# The 2021 UK Census

## Your first project: the UK Census

Let's start by importing some data into R. Because R is what is called an object-oriented programming language, we'll always take our information and give it a home inside a named object. There are many different kinds of objects, which you can specify, but usually R will assign a type that seems to fit best.

[If you'd like to explore this all in a bit more depth, you can find a very helpful summary in R for Data Science, chapter 8, ["data import"](https://r4ds.hadley.nz/data-import#reading-data-from-a-file).]{.aside}

In the example below, we're going to read in data from a comma separated value file ("csv") which has rows of information on separate lines in a text file with each column separated by a comma. This is one of the standard plain text file formats. R has a function you can use to import this efficiently called "read.csv". Each line of code in R usually starts with the object, and then follows with instructions on what we're going to put inside it, where that comes from, and how to format it:


```{r, results = 'hide'}
#| include: true
#| label: fig-polar
setwd("/Users/kidwellj/gits/hacking_religion_textbook/hacking_religion")
library(here)  |> suppressPackageStartupMessages()
library(tidyverse)  |> suppressPackageStartupMessages()
here::i_am("chapter_1.qmd")

# Set up local workspace:
if (dir.exists("data") == FALSE) {
  dir.create("data") 
}
if (dir.exists("figures") == FALSE) {
  dir.create("figures") 
}
if (dir.exists("derivedData") == FALSE) {
  dir.create("derivedData")
}

uk_census_2021_religion <- read.csv(here("example_data", "census2021-ts030-rgn.csv")) 
```

## Examining data:

What's in the table? You can take a quick look at either the top of the data frame, or the bottom using one of the following commands:

```{r .column-page}
head(uk_census_2021_religion)
```

This is actually a fairly ugly table, so I'll use an R tool called kable to give you prettier tables in the future, like this:

```{r}
knitr::kable(head(uk_census_2021_religion))
```

You can see how I've nested the previous command inside the `kable` command. For reference, in some cases when you're working with really complex scripts with many different libraries and functions, they may end up with functions that have the same name. You can specify the library where the function is meant to come from by preceding it with :: as we've done `knitr::` above. The same kind of output can be gotten using `tail`:

```{r}
knitr::kable(tail(uk_census_2021_religion))
```

## Parsing and Exploring your data

The first thing you're going to want to do is to take a smaller subset of a large data set, either by filtering out certain columns or rows. Now let's say we want to just work with the data from the West Midlands, and we'd like to omit some of the columns. We can choose a specific range of columns using `select`, like this:

You can use the `filter` command to do this. To give an example, `filter` can pick a single row in the following way:


```{r}
uk_census_2021_religion_wmids <- uk_census_2021_religion %>% filter(geography=="West Midlands")  
```

Now we'll use select in a different way to narrow our data to specific columns that are needed (no totals!).

[Some readers will want to pause here and check out Hadley Wickham's "R For Data Science" book, in the section, ["Data visualisation"](https://r4ds.hadley.nz/data-visualize#introduction) to get a fuller explanation of how to explore your data.]{.aside}

In keeping with my goal to demonstrate data science through examples, we're going to move on to producing some snappy looking charts for this data.


## Making your first data visulation: the humble bar chart

We've got a nice lean set of data, so now it's time to visualise this. We'll start by making a pie chart:

```{r}
uk_census_2021_religion_wmids <- uk_census_2021_religion_wmids %>% select(no_religion:no_response)
uk_census_2021_religion_wmids <- gather(uk_census_2021_religion_wmids)
```


There are two basic ways to do visualisations in R. You can work with basic functions in R, often called "base R" or you can work with an alternative library called ggplot:

### Base R

```{r}
df <- uk_census_2021_religion_wmids[order(uk_census_2021_religion_wmids$value,decreasing = TRUE),]
barplot(height=df$value, names=df$key)
``` 


