Wellbeing I: Introduction

2024-03-22T00:00:00+00:00

This is a variation of the famous trolley problem</a>, a thought experiment in ethics and moral psychology. The problem is designed to highlight the tension between two moral principles: the principle of utilitarianism, which states that the right action is the one that maximizes overall wellbeing, and the principle of deontology, which states that the right action is the one that respects individual rights and duties. In the trolley problem, the utilitarian answer is to kill the healthy person to save the five patients, while the deontological answer is to not kill the healthy person, as doing so would violate their right to life.
I am not a utilitarian. However, philosophies of justice (like liberalism) often become unhelpful in scenarios where wellbeing trade-offs are inevitable. For instance, consider now that the surgeon has healthy organs available, but only enough to save three of the five patients. What should the surgeon do in that case?
In the real world, policymakers often face decisions that entail such compromises in one way or another: for example, should we invest in a new hospital, or in a new school? Should a government implement regulations to protect consumers, which could increase the cost of goods and services and potentially stifle innovation? Should a city invest in improving public transportation, potentially reducing traffic and pollution, but at the cost of higher taxes? These are all questions that involve trade-offs between different individuals' wellbeing, and more often than not, they are not easy to answer.
One of the ultimate goals of policy should be to maximize wellbeing within constraints for the individuals that it serves to. Philosophies of justice can help us delimit those constraints, but they are not enough to guide us on how to identify wellbeing maximizing interventions and balance wellbeing trade-offs.
And this is where wellbeing research comes in.
This is the first post of a series about wellbeing measurement and policy.
Definitions 1</a> A variant of the former approach is to ask respondents to rate their life on a ladder, where the top of the ladder represents the best possible life for them, and the bottom of the ladder represents the worst possible life for them. This is commonly known as the Cantril ladder, which is used in many international household surveys such as the Gallup World Poll</a>.</li>
Hedonic approaches focus on measuring how people feel at a particular point in time. This is typically done by asking respondents to report their feelings at the moment of the survey, or to recall their feelings over a specific period of time. Experienced happiness is the most common measure of hedonic wellbeing, but negative emotions such as sadness, anger, and stress are also used.</li>
Eudaimonic measures are based on the Aristotelian tradition with the idea that wellbeing is not just about feeling good, but also about functioning well. This approach focuses on the extent to which people are able to realize their potential, and to live a life that is meaningful and fulfilling. This is typically measured by asking respondents about their sense of purpose, their autonomy, and their relationships with others.</li> </ul>
Tipically, evaluative and hedonic measures are the most widely used in policy research. Eudaimonic measures, although philosophically appealing, are harder to interpret – for example, when we talk about being virtuous, it is not clear whether that should be an outcome in itself or another determinant (among many) of wellbeing.
Figure 1</a> shows cross-country averages of life satisfaction, as measured by the Cantril ladder question from the Gallup World Poll. The figure shows that there is a wide variation in life satisfaction across countries, with the highest levels of life satisfaction being found in the Nordic countries, and the lowest levels of life satisfaction being found in sub-Saharan Africa and the Middle East. The difference between the country with the highest average life satisfaction (Finland) and the country with the lowest average life satisfaction (Afghanistan) is about 6 points on the 0-10 Cantril ladder scale.

Figure 1: Life satisfaction across countries </figcaption> </iframe> </figure>

It is (not) all about the money #</a></h2>
higher life expectancy</a>, better education systems</a>, well-functioning labor markets</a>, political stability, absence of violence...
...and wellbeing itself. Across countries, higher GDP per capita is associated with greater levels of life satisfaction. This finding is remarkably consistent across surveys, and it holds up irrespective of the definition of wellbeing we use. Figure 2</a> shows the relationship between GDP per capita and life satisfaction for a large number of countries, using data from the Gallup World Poll. High-income countries tend to have higher levels of life satisfaction than low-income countries, and the relationship is roughly log-linear. The latter is consistent with the idea of diminishing marginal utility of income: an extra dollar has more impact on wellbeing for the average Joe than for Jeff Bezos. Intuitive, right?

Figure 2: Self-reported life satisfaction vs. GDP per capita </figcaption> </iframe> </figure>

What is the catch then? Why not just stick to income as a proxy for wellbeing?
The Easterlin paradox
Figure 3</a>, which shows the evolution of GDP per capita and average life satisfaction (derived from large-scale longitudinal household surveys) for a sample of advanced economies. The figure shows that, despite considerable increases in within-country per capita income, the average level of life satisfaction has remained surprisingly constant over time. In the U.S., for example, GDP per capita has increased by more than 500% since 1950, but people are not any happier now than they were back then.

Figure 3: The Easterlin Paradox</figcaption> </iframe> </figure>
Source: Clark et al. (2018).</a>

In light of the above, the answer to the question on why should we not rely solely on income as a proxy for wellbeing is straightforward: because it is simply not a good standalone proxy in the long-run.
The social comparison hypothesis
Card et al. (2012)</a> run an experiment in which they randomly inform employees of the University of California about the salaries of their peers, and find that workers whose wage is below the median for their comparison group report lower job satisfaction and are more likely to start looking for a new job. Perez-Truglia (2020)</a> exploits a natural experiment whereby Norwegian tax records became publicly available for everyone to see, and shows that income transparency increases the gap in happiness between high- and low-income individuals by 29 percent. This occurs because people at the bottom (top) of the income distribution typically overestimate (underestimate) their relative position, and increased transparency makes them realize how much worse (better) off they are compared to others.
In a less clean way, one could test for the social comparison hypothesis using a very simple regression framework in which we regress wellbeing on both individual income and the average income of a comparison group:
$$Y_i = \alpha + \beta X_i + \gamma \bar{X} + \epsilon_i$$
where $Y_i$ is individual $i$'s wellbeing, $X_i$ is individual $i$'s income, $\bar{X}$ is the average income of a group of comparators, and $\epsilon_i$ is the error term. If the relative income hypothesis holds, we would expect $\gamma$ to be negative, as individual wellbeing would decrease as the average income of the group of comparators increases.
This is exactly what we find. Table 1</a> shows the results of such a regression for cross-sections of different countries, in which the group of comparators is defined by people of the same sex, age group, region, and survey year, and the outcome is life satisfaction (measured on a 0-10 scale). Individual wellbeing is positively associated with individual income, but negatively associated with the average income of the comparison group. Except for the United States, a 1% increase in the average income of comparators reduces wellbeing in the same amount or more than a 1% increase in individual income increases wellbeing.

Table 1: The social comparison hypothesis</figcaption>
</th> Britain</th> Germany</th> Australia</th> United States</th></tr></thead>

Own income</td> 0.16 (.01)</td> 0.26 (.01)</td> 0.16 (.01)</td> 0.31 (.01)</td></tr>
Comparator income</td> -0.23 (.07)</td> -0.25 (.04)</td> -0.17 (.06)</td> -0.19 (.03)</td></tr> </tbody></table>
Source: Layard and De Neve (2023)</a>, Table 13.4. Standard errors in parentheses. </div>

There are other potential explanations for the Easterlin paradox, such as the hedonic adaptation hypothesis. Roughly speaking, the idea here is that people adapt to their new income levels and their expectations increase (or decrease) accordingly. We will not discuss it in detail here, but there is equally interesting evidence on the role of adaptation in explaining the paradox.2</a>
Effect sizes
Lindqvist et al. (2020)</a> look at the long-run effects of lottery wealth on psychological wellbeing in Sweden, and find that an after-tax lottery prize of $100,000 increases overall life satisfaction by 0.037 standard deviations relative to the matched control group. To put it into perspective, that is about 70% of the difference in average life satisfaction between Denmark (the second happiest country in the world) and Iceland (the third happiest country in the world). In other words, this is a very small effect. Relatedly, the authors also explore the impact of lottery wealth on other outcomes, such as happiness and mental health, and in these cases the effects are not significantly different from zero. That is, lottery winners are not happier or less mentally distressed relative to the control group in the long-run.
As we saw in Figure 2</a>, the effects of income on wellbeing are also non-linear. You probably have come across media articles arguing that, beyond a certain income threshold, the marginal effect of income on wellbeing is virtually zero. This is pretty much based on the (now sort of old) paper by Kahneman and Deaton (2010)</a>, who use data for the U.S. and find that the relationship between income and emotional wellbeing is log-linear up to an annual income of $75,000, but flattens out beyond that point. The idea is that further increases in income do not necessarily improve the individual's ability to engage in activities that are most important for their emotional wellbeing, such as spending time with friends and family, or having better health.
However, a key thing to note is that here we are looking at hedonic measures of wellbeing. Specifically, they use the average of three dicotomous, binary measures of positive affect (happiness, enjoyment, and frequent smiling and laughter) and the average of two binary measures of negative affect or "blue" states (worry and sadness). In contrast, the authors find that the relationship between income and evaluative wellbeing (i.e., what we have been referring to as life satisfaction or general happiness) shows no sign of satiation, at least to an amount over $160,000 per year.
Things got interesting when Killingsworth (2021)</a> revisited this question and found that, conversely, hedonic wellbeing does rise with income well above the $75,000 per year threshold. The main argument of this paper is that the discrepancy with the original results might be explained by differences in the variable definitions: on top of just using binary variables (which are able to capture less variation), Kahneman and Deaton (2010)</a> results are based on retrospective survey responses, which are subject to recall bias. Instead, Killingsworth (2021)</a> uses large-scale real-time data collected through a smartphone app and a continuous scale to measure experienced wellbeing, finding that the latter increases linearly with log-income with no satiation point.
In a recent paper, the authors joined forces and reconciled the two previous findings (Killingsworth et al., 2023</a>). Here, they show that the flattening pattern in the relationship between income and experienced wellbeing depicted in the original study by Kahneman and Deaton does exist, but it only applies to a limited subset of the population. Using quantile regressions, they show that increases in income do not shift upwards the bottom 15-20% of the emotional wellbeing distribution (i.e., the least happy), but they do so across all other quantiles.
To illustrate this, see Figure 4</a>. The happiness-income gradient is positive and significant for the top 80% of the distribution, but the slope turns flat for the those below the 15th percentile after $120,000 per year.

