# Chaos in an Atmosphere Hanging on a Wall

This month marks the 50th anniversary of the 1963 publication of Ed Lorenz’s groundbreaking paper, Deterministic Nonperiodic Flow, by the Journal of Atmospheric Science. This seminal work, now cited more than 11,000 times, inspired a generation of mathematicians and physicists to bravely relax their linear assumptions about reality, and embrace the nonlinearity governing our complex world. Quoting from the abstract of his paper:

A simple system representing cellular convection is solved numerically. All of the solutions are found to be unstable, and almost all of them are nonperiodic.’

While many scientists had observed and characterized nonlinear behavior before, Lorenz was the first to simulate this remarkable phenomenon in a simple set of differential equations using a computer. He went on to demonstrate the limit of predictability of the atmosphere to be roughly 2 weeks, the time it takes for two virtually indistinguishable weather patterns to become completely different. No matter how accurate our satellite measurements get, no matter how fast our computers become, we will never be able to predict the likelihood of rain beyond 14 days. This phenomenon became known as the butterfly effect, popularized in James Gleick’s book Chaos.

Lorenz’s sketch of the attractor for his system.

Inspired by the work of Lorenz and colleagues, in our lab at the University of Vermont we’re using Computational Fluid Dynamics (CFD) simulations to understand the flow behaviors observed in a physical experiment. It’s a testbed for developing mathematical techniques to improve the predictions made by weather and climate models. Here you’ll find a brief video describing the experiment analogous to the model developed by Lorenz:

And below you’ll find a CFD simulation of the dynamics observed in the experiment:

What is most remarkable about Lorenz’s 1963 model is its relevance to the state-of-the-art in weather prediction today, despite the enormous advances that have been made in theoretical, observational, and computational studies of the Earth’s atmosphere. Every PhD student working in the field of weather prediction cuts their teeth testing data assimilation schemes on simple models proposed by Lorenz, his influence is incalculable.

In 2005, while I was a PhD student in Applied Mathematics at the University of Maryland, the legendary Lorenz visited my advisor Eugenia Kalnay in her office in the Department of Atmospheric & Oceanic Science. At some point during his stay, he penned the following on a piece of paper:

Chaos: When the present determines the future, but the approximate present does not approximately determine the future.’

Even near the end of his career, Lorenz was still searching for the essence of nonlinearity, seeking to describe this incredibly complicated phenomenon in the simplest of terms.

_______________________________________________________________

*Note: this post also appeared as part of the Mathematics of Planet Earth 2013 daily blog.

Taming Atmospheric Chaos with Big Data, a talk I gave at the 2011 UVM TEDx Conference Big Data, Big Stories:

How does food (or talking about food online) relate to how happy you are? This is part 3 of our series on the Geography of Happiness. Previously we’ve looked at how happiness varies across the United States (as measured from word frequencies in geotagged tweets), and then at how different socioeconomic factors relate to variations in happiness. Now we focus in on one particular important health factor that might influence happiness, obesity.

We looked at how happiness varied with obesity across the 190 largest metropolitan statistical areas in the United States, giving us the following scatter plot:

Each point represents one city; for example the city with both(!) lowest obesity and greatest happiness in this set is Boulder, CO, located at the top left. The red line is a linear trend through the data (a line of best fit). Again, for the mathematically minded onehappybird watchers, we show the Spearman correlation coefficient and its corresponding p-value at the lower left. We do this to convince you that there is, in fact, a statistically significant downward trend in the blob of points in the picture! The big story here is of course that as obesity goes up, happiness goes down.

The natural next question to ask is: are there any words which could be indicators of obesity? What foods are people in obese cities eating, or talking about? To answer this question we correlated word frequencies with obesity, and searched for the most strongly-correlating food-related words. Below are two examples: on the left, “mcdonalds”, and on the right, “cafe”.

