The human mind is an exceptional problem solving machine. Our ability to think broadly and make abstract connections in the world around us allows us to interpret nuances in information that are so far beyond the capabilities of machines. However, we are fallible. And it turns out there's been all sorts of research that's identified very specific ways in which we make mistakes, and there are a lot of them. Take a look at this diagram categorized by Buster Benson and arranged by JM3. Each one of these items is a specific way we alter information to deviate from reality. We call these phenomena cognitive biases, and most of us are susceptible to most, if not all, of them. Cognitive biases can impact the way we look at data and influence our interpretations of analysis. We'll be hard pressed to cover all this biases in one course let alone in one video. But what we will do is look at some of the most common and potentially damaging biases that play on to our analytical process. Specifically we'll look at confirmation bias, the framing effect, availability, or vividness bias, anchoring and fundamental attribution error. Let's start with confirmation bias. Confirmation bias is the tendency to favor information that confirms one's beliefs or hypotheses. There are really two ways that we can exhibit this bias. The first is by selectively gathering information, that is we only seek out data that would serve to support a hypothesis and fail to seek out data that might disprove the hypothesis. The second is by selectively interpreting information. This happens when we only focus on data that supports our hypothesis even when we have data that refutes it. Here are a couple of examples. Suppose we feel like our customer care centers are not doing a good job. We decide to send a survey to a sample of customers that have called in, but exclude customers who have received credits on their bills because these might be customers whose satisfaction has been bought. This would be an example of selectively seeking out information. Let's say that we were analyzing sales results for different locations in a region. We believe their performance is strong but when we look at the data we see that there are a couple of locations that are dragging down total sales below target. We decide there must be something wrong with the data and we exclude those locations from the analysis. This would be an example of selectively interpreting information. In business there is often pressure to show good results. And this can subconsciously play into how we gather and interpret information. People can also have strong beliefs about how things should be done, so you tend to go after information that supports their agenda. As data analysts, we should try our best to remain objective and avoid both of these traps. So that's confirmation bias. Let's move on to the framing effect. The framing effect is the tendency to draw different conclusions from information based on how it's presented. To illustrate this we'll take a classic example from research and recast it in the hypothetical business context. Let's say we've completed an analysis of how the restructure 600 bad business investments and we're presenting options to a set of decision makers. If we do nothing, all 600 investments will lose $100 each. If we do something to an investment, we have some chance it will make $100. Otherwise, it still loses $100. Let's say we outline the above scenario and present the options as follows. Option A, 200 investments will make $100 each. Option B, there's a one-third probability that all 600 investments will make $100 each and a two-thirds probability that all 600 will lose $100 each. Research suggests that most people would prefer option A, when the options are presented in this way. However, let's say that we presented the options like this. Option A, 400 investments will loose $100 each. Option B, there's a one-third probability that none of the investments will loose $100 and a two-thirds probability that all the investments will loose $100. In this case the resource suggest that most people now choose option B. Now, if you read closely, you'll see that both cases describe exactly the same choice. Option A is a certainty that loses $20,000 overall. And option B has a one-third chance of making $60,000 and a two-third chance of losing $60,000 with the expected outcome of the same net $20,000 loss. Rationally, decision makers should make the same choice in both scenarios but they don't. It turns out that framing things in a positive way can elicit different results than in framing them in a more negative way. As data analysts, we need to keep this in mind both as we look at options and make our own conclusions. As well as how we present options and recommend results to our decision maker. It turns out there are other ways in which how we are exposed to information influences our thinking. Availability or vividness bias is the tendency to believe recent or vivid events are more likely to occur. There are all sorts of common examples of this, like how people perceive the risk of a plane crash or getting attacked by a shark. Both events are exceedingly rare, but are often perceived to happen more frequently than they actually do. In the business context, when will we see this bias impact analysis is when small samples of input highlight touchy topics. For example, let's say we have a focus group of seven people, and two of