### GGPlot

```{r}
ggplot(uk_census_2021_religion_wmids, aes(x = key, y = value)) + # <1>
  geom_bar(stat = "identity") # <1>
ggplot(uk_census_2021_religion_wmids, aes(x= reorder(key,-value),value)) + geom_bar(stat ="identity") # <2>
```

1. First we'll plot the data using `ggplot` and then...
2. We'll re-order the column by size.

Let's assume we're working with a data set that doesn't include a "totals" column and that we might want to get sums for each column. This is pretty easy to do in R:

```{r}
uk_census_2021_religion_totals <- uk_census_2021_religion %>% select(no_religion:no_response) # <1>
uk_census_2021_religion_totals <- uk_census_2021_religion_totals %>%
   summarise(across(everything(), ~ sum(., na.rm = TRUE)))  # <2>
uk_census_2021_religion_totals <- gather(uk_census_2021_religion_totals) # <3>
ggplot(uk_census_2021_religion_totals, aes(x= reorder(key,-value),value)) + geom_bar(stat ="identity") # <4>
```

1. First, remove the column with region names and the totals for the regions as we want just integer data. 
2. Second calculate the totals. In this example we use the tidyverse library `dplyr()`, but you can also do this using base R with `colsums()` like this: `uk_census_2021_religion_totals <- colSums(uk_census_2021_religion_totals, na.rm = TRUE)`. The downside with base R is that you'll also need to convert the result into a dataframe for `ggplot` like this: `uk_census_2021_religion_totals <- as.data.frame(uk_census_2021_religion_totals)`
3. In order to visualise this data using ggplot, we need to shift this data from wide to long format. This is a quick job using gather()
4. Now plot it out and have a look!

You might have noticed that these two dataframes give us somewhat different results. But with data science, it's much more interesting to compare these two side-by-side in a visualisation. We can join these two dataframes and plot the bars side by side using `bind()` - which can be done by columns with cbind() and rows using rbind():

```{r}
uk_census_2021_religion_merged <- rbind(uk_census_2021_religion_totals, uk_census_2021_religion_wmids)
```

Do you notice there's going to be a problem here? How can we tell one set from the other? We need to add in something idenfiable first! This isn't too hard to do as we can simply create a new column for each with identifiable information before we bind them:

```{r}
uk_census_2021_religion_totals$dataset <- c("totals")
uk_census_2021_religion_wmids$dataset <- c("wmids")
uk_census_2021_religion_merged <- rbind(uk_census_2021_religion_totals, uk_census_2021_religion_wmids)
```

Now we're ready to plot out our data as a grouped barplot:

```{r}
ggplot(uk_census_2021_religion_merged, aes(fill=dataset, x= reorder(key,-value), value)) + geom_bar(position="dodge", stat ="identity")
```
If you're looking closely, you will notice that I've added two elements to our previous ggplot. I've asked ggplot to fill in the columns with reference to the `dataset` column we've just created. Then I've also asked ggplot to alter the `position="dodge"` which places bars side by side rather than stacked on top of one another. You can give it a try without this instruction to see how this works. We will use stacked bars in a later chapter, so remember this feature.

If you inspect our chart, you can see that we're getting closer, but it's not really that helpful to compare the totals. What we need to do is get percentages that can be compared side by side. This is easy to do using another `dplyr` feature `mutate`:

[It's worth noting that an alternative approach is to leave the numbers intact and simply label them differently so they render as percentages on your charts. You can do this with the `scales() library and the label_percent() function. The downside of this approach is that it won't transfer to tables if you make them.]{.aside}


```{r}
uk_census_2021_religion_totals <- uk_census_2021_religion_totals %>% 
  dplyr::mutate(perc = scales::percent(value / sum(value), accuracy = 0.1, trim = FALSE)) # <3>
uk_census_2021_religion_wmids <- uk_census_2021_religion_wmids %>% 
  dplyr::mutate(perc = scales::percent(value / sum(value), accuracy = 0.1, trim = FALSE)) # <3>
uk_census_2021_religion_merged <- rbind(uk_census_2021_religion_totals, uk_census_2021_religion_wmids)
ggplot(uk_census_2021_religion_merged, aes(fill=dataset, x=key, y=perc)) + geom_bar(position="dodge", stat ="identity")
```
Now you can see a very rough comparison, which sets bars from the W Midlands data and overall data side by side for each category. The same principles that we've used here can be applied to draw in more data. You could, for example, compare census data from different years, e.g. 2001 2011 and 2021. Our use of `dplyr::mutate` above can be repeated to add an infinite number of further series' which can be plotted in bar groups.