As obesity goes up, so does talk (at least on Twitter) about McDonalds, but talk about cafes follows the opposite trend! Does that mean that in order to lose weight we should spend more time sipping lattes in cafes? I wish.

Looking through the list of words, the top 5 food-related words that increase in frequency as obesity went up were:

1. mcdonalds
2. eat
3. wings
4. hungry
5. heartburn

We were surprised by ‘hungry’! On the other hand, the top food-related words which were used more as obesity went down were:

1. cafe
2. sushi
3. brewery
4. restaurant
5. bar

Perhaps unsurprisingly, these are words typically used by the high-socioeconomic group described in our previous post on city happiness, suggesting that better health correlates with higher socioeconomic status. You can find the complete list of how all words correlate with happiness here (page best viewed using Google Chrome). One surprising result was the observation that far more food-related words appeared in the low-obesity group than in the high-obesity group; in other words, food was being talked about more in the less-obese cities!

Summarizing: based on word usage, the Twitter diet consists of: breakfast at your favorite cafe, a delicious sushi lunch, dinner out at a fancy restaurant, with a nightcap at the best local bar or brewery. Thank you Twitter, don’t mind if I do.

All jokes aside, this sort of technique has great potential. Imagine being able to predict whether obesity was going to rise or fall in a city, or estimate changes in other demographics, just by analyzing the words people use online. Perhaps New York City Mayor Michael Bloomberg would find some early indicators of the success or failure of his war on soda!

And that’s all for this series of posts on the geography of happiness. More information on all of the results in this series can be found in our recently submitted arxiv paper. Please take a look at it and the accompanying online appendices, where you can look through all of the data yourself. As a special bonus feature, you can check out this video of me talking about this work at our recent TEDxUVM conference.  Thanks for reading!

# What makes a city happy?

Welcome back, onehappybird watchers! Wow, what a crazy week of coverage of our post about how happiness varies by city and state across the United States. Many, many people read, shared, and commented on the post, for which we are grateful. For the detailed explanation of the results, check out the full paper we recently submitted to PLoS ONE.

A number of readers wondered how variations in happiness relate to different underlying social and economic factors. To try to answer this question, we took data from the 2011 census (all helpfully available online on the Census Bureau’s American FactFinder website) and correlated it with our measure of happiness. Surprisingly, happiness generally decreases with the number of tweets per capita in a city (this doesn’t mean that tweeting more will make you less happy, it’s only a correlation):

Next, we grouped covarying demographic characteristics obtained from the census, and looked at how these clusters varied with happiness. For example, it might not surprise you that cities with a larger percentage of married couples also contain a larger percentage of children – this is what we mean by covarying demographics.  And you might or might not be surprised that more marriage is positively correlated with happiness.  There’s plenty of scatter but the connection is there:

Scatter plot of happiness vs. percentage of population married. Each dot represents one city, the rho and p-values reported are Spearman correlations.

We used an automated algorithm to bin the census data for us into eight groups and then compared the happiness of those groups, leading to the following figure:

Each point represents a characteristic from the census (for example, the % married/happiness plot above is now represented by one point in this figure), with the horizontal groupings representing covarying demographic characteristics. A point’s position on the vertical axis shows how that characteristic varies with happiness across all cities. A positive value means that happiness is higher in cities where that characteristic is higher, while a negative value means that happiness is lower in cities where that characteristic is higher. For example, the figure shows that as the percentage of married couples in a city increases, so does the average happiness of that city (no causality is implied).

Only two groupings (the colored dots on the far left and right) showed strong correlation (either positive or negative) with happiness. Looking at which characteristics make up these groups, it appears that the general story here is a socioeconomic one, and one that holds only at the extremes. With our peculiar Twitter-based lens, we see money statistically correlates with happiness, which is not quite as catchy as “money buys happiness” (see the debate over the Easterlin Paradox for more). You can delve into the data yourself – the correlations of all 432 characteristics of cities recorded by the census with happiness can be found here (page best viewed using Google Chrome).