them say they have a problem with our product. This input is vivid. Not only is there a problem but we're hearing about the problem through a personal and possibly passionate interaction with a customer. It's also two out of seven people. So we might jump to the conclusion that nearly 30% of our products have issues. This may or may not be that case. But cognitively we tend to over value these vivid examples. Here's a real example of something similar. An executive at a wireless carrier had spent an hour or so in a call center listening to interactions with customers. Of the ten calls he heard, two of them included complaints about dropped or blocked calls. He was already concerned about network quality, so he asked the analytics team to prove that high levels of dropped and blocked calls were leaving the customers cancelling their service. It turns out that not only was the actual rate of dropped and blocked calls extremely low, about 1%, even when viewed at the customer level. But, there was no discernible relationship to cancellation. In this example, not only was there availability bias involved, but also confirmation bias. The executive was quick to pick up on data that supported his belief that network quality was an issue. What's really interesting about this example is not that it happened, these type of analyses are pretty common. But the exact same request was made two more times by two different leaders, and despite the results of the initial analysis two more analyses were done only to come to the same conclusion. It turns out that some biases are hard to overcome. The good news is, we're actually able to use data to show the truth and prevent unnecessary actions that wouldn't have had any impact. Okay, let's talk about anchoring. Anchoring is our tendency to focus or rely too heavily on the first piece of information that's available to us. This bias is regularly exploited in areas like pricing and negotiation. For example, shoppers may respond more favorably to a product that was priced at $1,000 and marked down to $600 than one that was simply listed for $600. This is because the shopper initially anchors at $1,000 and relative to $1,000, $600 seems like a good deal. The way we look at data can also be subject to anchoring. Let's say that we're responsible for doing the monthly 12-month sales forecast. We make some assumptions and run our first forecast, which predicts average sales growth of 8%. The next month we run the same forecast and come up with an average sales growth of 14%. Even though we think our methodology is sound, because the estimate is so much higher than the first forecast, we back off some of our assumptions and take the forecast down to a more reasonable 12%. In this case, how we interpreted the second forecast was anchored by our initial forecast. As analysts, we need to question our analyses. But it's also important to question why we're questioning our analysis. And make sure we're asking ourselves the right questions. The last cognitive bias we'll discuss is called the fundamental attribution error. Fundamental attribution error impacts how we interpret things that we observe people doing. Specifically we have a tendency to focus on attributes or intentions of the person themselves versus the situation or the environment when explaining a person's behavior. Where this comes into play in analytics is often in the process of translating analytical observations into strategies and tactics for action. If we misinterpret why and how customers behave, we may end up taking action that is ineffective or counterproductive. One of the more interesting places I've seen this come up is in fraud. I'm using the air quotes here because the fraud itself is a negative term. It implies dishonesty and an active attempt to deceive. Let's suppose that we sent a 50% discount code out to a set of new customers to use on our website. Our intent with that it'd be for new customers only, and that the discount can only be used once. In fact, it clearly says just that in the terms and conditions of the offer. However, when we see the discount being used far more than expected, we look at the data and find that it's being used by some existing customers and in many cases more than once per customer. Very quickly, ideas are generated about how to stop the fraud and identify abusers, cancel any pending purchases and block them from the site. What just happened here? We've implicitly assumed anyone using the code we sent out in a way we don't intend are bad actors, and we moved right to punitive solutions. Are these really bad people or are they just acting rationally in the environment we created for them? In my experience it's surely wise to give people the benefit of the doubt and to focus on the structures we create for them in identifying problems and solutions. This helps avoid attribution error. Again, these are only a few of the many ways we as humans are prone to misinterpret information in situations. If you find this interesting I encourage you to explore some of the other biases we saw in the larger framework. But for now the key takeaway from this video is that even though we deal in hard data we can still make mistakes in how we interpret and communicate that data, as well as in how we frame options for taking action. By being aware of these biases and how they effect us, we can train ourselves to avoid them and minimize their influence on our work.