We'll draw this data into comparison with later sets in the next chapter. But the one glaring issue which remains for our chart is that it's lacking in really any aesthetic refinements. This is where `ggplot` really shines as a tool as you can add all sorts of things.

These are basically just added to our `ggplot` code. So, for example, let's say we want to improve the colours used for our bars. You can specify the formatting for the fill on the `scale` using `scale_fill_brewer`. This uses a particular tool (and a personal favourite of mine) called `colorbrewer`. Part of my appreciation of this tool is that you can pick colours which are not just visually pleasing, and produce useful contrast / complementary schemes, but you can also work proactively to accommodate colourblindness. Working with colour schemes which can be divergent in a visually obvious way will be even more important when we work on geospatial data and maps in a later chapter.

```{r}
ggplot(uk_census_2021_religion_merged, aes(fill=dataset, x=key, y=perc)) + geom_bar(position="dodge", stat ="identity") + scale_fill_brewer(palette = "Set1")
```
We might also want to add a border to our bars to make them more visually striking (notice the addition of `color` to the geom_bar below. I've also added `reorder()` to the x value to sort descending from the largest to smallest.

[You can find more information about reordering ggplots on the [R Graph gallery](https://r-graph-gallery.com/267-reorder-a-variable-in-ggplot2.html).]{.aside}

```{r}
uk_census_2021_religion_merged$dataset <- factor(uk_census_2021_religion_merged$dataset, levels = c('wmids', 'totals'))
ggplot(uk_census_2021_religion_merged, aes(fill=fct_reorder(dataset, value), x=reorder(key,-value),value, y=perc)) + geom_bar(position="dodge", stat ="identity", colour = "black") + scale_fill_brewer(palette = "Set1")
```
We can fine tune a few other visual features here as well, like adding a title with `ggtitle` and a them with some prettier fonts with `theme_ipsum()` (which requires the `hrbrthemes()` library). We can also remove the x and y axis labels (not the data labels, which are rather important).

```{r}
ggplot(uk_census_2021_religion_merged, aes(fill=fct_reorder(dataset, value), x=reorder(key,-value),value, y=perc)) + geom_bar(position="dodge", stat ="identity", colour = "black") + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the UK: 2021") + xlab("") + ylab("")
```
## Is your chart accurate? Telling the truth in data science

There is some technical work yet to be done fine-tuning the visualisation of our chart here. But I'd like to pause for a moment and consider an ethical question. Is the title of this chart truthful and accurate? On one hand, it is a straight-forward reference to the nature of the question asked on the 2021 census survey instrument. However, as you will see in the next chapter, large data sets from the same year which asked a fairly similar question yield different results. Part of this could be attributed to the amount of non-respose to this specific question which, in the 2021 census is between 5-6% across many demographics. It's possible (though perhaps unlikely) that all those non-responses were Sikh respondents who felt uncomfortable identifying themselves on such a survey. If even half of the non-responses were of this nature, this would dramatically shift the results especially in comparison to other minority groups. So there is some work for us to do here in representing non-response as a category on the census. 

It's equally possible that someone might feel uncertain when answering, but nonetheless land on a particular decision marking "Christian" when they wondered if they should instead tick "no religion. Some surveys attempt to capture uncertainty in this way, asking respondents to mark how confident they are about their answers, but the census hasn't capture this so we simply don't know. If a large portion of respondents in the "Christian" category were hovering between this and another response, again, they might shift their answers when responding on a different day, perhaps having just had a conversation with a friend which shifted their thinking. Even the inertia of survey design can have an effect on this, so responding to other questions in a particular way, thinking about ethnic identity, for example, can prime a person to think about their religious identity in a different or more focussed way, altering their response to the question. For this reason, some survey instruments randomise the order of questions. This hasn't been done on the census (which would have been quite hard work given that most of the instruments were printed hard copies!), so again, we can't really be sure if those answers given are stable. 

