So last week we covered a lot of installation and building up software and this week we're going to cover a little bit more about the conceptual ideas behind data science. So there are, a few different kinds of data science questions that I've listed up here in their approximate order of difficulty of actually achieving the goal of that analysis. So it starts with descriptive, and then it goes to exploratory, inferential, predictive, causal, and mechanistic. And I'm going to talk about each of these types of analysis over the next couple of slides. So the first is the descriptive analysis. So the goal here is just to describe a set of data. You're not trying to make any sort of decisions based on it or anything like that. It's the first kind of data analysis that was ever performed. And it's most commonly applied when you're talking about census data. The description and interpretation of this data are sort of different steps, so you gotta describe the data and then interpret what you've seen. Descriptions can usually not be generalized without additional statistical modeling. So in other words, you are describing what you are seeing in this data but you are not saying what that might be for the next person that might come along. So an example of a descriptive analysis is this US census. So this is a picture of the census twenty ten, a website that has collected a bunch of information about people in the United States and they are not necessarily analyzing it to predict something about, [INAUDIBLE]. Now other people or other countries they're just trying to describe the population. Another descriptive ex, analysis is the Google Ngram Viewer. So what this is is it's a collection of information about pairs or triplets of words. And so for example this is a plot over time of the observation of those, of the words Albert Einstein, Sherlock Holmes, and Frankenstein in books that have been scraped by Google. And so, again, this isn't trying to infer anything when making decisions. You could do that, but this is just purely a description of what's going on. The next is exploratory analysis. So here you're trying to look at some data and find relationships that you didn't know about previously, but not necessarily confirm them. So they're good for discovering new connections, and they're also useful to find, for defining future data science projects, where you're actually trying to confirm the exploration that you've performed. They're usually not the final say on any particular problem, and they usually shouldn't be used for generalizing or predicting. The important point is that you've probably heard before that correlation does not imply causation. So, you don't want to necessarily save it. You've discovered a relationship that is the critical relationship between two variables based on exploratory analysis alone. So here's an example of an exploratory analysis, where we're looking at brain images and trying to identify regions of the brain that lit up in response to a particular stimulus. So they explored this data, and they observed, you know, here's a region that lights up in response to that stimulus. And here's another re, region that lights up in response to the stimulus. And they haven't necessarily confirmed that that, what that means, but they've just discovered a new connection that they hadn't seen before they had this data. Another example that's actually hosted here at Johns Hopkin's is the Sloan Digital Sky Survey. So this is just terabytes or even more of data that has been collected looking at the night sky. So it's actually pictures of the night sky at different times and at different places that you can explore to try to discover new stars or try to discover how different things in the universe work together. And so, that data is actually used for exploration, but not necessarily for confirming anything that you discover. Then inferential analysis is a goal where you're actually trying to take a small amount of data, on a small number of observations, and sort of extrapolate that information, or generalize that information to a larger population. Inference is definitely the most common goal of most statistical models, and most statistics you may have heard about. It involves both estimating some quantity that you're interested in and also, more importantly maybe, the uncertainty of that quantity that you're interested in. And it depends heavily both on the population that you're looking at, the group of people or the group of objects that you care about, and a sampling you've discovered. So this is an example of, inferential analysis. So the idea is, here, you're trying to look at the, effect of air pollution control on life expectancy in the United States. So, what they've done is, they haven't analysed all of the counties in the United States, but they've analysed, so, sort of a subset of the counties in the United States. And they're using that to try to infer something about what's generally happening in the relationship between air pollution and life expectancy. Prediction analysis is a little bit different. The inferential analysis is that usually a little bit more challenging. So the idea is to use the data on some objects you collect the data on, to predict the values for another object for the next observation that comes to the door. An important thing to keep in mind is that even if x predicts y, it does not mean that x causes y. That's one of the main fallacies that you can run into when dealing with predictive analysis. Accurate prediction depends heavily on measuring the right variables. And although there are better and worse prediction models, it's pretty clear that more data and simple model tends to work really well. Prediction is very hard, especially about the future, and so I've, I, this is a, a funny quote that it's very accurate actually and I've listed some references here where you, or link to some references here where you can see other people who have said that same pre, thing over and over. So here's an example of a predictive analysis. So Nate Silver is one of the data scientists I mentioned earlier and FiveThrityEight is his blog where he tries to predict the outcome of U.S. elections. So what he does is he takes data from polling and he tries to predict what's going to happen in the next Presidential vote. And he was very accurate in doing that in both the two thousand and eight and two thousand and twelve elections. Here's another example, so it's another sort of bad example, if you were the teen in question. So, target figured out a teen girl was pregnant, before her father did, by looking at the purchases she made, and it sent her a flyer saying she was pregnant. Of course, the father and girl was a little upset, because he didn't know that she was pregnant. But it was example of how you could use data to predict characteristics of people including metadata. Causal analysis is a next level up. It's substantially harder to identify causal relationships from data. So the idea is, what happens if you change the values of one variable? What will, how will that change the value of another variable? The gold standard for doing this in general is using randomized studies or randomized controlled trials to identify causation. And you can try to do it from just observed data that you have saved in the database but it's a much harder sell. You have to make much stronger assumptions about the way that your model is work, [SOUND] working. Causal relationships are usually identified as average effects, in other words, on average, if we give this population a particular drug, then on average they will only have a little bit longer then if you didn't give them that drug. So these are usually the gold standard for data analysis in a particular for most scientific applications the goal is to end up with a causal relationship between variables. So here's an example of a causal analysis. So basically what they did was, they actually did Fecal transplants. It's quite a disgusting procedure of giving people transplanting fecal matter into different people so that the bacteria populate their colon and they recover better for recurrent infection of a particular type here. And so they were able to randomize people to get the fecal treatments and then they determined that his treatment was constantly associated with better outcomes. Mechanistic analysis is very rarely covered in data science, and so it's important to keep in mind just to keep a full handle on all types of analyses that could be done, but it's very rarely the goal of most analyses. The idea is to understand the exact changes and variables that lead to exact changes in other variables. So this is incredibly hard to infer if the NOIDs, data is noisy at all except in very simple situations or in situations that are very nicely modeled by a deterministic set of equations. The most common applications where this is possible is in the physical or engineering sciences where some more simplified models can describe a lot of the action that is happening. And generally the only random component when you're doing a mechanistic analysis is measurement error as opposed to any of the other types of variation you might see in data. [SOUND] So here's an example of a mechanistic analysis where the idea was to try to actually discover, what was the differences and changes that you would make in par, pavement design, that would directly lead to changes in, the functioning of that pavement. And so those with mechanistic analysis like I said, tend to end up in, engineering applications or in physics applications. That is a quick tour of the types of questions we will cover in data science. [SOUND]