A more interesting question might be how word usage varies with different demographics – to do this we correlated each word with each demographic characteristic across all 373 cities in our dataset, leading to a lot of data to sift through! (And you can too, by following the link in the above paragraph.) As an example, take a look at how the word “cafe” varies with the percentage of population with a college degree:

Each point in the figure represents one city, and broadly the trend is that the more “college-y” the city is, the more people talk about cafes online. (You can decide for yourself whether that’s surprising or not). The top 10 emotive words whose usage went up as percentage of population with a college degree went up turned out to be:

1. cafe
2. pub
3. software
4. yoga
5. grill
6. development
7. emails
8. wine
9. art
10. library

And the emotive words which went up as college degrees went down?

1. me
2. love
3. my
4. like
5. hate
6. tired
7. sleep
8. stupid
9. bored
10. you

We saw similar patterns of word use across many socioeconomic characteristics – emotive words and words about interpersonal relationships (‘me’ and ‘you’) at one end of the spectrum, and words about more complex social or intellectual themes at the other. Interestingly, we find more food-related words in this group as well.

Of course, all of this is open to interpretation. As many commenters last week pointed out, Twitter users (indeed, specifically those users who geotag their tweets using a mobile device) are a small, non-representative sample of the global population. Furthermore, our method is undeniably crude, and by breaking texts up into their constituent words ignores the context in which those words were used. That said, many of these results agree with our intuition (for example, many of the cities with low happiness scores also appeared on a list of America’s “most miserable cities” published late last week by Forbes), while some surprise us. There is certainly a lot to be learned by looking at what the data can tell us, and we encourage you to do so by exploring our website of supplementary data. Again, you can read the full technical details in our research paper here.

We’ll pick up on the theme of food again in our next post, which will focus on one important health factor relating to happiness – obesity.

# Where is the happiest city in the USA?

(Update: this work is now published at PLoS ONE)

Is Disneyland really the happiest place on Earth?* How happy is the city you live in? We have already seen how the hedonometer can be used to find the happiest street corner in New York City, now it’s time to let it loose on the entire United States.

We plotted over 10 million geotagged tweets from 2011 (all our results are in this paper, also on the arxiv), coloring each point by the average happiness of nearby words (detail on how we calculate happiness can be found in this article published in PLoS ONE):

As well as cities and the roads between them, we can make out many regions of higher and lower happiness, even within individual cities. As an example, check out this tweet-generated map of the city of Chicago:

Tweet-generated map of Chicago. Click to enlarge.

Notice the striking contrast between the relatively happy Central/North Side of the city, and the sadder South Side. You can also find a few airports in this map, and if you look very closely you might even be able to pick out happy and sad terminals!

To quantify this variation in happiness a bit better, let’s look at the average happiness of each state:

Southern states tend to produce sadder words than those in northern New England or out west. Hawaii emerges as the happiest state and Louisiana as the saddest, due to relative differences in the frequencies of happy and sad words used in each state. Here at onehappybird, we characterize such differences by “word shifts”, which are basically word clouds for grown-ups. You can find examples of these, as well as the full list of the average happiness of each state, here (page best viewed using Google Chrome).

Zooming in further to the level of cities, we produced a similar list for 373 cities in the lower 48 states (you can find the full list, as well as maps and word shifts for each city, here). With a score of 6.25, we found the happiest city to be Napa, CA, due to a relative abundance of such happy words as “restaurant”, “wine”, and even “cheers”, along with a lack of profanity.

At the other end of the spectrum, we found the saddest city to be Beaumont, TX, with a score of 5.82. In general, cities in the south tended to be less happy than those in the north, with a major contributing factor being the relative abundance of profanity used in those cities.

We can go even further than this, and group cities by similarities in word usage. Each square in the heatmap below represents the similarity (Spearman correlation for you mathematically minded onehappybird watchers) between word distributions for the largest cities in the US. Red squares mean that the corresponding cities use words in a similar fashion, while blue means that those cities tend to use different types of words with respect to each other. The colors in the tree diagram at the top signify clusters of cities exhibiting similar word usage (below a certain threshold).