Finally, researchers have also found that when people are asked to mark their religious affiliation, sometimes they can prefer to mark more than one answer. A person might consider themselves to be "Muslim" but also "Spiritual but not religious" preferring the combination of those identities. It is also the case that respondents can identify with more unexpected hybrid religious identities, such as "Christian" and "Hindu". The census only allows respondents to tick a single box for the religion category. It is worth noting that, in contrast, the responses for ethnicity allow for combinations. Given that this is the case, it's impossible to know which way a person went at the fork in the road as they were forced to choose just one half of this kind of hybrid identity. Finally, it is interesting to wonder exactly what it means for a person when they tick a box like this. Is it because they attend synagogue on a weekly basis? Some persons would consider weekly attendance at workship a prerequisite for membership in a group, but others would not. Indeed we can infer from surveys and research which aims to track rates of participation in weekly worship that many people who tick boxes for particular religious identities on the census have never attended a worship service at all.

What does this mean for our results? Are they completely unreliable and invalid? I don't think this is the case or that taking a clear-eyed look at the force and stability of our underlying data should be cause for despair. Instead, the most appropriate response is humility. Someone has made a statement which is recorded in the census, of this we can be sure. They felt it to be an accurate response on some level based on the information they had at the time. And with regard to the census, it is a massive, almost completely population level, sample so there is additional validity there. The easiest way to represent all this reality in the form of speaking truthfully about our data is to acknowledge that however valid it may seem, it is nonetheless a snapshot. For this reason, I would always advise that the best title for a chart is one which specifies the data set. We should also probably do something different with those non-responses:

```{r}
ggplot(uk_census_2021_religion_merged, aes(fill=fct_reorder(dataset, value), x=reorder(key,-value),value, y=perc)) + geom_bar(position="dodge", stat ="identity", colour = "black") + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the 2021 Census of England and Wales") + xlab("") + ylab("")
```


Change orientation of X axis labels
+ theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))

Relabel fields
Simplify y-axis labels
Add percentage text to bars (or maybe save for next chapter?)


## Making our script reproducible 

Let's take a moment to review our hacker code. I've just spent some time addressing how we can be truthful in our data science work. We haven't done much yet to talk abour reproducibility. 


## Multifactor Visualisation

One element of R data analysis that can get really interesting is working with multiple variables. Above we've looked at the breakdown of religious affiliation across the whole of England and Wales (Scotland operates an independent census), and by placing this data alongside a specific region, we've already made a basic entry into working with multiple variables but this can get much more interesting. Adding an additional quantative variable (also known as bivariate data) into the mix, however can also generate a lot more information and we have to think about visualising it in different ways which can still communicate with visual clarity in spite of the additional visual noise which is inevitable with enhanced complexity. Let's have a look at the way that religion in England and Wales breaks down by ethnicity.

::: {.callout-tip}
## What is Nomis?

For the UK, census data is made available for programmatic research like this via an organisation called NOMIS. Luckily for us, there is an R library you can use to access nomis directly which greatly simplifies the process of pulling data down from the platform. It's worth noting that if you're not in the UK, there are similar options for other countries. Nearly every R textbook I've ever seen works with USA census data, so you'll find plenty of documentation available on the tools you can use for US Census data. Similarly for the EU, Canada, Austrailia etc.

If you want to draw some data from the nomis platform yourself in R, have a look at the [companion cookbook repository](https://github.com/kidwellj/hacking_religion_cookbook/blob/main/nomis.R). 