Figure 4: Emotional wellbeing across the happiness distribution </figcaption> </iframe> </figure>
Source: Killingsworth et al. (2023).</a>

So, why Kahneman and Deaton (2010)</a> claimed that the satiation pattern applies to the entire population? As mentioned above, the answer is in the variable definition. Since they used dichotomous measures, the authors ended up with a very skewed distribution of happiness, especially among higher incomes, suggesting a "ceiling effect" whereby the measure could not capture differences between higher degrees of happiness. In fact, the average reported positive affect for the range of high incomes in Kahneman and Deaton (2010)</a> is 89% of a perfect score, but we know that the happy people are not all equally happy!
To see this, consider the example of a pass-or-fail math test, consisting of questions that most students can answer correctly. This test could be informative to identify the students who are doing badly, but it would not be able to differentiate between the students who are doing well because most of them would get the same pass grade. In a sense, the measures of affect in Kahneman and Deaton (2010)</a> are just like this test: they are good at identifying the people who are unhappy, but they have limited ability to discriminate among levels of happiness. Instead of "Happiness increases with income, but it does not further increase above $75,000 per year", it follows that the statement should have read "Unhappiness decreases with income, but it does not further decrease above $75,000 per year".
What's next?^2
For instance, see Di Tella et al. (2010)</a> and Kuhn et al. (2011)</a>. </div>

References #</a></h1>
Seth Stephens-Davidowitz</a>, former data scientist at Google and PhD in economics from Harvard. Everybody lies</a> is an incredibly fun read offering a compelling argument: that our online searches can reveal more about our true selves than we might like to admit. The book is not only a testament to the power of non-traditional data sources, but also an invitation to explore human behaviour through a different lens.
Stephens-Davidowitz's work shows that Google searches can be used to predict a wide range of socio-economic outcomes when conventional data sources are not available or are contaminated by social desirability bias. The latter occurs when people are reluctant to disclose their true preferences or beliefs in surveys, which is particularly important when studying sensitive topics such as racial animus, sexual orientation, or mental health. The idea is quite simple: if revealing what you really think leads to social stigma, then you would be more likely to confess to Google than to a survey interviewer. In a sense, the unfiltered truth of humanity lies in Google's autocomplete bar.
In one of his academic papers</a> published in the Journal of Public Economics, Stephens-Davidowitz uses Google searches to measure racial animus in the U.S. and its impact on the 2008 and 2012 presidential elections – specifically, he finds that racially charged Google search rates are a strong negative predictor of Obama's vote share. The point estimates in the preferred specification suggest that racial animus cost Obama around 4 percentage points in the 2008 election, which is significantly above previous estimates based on survey data.

Table 1: Google search terms and socio-economic outcomes </figcaption>
Term</th> Underlying variable</th> t-Stat</th> $R^2$</th></tr></thead>

God</td> Percent believe in god</td> 8.45</td> 0.65</td></tr>
Gun</td> Percent own gun</td> 8.94</td> 0.62</td></tr>
African American(s)</td> Percent Black</td> 13.15</td> 0.78</td></tr>
Hispanic</td> Percent Hispanic</td> 8.71</td> 0.61</td></tr>
Jewish</td> Percent Jewish</td> 17.08</td> 0.86</td></tr> </tbody></table>
Source: Stephens-Davidowitz (2015)</a>, Table 1. The t-stat and $R^2$ are from a regression with the normalized search volume of the word(s) in the first column as the independent variable and measures of the value in the second column as the dependent variable. The normalized search volume for all terms is from 2004 to 2007. All data are at the state level. </div>

But the potential of Google searches goes beyond measuring racial animus. In Table 1</a>, I show some of Stephen-Davidowitz's examples on how Google searches can be exploited to predict various population-level outcomes. For example, Google searches for the word "Gun" explain 62% of the variation in states' gun ownership rates, and Google searches for the word "Hispanic" explain 61% of the variation in the percentage of Hispanic population across U.S. states.
Can we do better? In this post, I will show you how we can leverage the popularity of Latin urban music artists in Google searches to predict state-level Hispanic population shares in the U.S.
Some background on Google Trends data #</a></h1>
Google Trends</a>, a tool that allows users to explore the popularity of search terms over time and across different regions within countries. The data is normalized to the time and location of a query, so that the popularity of a search term is measured relative to the total number of searches in a given region and time period. In other words, Google Trends gives measures of relative search term popularity, not an absolute measure of the number of searches for a given term.[^1] Relative popularity is measured on a scale from 0 to 100, where 100 is the most popular point in time or location for a given search term. For example, value of 50 means that the term is at half its peak observed popularity.
There are a bazillion online resources on how to use Google Trends, so I won't go into the details here. As a starting point, I would recommend you to check out the Google Trends basics page</a> and the search tips page</a>.
Predicting Hispanic population shares with Google Trends #</a></h1>
American Community Survey (ACS)</a> microdata for the year 2022.[^2]
First, we will use the search term "bad bunny" to predict Hispanic population shares. Bad Bunny</a> is arguably the most popular Latin artist worldwide nowadays – in 2022, he was Billboard's top artist of the year and the most streamed artist</a> in Spotify and Apple Music. He also has a huge following</a> on Instagram with over 40 million followers, and its popularity has played a crucial role in the reivindication of the Latino culture and identity in the U.S. (ahora todos quieren ser latinos</a>).
Let's do it. To start with, I show the Hispanic population shares across U.S. states in 2022 in Figure 1</a>. Unsurprisingly, the Hispanic population is concentrated in the South and West regions of the U.S. The states with the highest Hispanic population shares are New Mexico, California, and Texas, with shares of 50%, 40.3%, and 40.2%, respectively. The states with the lowest Hispanic population shares are West Virginia, Maine, and Vermont, with shares of around 2% each.

Figure 1</figcaption> </iframe> </figure>

Now, let's do the same for the popularity of the search term "bad bunny" in the U.S. in 2022:

Figure 2</figcaption> </iframe> </figure>

Similar, right?
We can visualize the relationship between the popularity of the search term "bad bunny" and the Hispanic population shares across U.S. states in a scatter plot (see Figure 3</a>). By running a simple OLS regression of the Hispanic population share on the popularity of the search term, we get that 76% of the variation in the Hispanic population share across U.S. states can be explained by the relative popularity of Bad Bunny in Google searches. This implies a 15 percentage point increase in the share of variation explained relative to just using the term "Hispanic" as a predictor.

Figure 3</figcaption> </iframe> </figure>

There is one state that is particularly off path: New Mexico (if we were to omit it, then the share of variation explained would go up to 89%). New Mexico has a Hispanic population share of 50%, but the popularity of Bad Bunny in Google searches is only 56% of the most popular point location for the term.
One particularity of New Mexico is that it has the highest population share with Mexican origin in the U.S. What if we focused on the popularity of a Mexican artist, and looked at Mexican population shares instead? I got curious about this, so I replicated the analysis exploiting the popularity of the search term "peso pluma". Peso Pluma</a> is a Mexican artist whose popularity has skyrocketed during the last couple of years, being Youtube's most viewed artist of 2023 in the U.S.</a>
The results are shown in Figure 4</a>. The fit is breathtaking: Peso Pluma's popularity in Google searches explains 94% of the variation in the Mexican population share across U.S. states.

Figure 4</figcaption> </iframe> </figure>

Measuring what matters
In Table 2</a>, I illustrate the comparison between how state-level suicide rates in the U.S. correlate with both Google search terms and self-reported measures of suicide risk. While Google searches are not perfect, they prove to be more reliable than the traditional survey-derived data, which yields statistically insignificant correlations.

Table 2: Correlations between Google search terms and suicide rates in the U.S. </figcaption>
Predictor</th> $r$</th> $p$</th></tr></thead>

Google search term</td> </td> </td></tr>
Suicide</td> -.06</td> .63</td></tr>
How to suicide</td> .49</td> < .001</td></tr>
How to kill yourself</td> .61</td> < .001</td></tr>
How to commit suicide</td> .63</td> < .001</td></tr>
Painless suicide</td> .59</td> < .001</td></tr>
Composite of all four suicidal terms</td> .67</td> < .001</td></tr>
Self-report measure</td> </td> </td></tr>
Suicidal thoughts</td> -.17</td> .22</td></tr>
Suicidal plans</td> -.14</td> .30</td></tr>
Suicidal attempts</td> -.03</td> .82</td></tr>
Composite of self-reported suicidality</td> -.16</td> .25</td></tr> </tbody></table>
Source: Ma-Kellams et al. (2016)</a>, Table 1. Data for 2008 and 2009. The column $r$ shows the correlation between suicidal death rates and the relevant Google search term at the state level. Self-reported measures are sourced from the National Survey on Drug Use and Health. </div>

We can also use Google searches to track economic activity in real time. For example, the OECD has developed a tool</a> that uses Google searches to nowcast GDP growth on a weekly basis. The algorithm gathers data related to various realms of economic activity, using search terms related to consumption (like "vehicles" and "household appliances"), labour market conditions (such as "unemployment benefits"), the housing sector (like "real estate agencies" and "mortgages"), business-related services, industrial activity, trade, and economic sentiment, among others. So, next time you are casually checking (again) whether rent prices are going up in your area, you might be contributing to a fancy economic forecast.
Wrapping up
here</a>). Could we predict vote shares for controversial candidates at scale with Google searches?</li>
Emergency situations and public safety: during emergencies, real-time data from Google Trends can provide insights into public sentiment or areas of concern, which might be faster or more accurate other methods. For instance, in the context of natural disasters, spikes in searches for emergency services or relief efforts can indicate the most affected areas or the public’s most urgent needs.</li> </ul>
I will leave it here. If you have worked with Google Trends data before and have some cool applications in mind you would like to share, please feel free to get in touch!