As we might expect for two cities that are geographically nearby, New Orleans and Baton Rouge are clumped together at the bottom right of the figure. On the other hand, New York and Seattle get clumped together as well, suggesting that similarities in language depend on more than just geographical proximity.

You can find more information about happiness and cities, as well as details on the methods used to produce these results, in our arxiv research article. In our next post, we’ll look at how these results are related to various underlying socioeconomic characteristics of cities. What makes a city happy or sad? Can we use Big Data to predict future changes in the demographics, health, or happiness of a city? How does happiness relate to the food you eat?

*By the way, to answer the question at the start of this post: According to this analysis Disneyland is not the happiest place on Earth; it isn’t even the happiest place in Southern California! See if you can find it in this tweet-generated map of LA! Or find your city here.

# Who will your friends be next week? The link prediction problem

Sitting in the student center of our university, I am surrounded by hundreds of students enjoying their lunch and socializing. They’re strengthening (and in some cases weakening) their social ties. Given the ability to observe this social network over time, we would see that some relationships flourish, while others disappear altogether.

This situation is not unique to university students. In fact, whether we’re studying the spreading of an infectious disease, or the growth of an organized crime network, the reality is that the relationships in these social networks are dynamic. They change. Being able to describe the current state of a network is important. Our goal is to predict the future state of the network by forecasting who will connect to whom. If we could make these sorts of predictions, then we may be able to better recommend or warn of future probable links, as well as come to understand mechanisms which may be driving the evolution of the network.

This brings us to the link prediction problem. Liben-Nowell and Kleinberg defined it as follows:

The link prediction problem asks, “Given a snapshot of a network at time t, can we predict new links which will occur at time t+1?”

In our work, we explore how topological similarity indices of a network can be combined with node specific data to develop a link prediction tool. Rather than pre-supposing that all similarity indices are of equal importance, we employ an evolutionary algorithm to evolve coefficients to be used in a linear combination of these similarity indices. Our approach has the advantage of being able to detect which similarity indices are more salient predictors and doesn’t require any knowledge of the type of network one may be working with.

To test our method, we begin by examining one of the largest social networks in existence, namely Twitter. Below, we visualize the network of reciprocal replies for users who interacted within a single week in late 2008.

Given information about interactions in a given week, we seek to predict links in a future week.

Our evolved predictors exhibit a thousand fold improvement over random link prediction, and a substantial improvement over individual indices used in isolation. The predictor also suggests possible factors which may be driving the evolution of Twitter reciprocal reply networks during the timespan of this study.

Our predictor reveals which topological or user specific indices are most important in link detection for a given a network. For example, index B is most often the top ranking index detected by the predictor, while indices E and I are also important. This type of output can be helpful in detecting the salient features which may be driving the network’s evolution over time.

Returning to our original question, we ask: Given a snapshot of a Twitter reciprocal reply network, is it possible to predict the links which will occur in the future?

One approach is to predict a link between all of the people in the network, but clearly this is not the approach we wish to take. For example, in the case of a modestly sized network (say N=30,000 nodes), the number of potential links is roughly half a million! If we’re trying to recommend books to a shopper or potential dates to a person using a dating service, we certainly would not want to suggest half a million people for potential dates. It would be nice if we could recommend 10 books (or dates) and have the majority of those suggested links be successful.

In the language of signal processing, we hope to have a true positive rate (e.g., the % of time you’re actually right about the links that you predict) that is greater than the false positive rate (e.g., the % of time you’ve issued a false alarm). This relationship can be captured by the Receiver Operating Curve (ROC) shown below.

The Receiver Operating Curve (ROC) compares the true positive rate (TPR) and the false positive rate (FPR). The ROC curve shown here depicts a classifier (our link predictor) for which TPR>FPR.

