The visualiization works best on Google Chrome.
People's perceptions of those living in extreme poverty are often overly simplistic. A common image of the extremely poor is that they spend almost all of their incomes on food, and therefore have few consumption choices to make.
Inspired by Esther Duflo & Abhijit Banerjee's 2007 paper 'The Economic Lives of the Poor', we want to challenge this narrative. We want to show that consumption patterns of those living in poverty are complex, and in many ways, reflect the same challenges and trade-offs as in richer societies (e.g., tasty vs. nutritious food).
India has one of the world's largest populations living in extreme poverty. Further, two of our team members have personal connections with India, so we decided to start our project there.
The India Human Development Survey 2011-12 is a nationally representative, multi-topic survey of 42,152 households in 1,503 villages and 971 urban neighborhoods across India. It is a joint project by the University of Maryland, College Park; the National Council of Applied Economic Research (NCAER) in Delhi, and the Indiana University, and the University of Michigan.
Typically, the Consumer Expenditure Survey (CES, conducted by the Indian National Statistical Office) is used to calculate national poverty statistics. The latest round of CES has not yet been published officially due to concerns regarding data quality. However, some argue that the reasons are political.
This makes IHDS 2011-12 and the CES from 2011-12 the latest nationally representative surveys with detailed data on household consumption patterns. Since the IHDS is better documented and includes data on additional household characteristics that we could use in future iterations, we decided to use it.
We did some basic data exploration and extensive preprocessing of the data in R.
Data pre-processing included the following steps:
- Filter data to select households that are poor according to the national poverty line
- Merge labels of consumption items to be displayed in the graph
- Categorize consumption items into groups
- Reshape household-level data to long format
- Harmonize reference periods of consumption items (i.e., rescale consumption of every item to Rupees spent over last 30 days)
- Estimate means using household survey weights
We ended up using the national poverty line as a threshold. However, we had also considered using the international poverty line of 1.90 International USD PPP per day. For this purpose, we calculated the total individual consumption value in Rupees per day and multiplied it with the PPP conversion factor provided by the World Bank and OECD. However, we found that the resulting poverty estimates were different from what was reported from other surveys. Using the poverty indicator already included in the dataset therfore seemed the safer choice.
As we wanted people to reflect on their (mis-)perceptions of the spending patterns of the poor, we decided to use a narrative technique that promotes this kind of self-reflection: we first ask users to guess the share of each consumption category, and then juxtapose users' estimate to the real consumption statistics.
We iterated over several methods to collect user input. We started with simple sliders but found it took users too long to allocate 100%. Also, asking users to make estimates on the percent level seemed to invite false precision (e.g., users likely have no strong opinion on whether food consumption is 63% or 64%, so asking them to make such a precise estimate may be counter-productive). We then shifted to text input, paired with buttons that increment by 5 percentage points. We also added a doughnut chart that updates as users make input so that users have immediate visual feedback of their input. Providing default values for the categories may anchor users' estimates but also helps them to quickly get to the real data. Other ideas were to let users drag-and-drop money icons on different consumption categories or let them build the pie chart themselves but these solutions would have increased the development efforts significantly.
To facilitate the comparison of the user input and reality, we aligned the doughnut charts concentrically and made the color-coding consistent. We also added the exact values as text next to users' estimates.
We also wanted to give users a chance to explore the granularity of our dataset. Therefore, another design choice we made was to use a Zoomable Sunburst chart, building on top of the D3 sample. The users can click on a specific category to 'expand' on the different expenditures within that category, while a click in the center of the donut takes them back to the complete view.
The visualization still has some limitations. The concentric organization of the doughnut/sunbursts facilitates the comparison between user estimates and reality, but eventually, doughnut charts may not be the most effective encoding for visual comparisons (e.g., as discussed in class, angle and area are generally less effective encodings than length and position on a common scale).
Further, collecting user input is still not as intuitive as it could be. For example, users have to do some cognitive work to get to 100% if they change any of the default values, and instructions for how to interact with the visualization could potentially be clearer.
Moises took the lead on the D3 visualization and CSS, Akshit took the lead on user interactions and project management, and Lucas took the lead on data exploration & processing, presentation, and write-up.
We estimate that in total we spent approximately 40 people hours on this assignment.