:::


```{r}
# Get table of Census 2011 religion data from nomis
z <- readRDS(file = (here("example_data", "z.rds")))

# Filter down to simplified dataset with England / Wales and percentages without totals
uk_census_2011_religion <- filter(z, GEOGRAPHY_NAME=="England and Wales" & RURAL_URBAN_NAME=="Total" & C_RELPUK11_NAME != "All categories: Religion")
# Drop unnecessary columns
uk_census_2011_religion <- select(uk_census_2011_religion, C_RELPUK11_NAME, OBS_VALUE)
# Plot results
plot1 <- ggplot(uk_census_2011_religion, aes(x = C_RELPUK11_NAME, y = OBS_VALUE)) + geom_bar(stat = "identity") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
# ggsave(filename = "plot.png", plot = plot1)

# grab daata from nomis for 2001 census religion / ethnicity
z0 <- readRDS(file = (here("example_data", "z0.rds")))

# select relevant columns
uk_census_2001_religion_ethnicity <- select(z0, GEOGRAPHY_NAME, C_RELPUK11_NAME, C_ETHHUK11_NAME, OBS_VALUE)
# Filter down to simplified dataset with England / Wales and percentages without totals
uk_census_2001_religion_ethnicity <- filter(uk_census_2001_religion_ethnicity, GEOGRAPHY_NAME=="England and Wales" & C_RELPUK11_NAME != "All categories: Religion")
# Simplify data to only include general totals and omit subcategories
uk_census_2001_religion_ethnicity <- uk_census_2001_religion_ethnicity %>% filter(grepl('Total', C_ETHHUK11_NAME))

# grab data from nomis for 2011 census religion / ethnicity table
z1 <- readRDS(file = (here("example_data", "z1.rds")))

# select relevant columns
uk_census_2011_religion_ethnicity <- select(z1, GEOGRAPHY_NAME, C_RELPUK11_NAME, C_ETHPUK11_NAME, OBS_VALUE)
# Filter down to simplified dataset with England / Wales and percentages without totals
uk_census_2011_religion_ethnicity <- filter(uk_census_2011_religion_ethnicity, GEOGRAPHY_NAME=="England and Wales" & C_RELPUK11_NAME != "All categories: Religion" & C_ETHPUK11_NAME != "All categories: Ethnic group")
# Simplify data to only include general totals and omit subcategories
uk_census_2011_religion_ethnicity <- uk_census_2011_religion_ethnicity %>% filter(grepl('Total', C_ETHPUK11_NAME))

# grab data from nomis for 2021 census religion / ethnicity table
z2 <- readRDS(file = (here("example_data", "z2.rds")))

# select relevant columns
uk_census_2021_religion_ethnicity <- select(z2, GEOGRAPHY_NAME, C2021_RELIGION_10_NAME, C2021_ETH_8_NAME, OBS_VALUE)
# Filter down to simplified dataset with England / Wales and percentages without totals
uk_census_2021_religion_ethnicity <- filter(uk_census_2021_religion_ethnicity, GEOGRAPHY_NAME=="England and Wales" & C2021_RELIGION_10_NAME != "Total" & C2021_ETH_8_NAME != "Total")
# 2021 census includes white sub-groups so we need to omit those so we just have totals:
uk_census_2021_religion_ethnicity <- filter(uk_census_2021_religion_ethnicity, C2021_ETH_8_NAME != "White: English, Welsh, Scottish, Northern Irish or British" & C2021_ETH_8_NAME != "White: Irish" & C2021_ETH_8_NAME != "White: Gypsy or Irish Traveller, Roma or Other White")

ggplot(uk_census_2011_religion_ethnicity, aes(fill=C_ETHPUK11_NAME, x=C_RELPUK11_NAME, y=OBS_VALUE)) + geom_bar(position="dodge", stat ="identity", colour = "black") + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the 2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))