[^1] However, with some assumptions and creativity, it is possible to come up with proxies for absolute Google search volumnes. See Appendix C in Stephens-Davidowitz (2015)</a> for more details.
[^2] I identify people with Hispanic/Latino origin in the ACS using the variable HISPAN</code>. More details can be found here</a>.
[^3] One further caveat is that the Google Trends metro areas are custom aggregates and do not correspond to standard metropolitan statistical areas (MSAs) or metropolitan divisions. Due to this, there has been some attempts to map Google Trends metro areas to more granular geographic units, such as counties in the U.S</a>.
References#</a></h1> Stephens-Davidowitz, S. (2017). Everybody lies: Big data, new data, and what the internet can tell us about who we really are. Bloomsbury Publishing.</li> Stephens-Davidowitz, S. (2015). The cost of racial animus on a black candidate: Evidence using Google search data. Journal of Public Economics, 118, 26-40.</li> Ma-Kellams, C., Or, F., Baek, J. H., & Kawachi, I. (2016). Rethinking suicide surveillance: Google search data and self-reported suicidality differentially estimate completed suicide risk. Clinical Psychological Science, 4(3), 480-484.</li> </ul> A modern IDE for Stata with VSCode + Quarto 2024-02-27T00:00:00+00:00 Update Aug 2024 I only used this setup for a couple of weeks before switching to a non-interactive integration that allows for running .do</code> files within VSCode. The lack of a data browser and the variable explorer in the Jupyter kernel was a dealbreaker for me (plus I have grown to find interactive notebooks a bit annoying except for very specific tasks). To set this up, follow the instructions here</a> under the "way one" section. Moreover, I have increasingly found myself using R for most of my work over the last year, and I have to stay that I no longer agree with some of the points I made in this post. Most of the time, the so-claimed "superiority" of Stata for reduced-form empirical work is just a skill issue. My advice would be to invest time in mastering R or Python, and then decide whether you really need to use Stata (you probably won't). </div> </blockquote> Over the last few decades, applied microeconomic research has witnessed what is widely termed as the credibility revolution</a>, a paradigm shift that emphasized the importance of experimental and quasi-experimental methods to credibly identify causal relationships between variables. This transformative period has not only influenced the methodologies employed in academic papers but also strengthened the field's impact on policy-making, by providing more solid and rigorous evidence upon which to base decisions. This emphasis on methodologically robust empirical approaches has resulted in bigger, more complex projects, naturally elevating the importance of data skills within the field. However, despite the increasing importance of coding in economics, many economists lag in adopting modern coding practices and development environments. One of the most widely used tools in the profession, Stata</a>, is not even a programming language but a purpose-built statistical package.1</a> Stata is also commercial software, meaning that it does not have the benefits of open-source languages like Python or R, which have amassed extensive communities contributing to their development and offering support. Due to its nature, it has traditionally been hard to integrate Stata with modern IDEs, resulting in a less efficient workflow when one has to write code in multiple languages or use the latest tools for version control, code completion, and so on. This post provides a quick tutorial on setting up a modern development environment for Stata with Visual Studio Code and Quarto. It is by no means perfect and it comes with its own caveats, but it is a step in the right direction for those who want to modernize a bit their workflow without having to give up on Stata. Why Stata#</a></h1> I know what you are thinking. jUsT dOn't usE sTAta BrO. I get it. I code in multiple languages and it is indeed not a good idea to only use Stata, simply because there are lots of cool things (web scraping, geospatial stuff, ML) that you would miss out on by doing so. However, the truth is that if you do applied microeconomics for a living, it would be very stupid not to use Stata sometimes. Whoever tells you otherwise is i) not an applied microeconomist (good for them!) or ii) underestimating how important Stata is in the profession: Stata can be found in more than 70% of replication packages</a> for published papers in economics. It is followed by Matlab (which we deliberately ignore since it is aimed at structural work), and R comes in third with less than 10%. I am not going to argue whether this is a good or a bad equilibrium but it is an equilibrium, and if you want collaborate with others and be able to replicate their work, you will have to use Stata </li> For reduced-form econometrics, Stata is unparalleled. Period. I have been writing code for +3.5 years and it still amazes me how ridiculously cumbersome R and Python can be relative to Stata when it comes to running regressions. For some specifications, it is simply not possible to use something else unless you want to code the estimator from scratch (good luck running modern difference-in-differences estimators for heterogeneous treatment effects in Python, for example) </li> Economists love regression tables. And again, the functionalities and flexibility of esttab</code> to export fully formatted and reproducible LaTeX tables in Stata are light years ahead of the alternatives in other languages (e.g., stargazer</code> in R) </li> For most people doing applied micro research or policy work, coding will hardly ever be their comparative advantage. Instead, your value lies in the quality of your ideas and how well you execute them – coding is only useful insofar as it helps in that purpose. Some folks (myself included) enjoy venturing off the beaten path, but most researchers would prioritize research productivity if the marginal cost of learning new data skills is sufficiently high. In this sense, Stata has a quite shallow and gentle learning curve, so usually people learn it as their first "language" and then just stick to it. I am not saying this is good (I think is bad!), but in the end people respond to incentives </li> </ul> Why VSCode#</a></h1> VSCode is Microsoft's open-source code editor. It supports a wide array of programming languages and comes with the usual IDE features like syntax highlighting, code completion, Copilot</a>, and version control integration.2</a> More importantly, turns out that it is fairly straightforward to integrate VSCode with Stata via plug-ins. VSCode also has nice LaTeX extensions, which in my opinion makes it the only real alternative to Overleaf. This means that you can seamlessly integrate all your workflow into a single code editor, from data cleaning to drafting. Why Quarto#</a></h1> Quarto is an open-source scientific and technical publishing system that allows users to create dynamic documents, slide decks, and reports in various formats from plain text files. Quarto files support multiple programming languages and can be rendered into interactive HTML documents, which allows the user to combine code execution with high-quality document production. The main advantage of using a notebook-based interface relative to the native Stata editor is that it allows you to create and present exploratory analysis combining code, output, and text (including comments, LaTeX equations, and more), all in one convenient place. And last but not least, Quarto outputs are beautiful</a>. Setup#</a></h1> Step 1: Install Quarto#</a></h2> Go to https://quarto.org/</a>, click on "Get Started", download the latest Quarto version available, and install it on your machine. Step 2: Install VSCode#</a></h2> Go to https://code.visualstudio.com/download</a>, download the latest VSCode version available, and install it on your machine. Step 3: Install VSCode extensions#</a></h2> You will have to install two VSCode extensions: To get nice Stata syntax highlighting and code autocompletion, you will have to install the Stata Enhanced</a> package. Open VSCode, go to the extensions tab (highlighted in red in the screenshot), search for the extension, and install it.</li> For integrated render and preview for Quarto documents, repeat the steps above and install the Quarto</a> extension.