Atish Mehta presenting Happy Factor at the QS Show&Tell; video by Paul Lundahl.
This long post is an attempt to provoke QS readers to approach the question of measuring mood, and to use their unique combinations of skills to advance the cause in a practical way. It is inspired by the Facebook app Happy Factor, presented by Atish Mehta at a recent QS Show&Tell. If you watch the video above, you will see a the presentation and the very good discussion that followed. Since Atish has already built an app to measure happiness, while others here have expert knowledge of how to analyze data, and others know computer science, and still others are social scientists, journalists and ethnographers, it seems promising to drop a few annotated references on you and see if anything happens.
There is also an interesting story to tell. As self-measurers, you won’t be surprised to discover that measuring mood has been the subject of controversy. (Those of you who are also academics won’t be surprised to know that the controversy has at times become hostile.) But in it are clues to some of the problems immediately confronting anybody who is trying to track how they feel. Let’s start the story in the 1980′s, when psychologists began to increasingly use a two dimensional model to describe and measure mood. The key paper is James A. Russell’s “A Circumplex Model of Affect,” published in 1980 in the Journal of Personality and Social Psychology. [Abstract; subscription required]
Think of your emotional space as a two dimensional grid. On the x-axis is “pleasantness/unpleasantness,” sometimes called “valence.” On the y-axis is “arousal,” or “activation.” Arranged in a rough circle around this two dimensional space are the varieties of human feeling, like this:
There were lots variations on this circumplex. Here is another one:
Here is another one:
The nice thing about this model is that you can track mood with a mood checklist containing any number of terms, but whether you use more terms or fewer, this circumplex pattern is generally going to emerge. If you are only interested in the high level constructs, rather than in all the component descriptors you should be able to approximate your mood with two questions: How happy do you feel? How energetic do you feel?
Or maybe not. In 1988, David Watson, Lee Anna Clark, and Auke Tellegen published a paper introducing the Positive and Negative Affect Schedule (PANAS), based on the idea that positive and negative affect should be separately tracked because they vary independently. In other words, it is possible to feel good and bad at the same time. To force mood tracking onto a bipolar two-dimension grid, when the supposed poles are not really opposites, is to risk screwing up badly. You might get results which reflect the architecture of the model, rather than your actual moods. So, better not use the old circumplex. Use the PANAS instead.
Here is an example of a PANAS checklist, from a survey designed to measure emotions. You are asked to report to what extent you have felt this way during the time period being measured (right now, past few hours, past week, etc.).
1 2 3 4 5
very slightly a little moderately quite a bit extremely
or not at all
______ cheerful ______ sad ______ active ______ angry at self
______ disgusted ______ calm ______ guilty ______ enthusiastic
______ attentive ______ afraid ______ joyful ______ downhearted
______ bashful ______ tired ______ nervous ______ sheepish
______ sluggish ______ amazed ______ lonely ______ distressed
______ daring ______ shaky ______ sleepy ______ blameworthy
______ surprised ______ happy ______ excited ______ determined
______ strong ______ timid ______ hostile ______ frightened
______ scornful ______ alone ______ proud ______ astonished
______ relaxed ______ alert ______ jittery ______ interested
______ irritable ______ upset ______ lively ______ loathing
______ delighted ______ angry ______ ashamed ____ confident
______ inspired ______ bold ______ at ease ______ energetic
______ fearless ______ blue ______ scared ______ concentrating
______ disgusted ______ shy ______ drowsy ______ dissatisfied with self
Thus the great bipolarity controversy was launched. The details are highly technical. (Technical in a good way, I’m sure, as long as you don’t fear getting lost in the byways of this debate for weeks longer than you originally intended: you have been warned.) For a short taste of the polemics you can read Russell and James M Carroll’s The Phoenix of Bipolarity: Reply to Watson and Tellegen (1999) [PDF]. In this short essay, which served as a final salvo on the first stage of the controversy, the Forces of the Circumplex seemed to achieve victory. It turned out that under widely varying circumstances data from the PANAS showed – when analyzed – very strong evidence of bipolarity. Items from the poles of the circumplex descriptors had strong correlations (positive or negative, as the case may be), and the descriptors arranged themselves roughly around the circle. The opposition between happiness and sadness was not an artifact of the model, in general. Watson and Tellegen appeared to concede the most controversial points, and bipolarity was saved. Bring back the circumplex!