The area under the curve (AUC) provides one way of measuring the relative success of one’s method. An AUC of .50 suggests that the true positive rate is equal to the false positive rate, while an AUC > 0.50 indicates that the true positive rate is greater than the false positive rate. As shown in the picture above, our AUC is well above 0.50, in fact it is approximately 0.72, which is good!

These results are exciting! We’ve put together a research paper in which we describe our analysis and algorithm in more detail (http://arxiv.org/abs/1304.6257). Although we focused on Twitter in this investigation, our methodology is general and may apply to link prediction in many other types of time varying networks, such as disease networks or crime. It could also improve the “friends you may know” feature offered by many social networking services.

Here is a short video summarizing our work on link prediction:

# The Daily Unraveling of the Human Mind

Each morning we find ourselves in wide flung arms of drowsy possibilites. Cradled by the warm embrace of our beds, we begin our day, rebooted and rejuvenated. Having not eaten for a full eight hours, we can enjoy a guilt free breakfast, setting a blissful tone for the day.

Hourly frequency of meal references on twitter.
See figure 1 page 3 of our paper for details.

Last night’s dreams of victory and triumph bolster our delusions of adequacy, preparing us to surmount any of life’s challenges. But the moment we step outside, reality commences its slow and insidious descent. Its weight, compressing our spine, crushing our dreams, alters the course of the day completely.  The soul crushing litany of work, interacting with people, and generally living our lives takes its toll. As our sanity unravels, apathy takes root. The profane becomes our standard of expression. In the throes of despair, we swear just to feel something. We swear increasingly as we realize the inevitability of repeating this all again tomorrow.

F***, that’s a terrifying thought.

This ephemeral pattern is reflected in our tweets, our spontaneous burst of being. Below, we see our happiness peaks during the early hours of the day, and degrades as the hours progress (yellow circles). The proportion of profanity in our tweets, however, follows a reverse cycle. Profanity appears in a smaller percent of tweets at the start of each day, and increases gradually as time wears on.

Daily Unraveling
See figure 10 page 15 of our paper for details.

Remarkably, the relative frequency of these five expressions of frustration (a******,  b****, s***, f***, m***********) are quite similar.

Well done, humans.

To avoid experiencing the daily unraveling, we recommend eating organic, local dark chocolate all day long.

# What’s the Most Important Theorem?

Mathematical truths are organized in an incredibly structured manner. We start with the basic properties of the natural numbers, called axioms, and slowly, painfully work our way up, reaching the real numbers, the joys of calculus, and far, far beyond. To prove new theorems, we make use of old theorems, creating a network of interconnected results—a mathematical house of cards.

So what’s the big picture view of this web of theorems? Here, we take a first look at a part of the Theorem Network’, and uncover surprising facts about the ones that are important.  This is blatantly fun for us. Really.

Let’s go through an example starting with the real numbers.  Mathematicians like to write these numbers as $mathbb{R}$, and here we’ll start by bravely assuming that they exist. One result that follows from the existence of $mathbb{R}$: Given a real number $a$ belonging to $mathbb{R}$, we can find a natural number $n$ (e.g. 1, 2, 3 …) such that $n>a$. This is known as the Archimedean property.  To visualize this relationship, we draw an arrow from the existence of $mathbb{R}$ to the Archimedean property:

Now, the fact that real numbers satisfy the Archimedean property tells us something about sets that contain them. For example, more than a century ago, two guys named Heine and Borel used the Archimedean property to help prove their glorious, eponymous theorem.  We’ll now add an arrow leading from the Archimedean Property to the Heine Borel theorem, and we’ll include the one other component Heine and Borel needed:

All right: who is this De Morgan and what are his laws?  Back in the mid 1800′s, Augustus De Morgan dropped this bit of logical wizardry on the masses: “the negation of a conjunction is the disjunction of the negation.” We know, really exciting words.  If it’s not true that both A and B are true, then this is the same as saying either A or B or both are not true.  Better?