</li> </ol> Step 4: Install nbstata</code>#</a></h2> nbstata</code></a> is a Jupyter kernel for Stata, built on top of the functionalities of the pystata</code> Python package introduced in Stata version 17. This means that you will need Stata 17 or higher to install it.3</a> You can find the full user and installation guide here</a>. For the curious, I encourage you to go through the author's guide. For the lazy ones, the short version goes as follows: To install nbstata</code>, you need an installation of Python 3.7 or higher. For first-time Python users, I would follow the suggestion from the author of nbstata</code> and just install the Anaconda distribution</a>. If you want a lighter, minimalistic version, you can go for Miniconda</a>.4</a></li> Once you have installed Anaconda, open the Anaconda prompt and type</li> </ol> pip install nbstata</code></pre> Finally, install nbstata</code> by running</li> </ol> python -m nbstata.install</code></pre> And that is it! Your machine should be now ready to render Quarto Markdown (.qmd</code>) files in VSCode using the nbstata</code> kernel. An example#</a></h2> To begin, create a .qmd</code> file. At the very start of this file, incorporate a YAML header to specify the Jupyter kernel you wish to use. In our case, to use nbstata</code> in your Quarto document, add jupyter: nbstata</code> within the YAML header. The YAML header is a section located at the top of Markdown files, designed to hold metadata such as configuration settings and document properties. This metadata is structured in a readable format and is essential for defining various aspects of your document's processing and presentation. A sample of YAML header could look as follows: --- title: "Some dataviz with Stata and Quarto" date: 2024-24-01 author: "Pablo" jupyter: nbstata format: html: self-contained: true toc: true toc-depth: 3 toc-float: collapsed: false smooth-scroll: true width: 20% ---</code></pre> You can now create code chunks as you would do in any other notebook-based interface, specifying stata</code> in your code chunk header. For our example, we will create a chart to visualize gender gaps in suicide rates across low-income countries. To load the dataset into our environment, we will use the Stata wrapper for the World Bank's WDI database API, wbopendata</code></a>: ```{stata load-data} wbopendata, indicator(SH.STA.SUIC.FE.P5; SH.STA.SUIC.MA.P5) clear long desc ```</code></pre>Metadata for indicator SH.STA.SUIC.FE.P5 ------------------------------------------------------------------------------- Name: Suicide mortality rate, female (per 100,000 female population) ------------------------------------------------------------------------------- Collection: 2 World Development Indicators ------------------------------------------------------------------------------- Description: Suicide mortality rate is the number of suicide deaths in a year per 100,000 population. Crude suicide rate (not age-adjusted). ------------------------------------------------------------------------------- Note: World Health Organization, Global Health Observatory Data Repository (http: apps.who.int ghodata ). ------------------------------------------------------------------------------- Topic(s): 8 Health ------------------------------------------------------------------------------- Metadata for indicator SH.STA.SUIC.MA.P5 ------------------------------------------------------------------------------- Name: Suicide mortality rate, male (per 100,000 male population) ------------------------------------------------------------------------------- Collection: 2 World Development Indicators ------------------------------------------------------------------------------- Description: Suicide mortality rate is the number of suicide deaths in ... sh_sta_suic_m~5 float %8.0g SH.STA.SUIC.MA.P5 ------------------------------------------------------------------------------- Sorted by: countrycode year Note: Dataset has changed since last saved. Output is truncated. View as a scrollable element or open in a text editor. Adjust cell output settings...</code></pre> Let's create a simple dumbbell chart by sex: ```{stata gen-chart} * Prepare data cap ren sh_sta_suic_fe_p5 fem_rate cap ren sh_sta_suic_ma_p5 male_rate keep if !inlist(.,fem_rate,male_rate) // keep non-missing pairs of female and male suicide rates bys countrycode (year): keep if _n == _N // keep latest country-year obs. drop if region == "" // drop regional aggregates keep if incomelevelname == "Low income" // keep low-income countries * Chart settings distinct countrycode local ub = `r(ndistinct)' // N of categories for the x-axis sort male_rate cap g order = _n // order of x-axis categories labmask order, val(countryname) su male_rate, d local x_max = round(r(max),0.1) // max value for y-axis * Export chart #d; twoway (rspike male_rate fem_rate order, lcolor(gs14%90)) (scatter male_rate order, mcolor(midblue)) (scatter fem_rate order, mcolor(pink)), ytitle("Suicide mortality rate (per 100,000 population)") xtitle("") ylabel(0(5)`x_max') xlabel(1(1)`ub', angle(45) labsize(vsmall) valuelabel) legend(order(2 "Males" 3 "Females") pos(11) row(1) size(3)) title(Gender gaps in suicide rates across low-income countries, pos(11) size(medium)); #d cr ```</code></pre> Caveats#</a></h1> Sadly, the integration of Stata with VSCode and Quarto is far from perfect, specially if you are used to the native Stata editor and its functionalities. Here are a couple of issues that bother me: You will be missing the data browser and the variable explorer. This is a big one. The Stata data browser is also incredibly smooth and fast compared to the alternatives in R and Python, so it is a shame to lose one of its comparative advantages. Although nbstata</code> comes with Jupyter magics</a> that allow for browsing, such as %browse</code> for the main dataset and %frbrowse</code> for frames, so far I have not figured out a way to expand the cell output in VSCode (nor in Jupyter itself), so this is not useful. If you know of a solution, please reach out! </li> You will not be able to change the font for your charts within VSCode. Not a big deal, you can always change it by opening Stata itself and changing the defaults with the graph set</code> commands, but it is a bit annoying </li> </ul> Troubleshooting#</a></h1> In case you are finding problems with your setup, you should verify your installations and ensure that the Jupyter kernel is correctly set up and recognnized by your system. To check if Quarto is installed and determine its version, open your command prompt (or terminal on macOS and Linux) and enter the following command: quarto check</code></pre> If Quarto is not found, ensure it is installed and that its installation directory is added to your system's PATH environment variable. Reinstalling may also help. To verify that Python is installed on your system and to check its version, run: python --version</code></pre> To check your Python installation directory, run where python</code></pre> This command should return the version of Python installed on your system. If Python is not installed or if there are issues with the installation, this command may return an error or indicate that Python is not recognized. Similar to Quarto, ensure Python's installation directory is in your PATH. If you have multiple Python versions installed, ensure the correct version's path is prioritized. To check which Jupyter kernels are installed on your system, use: jupyter kernelspec list</code></pre> This command will list all available kernels, including their names and installation paths. If you've installed a kernel (such as nbstata</code>) but it doesn't appear in this list, there may be an issue with how the kernel was installed or registered with Jupyter. Common issue: Quarto does not recognize the Python installation#</a></h2> If you have problems with Quarto finding the correct Python installation, you can try the following steps (here I have assumed you have installed Anaconda, but the same logic applies to Miniconda or any other Python distribution): Disable app execution aliases for Python: Open Settings, go to Windows Settings > Apps > Apps & features, and manage execution aliases to disable Python aliases (i.e., turn off the toggle(s) that point to the Microsoft Store app, python.exe</code> and python3.exe</code>). By disabling these, you prevent Windows from redirecting Python commands to the Microsoft Store app, allowing your system to use the Anaconda Python executable instead.</li> Adjusting the PATH order: go to System Properties > Advanced > Environment Variables, then select the "Path" variable under User variables or System variables and click "Edit". In the Edit Environment Variable window, find the entry for C:\Users\[your-username]\AppData\Local\Microsoft\WindowsApps</code>. You can either remove this entry or move it down the list so that the paths to your Anaconda installation appear above it. This changes the order in which Windows searches these paths for executable files. To check the order of your PATH environment variable, type where python</code> in the Anaconda prompt (usually, it should be installed in C:\ProgramData\anaconda3</code>). Then, add this path to the PATH environment variable and move it to the top of the list, adding also the Scripts</code> folder (usually, C:\ProgramData\anaconda3\Scripts</code>).</li> </ol> ^{1 Stata though has its own matrix-based programming language, Mata. Some of the most popular user-written Stata packages, like reghdfe</code> for regressions with high dimensional fixed effects, are largely written in Mata. </div>}^{2 For open-source languages, like R or Python, Copilot is a game-changer. For Stata, it is less useful, although I have to say that I am pleasantly surprised with the quality of its suggestions relative to ChatGPT 4. </div>}^3For earlier Stata versions, an alternative would be using the stata_kernel</code> Jupyter kernel. I have not tested it, but I have good references from colleagues. You can find the installation guide here</a>. </div> ^{4 nbstata</code> is a Jupyter kernel. If you install Anaconda, you will also get JupyterLab installed, so you will be able to run Stata code from Jupyter notebooks as well. I personally hate them, but if you don't, it is an alternative. </div>}