Or maybe not. Let’s bring this subject into the realm of anecdote, where journalists are comfortable. I can think of many situations in which I feel both positive and negative feelings. A good day comes to an end. A good night comes to an end. I speak a foreign language with happy enthusiasm mixed with sad awareness that lack of practice has insured my incompetence. We are having inspiring conversation which is pleasing me extremely, but the whole time I’m thinking: it’s tragic I can’t give it my whole attention because I just heard a sad story from a friend I’m worried about. I like a devastating intellectual critique as much as the next bystander; and Russell and Carroll’s reply to Watson and Tellegen seemed definitive; but in me, at least, positive and negative feelings coexist. And since we are talking about subjective experience, don’t the subjects get to have the last word?
In fact, this problem with the circumplex model has also been noticed by professionals, and looked into, and confirmed. Among the leading researchers to insist upon the independence of positive and negative feelings is John T. Cacioppo. The model Cacioppo recommends is called the Evaluative Space model; you can read more about it in his important paper: Relationship Between Attitudes and Evaluative Space; A Critical Review, With Emphasis on the Sparability of Positive and Negative Substrates. [PDF]. Cacioppo proposed an architecture of the emotional space that looks like this. Please start redesigning your mood trackers now.
Then, in 2001 paper called Can People Feel Happy and Sad at the Same Time? [PDF], Jeff T. Larsen, A. Peter McGraw, and Cacioppo used some experiments to demonstrate clearly that happiness and sadness can indeed co-occur. The frequency with which this happens is obviously something that could be further researched, but the key point is that if happiness and sadness usually correlate negatively, but sometimes correlate positively, then they cannot be polar opposites. Another researcher, Eshkol Rafaeli (whose web site contains many interesting references), has put the label affective synchrony on the phenomenon of simultaneous happiness and sadness, and has done some work which, while limited in scope, clearly suggest that all people are not alike when it comes to affective synchrony. Some people experience happiness and sadness as bipolar opposites; others have more experiences of mixed emotions. (See Rafaeli’s Affective Synchrony: Individual Differences in Mixed Emotions. [PDF]) Down with the circumplex!
Of course in your self-tracker you are welcome to measure anything you want. The controversy is relevant because what you choose to measure defines your experiment. If the circumplex is valid, you can simply measure “mood” on a simple grid. If it is not valid, you need to think more about what descriptors to present.
At this point, I hope you are now both happy and sad. Happy, that so much research on this question has been done for you. Sad, because this research remains ambiguous where it is not merely confusing. What would a good mood tracker look like? Is there a clear, promising direction? If you only look at one of the references attached to this post, I recommend Lisa Feldman Barrett’s Solving the Emotion Paradox: Categorization and the Experience of Emotion (PDF). Barrett offers both a wide-ranging review of the state of the art in modeling human emotion; more importantly, her theory of emotion offers some clear and plausible guidance for self-trackers.
Barrett divides emotion into two parts. Core affect is the totality of a person’s state that is available for emotional processing. The physiology of core affect is shared among mammals. All of us carry out some degree of assessment along the lines of good/bad; and all of us vary in our level of energy. (The traits evolved long before mammels, of course, but let’s try to keep our focus.) The fact that human descriptions of emotion clump together along the two dimensions of the circumflex reflects the structure of our core affect. This suggests that the circumplex is perfectly valid for many potential self-tracking experiments.
Let’s say we want to quit smoking, but we know that the process of quitting will make us tense and unhappy. We have been warned not to make any important life decisions during the traumatic period of our quit attempt, we know from reading the literature that the effects of smoking and smoking withdrawal on mood will show strong diurnal rhythms. Perhaps it would be prudent to have a mood meter that, based on our self-ratings of mood on a simple bipolar, two dimensional scale, can predict what times of day we ought to take ourselves off-line, as well as show us the rate of improvement and projected “finish line” after which the acute effects of quitting will be negligible. All of these are easily envisioned applications of the circumplex model. We are concerned about some basic elements of our emotional life – valence /arousal or pleasantness/activation – in which our categories do not have to do very much subtle work. “These feelings” writes Feldman, “are primitive (psychologically irreducible) and universal…” Go circumplex!
But there are other situations in which the circumplex will fail us. Wherever the aim of our experiment is to understand and/or alter the structure of our mood, we shouldn’t start with a system that assumes a structure and incorporates into its measurement protocols.
