- Candidates running for office often send mailers  to potential voters, encouraging them to go out and vote. These letters or postcards usually  include information about both the candidates and how to vote. Let's say we wanna know  whether these mailers increase the chances that individuals vote. What kind of  research design might we wanna use? Well, in this video we'll introduce you to  a very important type of research design: the experiment. After watching this  video, you should be able to identify a randomized experiment and articulate how  it is different from an observational study; to find treatment and identify the treatment  and control groups in a given experiment; and explain how randomized experiments  mitigate our concerns about confounds. First, let's quickly remember why simple polling  wouldn't be sufficient to answer our question. And by simple polling, I just mean you go out and  you conduct a survey of a random sample of people to see, first, whether they received a mailer for  the last election, and second, whether they voted. The reason is that our analysis might  be influenced by confounding variables. In this video, I'll provide a brief  recap of confounding variables but if this concept seems new or confusing  I encourage you to revisit the module in confounding variables and  then come back to this video. Okay, so what are confounding variables?  Briefly, confounding variables are variables that influence our dependent variable and  are correlated with our independent variable, and they may cause us to misinterpret our data if  we don't take them into account. For example, it could be the case that as X increases Y increases.  So we're tempted to conclude that X causes Y. But in reality, a third variable caused  both the increase in X and in Y. And if we don't realize this, we could come  to faulty conclusions about our data. In this video, we'll talk about one way to  deal with confounding variables: experiments. Let's go back to our turnout example. We wanna  know whether sending potential voters postcards with information about the election will  increase the chances they vote. An initial poll suggests that they might. People who  receive postcards were more likely to vote. But can you think of any potential confounds? Well, it could be that political parties are  targeting specific types of voters. For example, consider a candidate that doesn't have much  money to spend on his re-election campaign. This candidate might want to send postcards to voters  that live in districts he won in the last election because he thinks that they're more likely to vote  for him in this election. But these voters were already more likely to vote. We know this because  they voted in the last election. So our poll might just be showing us that people who voted  last time were more likely to vote this time. Consider another example. Maybe a  politician wants to send postcards to the people who didn't vote last  time because the politician thinks that people who voted for him last time  will turn out to vote without a reminder. This might actually decrease the apparent  relationship between postcards and voting. In other words, the effect of postcards on voting  might actually be bigger than we see in our poll. So what do we do about confounds? Well, in this  video series, we'll talk about two main types of solutions. First, we can fix the problem before we  collect data using research design. Or we can do it after we collect data using statistics. In this  video, we'll just talk about the first solution. What would be the ideal evidence to show us that  receiving an informational postcard really does increase the chances that an individual turns out  to vote? Well, you might think that we would wanna select a group of people, send them informational  postcards with information about where to vote, the day of the election, et cetera, and  then observe whether they decide to vote. But to be really sure that the postcards have an  effect, we would also wanna know what those same people would have done without any interference  from us, meaning we hadn't sent them the postcard. Ideally, we'd be able to use a time machine to  go back in time before we sent the postcards and record whether they vote anyways without  the postcards. Then we could compare how many people voted in each of these alternate  realities. Did more people vote in the reality in which we sent the postcards compared to the  reality in which we didn't? If they did then we could conclude that sending the postcards  really did increase turnout in the election. So why is this the ideal solution? This  is worth taking a moment to think about. This is the ideal research design  because we know that everything in our two alternate realities is exactly the  same: the potential voters, the economy, the important issues in the election, and  the candidates. Everything's the same, except for one very important factor: whether  the potential voters receive postcards or not. This allows us to conclude that any differences  in turnout must be a result of the postcards. But obviously, we don't all have time machines  laying around or the ability to observe alternate realities. So what do we do in real life? What  we do is compare two very similar groups, think of them as our alternate realities or our pre-time  machine voters and our post-time machine voters. It's important that these two groups are  comparable in all ways, except for whether they receive the postcards. By "comparable," I  just mean that we want the two groups to be as similar as possible to make sure they approximate  the two alternate realities as much as we can. Note that the two groups don't have to  be exactly similar in every single way. We just need them to be what  we call similar on average. And we can get these two comparable groups simply  by randomly assigning each potential voter to one group or the other. Once we have the two groups,  we send one group postcards but we don't send anything to the second group. We call the first  group the treatment group because they receive the treatment, in this case the postcards, and we call  the second group - the group that didn't receive anything in the mail - the control group. Finally,  we observe how many people in each group votes. If more people voted in the treatment group, then we  can conclude that the postcards increased turnout. Let's take a step back and make sure  that we understand how that worked. We wanna understand the effects of informational  mailers on voting, but we were worried that maybe politicians only send these mailers to  certain types of voters. In other words, we were worried that receiving a mailer would be  correlated with some other characteristic of the voter, like the voter's baseline propensity  to vote. If the voter's baseline propensity to vote influences whether they voted in  this election - our dependent variable - and is correlated with receiving a mailer, our  independent variable is a confounding variable. When we randomly assign some people  to get postcards and others not to, we eliminate the possibility of a correlation  between receiving a postcard and the baseline propensity to vote. In other words, we mitigate  the problem of confounding variables. Note that for this to work, the process of splitting the  groups, so deciding who gets postcards in this case, has to be random. That's how we know that  the groups are comparable or the same on average. While we won't get into the math behind why  this is the case, the cool thing about randomly assigning units to treatment and control  groups is that it creates two groups that are very similar across both observable and  unobservable characteristics. In contrast, when candidates are choosing to send postcards  only to voters they think will vote for them, they're probably not sending  the postcards randomly. These groups could be different  in a bunch of different ways. They could belong to different  political parties for example, or one group may be more interested in politics.  So we really like randomization. In fact, many researchers consider randomized experiments  to be the gold standard of research designs. However, it's important to note that randomized  experiments are not the perfect solution for every research question. For example, consider  studying the effects of war on polarization. Can we randomly assign some countries  to go to war and some not to go to war and then measure how polarized the country is? No, there are obvious practical and ethical  reasons why we should not attempt to do this. So to conclude, let's summarize what  we've discussed. First, understanding how to deal with the confounding variables is  a really important task for all researchers. Without understanding how to do this, we can't be  confident in the relationships we see in our data. In this video, we wondered whether sending  potential voters informational postcards increases voter turnout and briefly discussed what  ideal evidence to test what this hypothesis would look like: the time machine study. Since this  type of study is impossible in the near future, we then considered it a hypothetical  experiment to test our hypothesis. We identified the treatment, the  treatment group, and the control group, and we explained that using randomization to  create two similar groups where one receives the treatment and the other doesn't, allows us  to simulate the time machine research design. We also noted that experiments aren't always  practical or ethical for researchers to use, and briefly mentioned that there are other  ways to decrease our concerns about confounds. While experiments reduce concerns  about confounds before we collect data, we can also use statistical tools to  control for confounds after we collect data. And that's our introduction to experiments.  Before moving on to other videos, I encourage you to check your understanding  using the quiz questions in this module.