Before diving into a larger network, let’s think some more about these links.  One could prove the Fundamental Theorem of Calculus (which sounds important but could be just good branding) with nothing more than the axioms of ZFC set theory. But such a proof would be so long and tedious that any hope of conveying a clear understanding  would be lost.  Imagine taking all the atoms that make up a duck and trying to stick them together to create a duck; this would be the worst Lego kit ever.  And so in any mathematical analysis textbook, the theorems contain small stories of logic that are meaningful to mathematicians, and theorems that are connected are neither too close or too far apart.

For this post, what we’ve done is to take all of the theorems contained in the third edition of Walter Rudin’s Principles of Mathematical Analysis, and displayed them as nodes in a network. As for our simple networks above, directed edges are drawn from Theorem $A$ to Theorem $B$ if the proof of $B$ relied on $A$ explicitly. Here’s the full network:

###### Node size weighted by total incoming degree, colored by chapter, and laid out by Gephi’s Force Atlas.

We find that Lebesgue theory (capstoned by Lebesgue Dominated Convergence) lives on the fringe, not nearly as tied up with the properties of the real numbers as the Riemann-Stieltjes integral or the integration of differential forms. Visually, it appears that the integration of differential forms and functions of several variables rely the most on prior results. Over on the right, we’ve got things going on with sequences and series, where the well-known Cauchy Convergence criterion is labeled. By sizing the nodes proportional to their outgoing degree (i.e., the number of theorems they lead to), we observe that the basic properties of $mathbb{R}$, of sets, and of topology (purple) lie at the core.

By considering the difference between outgoing and incoming degrees, we can find the most fundamental result (highest differential in outgoing and incoming degree, or net outgoing degree), and the most important or “end of the road” result (highest differential in incoming and outgoing degrees, or net incoming degree).  In Rudin’s text, the most fundamental result is De Morgan’s Laws, and the most important result is Multivariate Change of Variables in Integration Theorem (MCVIT, that’s a mouthful).

So the Fundamental Theorem of Calculus falls short of the mark with a net incoming degree 19, not even half of MCVIT’s net incoming degree of 45. And it is not the axioms of the real numbers that are the most fundamental, with the Existence of $mathbb{R}$ having a net outgoing degree of 94, but instead the properties of sets shown by De Morgan with a whopping net outgoing degree of 122. Larry Page’s PageRank (the original algorithm behind Google) and Jon Kleinberg’s HITS algorithm also both rate the MCVIT as the most important result.

Would you agree that MCVIT is the most “important” result in Rudin’s text? It could just be the most technical.  We have only used a few lenses through which one might choose to evaluate the importance of theorems, so let us know what you think, or give it a try. Here’s a link to the Gephi files, containing all of the data used here.

Lastly, the network itself can be built differently by changing which theorems are included, or which are used in proofs. The resulting structures combine historical development with the author’s understanding. The goal of new textbooks is, in part, to organize the results in the most understandable fashion. With this view, we can start to think of the Theorem Network as the natural structuring of complex mathematical ideas for the human mind.

Now, one might idly think of extending this analysis to all of human knowledge. In that direction, Griff over at Griff’s Graphs has been making some very nice pictures leveraging the work of all those who edit Wikipedia.

# If you’re happy and we know it … are your friends?

Do your friends influence your behavior?  Of course they do.  But it’s hard to actually measure their influence.  Social contagion is difficult to distinguish from homophily, the tendency we have to seek relationships with people like ourselves.

In response to the “happiness is contagious” phenomenon promoted by Nicholas Christakis and James Fowler, we here at onehappybird were wondering whether happy Twitter users were more likely to be connected to each other.  In other words, is happiness assortative in the Twitter social network?  (See related work here.)

In the image below, each circle represents a person in the social network of the center node.  We color nodes by the happiness of their tweets during a single week.  Pink colors are happier, gray colors are sadder, and nodes depicted with the color black did not meet our thresholding criteria (50 labMT words).

