4. Gain an understanding of users

The majority of the process of creating personas is usually spent interviewing and observing people. Collaborate with another team member so that they capture anything in the interview or observation session that you might miss. You and your colleague should be physically present with the research participants, in a location that makes the most sense for the project. In a suitable location, participants tend to provide more detailed and accurate information and are less likely to behave unnaturally.

Observing and interviewing in person and in a location relevant to the project will yield the most useful data. Also, capturing information in multiple ways, such as by taking notes and recording audio, will ensure accuracy. (Image: Interactions⁶)

Direct, unfiltered interaction with participants is critical at this stage in order to gain empathy with the people you will be designing for. Being there in person is not always feasible, but try to do it whenever possible. If you can’t do that, then consider the next best thing: interviewing people who have had direct interaction with users. Personas created from secondhand information are called provisional personas⁷. The perspectives will be secondhand, filtered and biased, but they’re better than nothing and you’ll have somewhere to start your research.

Empathy is important to understanding users holistically, and direct interaction is the only way to inform the heart as well as the mind. I’m not one to get touchy-feely for its own sake, but when you are able to empathize with users firsthand, then you’re better able to tap into their intuition and act on inklings that you otherwise would not have. Firsthand exposure is paramount.

5. Analyze the data

The analysis stage is the most complicated because you must compare multiple variables of behavior and attitudes among many research participants. It requires some practice, and it gets much easier with experience. To find out more in depth, look at the descriptions in the Fluid project⁸, or consider reading Designing for the Digital Age⁹.

After learning about research participants directly or indirectly, you will need to make sense of it all by finding patterns in the data. In short, for each group of participants with the same role (for example, a group of doctors), you would rank each participant against a number of attributes of behavior and attitude, determining which participants have similar attribute rankings in order to discover common traits among them. Each group of similar participants would then serve as the source of a persona.

Personas are not just “roles,” although they might seem to be at first. Roles are great for segmenting and grouping similar users together for analysis, but roles are not personas. Roles are defined largely by the tasks people perform, rather than by how they perform those tasks or how they feel about accomplishing the tasks. Usually, two or more personas are required to represent the range of behavior within a role.

For instance, if you are designing medical software to be used by doctors, nurses, technicians and patients, then you would want to interview people in all four roles; however, when analyzing their responses, you would compare doctors to doctors, not doctors to technicians. Their roles might overlap somewhat, but don’t muddy the water by combining data from multiple roles. Compare participants with the same role; otherwise, large differences will obscure smaller differences between participants.

You can represent most of the observed behaviors and attitudes as a spectrum (for example, from low to high or sad to happy). Each spectrum should be discrete and divided into four levels.

In this sample spectrum, low and negative attributes are on the left, and high and positive attributes are on the right.

On each spectrum, participants can be given a score of 1, 2, 3 or 4 (similar to the Likert scale¹⁰). The even number prevents a neutral score (for example, a score of 3 out of 5 would provide little value in creating a persona). Some variables are difficult or impossible to represent on a spectrum; in this case, don’t force it, and instead express the variable as a multiple-choice question.

To determine how many spectra to analyze, go over the responses to the interview questions and note any distinct behavior or differences in attitude. Fewer than 5 spectra is usually too few, and over 20 is too many, so aim for somewhere in between. If you have trouble deciding which spectra to use, look back at the categories of questions in the interview script shown earlier, and determine which categories most effectively group and differentiate research participants. Motivations and goals, frequency and duration of tasks, and attitudes towards tasks are good places to start.

After you’ve listed 5 to 20 variables for a role, place each research participant on each spectrum, as shown below.

When placing each participant on a spectrum, keep in mind that the data is all relative at this point. The fact that Whitney does more surgeries than Doug, who does fewer than Elliot, is more important than who is a 2 rather than a 3 or 4 on the scale.

Once you’ve mapped all research participants on the spectra, it is time to identify patterns. This part might seem overwhelming, especially if you have researched many people. Start small by finding two people who tend to have the same scores on various spectra. This pattern analysis is similar to that of a semantic differential¹¹ or a multivariate affinity diagram¹².

Of the five spectra below that describe the doctors interviewed, Tiffany and Sam appear in the same place three times, in similar places one time, and in disparate places one time. As far as we’re concerned, Tiffany and Sam are similar people and should be represented by one persona. Looking further, Dan and Elliot also have similar or the same placement on the spectra, too, but their placement differs from that of Tiffany and Sam. Dan and Elliot should be represented by a different persona from Tiffany and Sam.

(View large version¹³)

This researcher has created five variables to compare the doctors who were interviewed and has mapped all of the participants to those spectra. If Tiffany and Sam tend to appear in the same or similar places in a lot of the spectra, then they probably constitute a pattern that will eventually help to describe a persona. Most of the time, the patterns that you find among participants will not apply to every single one of the spectra, because almost all patterns are imperfect.

This is not a problem; as long as participants match on a majority of the spectra, the pattern is valid. Repeat this process until you have discovered all of the patterns that represent groups of similar users. If done correctly, you will have compared a lot of the participants to each other, and the result should look something like the graph above (the color blue denotes the first pattern found).

Grouping similar participants together can be laborious, but it is worth every minute. Consider collaborating with other researchers to benefit from multiple perspectives and to discuss points of ambiguity or confusion. (Image: Vijay Kumar¹⁴)

This is where the difference between roles and personas comes into play. Even though the researcher interviewed many people who had the role of doctor, the participants had distinct patterns of responses, which will result in multiple personas with the role of doctor. Each unique pattern of behaviors and attitudes among participants should be represented by a unique persona.