As the year 2016 draws to a close, it is the perfect opportunity to look back at what we have accomplished and what lies ahead of us. New Year’s resolutions are a popular way to set ambitious goals. Yet, a new year does not magically bestow us with the willpower and persistence to succeed. This is why such resolutions have garnered a bad reputation in the press lately as if we were set to fail. However, the fate of our resolutions is barely more remarkable than the fate of the many good intentions that we fail to put into practice every day. Regardless of the season, intention is simply a bad predictor of long-term action. No need to heap blame on the New Year.
But why is it so hard to predict what we are actually going to do once we have settled on our goals? I would love to add in some simple-minded neurosciency explanations, you know, the dopamine/serotonin/oxytocin makes you (fail to) do stuff type of appealing narrative. Such neuro buzzwords invoke a sense of deep molecular understanding of the brain, but they can hardly conceal that our daily actions are determined by many unknows. Even massive data on brain function does not really reduce this uncertainty. Without “habits”, it would be impossible to carry out our everyday life nearly as efficient as we do. The cost of intentional control is so substantial that we have developed a sophisticated automation scheme. Thus, breaking the habit is difficult and the most effective way might be to settle for a refined habit. However, it is not only the heavy weight of intentional control on our shoulders that is difficult to handle and mandatory to get rid off.
#1 resolution: Lose weight
According to statisticbrain.com, losing weight tops the 2015 list of New Year’s resolutions. Since I am doing research on the neural mechanisms subserving body weight control, I was wondering how this affects what we are trying to capture in our studies. The answer is more complex than I initially thought it might be. The complications arise because when we measure intentions in the lab, we typically use standardized questionnaires. How well this is in line with more naturalistic measurements of intention in daily life, is difficult to say unless you sample ecologically valid behavior extensively. Furthermore, I am not aware of many attempts to use intention to participate in a study to statistically model the outcome (as detailed in recruitment modeling, for example). Suppose that we are advertising a weight loss trial. It is reasonable to assume that an intention to lose weight would modulate the likelihood that a person signs up for a study. So far, we are mostly ignoring this aspect, but it makes it more difficult to say something about the population. These considerations become even more important when intentions are not fluctuating randomly throughout the year. But is there a proxy for intentions at the level of the population?
The answer to any question: google
Of course, I instantly had to turn to google in search of an answer. At first, I failed to locate good databases on statistics that could provide the much needed data. My second thought was to turn directly to google search traffic as primary source. If someone would like to reconstruct my momentary intentions, my google searches would be a pretty good point to start off. Likewise, google Trends is a great resource for this purpose. There has been a lot of hype in the scientific community about using google Trends data to predict the progression of the flu (not without controversy about the true scientific value of course, but we’ll leave this aside at the moment) so maybe it can speak to the question. I entered “lose weight” as key term and downloaded the data starting from Jan 2004. This is what I got:
There are at least two striking patterns. First, we have a pronounced increase in use of the keywords starting in ~2008. Second, on top of the rise by the end of the last decade, we can spot a pattern within each year across the months. These monthly fluctuations also appear to be more prominent in this decade. How can we test these visual suspicions about the search activity? I subjected the data to a full mixed effects regression analysis. Here, search activity (the outcome) is measured repeatedly and aggregated by month and I assumed that these measures are nested within a given year.
If we plot the data by putting month on the x-axis, we see that the max is observed in January and the min is observed in December. Interestingly, there is a second trend in the data: a local peak in July, perhaps the correlate of the swimsuit season. Once the peak temperatures have passed, interest in losing weight seems to vanish only to be replenished by the optimism entailed in New Year’s eve. Awesome.
Is losing weight only a recent top resolution?
In addition to the seasonal patterns, we can also investigate the annual changes in search activity. To this end, I added separate lines for each year and color coded them such that recent years appear in hotter colors.
In this plot, it looks as if the decrease in interest throughout the year is a rather recent phenomenon. To explicate this question, I entered linear, quadratic, and cubic predictors for month and looked at the freely fitted coefficients for each year. We get significant linear (p = .012) and quadratic components (p = < .001) in search activity, but the cubic component is not significant (p = .38). Still, there was some indication of significant variability across the years, which is why I left the cubic component in the model (caution is warranted though because we have a low number of years).
When we inspect the fitted coefficients, it is evident that the decrease in search activity within one year became stronger recently (r = -.68). Similarly, the quadratic component became more negative (r = -.85), which basically indicates that the swimsuit season had a more pronounced effect on search activity in recent years. Lastly, we can look at the performance of our model of the data:
In this plot, I have mapped the multivariate (Mahalanobis) distance onto the dot size and the natural logarithm of the variance in the residuals within each year onto the y-axis. In other words, the bigger the dots, the more unusual was search activity compared to the other years and the higher the dot is located on the y-axis, the stronger were the monthly fluctuations in search activity even after accounting for linear to cubic trends in the search data. What we can spot here is the cluster of atypical years in the middel indicating a sharp transition in search activity that has stabilized in recent years again. Interestingly, the new plateau also coincides with the more pronounced linear and quadratic effects in search trends.
The plot thickens
Google trends indeed provided some interesting insights that back up the initial suspicions. We can spot a linear decline in search activity that is congruent with the report of a common New Year’s resolution. Moreover, we might even speculate that this is a rather recent phenomenon added on top of the overall interest in losing weight. At this stage, this is still not more than a data analysis exercise, but it was enough to convince myself that google search activity could be a useful proxy of intentions at the level of the population. Arguably, any inference based on one group of keywords with just a few years of observations is limited, in particular when it comes to annual trends. At the same time, I suspect that such data provides unique insights complementary to classic intention questionnaires. My New Year’s resolution is to pay more attention to the rich information that is now available at an unprecedented scale in order to narrow the gap between intentions and implementations of actions. We need to pay more attention to how we draw samples in neuroscience and how we can connect the information that we gather in the lab to what is out there already. There is good reason for excitement about the future of human neuroscience once we get to integrate real life behavior in a much more comprehensive way than today.