Data can be found or observed at many places. An obvious data source is the Internet. Intel (2013) presented the following iconographic, showing the massive amount of data collected by different Internet services. In 2013, digital devices created four zettabytes (1021 = billion terabytes) of data. In 2017, it is expected that the number of connected devices will reach three times the number of people on earth; hence, the amount of data generated and collected will increase even further:

To get the data from the Internet, there are multiple options, as shown in the following:

The main drawbacks of found data are that it takes time and space to accumulate the data; they cover only what happened, for instance, intentions, motivations, or internal motivations are not collected. Finally, such data might be noisy, incomplete, inconsistent, and may even change over time.

Another option is to collect measurements from sensors such as inertial and location sensors in mobile devices, environmental sensors, and software agents monitoring key performance indicators.

An alternative approach is to generate the data by yourself, for example, with a survey. In survey design, we have to pay attention to data sampling, that is, who are the respondents answering the survey. We only get data from the respondents who are accessible and willing to respond. Also, respondents can provide answers that are in line with their self-image and researcher's expectations.

Next, the data can be collected with simulations, where a domain expert specifies behavior model of users at a micro level. For instance, crowd simulation requires specifying how different types of users will behave in crowd, for example, following the crowd, looking for an escape, and so on. The simulation can be then run under different conditions to see what happens (Tsai et al. 2011). Simulations are appropriate for studying macro phenomena and emergent behavior; however, they are typically hard to validate empirically.

Furthermore, you can design experiments to thoroughly cover all the possible outcomes, where you keep all the variables constant and only manipulate one variable at a time. This is the most costly approach, but usually provides the best quality of data.

Data collection may involve many traps. To demonstrate one, let me share a story. There is supposed to be a global, unwritten rule for sending regular mail between students for free. If you write student to student to the place where the stamp should be, the mail is delivered to the recipient for free. Now, suppose Jacob sends a set of postcards to Emma, and given that Emma indeed receives some of the postcards, she concludes that all the postcards are delivered and that the rule indeed holds true. Emma reasons that as she received the postcards, all the postcards are delivered. However, she does not possess the information about the postcards that were sent by Jacob, but were undelivered; hence, she is unable to account this into her inference. What Emma experienced is survivorship bias, that is, she drew the conclusion based on the survived data only. For your information, the postcards that are being sent with student to student stamp get a circled black letter T stamp on them, which means postage is due and that receiver should pay it, including a small fine. However, mail services often have higher costs on applying such fee and hence do not do it (Magalhães, 2010).

Another example is a study, which found that the profession with the lowest average age of death was student. Being a student does not cause you to die at an early age, being a student means you are young. This is what makes the average of those that die so low (Gelman and Nolan, 2002).

Furthermore, a study that found that only 1.5% of drivers in accidents reported they were using a cell phone, whereas 10.9% reported another occupant in the car distracted them. Can we conclude that using a cell phone is safer than speaking with another occupant (Uts, 2003)? To answer this question, we need to know the prevalence of the cell phone use. It is likely that a higher number of people talked to another occupant in the car while driving than talking on the cell during the period when the data was collected.