```
The trouble with using grouped bars here, as you can see, is that there are quite sharp disparities which make it hard to compare in meaningful ways. We could use logarithmic rather than linear scaling as an option, but this is hard for many general public audiences to apprecaite without guidance. One alternative quick fix is to extract data from "white" respondents which can then be placed in a separate chart with a different scale.

```{r}
# Filter down to simplified dataset with England / Wales and percentages without totals
uk_census_2011_religion_ethnicity_white <- filter(uk_census_2011_religion_ethnicity, C_ETHPUK11_NAME == "White: Total")
uk_census_2011_religion_ethnicity_nonwhite <- filter(uk_census_2011_religion_ethnicity, C_ETHPUK11_NAME != "White: Total")

ggplot(uk_census_2011_religion_ethnicity_nonwhite, aes(fill=C_ETHPUK11_NAME, x=C_RELPUK11_NAME, y=OBS_VALUE)) + geom_bar(position="dodge", stat ="identity", colour = "black") + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the 2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))

```

This still doesn't quite render with as much visual clarity and communication as I'd like. For a better look, we can use a technique in R called "faceting" to create a series of small charts which can be viewed alongside one another.

```{r}
ggplot(uk_census_2011_religion_ethnicity_nonwhite, aes(x=C_RELPUK11_NAME, y=OBS_VALUE)) + geom_bar(position="dodge", stat ="identity", colour = "black") + facet_wrap(~C_ETHPUK11_NAME, ncol = 2) + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the 2011 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
```

For our finale chart, I'd like to take the faceted chart we've just done, and add in totals for the previous two census years (2001 and 2011) so we can see how trends are changing in terms of religious affiliation within ethnic self-identification categories. We'll draw on some techniques we're already developed above using `rbind()` to connect up each of these charts (after we've added a column identifying each chart by the census year). We will also need to use one new technique to change the wording of ethnic categories as this isn't consistent from one census to the next and ggplot will struggle to chart things if the terms being used are exactly the same. We'll use `mutate()` again to accomplish this with some slightly different code.

```{r}
# First add column to each dataframe so we don't lose track of the census it comes from:
uk_census_2001_religion_ethnicity$dataset <- c("2001")
uk_census_2011_religion_ethnicity$dataset <- c("2011")
uk_census_2021_religion_ethnicity$dataset <- c("2021")

# Let's tidy the names of each column:

names(uk_census_2001_religion_ethnicity) <- c("Geography", "Religion", "Ethnicity", "Value", "Year")
names(uk_census_2011_religion_ethnicity) <- c("Geography", "Religion", "Ethnicity", "Value", "Year")
names(uk_census_2021_religion_ethnicity) <- c("Geography", "Religion", "Ethnicity", "Value", "Year")

# Next we need to change the terms using mutate()
uk_census_2001_religion_ethnicity <- uk_census_2001_religion_ethnicity %>% 
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^White: Total$", replacement = "White")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Mixed: Total$", replacement = "Mixed")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Asian: Total$", replacement = "Asian")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Black or Black British: Total$", replacement = "Black")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Chinese or Other ethnic group: Total$", replacement = "Other"))
  
uk_census_2011_religion_ethnicity <- uk_census_2011_religion_ethnicity %>% 
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^White: Total$", replacement = "White")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Mixed/multiple ethnic group: Total$", replacement = "Mixed")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Asian/Asian British: Total$", replacement = "Asian")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Black/African/Caribbean/Black British: Total$", replacement = "Black")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Other ethnic group: Total$", replacement = "Other"))

uk_census_2021_religion_ethnicity <- uk_census_2021_religion_ethnicity %>% 
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^White: Total$", replacement = "White")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Mixed or Multiple ethnic groups$", replacement = "Mixed")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Asian, Asian British or Asian Welsh$", replacement = "Asian")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Black, Black British, Black Welsh, Caribbean or African$", replacement = "Black")) %>%
  mutate(Ethnicity = str_replace_all(Ethnicity, 
            pattern = "^Other ethnic group$", replacement = "Other"))

# Now let's merge the tables:

uk_census_merged_religion_ethnicity <- rbind(uk_census_2021_religion_ethnicity, uk_census_2011_religion_ethnicity)

uk_census_merged_religion_ethnicity <- rbind(uk_census_merged_religion_ethnicity, uk_census_2001_religion_ethnicity)

# As above, we'll split out non-white and white:

uk_census_merged_religion_ethnicity_nonwhite <- filter(uk_census_merged_religion_ethnicity, Ethnicity != "White")

# Time to plot!

ggplot(uk_census_merged_religion_ethnicity_nonwhite, aes(fill=Year, x=Religion, y=Value)) + geom_bar(position="dodge", stat ="identity", colour = "black") + facet_wrap(~Ethnicity, ncol = 2) + scale_fill_brewer(palette = "Set1") + ggtitle("Religious Affiliation in the 2001-2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
```
 