We established a friendship link between two users if they both replied directly to the other at least once during the week.

As users are added to this network, it quickly becomes difficult to tell whether pink nodes are disproportionately connected to each other, so instead we look at the correlation of their happiness scores.  The plot below shows the Spearman correlation coefficient of the happiness ranks for roughly 100,000 people, with blue squares and green diamonds indicating different word thresholds, and red circles representing the same network but with randomly shuffled happiness scores.

The larger correlation for friends indicates that happy users are likely to be connected to each other, as are sad users. Moving further away from one’s local social neighborhood to friends of friends, and friends of friends of friends, the strength of assortativity decreases as expected.

We also looked at the average happiness of users as a function of their number of friends (degree k). Happiness increases gradually with popularity, with large degree nodes demonstrating a larger average happiness than small degree nodes.

The most popular users used words such as “you,” “thanks,” and “lol” more frequently than small degree nodes, while the latter group used words such as “damn,” “hate,” and “tired” more frequently.  The transition appears to occur near Dunbar’s number (around 150), demonstrating a quantitative difference between personal and professional relationships.

Finally, here we show a visualization of the reciprocal-reply network for the day of October 28, 2008.

The size of the nodes is proportional to their degree, and colors indicate communities detected by Gephi’s community detection algorithm.

For more details, see the publication:

C. A. Bliss, I. M. Kloumann, K. D. Harris, C. M. Danforth, P. S. Dodds.  Twitter Reciprocal Reply Networks Exhibit Assortativity with Respect to Happiness. Journal of Computational Science. 2012. [pdf]

Abstract: Based on nearly 40 million message pairs posted to Twitter between September 2008 and February 2009, we construct and examine the revealed social network structure and dynamics over the time scales of days, weeks, and months. At the level of user behavior, we employ our recently developed hedonometric analysis methods to investigate patterns of sentiment expression. We find users’ average happiness scores to be positively and significantly correlated with those of users one, two, and three links away. We strengthen our analysis by proposing and using a null model to test the effect of network topology on the assortativity of happiness. We also find evidence that more well connected users write happier status updates, with a transition occurring around Dunbar’s number. More generally, our work provides evidence of a social sub-network structure within Twitter and raises several methodological points of interest with regard to social network reconstructions.

# Question: Where is the happiest place in New York City?

1. Immediately adjacent to any hot dog stand.
2. Madison Square Garden during moments of Linsanity.
3. Tim Tebow’s new apartment building.

No really though, let’s measure some stuff.

Facts: (1) New York City is the most populous city in the US and (2) Manhattan streets are arranged on a rectangular grid. We have already seen how cities, airports, and even streets can be identified using geotagged tweets – here we use more than a half million messages from 2011 to investigate the happiness of NYC streets and avenues (clearly visible in the image below, as is Central Park).

Binning tweets by avenue and street, we use the labMT word list to measure happiness in tweets as a function of avenue and street number:

The results suggest that the west side is slightly happier than the east side, and that happiness actually declines as one moves further uptown. Next we bin by intersection and plot a heat map showing the distribution of happiness over all of the street corners in Manhattan:

The happiest “corner” is actually just inside the western edge of Central Park, where the intersection of 7th and 77th would be (this is just north of the lake and east of the Hayden Planetarium)*. This corner elicits tweets with a relatively high abundance of the positive words “loves” and “sky”, and proportionally less negative words like “not”, “fear” and “no”. Many of the happiest locations actually fall within Central Park!

* Please note that the results reported in this post have not been vetted through panels of experts, statistical tests of significance, or scientific peer review.  They are intended to be a fun and lighthearted exploration of our more formal research interests.

