Analyzing the Pew Research Center’s Religious Landscape Study (2023-2024)

Background

I was interested in analyzing data about Americans’ faith communities and religious practices over time, and discovered that the Pew Center systematically tracks this data. The most recent survey instrument, involving n=36,908 respondents, was the 2023-24 Religious Landscape Study (RLS), administered on July 17, 2023 – March 4, 2024. Two similar surveys had been administered in 2007 and 2014. While I was hoping to analyze trends over time using all three of these data sets, I ended up focusing on the most recent survey, as it was very time-consuming to translate the raw data into something useful. (So basically, I ran out of time and steam!)

When I first opened up the downloaded file from the Pew website, the data seemed completely opaque. The spreadsheet headers made no sense (e.g. CHNG_A, DIVRELPOP), nor was it clear what the numbers within each cell referred to (Fig 1). My first order of business – before analyzing the survey results – was to figure out how to “read” the data.

 It was obvious that the 30,000+ rows referred to individual respondents, but I needed to crack the code on the headers and values. Fortunately, Pew’s website provided a “codebook” to do this (Fig 2).

With the codebook, I was able to interpret the headers and to translate the numerical data into survey responses. However, I realized that I would need a tool to convert the numerical data back into the survey responses, because that’s what I would want to present in Tableau. It would have taken me forever to manually convert the cells from numerical data to actual survey responses, so I turned to OpenRefine to help me with this task. 

When I started using OpenRefine, it was taking a very long time to load and work on the data, so I realized that I needed to pare down the data set (do I just need more computer memory?). I found the original survey and identified the questions I was most interested in analyzing; I retained those columns of data in the dataset and removed the others. This paring down allowed me to manipulate the data in OpenRefine without being slowed down. I then launched ChatGPT, uploaded the codebook file, and asked for specific formulas to map survey responses onto the numerical data. I spent about an hour trying to implement ChatGPT’s  instructions, getting nowhere and getting increasingly agitated. Just for fun, here are some snippets of my interaction with the incredibly-chirpy AI (Fig 3):

Finally, ChatGPT asked me to upload the raw data file and the codebook file, did some magic, and sent me back a revised data file. I was then able to upload the data to Tableau and get to work!

I decided to focus my initial analysis on two areas: (1) The “popularity” of different religions in the U.S., based on the number of members of each group, and (2) How “religious” (faithful in their beliefs and ritual practices) were members of each religion, generally.  What follows are two examples of my process – and the evolving data visualizations – exploring these two topics. 

Example #1: Presenting the Relative Size of Religious Groups in the U.S.

My first analysis involved looking at the relative population sizes of different religious groups in the U.S. Assuming that the survey methodology was designed to reflect the U.S. population as a whole (parenthetically, I would need to research their survey methodology more closely before officially presenting any of this work), I began by creating a bar chart of the entire survey population (n=36,908) (Fig 4.1). 

The chart had two main problems:

1. It was sorted arbitrarily (actually, based on the structure of the survey), rather than from smallest to largest. This was easy to fix by just manually resorting the bars.
2. There was an excess of white space. One small way I could fix this would be to combine some of the groups (assuming they were sufficiently alike). Here is the way I chose to do this, using the Group function (FIg 4.2)

I combined “Orthodox Christian” and “Other Christian” into one group labeled “Orthodox Christian & Other Christian”, even though this might be a bit of apples and oranges. I basically traded off aesthetics for logical groupings, admittedly an imperfect decision. I also combined “Other faiths” and “Other world religions”, which felt totally legitimate. 

The new sorting and combining of groups made the chart more readable (Fig 4.3), but there was still way too much white space. More problematically, the x-axis (# of respondents) drew attention to the number of survey respondents, whereas I wanted the focus to be on the relative sizes of the groups. 

To solve for this, I played around with alternative visualization. I experimented with two different approaches – a pie chart (Fig 4.4), not particularly successful, and a tree chart (Fig 4.5), a bit better.

Of the two, the tree chart did a better job of giving the viewer a real sense of the relative size of each religious community. However, to improve the readability, I thought it might be helpful to cluster similar groups (e.g., the variety of Christian denominations) visually. I tried to achieve this by using similar colors (Fig 4.6).

I thought that the chart could be further improved by clustering all the Christian denominations in the same general vicinity, not just by using similar colors. However, I couldn’t figure out a way to do this in Tableau, so I was stuck using the layout automatically generated.

Process Example #2: Religious Observance by Faith Community

My second analysis involved showing which faith tradition’s adherents were “most religious”. The survey question poses the following question: “To what extent do you consider yourself a religious person?” (The question was left purposefully broad, without qualifications, and it left me wondering how different people might interpret that quality of “being religious”.)

After playing around with the measurement and other settings, I arrived at a basic bar chart (Fig 5.1). I then realized that two of the categories – “Refused to answer the question” and “Religiously unaffiliated” – were superfluous for my original question, so I reworked the chart to remove them (Fig 5.2)

Because the labels were cut off at the bottom, I chose to reorient the chart so that the full labels were visible (Fig 5.3).

Conclusion

This assignment was a great wake-up call regarding the amount of effort it takes (at least for a novice!) to produce an effective dataviz. I hadn’t previously appreciated the amount of time it would take to simply clean up the data (even very well vetted data from sources like the Pew Foundation) to the point where it was usable. I also realized that while Tableau is incredibly powerful, there are still some quirks and limitations to being able to present data in exactly the way you’d like (for example, not being able to change the placement of different elements in the tree chart). Having said that, I walked away with a real appreciation for the power of Tableau, and, given the time, I’d love to explore other parts of the Pew dataset (e.g. how members of different religious groups feel about controversial political and cultural topics) using this tool.