There are a few formatting issues which remain. Our y-axis number labels are in scientific format which isn't really very easy to read. You can use the very powerful and flexible `scales()` library to bring in some more readable formatting of numbers in a variety of places in R including in ggplot visualizations. 

```{r}
library(scales) |> suppressPackageStartupMessages()
ggplot(uk_census_merged_religion_ethnicity_nonwhite, aes(fill=Year, x=Religion, y=Value)) + geom_bar(position="dodge", stat ="identity", colour = "black") + facet_wrap(~Ethnicity, ncol = 2) + scale_fill_brewer(palette = "Set1") + scale_y_continuous(labels = unit_format(unit = "M", scale = 1e-6), breaks = breaks_extended(8)) + ggtitle("Religious Affiliation in the 2001-2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))

# https://ggplot2-book.org/scales-position#sec-position-continuous-breaks

```
This chart shows an increase in almost every category, though it's a bit hard to read in some cases. However, this information is based on the increase in raw numbers. It's possbile that numbers may be going up, but in some cases the percentage share for a particular category has actually gone down. Let's transform and visualise our data as percentages to see what kind of trends we can actually isolate:

```{r}
uk_census_merged_religion_ethnicity <- uk_census_merged_religion_ethnicity %>%
  group_by(Ethnicity, Year) %>%
  dplyr::mutate(Percent = Value/sum(Value))

ggplot(uk_census_merged_religion_ethnicity, aes(fill=Year, x=Religion, y=Percent)) + geom_bar(position="dodge", stat ="identity", colour = "black") + facet_wrap(~Ethnicity, scales="free_x") + scale_fill_brewer(palette = "Set1") + scale_y_continuous(labels = scales::percent) + ggtitle("Religious Affiliation in the 2001-2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))
```
Now you can see why this shift is important - the visualisation tells a completely different story in some cases across the two different charts. In the first, working off raw numbers we see a net increase in Christianity across all categories. But if we take into account the fact that the overall share of population is growing for each of these groups, their actual composition is changing in a different direction. The proportion of each group is declining across the three census periods (albeit with an exception for the "Other" category from 2011 to 2021).

To highlight a few features of this final plot, I've used a specific feature within `facet_wrap` `scales = "free_x"` to let each of the individual facets adjust the total range on the x-axis. Since we're looking at trends here and not absolute values, having correspondence across scales isn't important and this makes for something a bit more visually tidy. I've also shifted the code for `scale_y_continuous` to render values as percentages (rather than millions).

In case you want to print this plot out and hang it on your wall, you can use the ggsave tool to render the chart as an image file:

```{r}
plot1 <- ggplot(uk_census_merged_religion_ethnicity, aes(fill=Year, x=Religion, y=Percent)) + geom_bar(position="dodge", stat ="identity", colour = "black") + facet_wrap(~Ethnicity, scales="free_x") + scale_fill_brewer(palette = "Set1") + scale_y_continuous(labels = scales::percent) + ggtitle("Religious Affiliation in the 2001-2021 Census of England and Wales") + xlab("") + ylab("") + theme(axis.text.x = element_text(angle = 90, vjust = 0.5, hjust=1))

ggsave("chart.png", plot=plot1, width = 8, height = 10, units=c("in"))
```

# References {.unnumbered}

::: {#refs}
:::