Why should we be interested in the structure our moods? Here is where the second part of Barrett’s paper becomes interesting. In her model of emotion, core affect is the material on which emotion works, but the experience of emotion – the inner experience, as well as most of the repertoire of outwardly emotional behavior – comes from the act of categorizing core affect, giving it a label such as “anger,” “sadness” or “fear.” This does not mean that the emotion is not experienced until you are conscious of putting a name on it. You don’t have to quietly mutter “anger” in order to feel anger. But it does suggest that anger is a concept that you begin learning in fancy and may continue to extend and revise throughout life. The repeated experience of labeling a combination of core affect and the context in which it occurs as “anger” trains you in how to be angry and how to recognize anger. Barrett describes emotions as simulations, in the sense that they take an experience of core affect, plus the situation in which it occurs, and compute an appropriate result:
“….conceptual knowledge about emotion constitutes expertise about how to deal with your own internal state – experienced as “an emotion” – and the situation or event that you believe caused that emotion in the first place. In this sense, emotional categorization is functional. Situated conceptualizations may be thought of as an inference about what will make for successful self-regulation or goal achievement…”
This is a theory of emotion that could be articulated by a robot, and it strikes me as entirely plausible. Here are the practical implications:
“…conceptualizing core affect as emotion, like conceptualizing in general, is a skill… This skill for wielding conceptual knowledge about emotion might be considered a core aspect of emotional intelligence… It is a skill to simulate the most appropriate or effective representation, or even to know when to inhibit a simulated conceptualization that has been incidentally primed. Presumably, this skill not only can be measured, it can also be trained.”
This suggests that we can revise our emotional architecture through experiments in description.
When I was just out of my teens I worked in a very fancy restaurant under the mad tutelage of the chef-owner. He was proud of his wine list, and asked us to be prepared to sell it to patrons; to this end he let us taste everything and trained us in the vocabulary of oenology. The obscurity of the distinctions permissible in describing wine are notorious, but the chef only tolerated my faux naïve remarks for the first five minutes of the first lesson, after which he said: “You are being asked to remember these wines. These adjectives are your labels. You are welcome to make up your own, but then I can’t instruct you, and nobody will understand you.” Suddenly, I saw that the absurd refinement of descriptors might have a useful purpose. Like emotions in the scheme Barrett describes, such phrases are simulations, bringing to mind previous experiences, triggering a cascade of associations. They were a form of knowledge, a social prediction machine; they allowed me to meet the demands of the situation, assess possibilities, invite cooperation.
Barrett’s theory of emotion opens the door for another type of self-tracking than is permitted by the circumplex, tracking that asks questions like: How many emotions do I have? What is the range of my emotions? Do I meet various experiences with well tuned emotional responses, or have my feelings become rigid and stereotyped? Her paper suggests that we can improve our emotional structure, increasing the granularity of emotional experiences by enriching our vocabulary and learning to apply it to previously unnoticed patterns in affect and context. (I am assuming for the moment that a more complex structure of emotion is a good thing. This could be questioned. But the first step in any case is mapping our emotional architecture.)
I’ll end with an idea about how this second type of mood tracking could work. First, let’s revisit two projects already underway: Ka-Ping Yee’s time allocation diary; and Atish Mehta’s Happy Factor. At the first QS Show&Tell, Ping showed us his time-allocation diary, which he keeps using a widget that stays open on his screen. He can enter some text into the box, where it automatically gets a time/date stamp. He often adds a keyword, so that he can graph his activities by category. Such a system is simple and flexible, and is not dependent on fixed categories; Ping can always start and maintain a new category simply by using a new word in his short entries.
Meanwhile, Atish has created a Facebook app that randomly queries users with a text message and asks them to rate their happiness. The ratings get a time/date stamp, and allow for the entry of a short note.
A mood tracking system to investigate emotional architecture might fruitfully combine these two methods. The ability to perform randomly timed queries is powerful. (For more on this, see “The Descriptive Experience Sampling Method” (PDF) by Russell T. Hurlburt and Sarah A. Akhter. I also discuss this method in a review of Describing Inner Experience by Hurlburt and philosopher Eric Schwitzgebel.) Right now Happy Factor asks only about happiness. Asking a constrained question such as “how happy are you” is only useful in the context of other data; to make use of Happy Factor as currently designed requires exporting the data and combining it with data about some other dimension of your life, in order to give it meaning. But think about if the question were unconstrained: “what emotion are you experiencing right now?” Suddenly, the descriptive landscape our mood becomes accessible. [Starting at about minute 12:00 in the video above, some of these ideas are batted around in the discussion, along with other interesting prospects.]
Once the structure of our moods became accessible for visualization, experiments and interventions become possible. The type of experiments or interventions that might be interesting is left as an exercise for the reader.
I have used mood, feeling, and emotion interchangeably. Though they are not the same, similar questions of measurement apply. A good recent summary of the controversy over the circumplex is given in: Causes and Consequences of Feelings (Studies in Emotion and Social Interaction) (Paperback) by Leonard Berkowitz. The book is expensive but I’m happy to share the key pages if anybody needs them.