Population group</th>	Change equation</th>	Newborns</th></tr></thead>
Native-born with foreign-born mothers (NBF)</td>	$NBF(a+1,t+1) = NBF(a,t) \times s(a,t) + M_{NBF}(a,t)$</td>	$NBF(0,t+1) = B(t) \times s_{infant}(p,t) \times \Phi_{FB}(t) + M_{NBF}(0,t)$</td></tr>
Native-born with native-born mothers (NBNM)</td>	$NBNM(a+1,t+1) = NBNM(a,t) \times s(a,t) + M_{NBNM}(a,t)$</td>	$NBNM(0,t+1) = B(t) \times s_{infant}(p,t) \times (1 - \Phi_{FB}(t)) + M_{NBNM}(0,t)$</td></tr>
Foreign-born population (FB)</td>	$FB(a+1,t+1) = FB(a,t) \times s(a,t) + M_{FB}(a,t)$</td>	$FB(0,t+1) = M_{FB}(0,t)$</td></tr> </tbody></table> In the table above, $\Phi_{FB}(t)$ is the share of births to foreign-born mothers in year $t$, and $M_{NBF}(a,t)$, $M_{NBNM}(a,t)$, and $M_{FB}(a,t)$ are the migration flows for each population group.</p> Notice that $s_{infant}(p,t) \neq s(p,0,t)$. The intuition is the following. Demographic projections typically reference population stocks as of January 1st of each year, which means that the age 0 population at time $t+1$ consists of children born during year $t$ who survived to January 1st. Since births occur throughout year $t$, newborns are exposed to mortality risk for only part of the year on average. To account for this partial-year exposure, I use an adjusted infant survival probability that differs from standard age-specific survival rates:</p> $$s_{infant}(p,t) = 1 - 0.5 \times m(0,p,t)$$</p> where $m(0,p,t)$ is the mortality rate for age 0 in province $p$ during year $t$. The 0.5 multiplier reflects that, on average, newborns are exposed to mortality risk for approximately half a year before reaching the January 1st reference date.</p> Modeling births to foreign-born mothers Handling migrationResults #</a></h1> Figure 1</a>. The share of native-born individuals with native-born mothers will decrease from 77.1% in 2024 to just 62.3% by 2039. In just 15 years, the traditionally dominant demographic group will lose around 15 percentage points of its population. Meanwhile, first-generation immigrants will surge from 18.1% to 28.6% of the population, while second-generation immigrants (native-born with foreign-born mothers) will grow from 4.8% to 9.1%.</p> Put these numbers together, and you get a striking conclusion: by 2039, nearly 4 in 10 Spanish residents will be either immigrants themselves or the children of immigrants</strong>. This represents one of the most rapid demographic transformations in modern European history.</p> Figure 1 </figcaption> </iframe> </figure> However, the national picture masks a profound regional divergence. As Figure 2</a> reveals, Spain will fracture into two distinct demographic realities.</p> In Southern provinces like Córdoba, Jaén, and Badajoz — areas with traditionally agrarian economies and fewer economic opportunities — the demographic composition will remain relatively stable. Here, native-born individuals with native-born mothers will still comprise close to 90% of the population by 2039, preserving much of the traditional demographic character. But along the Mediterranean coast and in economic powerhouses like Madrid and Barcelona, a dramatically different Spain is emerging: in these regions, immigrant and second-generation populations combined will approach or even exceed 50% of the total population by 2039</strong>.</p> In Spain's most economically vital and populous areas, the native-origin population will be close to becoming a minority in less than 15 years. Absent any major shifts in migration patterns, the municipalities of Madrid and Barcelona will transform into multicultural hubs more reminiscent of London or New York than the Spain of previous generations.2</a></sup></p> Figure 2 </figcaption> </iframe> </figure> The scale of this shift becomes even more striking when we consider the age distribution. By 2039, the immigrant-origin population (first and second generation combined) will be substantially overrepresented among Spain's working-age population, while the native-origin population will be disproportionately concentrated among older age groups.</p> We can do a simple back-of-the-envelope calculation to illustrate this point. In 2024, foreign-born individuals constituted 23.1% of Spain's working-age population (ages 15-64), despite representing only 18.2% of the total population—an overrepresentation factor of 1.27. Projecting this pattern forward, by 2039, foreign-born individuals will comprise approximately 36.3% of Spain's working-age population (28.6% $\times$ 1.27).</p> Among the remaining native-born segment (63.7% of the working-age population), second-generation immigrants will represent about 10.3%, translating to 6.6% of the total working-age population. When combined, these figures imply that, by 2039, approximately 43% of Spain's workforce — over one in four working-age individuals — will be either first or second-generation immigrants.</strong></p> Some external validationhere</a>). In 2024, Eurostat reports that native-born individuals with at least one foreign-born parent (they don’t break out figures for foreign-born mothers alone) constitute 3.80% of Spain's working-age population. To be extremely tedious, I recalculated my projections using the same definition — either parent foreign-born rather than just mothers — and obtained 3.75% for the same year. This is a difference of only 0.05 percentage points relative to official counts, which is well within the margin of error.</p> Conclusion ^1</sup> http://dedonde.es/</a>.</p> </div> ^{2</sup> With the caveat that, compared to London or New York, origin countries are much more concentrated in Spanish cities. For example, individuals born in Latin America accounted for roughly 48% of the total foreign-born population stock at the national level in 2024.</p> </div>} Montoya va donde brilla (in praise of not belonging) 2025-02-09T00:00:00+00:00 Let’s start with the circus. La Isla de las Tentaciones</a></em> — Spain’s answer to Temptation Island</a> — is not a reality TV show. It’s a social experiment designed to strip love down to its most pathetic, voyeuristic core. The premise is simple: take a handful of couples, separate them from their partners on a tropical paradise, and throw in attractive singles to test their loyalty. The result? A spectacle of jealousy, cheating, and the kind of emotional carnage that makes you wonder why anyone would willingly sign up for this</strong>.</p> Who is Montoya?In Season 8, José Carlos Montoya</a> became the lab rat we deserved. His girlfriend, Anita, did what anyone would do under fluorescent lights and the gaze of a million strangers: she fucked someone else</strong>. Montoya’s response? He ran</a>. Not toward dignity, but into his own humiliation, slapping sand as producers barked Montoya, por favor!</em> like they were trying to stop a drunk friend from doing something stupid.</p> The cameras loved it. We loved it. The clip went viral. Of course it did. The editing was exquisite – Spielberg himself could not have staged it better. The camera angles, the lighting, the rhythmic cuts between Montoya's crumbling face and his girlfriend's... alternative activities. Pure cinema</strong>.</p> </figure> But before this pivotal moment, Montoya had already established himself as the show's main character through a series of fantastic one-liners, performances, and a general air of tragicomedy. One of his most memorable moments was when, during a breakdown, he declared with the solemnity of a Shakespearean actor that *"Montoya va donde brilla"</em></strong></a> ("Montoya goes where he shines").</p>* At first, it seemed like a linguistic flub — an unintended inversion of the phrase "Montoya brilla donde va"</em></strong> ("Montoya shines wherever he goes", which is a common Spanish idiom to say that someone is able to thrive in any environment). But as the meme spread, I realized that this accidental catchphrase was never meant to be a joke. It was a manifesto</strong>.</p> The accidental prophet This is a variation of the famous trolley problem</a>, a thought experiment in ethics and moral psychology. The problem is designed to highlight the tension between two moral principles: the principle of utilitarianism</strong>, which states that the right action is the one that maximizes overall wellbeing, and the principle of deontology</strong>, which states that the right action is the one that respects individual rights and duties. In the trolley problem, the utilitarian answer is to kill the healthy person to save the five patients, while the deontological answer is to not kill the healthy person, as doing so would violate their right to life.</p> I am not a utilitarian. However, philosophies of justice (like liberalism) often become unhelpful in scenarios where wellbeing trade-offs are inevitable. For instance, consider now that the surgeon has healthy organs available, but only enough to save three of the five patients. What should the surgeon do in that case?</p> In the real world, policymakers often face decisions that entail such compromises in one way or another: for example, should we invest in a new hospital, or in a new school? Should a government implement regulations to protect consumers, which could increase the cost of goods and services and potentially stifle innovation? Should a city invest in improving public transportation, potentially reducing traffic and pollution, but at the cost of higher taxes? These are all questions that involve trade-offs between different individuals' wellbeing, and more often than not, they are not easy to answer.</p> One of the ultimate goals of policy should be to maximize wellbeing within constraints</em> for the individuals that it serves to. Philosophies of justice can help us delimit those constraints, but they are not enough to guide us on how to identify wellbeing maximizing interventions and balance wellbeing trade-offs.</p> *And this is where wellbeing research comes in.</p>* *This is the first post of a series about wellbeing measurement and policy.</p>* *Definitions* 1</a></sup> A variant of the former approach is to ask respondents to rate their life on a ladder, where the top of the ladder represents the best possible life for them, and the bottom of the ladder represents the worst possible life for them. This is commonly known as the Cantril ladder</em>, which is used in many international household surveys such as the Gallup World Poll</a>.</li> Hedonic</strong> approaches focus on measuring how people feel at a particular point in time. This is typically done by asking respondents to report their feelings at the moment of the survey, or to recall their feelings over a specific period of time. Experienced happiness is the most common measure of hedonic wellbeing, but negative emotions such as sadness, anger, and stress are also used.</li> Eudaimonic</strong> measures are based on the Aristotelian tradition with the idea that wellbeing is not just about feeling good, but also about functioning well. This approach focuses on the extent to which people are able to realize their potential, and to live a life that is meaningful and fulfilling. This is typically measured by asking respondents about their sense of purpose, their autonomy, and their relationships with others.</li> </ul> Tipically, evaluative and hedonic measures are the most widely used in policy research. Eudaimonic measures, although philosophically appealing, are harder to interpret – for example, when we talk about being virtuous, it is not clear whether that should be an outcome in itself or another determinant (among many) of wellbeing.</p> Figure 1</a> shows cross-country averages of life satisfaction, as measured by the Cantril ladder question from the Gallup World Poll. The figure shows that there is a wide variation in life satisfaction across countries, with the highest levels of life satisfaction being found in the Nordic countries, and the lowest levels of life satisfaction being found in sub-Saharan Africa and the Middle East. The difference between the country with the highest average life satisfaction (Finland) and the country with the lowest average life satisfaction (Afghanistan) is about 6 points on the 0-10 Cantril ladder scale.</p> Figure 1: Life satisfaction across countries </figcaption> </iframe> </figure> It is (not) all about the money #</a></h2> higher life expectancy</a>, better education systems</a>, well-functioning labor markets</a>, political stability, absence of violence...</p> ...and wellbeing itself</strong>. Across countries, higher GDP per capita is associated with greater levels of life satisfaction. This finding is remarkably consistent across surveys, and it holds up irrespective of the definition of wellbeing we use. Figure 2</a> shows the relationship between GDP per capita and life satisfaction for a large number of countries, using data from the Gallup World Poll. High-income countries tend to have higher levels of life satisfaction than low-income countries, and the relationship is roughly log-linear. The latter is consistent with the idea of diminishing marginal utility of income: an extra dollar has more impact on wellbeing for the average Joe than for Jeff Bezos. Intuitive, right?</p> Figure 2: Self-reported life satisfaction vs. GDP per capita </figcaption> </iframe> </figure> What is the catch then? Why not just stick to income as a proxy for wellbeing?