[Part 2] Beyond average: How to visualise polarisation in data

Unlocking insights in how polarisation changes over time

In the second case of this series, we dive into customer data to understand how polarising attitudes about different themes changed over 10 years across different countries, achieving the goal to identify relatively stable behaviours towards which businesses can more safely plan in the long-term.

Polarised opinions often change over time, and investigating those changes is key to understanding some human behaviours. In personal relationships, it may seem that when you first became friends or romantical partners with someone, most of your opinions were similar. Over time people sometimes grow apart, realising that their opinions become more polarised, making it difficult to find agreements and live in harmony. When it comes to public opinions, History has shown how growing dissent among ideological lines can cause unrest and instability, but can also fuel revolutions and drive important societal changes.

The second part of this study about visualising polarisation in data adds one more level of complexity: the passage of time. The examples use data from surveys that have asked the same questions over the years.

Case 1-B: Polarisation over time visualised in stacked bars

In Part 1 of this article, case 1-A showed how results from a Diversity & Inclusion survey were visualised to understand diverging opinions. Stacked bars with a divergent colour scale were used:

Data altered from a Diversity & Inclusion survey. The data has been slightly altered and the questions have been changed into a context of colours, so that the charts could still be used as examples while still respecting the respondents’ privacy. The original data involved questions such as “to what extent do you feel welcome/safe at the company?”.

To track the progress of the initiatives implemented in the company, the Benelux team from dentsu international that realised the survey had to repeat the same questions in yearly Diversity & Inclusion surveys. The new visualisation challenge was understanding how opinions changed over time. A second version of this chart was created, including the results of previous surveys in the faded bars identified as S2 (second survey) and S1 (first survey) under each question.

Data altered from Diversity & Inclusion surveys, for privacy reasons.

Besides that, two indicator columns were added at the right side:

1. The percentage of respondents that answered favourably (4 or 5);
2. The variation of that percentage between the second and third surveys in percentage points, showing how much opinions changed over the years.

The two new columns were quite relevant considering that the company had set percentage targets for some of those survey questions. They simplified the main message from the analysis and made the chart not only good for diagnosing what was going on, but also to present the results and encourage action.

With these changes, this chart helped the people involved in Diversity & Inclusion at the company to analyse whether their initiatives, actions and programs were having a quantifiable impact on the feelings of belonging at the company (check Part 1 of this article for more details about this survey). It also helped justify how important it was to keep investing in Diversity & Inclusion at the company and ensure support from the stakeholders.

Case 2: Polarisation over time visualised in stacked columns

This second case had a different approach to polarisation over time, for two reasons:

1. It was relevant to monitor a measure of central tendency like average or median;
2. There were many more data points, specifically 11 years of data.

The original data came from the Customer Connection Systems (CCS), a proprietary survey from dentsu international that aims to deeply understand customer behaviours regarding different brands and media channels, and customer attitudes towards different themes. One of the survey sections shows several attitude statements and respondents must answer in an agreement scale that goes from “strongly disagree” to “strongly agree”. The attitude statements cover topics like advertising, health, leisure, money, fashion, technology, politics/religion/nationality, and work & ambition. Here’s an example of attitude statements asked about health:

1. “I’m always on the look-out for healthier food/drink alternatives”
2. “Organic products are healthier”
3. “I make sure that I eat well-balanced/nutritious meals”

The choice of vertical bars (columns) instead of horizontal bars intended to facilitate interpretation of the changes across multiple years, because that made possible to use a combination of stacked columns to represent the polarisation and a line to represent the average. In general, line charts are more effective than columns to represent changes over time.

In this case, polarisation meant how much more or less neutral the answers about a particular attitude statement became over the years. The following weights were attributed to the answer options: “strongly disagree” (1), “disagree” (2), “neither agree nor disagree” (3), “agree” (4), “strongly agree”(5). The mean or average score was the central tendency of the score for a certain statement, calculated based on the weights.

Besides the average, it was possible to calculate how much more or less polarised a statement became between the earliest and the latest year, in percentage points (pp). A positive value meant that a statement became more polarised (less neutral), and negative values meant that a statement became less polarised (more neutral). The polarisation was simply the neutral percentage (people who responded “neither agree nor disagree” divided by the total of respondents) of the latest year minus the neutral percentage of the earliest year with available data.

Interactive version of the chart showing polarisation over time, created using Google Looker Studio. It was important to democratise this data across strategy teams from multiple countries to understand not only the variations in polarisation, but what statements remained stable over the years. Analysing marketing & media data, we most often focus on the variations, but more long-term planning can be enabled with attitude statements that remain fairly stable over a decade.

The stacked column chart shows one column per year, and each column is split in 3: agree (green), disagree (red), and neutral (grey). The percentage of agreement included respondents that answered “agree” or “strongly agree” with a certain statement, and the percentage of disagreement included respondents that answered “disagree” or “strongly disagree”.

If a statement becomes less polarised, the grey section of the column will increase when compared to previous years. If it becomes more polarised, the grey section will decrease, meaning that more respondents chose the answers in the extremes.

This is a version of the chart shown previously, that highlights the year 2020. Notice how in that year, when most countries were struggling with the global covid-19 pandemic, the grey block representing the percentage of neutral opinions was narrower than any of the other years analysed. Variations over time on the polarisation of opinions can be highly influenced by local and global events.

An alternative approach to calculate polarisation with similar data would be to consider the “extreme” answers only the “strongly agree” and “strongly disagree”, instead of any agree (“agree” & “strongly agree”) and any disagree (“disagree” and “strongly disagree”).

To replicate this kind of chart, it is important to notice that different platforms may require different ways of structuring the data.

How the data was structured to create a version of the stacked column chart in Excel or Powerpoint. There is one year for each agreement type, and there is only one row for each year.

Other platforms like Google Looker Studio may require a different structure to feed a similar stacked column chart. There may be multiple rows for the same year, with one column to define the agreement type and one column with the agreement percentages.

Conclusions

The passage of time adds one level of complexity to the analysis of polarisation. This welcomes different approaches, depending on the amount of data points, the importance of measures of central tendency, the existence of targets or the way the variation over the years will be used to support decisions.

In Part 3 of this article, another level of complexity will be added: multiple categories. Although stacked bars and stacked columns are great to show how divergent opinions change over time, a different approach is needed when the polarisation changes need to be compared across categories.