# Does QWERTY Affect Happiness?

Last week, news broke of a paper published in the Psychonomic Bulletin and Review by Kyle Jasmin and Daniel Casasanto claiming to observe a positive relationship between the “right-handedness” of a word and its emotional valence. This is being called the ‘QWERTY effect’. (You may recall that ‘valence’ is psych-speak for ‘happiness’ associated with words.  What I called right-handedness they call the right side advantage of a word, $\text{RSA} = (\text{\# of right side letters}) - (\text{\# of left side letters})$ when typed using the ubiquitous QWERTY keyboard. )  You can read the original paper here, and there’s a Wired article that explains their conclusions.

Particularly interesting for the group here in Vermont was Jasmin and Casasanto’s use of the Affective Norms for English Words (ANEW, from Bradley and Lang (1999)) dataset, along with comparable data for Spanish and Dutch, in their analysis. The hedonometric work we’ve done on blogs, music lyrics, Twitter, etc. was initially based on the happiness scores from the ANEW study. The 1034 ANEW words were handpicked to represent the emotional spectrum, and as such don’t represent a uniform selection of words found in English-language texts. We merged the 5,000 most common words from 4 corpora (Twitter, Google Books, the New York Times, and music lyrics) and had Mechanical Turk users evaluate their valence in the same way as was done for ANEW, producing a list of ~10,000 words and their associated happiness scores. We’re calling this dataset LabMT-1.0, for Language assessment by Mechanical Turk. Since LabMT words were picked by frequency of usage, they provide much better coverage (i.e. the percent of words identified in a text) than ANEW.

When Jasmin and Casasanto’s paper appeared and achieved the impressive press coverage that it did, it also attracted the scrutiny of other language researchers who weren’t so sure of the significance of the QWERTY effect. A public debate has taken place between Mark Liberman of the Language Log blog and the authors of the study. See post1, post2, the response from J&C, and the response back. After being informed by (our) Peter Dodds of the LabMT data, Liberman made the second post, in which he calculated the RSA of our LabMT words but continued to find no or little QWERTY effect.

As we’ve explored in our hedonometrics papers, the hedonometer can be thought of as a tunable instrument when you remove neutral-valence stop words, effectively increasing the sensitivity of happiness measurements for texts. I wanted to see if tuning $\Delta h_{\text{avg}}$ changed anything. In the process, I also repeated Liberman’s analysis of the LabMT data and am making available the R scripts and data that went into that.

analyze_rsa_labmt.R – script for the analysis and plots
labMT.rsa.txt – Liberman’s computation of RSA for the subset of LabMT-1.0 words containing only alphabetic characters

We haven’t seen any more evidence than Liberman did when looking simply at the relationship between RSA and $h_{\text{avg}}$. If the QWERTY effect is real, then it is exceedingly small, but the above data point to it being indistinguishable from zero.  It’s useful to look at the raw data, binned in both variables.

Raw data binned (RSA spacing of 1, $h_text{avg}$ spacing of 0.1) and plotted.

There is not any obvious, visually distinguishable correlation.

Now, if you take the average happiness of words for each RSA value, you can do a linear regression on that data, weighting each point by the number of words for that RSA value.

Data binned by RSA with the line indicating the linear regression weighted by the number of words for that RSA. Note that this is the same as a linear fit of the unbinned data, but the resulting plot is less cluttered and easier to read.

The trend actually runs in the negative direction, but with a p-value of 0.74, meaning there is no effect. Jasmin and Casasanto controlled for more variables in a different dataset, and independent evaluation of the significance of the correlations they observed, controlling for these other attributes, would be possible if all the original data were released. Sure, the data sources are listed, but it would be a significant effort to recreate the entire set. I’d also be curious to see if similar correlations could be observed in the other affective variables measured in ANEW (arousal and dominance).

Final note: Changing $\Delta h_{\text{avg}}$, our tuning knob, does change the magnitude of the correlation. (Imagine removing a horizontal band from the binned plot above; this changes the correlation.) However, it is still impossible to conclude that the effect is significant. Also, analyzing positive and negative words separately shows opposite trends for $\Delta h_{\text{avg}} = 1$. The code for this is all included in the script above.