</p> The Easterlin paradox Figure 3</a>, which shows the evolution of GDP per capita and average life satisfaction (derived from large-scale longitudinal household surveys) for a sample of advanced economies. The figure shows that, despite considerable increases in within-country per capita income, the average level of life satisfaction has remained surprisingly constant over time. In the U.S., for example, GDP per capita has increased by more than 500% since 1950, but people are not any happier now than they were back then.</p> Figure 3: The Easterlin Paradox</figcaption> </iframe> </figure> Source: Clark et al. (2018).</a></p> In light of the above, the answer to the question on why should we not rely solely on income as a proxy for wellbeing is straightforward: because it is simply not a good standalone proxy in the long-run.</p> The social comparison hypothesis Card et al. (2012)</a> run an experiment in which they randomly inform employees of the University of California about the salaries of their peers, and find that workers whose wage is below the median for their comparison group report lower job satisfaction and are more likely to start looking for a new job. Perez-Truglia (2020)</a> exploits a natural experiment whereby Norwegian tax records became publicly available for everyone to see, and shows that income transparency increases the gap in happiness between high- and low-income individuals by 29 percent. This occurs because people at the bottom (top) of the income distribution typically overestimate (underestimate) their relative position, and increased transparency makes them realize how much worse (better) off they are compared to others.</p> In a less clean way, one could test for the social comparison hypothesis using a very simple regression framework in which we regress wellbeing on both individual income and the average income of a comparison group:</p> $$Y_i = \alpha + \beta X_i + \gamma \bar{X} + \epsilon_i$$</p> where $Y_i$ is individual $i$'s wellbeing, $X_i$ is individual $i$'s income, $\bar{X}$ is the average income of a group of comparators, and $\epsilon_i$ is the error term. If the relative income hypothesis holds, we would expect $\gamma$ to be negative, as individual wellbeing would decrease as the average income of the group of comparators increases.</p> This is exactly what we find. Table 1</a> shows the results of such a regression for cross-sections of different countries, in which the group of comparators is defined by people of the same sex, age group, region, and survey year, and the outcome is life satisfaction (measured on a 0-10 scale). Individual wellbeing is positively associated with individual income, but negatively associated with the average income of the comparison group. Except for the United States, a 1% increase in the average income of comparators reduces wellbeing in the same amount or more than a 1% increase in individual income increases wellbeing.</p> Table 1: The social comparison hypothesis</figcaption> </th> Britain</th> Germany</th> Australia</th> United States</th></tr></thead> Own income</td> 0.16 (.01)</td> 0.26 (.01)</td> 0.16 (.01)</td> 0.31 (.01)</td></tr> Comparator income</td> -0.23 (.07)</td> -0.25 (.04)</td> -0.17 (.06)</td> -0.19 (.03)</td></tr> </tbody></table> Source: Layard and De Neve (2023)</a>, Table 13.4. Standard errors in parentheses.</p> </div> There are other potential explanations for the Easterlin paradox, such as the hedonic adaptation hypothesis</strong>. Roughly speaking, the idea here is that people adapt to their new income levels and their expectations increase (or decrease) accordingly. We will not discuss it in detail here, but there is equally interesting evidence on the role of adaptation in explaining the paradox.2</a></sup></p> Effect sizes Lindqvist et al. (2020)</a> look at the long-run effects of lottery wealth on psychological wellbeing in Sweden, and find that an after-tax lottery prize of $100,000 increases overall life satisfaction by 0.037 standard deviations relative to the matched control group. To put it into perspective, that is about 70% of the difference in average life satisfaction between Denmark (the second happiest country in the world) and Iceland (the third happiest country in the world). In other words, this is a very small effect. Relatedly, the authors also explore the impact of lottery wealth on other outcomes, such as happiness and mental health, and in these cases the effects are not significantly different from zero. That is, lottery winners are not happier or less mentally distressed relative to the control group in the long-run.</p> As we saw in Figure 2</a>, the effects of income on wellbeing are also non-linear. You probably have come across media articles arguing that, beyond a certain income threshold, the marginal effect of income on wellbeing is virtually zero. This is pretty much based on the (now sort of old) paper by Kahneman and Deaton (2010)</a>, who use data for the U.S. and find that the relationship between income and emotional wellbeing</em> is log-linear up to an annual income of $75,000, but flattens out beyond that point. The idea is that further increases in income do not necessarily improve the individual's ability to engage in activities that are most important for their emotional wellbeing, such as spending time with friends and family, or having better health.</p> However, a key thing to note is that here we are looking at hedonic</em> measures of wellbeing. Specifically, they use the average of three dicotomous, binary measures of positive affect (happiness, enjoyment, and frequent smiling and laughter) and the average of two binary measures of negative affect or "blue" states (worry and sadness). In contrast, the authors find that the relationship between income and evaluative</em> wellbeing (i.e., what we have been referring to as life satisfaction or general happiness) shows no sign of satiation, at least to an amount over $160,000 per year.</p> Things got interesting when Killingsworth (2021)</a> revisited this question and found that, conversely, hedonic wellbeing does</em> rise with income well above the $75,000 per year threshold. The main argument of this paper is that the discrepancy with the original results might be explained by differences in the variable definitions: on top of just using binary variables (which are able to capture less variation), Kahneman and Deaton (2010)</a> results are based on retrospective survey responses, which are subject to recall bias. Instead, Killingsworth (2021)</a> uses large-scale real-time data collected through a smartphone app and a continuous scale to measure experienced wellbeing, finding that the latter increases linearly with log-income with no satiation point.</p> In a recent paper, the authors joined forces and reconciled the two previous findings (Killingsworth et al., 2023</a>). Here, they show that the flattening pattern in the relationship between income and experienced wellbeing depicted in the original study by Kahneman and Deaton does exist</em>, but it only applies to a limited subset of the population. Using quantile regressions, they show that increases in income do not shift upwards the bottom 15-20% of the emotional wellbeing distribution (i.e., the least happy), but they do so across all other quantiles.</p> To illustrate this, see Figure 4</a>. The happiness-income gradient is positive and significant for the top 80% of the distribution, but the slope turns flat for the those below the 15th percentile after $120,000 per year.</p> Figure 4: Emotional wellbeing across the happiness distribution </figcaption> </iframe> </figure> Source: Killingsworth et al. (2023).</a></p> So, why Kahneman and Deaton (2010)</a> claimed that the satiation pattern applies to the entire population? As mentioned above, the answer is in the variable definition. Since they used dichotomous measures, the authors ended up with a very skewed distribution of happiness, especially among higher incomes, suggesting a "ceiling effect" whereby the measure could not capture differences between higher degrees of happiness. In fact, the average reported positive affect for the range of high incomes in Kahneman and Deaton (2010)</a> is 89% of a perfect score, but we know that the happy people are not all equally happy!</p> To see this, consider the example of a pass-or-fail math test, consisting of questions that most students can answer correctly. This test could be informative to identify the students who are doing badly, but it would not be able to differentiate between the students who are doing well because most of them would get the same pass grade. In a sense, the measures of affect in Kahneman and Deaton (2010)</a> are just like this test: they are good at identifying the people who are unhappy, but they have limited ability to discriminate among levels of happiness. Instead of "Happiness increases with income, but it does not further increase above $75,000 per year"</em>, it follows that the statement should have read "Unhappiness decreases with income, but it does not further decrease above $75,000 per year"</em>.</p> What's next?^2</sup> For instance, see Di Tella et al. (2010)</a> and Kuhn et al. (2011)</a>.</p> </div> References #</a></h1> Seth Stephens-Davidowitz</a>, former data scientist at Google and PhD in economics from Harvard. Everybody lies</em></a> is an incredibly fun read offering a compelling argument: that our online searches can reveal more about our true selves than we might like to admit. The book is not only a testament to the power of non-traditional data sources, but also an invitation to explore human behaviour through a different lens.</p> Stephens-Davidowitz's work shows that Google searches can be used to predict a wide range of socio-economic outcomes when conventional data sources are not available or are contaminated by social desirability bias. The latter occurs when people are reluctant to disclose their true preferences or beliefs in surveys, which is particularly important when studying sensitive topics such as racial animus, sexual orientation, or mental health. The idea is quite simple: if revealing what you really think leads to social stigma, then you would be more likely to confess to Google than to a survey interviewer. In a sense, the unfiltered truth of humanity lies in Google's autocomplete bar.</p> In one of his academic papers</a> published in the Journal of Public Economics</em>, Stephens-Davidowitz uses Google searches to measure racial animus in the U.S. and its impact on the 2008 and 2012 presidential elections – specifically, he finds that racially charged Google search rates are a strong negative predictor of Obama's vote share. The point estimates in the preferred specification suggest that racial animus cost Obama around 4 percentage points in the 2008 election, which is significantly above previous estimates based on survey data.</p> Table 1: Google search terms and socio-economic outcomes </figcaption> Term</th> Underlying variable</th> t-Stat</th> $R^2$</th></tr></thead> God</td> Percent believe in god</td> 8.45</td> 0.65</td></tr> Gun</td> Percent own gun</td> 8.94</td> 0.62</td></tr> African American(s)</td> Percent Black</td> 13.15</td> 0.78</td></tr> Hispanic</td> Percent Hispanic</td> 8.71</td> 0.61</td></tr> Jewish</td> Percent Jewish</td> 17.08</td> 0.86</td></tr> </tbody></table> Source: Stephens-Davidowitz (2015)</a>, Table 1. The t-stat and $R^2$ are from a regression with the normalized search volume of the word(s) in the first column as the independent variable and measures of the value in the second column as the dependent variable. The normalized search volume for all terms is from 2004 to 2007. All data are at the state level.</p> </div> But the potential of Google searches goes beyond measuring racial animus. In Table 1</a>, I show some of Stephen-Davidowitz's examples on how Google searches can be exploited to predict various population-level outcomes. For example, Google searches for the word "Gun" explain 62% of the variation in states' gun ownership rates, and Google searches for the word "Hispanic" explain 61% of the variation in the percentage of Hispanic population across U.S. states.</p> Can we do better? In this post, I will show you how we can leverage the popularity of Latin urban music artists in Google searches to predict state-level Hispanic population shares in the U.S.</p> Some background on Google Trends data #</a></h1> Google Trends</a>, a tool that allows users to explore the popularity of search terms over time and across different regions within countries. The data is normalized to the time and location of a query, so that the popularity of a search term is measured relative to the total number of searches in a given region and time period. In other words, Google Trends gives measures of relative search term popularity, not an absolute measure of the number of searches for a given term.[^1] Relative popularity is measured on a scale from 0 to 100, where 100 is the most popular point in time or location for a given search term. For example, value of 50 means that the term is at half its peak observed popularity.</p> There are a bazillion online resources on how to use Google Trends, so I won't go into the details here. As a starting point, I would recommend you to check out the Google Trends basics page</a> and the search tips page</a>.</p> Predicting Hispanic population shares with Google Trends #</a></h1> American Community Survey (ACS)</a> microdata for the year 2022.[^2]</p> First, we will use the search term "bad bunny" to predict Hispanic population shares. Bad Bunny</a> is arguably the most popular Latin artist worldwide nowadays – in 2022, he was Billboard's top artist of the year and the most streamed artist</a> in Spotify and Apple Music. He also has a huge following</a> on Instagram with over 40 million followers, and its popularity has played a crucial role in the reivindication of the Latino culture and identity in the U.S. (ahora todos quieren ser latinos</em></a>).