After you prepare for your analysis by clearly formulating your goal and KPIs, it's time to take inventory of your data and data sources. What data is available to you and what data will you need to collect? A business can receive data from many different sources. Here are some examples. You can receive data from tests and experiments that you set up, like Facebook brand lift or conversion lift tests, for instance, or A/B tests. You can also receive data related to marketing performance through Facebook's as manager. Or you may be working with a partner who provides you with marketing mix models and their results. You may also have access to data on actions people take on a website, in a mobile app, or in stores. That data may have been collected through a pixel on a website or through an SDK for mobile apps, or through direct monitoring of offline events like purchases. You may also have access to data resulting from a survey. As you've seen throughout this program, there are many different data sources. We've made a distinction between first-party and third-party data. First-party data refers to data resulting from a direct relationship between the company that collects the data and its customers. So it's data you collect about your customers directly from your customers. Third-party data is data collected by a third-party that has no direct relationship to users it's collecting data about. First-party data are always preferred as they are collected under a privacy agreement that you made with your customers. With third-party data, it's much less clear to the user that data is collected about them. And online, we see that third-party data collection is increasingly blocked. Once you decide which data sources you're going to work with, it's important to reflect on the data and the limitations that some measurement methods may have. Data quality and granularity can vary quite a bit, and some collection methods are better than others. Let's look at the limitations you should keep in mind for some of the measurement methodologies we've discussed in this program and that you will often use as a marketing analyst. Let's start with A /B tests. Remember that A/B tests are a form of experiments that allows you to compare multiple assets by randomly splitting up audiences in non-overlapping groups. A/B tests are a great way to optimize and fine tune campaigns, but you need to keep in mind that A/B tests do not assess incremental impact on business outcomes. For instance, you may find which ad is the better ad through an A/B test, but you cannot use the results of the test to evaluate how much effect your ads had on sales. To do that, you need a control group, a group not exposed to your ad, which A/B tests don't include. Also, for A/B tests to be reliable, you need a confidence interval of at least 75%, and ideally, your confidence level should be 90%. In other words, you want to have a test result that has a 90% chance of being replicated if you were to run the exact same test again. Next, let's look at the limitations of randomized controlled trials, or RCTs. We looked at two examples, the brand lift test and the conversion lift test. RCTs test a hypothesis by introducing a treatment and determining the impact of that treatment on business outcomes. In brand lift tests, we study the effect of advertising on brand metrics, like brand awareness. And in conversion lift studies, we study the effect of advertising on conversions. The RCTs use an experimental or treatment group that is exposed to your ad and a control group that doesn't see your ad. RCTs can show you a causal relationship between your advertising and desired business outcome. Whether that's an increase in brand awareness or an increase in conversion. RCTs require larger budgets and a longer campaign interval to achieve statistical power, or in other words, to be reliable. It can also be a bit harder to control for all variables and isolate the treatment variable to make sure that people who are in your treatment group see your add, and people in the control group don't. It means you have to control all your marketing channels where possible exposure may happen, and that isn't always easy to do. So running a clean RCT can be a bit of a challenge. Also, it may not always be possible to capture all the effects of the treatment. While you may capture their initial interactions, like a click on an ad, it's possible that a sale happens quite a bit later and maybe even after the end of your test. And you cannot capture the unknown. Some people may buy as a result of your ad, but they may do so offline, with cash, for instance, and that sale may not be recorded. And, as in any test, your test and control groups may have outliers, or people who behave quite atypical and therefore skew your results. As a result of all this, the outcomes of these tests may sometimes be difficult to replicate. But even with these limitations, RCTs are powerful tests and they're a great way to assess the effectiveness of advertising campaigns. Finally, we also discussed how observational methods can also be used to evaluate advertising effectiveness, although they are less powerful and precise. In an observational method, you observe the effect of ads on people without changing who is exposed to the ads. Observational methods are not experimental. You don't have a control group, so it makes it harder to make causal inferences. In other words, it's hard to prove that any change in behavior that you observe is caused by your ads as you don't know what would have happened if people weren't exposed to the ad. What you would otherwise evaluate with a control group. And while researchers will try to identify natural groups of people who were or were not exposed to their advertising, it's very hard to perform comparisons in a controlled way as there are usually too many external variables that you can't control. Observational methods often don't take other contextual variables into account, variables that may have an effect on the outcomes. Say, for instance, you may try to compare conversions for people who were exposed to your ads on Facebook with people who were not exposed to your ads. But the people who were not exposed to your ad may be people that don't use Facebook much at all, and they may actually have different buying patterns. Unless you compare two groups where people have been randomly assigned to a group, there may be some variables that differ between the groups you are comparing, variables you can't control. As a result, we often see biased outcomes from observational methods. So it's good to evaluate whether any of these limitations affect the data set you are evaluating. If you completed the second course in this program, you learned about the awesome framework as a methodological approach to data analytics. You first obtain data, you scrub the data, then you explore and model, and then you interpret. Assessing your data and evaluating the quality and limitations is a crucial part of the obtain face of this cycle. And once you obtain the data, you will scrub or clean them to make sure you have no duplicate, missing values, etc. These are important steps in your data analysis. Once you've obtained and scrubbed the data, it's time to formulate your hypothesis that will guide you through your analysis. Let's look at that next.