</p> Let's do it. To start with, I show the Hispanic population shares across U.S. states in 2022 in Figure 1</a>. Unsurprisingly, the Hispanic population is concentrated in the South and West regions of the U.S. The states with the highest Hispanic population shares are New Mexico, California, and Texas, with shares of 50%, 40.3%, and 40.2%, respectively. The states with the lowest Hispanic population shares are West Virginia, Maine, and Vermont, with shares of around 2% each.</p> Figure 1</figcaption> </iframe> </figure> Now, let's do the same for the popularity of the search term "bad bunny" in the U.S. in 2022:</p> Figure 2</figcaption> </iframe> </figure> Similar, right?</p> We can visualize the relationship between the popularity of the search term "bad bunny" and the Hispanic population shares across U.S. states in a scatter plot (see Figure 3</a>). By running a simple OLS regression of the Hispanic population share on the popularity of the search term, we get that 76% of the variation in the Hispanic population share across U.S. states can be explained by the relative popularity of Bad Bunny in Google searches. This implies a 15 percentage point increase in the share of variation explained relative to just using the term "Hispanic" as a predictor.</p> Figure 3</figcaption> </iframe> </figure> There is one state that is particularly off path: New Mexico (if we were to omit it, then the share of variation explained would go up to 89%). New Mexico has a Hispanic population share of 50%, but the popularity of Bad Bunny in Google searches is only 56% of the most popular point location for the term.</p> One particularity of New Mexico is that it has the highest population share with Mexican origin in the U.S. What if we focused on the popularity of a Mexican artist, and looked at Mexican population shares instead? I got curious about this, so I replicated the analysis exploiting the popularity of the search term "peso pluma". Peso Pluma</a> is a Mexican artist whose popularity has skyrocketed during the last couple of years, being Youtube's most viewed artist of 2023 in the U.S.</a></p> The results are shown in Figure 4</a>. The fit is breathtaking: Peso Pluma's popularity in Google searches explains 94% of the variation in the Mexican population share across U.S. states.</p> Figure 4</figcaption> </iframe> </figure> Measuring what matters In Table 2</a>, I illustrate the comparison between how state-level suicide rates in the U.S. correlate with both Google search terms and self-reported measures of suicide risk. While Google searches are not perfect, they prove to be more reliable than the traditional survey-derived data, which yields statistically insignificant correlations.</p> Table 2: Correlations between Google search terms and suicide rates in the U.S. </figcaption> Predictor</th> $r$</th> $p$</th></tr></thead> Google search term</strong></td> </td> </td></tr> Suicide</td> -.06</td> .63</td></tr> How to suicide</td> .49</td> < .001</td></tr> How to kill yourself</td> .61</td> < .001</td></tr> How to commit suicide</td> .63</td> < .001</td></tr> Painless suicide</td> .59</td> < .001</td></tr> Composite of all four suicidal terms</td> .67</td> < .001</td></tr> Self-report measure</strong></td> </td> </td></tr> Suicidal thoughts</td> -.17</td> .22</td></tr> Suicidal plans</td> -.14</td> .30</td></tr> Suicidal attempts</td> -.03</td> .82</td></tr> Composite of self-reported suicidality</td> -.16</td> .25</td></tr> </tbody></table> Source: Ma-Kellams et al. (2016)</a>, Table 1. Data for 2008 and 2009. The column $r$ shows the correlation between suicidal death rates and the relevant Google search term at the state level. Self-reported measures are sourced from the National Survey on Drug Use and Health.</p> </div> We can also use Google searches to track economic activity in real time. For example, the OECD has developed a tool</a> that uses Google searches to nowcast GDP growth on a weekly basis. The algorithm gathers data related to various realms of economic activity, using search terms related to consumption (like "vehicles" and "household appliances"), labour market conditions (such as "unemployment benefits"), the housing sector (like "real estate agencies" and "mortgages"), business-related services, industrial activity, trade, and economic sentiment, among others. So, next time you are casually checking (again) whether rent prices are going up in your area, you might be contributing to a fancy economic forecast.</p> Wrapping up here</a>). Could we predict vote shares for controversial candidates at scale with Google searches?</li> Emergency situations and public safety</strong>: during emergencies, real-time data from Google Trends can provide insights into public sentiment or areas of concern, which might be faster or more accurate other methods. For instance, in the context of natural disasters, spikes in searches for emergency services or relief efforts can indicate the most affected areas or the public’s most urgent needs.</li> </ul> I will leave it here. If you have worked with Google Trends data before and have some cool applications in mind you would like to share, please feel free to get in touch!</p> [^1] However, with some assumptions and creativity, it is possible to come up with proxies for absolute Google search volumnes. See Appendix C in Stephens-Davidowitz (2015)</a> for more details.</p> [^2] I identify people with Hispanic/Latino origin in the ACS using the variable `HISPAN</code>. More details can be found` `here</a>.</p>` `[^3] One further caveat is that the Google Trends metro areas are custom aggregates and do not correspond to standard metropolitan statistical areas (MSAs) or metropolitan divisions. Due to this, there has been some attempts to map Google Trends metro areas to more granular geographic units, such as counties in the U.S</a>.</p>` References#</a></h1> Stephens-Davidowitz, S. (2017). Everybody lies: Big data, new data, and what the internet can tell us about who we really are</em>. Bloomsbury Publishing.</li> Stephens-Davidowitz, S. (2015). The cost of racial animus on a black candidate: Evidence using Google search data. Journal of Public Economics</em>, 118, 26-40.</li> Ma-Kellams, C., Or, F., Baek, J. H., & Kawachi, I. (2016). Rethinking suicide surveillance: Google search data and self-reported suicidality differentially estimate completed suicide risk. Clinical Psychological Science, 4(3), 480-484.</li> </ul> A modern IDE for Stata with VSCode + Quarto 2024-02-27T00:00:00+00:00 Update Aug 2024</strong></p> I only used this setup for a couple of weeks before switching to a non-interactive integration that allows for running .do</code> files within VSCode. The lack of a data browser and the variable explorer in the Jupyter kernel was a dealbreaker for me (plus I have grown to find interactive notebooks a bit annoying except for very specific tasks). To set this up, follow the instructions here</a> under the "way one" section.</p> Moreover, I have increasingly found myself using R for most of my work over the last year, and I have to stay that I no longer agree with some of the points I made in this post. Most of the time, the so-claimed "superiority" of Stata for reduced-form empirical work is just a skill issue. My advice would be to invest time in mastering R or Python, and then</em> decide whether you really need to use Stata (you probably won't).</p> </div> </blockquote> Over the last few decades, applied microeconomic research has witnessed what is widely termed as the credibility revolution</a>, a paradigm shift that emphasized the importance of experimental and quasi-experimental methods to credibly identify causal relationships between variables. This transformative period has not only influenced the methodologies employed in academic papers but also strengthened the field's impact on policy-making, by providing more solid and rigorous evidence upon which to base decisions.</p> This emphasis on methodologically robust empirical approaches has resulted in bigger, more complex projects, naturally elevating the importance of data skills within the field. However, despite the increasing importance of coding in economics, many economists lag in adopting modern coding practices and development environments. One of the most widely used tools in the profession, Stata</a>, is not even a programming language but a purpose-built statistical package.1</a></sup> Stata is also commercial software, meaning that it does not have the benefits of open-source languages like Python or R, which have amassed extensive communities contributing to their development and offering support. Due to its nature, it has traditionally been hard to integrate Stata with modern IDEs, resulting in a less efficient workflow when one has to write code in multiple languages or use the latest tools for version control, code completion, and so on.</p> This post provides a quick tutorial on setting up a modern development environment for Stata with Visual Studio Code and Quarto. It is by no means perfect and it comes with its own caveats, but it is a step in the right direction for those who want to modernize a bit their workflow without having to give up on Stata.</p> Why Stata#</a></h1> I know what you are thinking.</p> jUsT dOn't usE sTAta BrO</strong>.</p> I get it. I code in multiple languages and it is indeed not a good idea to only</em> use Stata, simply because there are lots of cool things (web scraping, geospatial stuff, ML) that you would miss out on by doing so.</p> However, the truth is that if you do applied microeconomics for a living, it would be very stupid not to use Stata</strong> sometimes. Whoever tells you otherwise is i) not an applied microeconomist (good for them!) or ii) underestimating how important Stata is in the profession:</p> Stata can be found in more than 70% of replication packages</a> for published papers in economics. It is followed by Matlab (which we deliberately ignore since it is aimed at structural work), and R comes in third with less than 10%. I am not going to argue whether this is a good or a bad equilibrium but it is an equilibrium</em>, and if you want collaborate with others and be able to replicate their work, you will have to use Stata</p> </li> For reduced-form econometrics, Stata is unparalleled. Period. I have been writing code for +3.5 years and it still amazes me how ridiculously cumbersome R and Python can be relative to Stata when it comes to running regressions. For some specifications, it is simply not possible to use something else unless you want to code the estimator from scratch (good luck running modern difference-in-differences estimators for heterogeneous treatment effects in Python, for example)</p> </li> Economists love regression tables. And again, the functionalities and flexibility of esttab</code> to export fully formatted and reproducible LaTeX tables in Stata are light years ahead of the alternatives in other languages (e.g., stargazer</code> in R)</p> </li> For most people doing applied micro research or policy work, coding will hardly ever be their comparative advantage</strong>. Instead, your value lies in the quality of your ideas and how well you execute them – coding is only useful insofar as it helps in that purpose. Some folks (myself included) enjoy venturing off the beaten path, but most researchers would prioritize research productivity if the marginal cost of learning new data skills is sufficiently high. In this sense, Stata has a quite shallow and gentle learning curve, so usually people learn it as their first "language" and then just stick to it. I am not saying this is good (I think is bad!), but in the end people respond to incentives</p> </li> </ul> Why VSCode#</a></h1> VSCode is Microsoft's open-source code editor. It supports a wide array of programming languages and comes with the usual IDE features like syntax highlighting, code completion, Copilot</a>, and version control integration.2</a></sup></p> More importantly, turns out that it is fairly straightforward to integrate VSCode with Stata via plug-ins. VSCode also has nice LaTeX extensions, which in my opinion makes it the only real alternative to Overleaf. This means that you can seamlessly integrate all your workflow into a single code editor, from data cleaning to drafting.</p> Why Quarto#</a></h1> Quarto is an open-source scientific and technical publishing system that allows users to create dynamic documents, slide decks, and reports in various formats from plain text files. Quarto files support multiple programming languages and can be rendered into interactive HTML documents, which allows the user to combine code execution with high-quality document production.</p> The main advantage of using a notebook-based interface relative to the native Stata editor is that it allows you to create and present exploratory analysis combining code, output, and text (including comments, LaTeX equations, and more), all in one convenient place.</p> And last but not least, Quarto outputs are beautiful</a>.</p> Setup#</a></h1> Step 1: Install Quarto#</a></h2> Go to https://quarto.org/</a>, click on "Get Started", download the latest Quarto version available, and install it on your machine.</p> </p> Step 2: Install VSCode#</a></h2> Go to https://code.visualstudio.com/download</a>, download the latest VSCode version available, and install it on your machine.</p> </p> Step 3: Install VSCode extensions#</a></h2> You will have to install two VSCode extensions:</p> To get nice Stata syntax highlighting and code autocompletion, you will have to install the Stata Enhanced</a> package. Open VSCode, go to the extensions tab (highlighted in red in the screenshot), search for the extension, and install it.</li> For integrated render and preview for Quarto documents, repeat the steps above and install the Quarto</a> extension.</li> </ol> </p> Step 4: Install nbstata</code>#</a></h2> nbstata</code></a> is a Jupyter kernel for Stata, built on top of the functionalities of the pystata</code> Python package introduced in Stata version 17. This means that you will need Stata 17 or higher to install it.3</a></sup></p> You can find the full user and installation guide here</a>. For the curious, I encourage you to go through the author's guide. For the lazy ones, the short version goes as follows:</p> To install nbstata</code>, you need an installation of Python 3.7 or higher. For first-time Python users, I would follow the suggestion from the author of nbstata</code> and just install the Anaconda distribution</a>. If you want a lighter, minimalistic version, you can go for Miniconda</a>.4</a></sup></li> Once you have installed Anaconda, open the Anaconda prompt and type</li> </ol> pip install nbstata</span></span></code></pre> Finally, install nbstata</code> by running</li> </ol> python -m nbstata.install</span></span></code></pre> And that is it! Your machine should be now ready to render Quarto Markdown (.qmd</code>) files in VSCode using the nbstata</code> kernel.</p> An example#</a></h2> To begin, create a .qmd</code> file. At the very start of this file, incorporate a YAML header to specify the Jupyter kernel you wish to use. In our case, to use nbstata</code> in your Quarto document, add jupyter: nbstata</code> within the YAML header. The YAML header is a section located at the top of Markdown files, designed to hold metadata such as configuration settings and document properties. This metadata is structured in a readable format and is essential for defining various aspects of your document's processing and presentation. A sample of YAML header could look as follows:</p> ---</span></span> title: "Some dataviz with Stata and Quarto"</span></span> date: 2024-24-01</span></span> author: "Pablo"</span></span> jupyter: nbstata</span></span> format: </span></span> html:</span></span> self-contained: true</span></span> toc: true</span></span> toc-depth: 3</span></span> toc-float:</span></span> collapsed: false</span></span> smooth-scroll: true</span></span> width: 20%</span></span> ---</span></span></code></pre> You can now create code chunks as you would do in any other notebook-based interface, specifying stata</code> in your code chunk header.</p> For our example, we will create a chart to visualize gender gaps in suicide rates across low-income countries. To load the dataset into our environment, we will use the Stata wrapper for the World Bank's WDI database API, wbopendata</code></a>:</p> `</span>`</span>`</span>{</span>stata load-data}</span></span> wbopendata</span>,</span> indicator(SH.STA.SUIC.FE.P5; SH.STA.SUIC.MA.P5) clear long</span></span> desc</span></span> ```</span></span></code></pre>Metadata for indicator SH.STA.SUIC.FE.P5</span></span> -------------------------------------------------------------------------------</span></span> Name: Suicide mortality rate, female (per 100,000 female population)</span></span> -------------------------------------------------------------------------------</span></span> Collection: 2 World Development Indicators</span></span> -------------------------------------------------------------------------------</span></span> Description: Suicide mortality rate is the number of suicide deaths in</span></span> a year per 100,000 population. Crude suicide rate (not age-adjusted).</span></span> -------------------------------------------------------------------------------</span></span> Note: World Health Organization, Global Health Observatory Data</span></span> Repository (http: apps.who.int ghodata ).</span></span> -------------------------------------------------------------------------------</span></span> Topic(s): 8 Health</span></span> -------------------------------------------------------------------------------</span></span> </span> </span> </span> Metadata for indicator SH.STA.SUIC.MA.P5</span></span> -------------------------------------------------------------------------------</span></span> Name: Suicide mortality rate, male (per 100,000 male population)</span></span> -------------------------------------------------------------------------------</span></span> Collection: 2 World Development Indicators</span></span> -------------------------------------------------------------------------------</span></span> Description: Suicide mortality rate is the number of suicide deaths in</span></span> ...</span></span> sh_sta_suic_m~5 float %8.0g SH.STA.SUIC.MA.P5</span></span> -------------------------------------------------------------------------------</span></span> Sorted by: countrycode year</span></span> Note: Dataset has changed since last saved.</span></span> Output is truncated. View as a scrollable element or open in a text editor. Adjust cell output settings...</span></span></code></pre> Let's create a simple dumbbell chart by sex:</p> `</span>`</span>`</span>{</span>stata gen-chart}</span></span> </span> Prepare data</span></span> cap ren sh_sta_suic_fe_p5 fem_rate </span></span> cap ren sh_sta_suic_ma_p5 male_rate </span></span> </span> keep if !inlist(.,fem_rate,male_rate) // keep non-missing pairs of female and male suicide rates</span></span> bys countrycode (year): keep if _n == _N // keep latest country-year obs.</span></span> drop if region == "" // drop regional aggregates</span></span> keep if incomelevelname == "Low income" // keep low-income countries</span></span> </span> Chart settings </span></span> distinct countrycode</span></span> local ub = `r(ndistinct)</span>'</span> // N of categories for the x-axis</span></span> </span> sort male_rate </span></span> cap g order = _n // order of x-axis categories</span></span> labmask order, val(countryname)</span></span> </span> su male_rate, d</span></span> local x_max = round(r(max),0.1) // max value for y-axis</span></span> </span> * Export chart</span></span> #d;</span></span> twoway (rspike male_rate fem_rate order, lcolor(gs14%90)) </span></span> (scatter male_rate order, mcolor(midblue)) </span></span> (scatter fem_rate order, mcolor(pink)), </span></span> ytitle("Suicide mortality rate (per 100,000 population)") </span></span> xtitle("") </span></span> ylabel(0(5)`x_max</span>'</span>)</span> </span></span> xlabel(1(1)`ub</span>'</span>,</span> angle</span>(</span>45</span>)</span> labsize</span>(</span>vsmall</span>)</span> valuelabel</span>)</span> </span></span> legend</span>(</span>order</span>(</span>2</span> "</span>Males</span>"</span> 3</span> "</span>Females</span>"</span>)</span> pos</span>(</span>11</span>)</span> row</span>(</span>1</span>)</span> size</span>(</span>3</span>)</span>)</span> </span></span> title</span>(</span>Gender</span> gaps</span> in</span> suicide</span> rates</span> across</span> low</span>-</span>income</span> countries</span>,</span> pos</span>(</span>11</span>)</span> size</span>(</span>medium</span>)</span>)</span>;</span></span> #</span>d</span> cr</span></span> `</span>`</span>`</span></span></code></pre> </p> Caveats#</a></h1> Sadly, the integration of Stata with VSCode and Quarto is far from perfect, specially if you are used to the native Stata editor and its functionalities. Here are a couple of issues that bother me:</p> You will be missing the data browser and the variable explorer. This is a big one. The Stata data browser is also incredibly smooth and fast compared to the alternatives in R and Python, so it is a shame to lose one of its comparative advantages. Although nbstata</code> comes with Jupyter magics</a> that allow for browsing, such as %browse</code> for the main dataset and %frbrowse</code> for frames, so far I have not figured out a way to expand the cell output in VSCode (nor in Jupyter itself), so this is not useful. If you know of a solution, please reach out!</p> </li> You will not be able to change the font for your charts within VSCode. Not a big deal, you can always change it by opening Stata itself and changing the defaults with the graph set</code> commands, but it is a bit annoying</p> </li> </ul> Troubleshooting#</a></h1> In case you are finding problems with your setup, you should verify your installations and ensure that the Jupyter kernel is correctly set up and recognnized by your system.</p> To check if Quarto is installed and determine its version, open your command prompt (or terminal on macOS and Linux) and enter the following command:</p> quarto check</span></span></code></pre> If Quarto is not found, ensure it is installed and that its installation directory is added to your system's PATH environment variable. Reinstalling may also help.</p> To verify that Python is installed on your system and to check its version, run:</p> python --version</span></span></code></pre> To check your Python installation directory, run</p> where python</span></span></code></pre> This command should return the version of Python installed on your system. If Python is not installed or if there are issues with the installation, this command may return an error or indicate that Python is not recognized. Similar to Quarto, ensure Python's installation directory is in your PATH. If you have multiple Python versions installed, ensure the correct version's path is prioritized.</p> To check which Jupyter kernels are installed on your system, use:</p> jupyter kernelspec list</span></span></code></pre> This command will list all available kernels, including their names and installation paths. If you've installed a kernel (such as nbstata</code>) but it doesn't appear in this list, there may be an issue with how the kernel was installed or registered with Jupyter.</p> Common issue: Quarto does not recognize the Python installation#</a></h2> If you have problems with Quarto finding the correct Python installation, you can try the following steps (here I have assumed you have installed Anaconda, but the same logic applies to Miniconda or any other Python distribution):</p> Disable app execution aliases for Python</strong>: Open Settings, go to Windows Settings > Apps > Apps & features, and manage execution aliases to disable Python aliases (i.e., turn off the toggle(s) that point to the Microsoft Store app, python.exe</code> and python3.exe</code>). By disabling these, you prevent Windows from redirecting Python commands to the Microsoft Store app, allowing your system to use the Anaconda Python executable instead.</li> Adjusting the PATH order</strong>: go to System Properties > Advanced > Environment Variables, then select the "Path" variable under User variables or System variables and click "Edit". In the Edit Environment Variable window, find the entry for C:\Users\[your-username]\AppData\Local\Microsoft\WindowsApps</code>. You can either remove this entry or move it down the list so that the paths to your Anaconda installation appear above it. This changes the order in which Windows searches these paths for executable files. To check the order of your PATH environment variable, type where python</code> in the Anaconda prompt (usually, it should be installed in C:\ProgramData\anaconda3</code>). Then, add this path to the PATH environment variable and move it to the top of the list, adding also the Scripts</code> folder (usually, C:\ProgramData\anaconda3\Scripts</code>).</li> </ol> ^{1</sup> Stata though has its own matrix-based programming language, Mata</em>. Some of the most popular user-written Stata packages, like reghdfe</code> for regressions with high dimensional fixed effects, are largely written in Mata.</p> </div>}^{2</sup> For open-source languages, like R or Python, Copilot is a game-changer. For Stata, it is less useful, although I have to say that I am pleasantly surprised with the quality of its suggestions relative to ChatGPT 4.</p> </div>}^3</sup>For earlier Stata versions, an alternative would be using the stata_kernel</code> Jupyter kernel. I have not tested it, but I have good references from colleagues. You can find the installation guide here</a>.</p> </div> ^{4</sup> nbstata</code> is a Jupyter kernel. If you install Anaconda, you will also get JupyterLab installed, so you will be able to run Stata code from Jupyter notebooks as well. I personally hate them, but if you don't, it is an alternative.</p> </div>}

Methodology #</a></h1>
The fundamentals

The fundamentals

Montoya va donde brilla (in praise of not belonging)

The accidental prophet

Wellbeing I: Introduction

Predicting socio-economic outcomes with Google Trends

A modern IDE for Stata with VSCode + Quarto

Why Quarto
beautiful</a>.</p>

Step 1: Install Quarto #</a></h2>
Go to https://quarto.org/</a>, click on "Get Started", download the latest Quarto version available, and install it on your machine.</p>
</p>

Step 2: Install VSCode #</a></h2>
Go to https://code.visualstudio.com/download</a>, download the latest VSCode version available, and install it on your machine.</p>
</p>

pablo garcía-guzmán

No country for young men: age-earnings profiles by birth cohort and sex in Spain

A demographic projection of Spain's native-born population: quantifying the stock of second-generation immigrants