Multivariate Data

class: middle, center, title-slide

# Multivariate Data
## STAE04: Data Visualization
### Johan Larsson
### The Department of Statistics, Lund University

---

## Visualizing Multivariate Data

visualizations often most intriguing when they show multiple variables

<div class="figure" style="text-align: center">
<img src="images/minard-migration.jpg" alt="The Emigrants of the World, Charles Joseph Minard, 1858." width="78%" />
<p class="caption">The Emigrants of the World, Charles Joseph Minard, 1858.</p>
</div>

---

## 3D

.pull-left-40[
seems like a logical choice for a third continuous variable

unfortunately, 3D visualizations are notoriously hard to read

as a thumb rule, **avoid 3D visualizations**
]

.pull-right-60[
<div class="figure" style="text-align: center">
<img src="10-multivariate-data_files/figure-html/unnamed-chunk-2-1.png" alt="A 3D scatter plot." width="504" />
<p class="caption">A 3D scatter plot.</p>
</div>
]

---

<div class="figure" style="text-align: center">
<img src="10-multivariate-data_files/figure-html/unnamed-chunk-3-1.png" alt="A particularly bad example of a 3D plot. 3D bar charts are never a good idea. Data shows death rates in Virginia in 1940." width="576" />
<p class="caption">A particularly bad example of a 3D plot. 3D bar charts are never a good idea. Data shows death rates in Virginia in 1940.</p>
</div>

---

### 3D Visualizations Are Not All Bad!

3D visualizations may be useful in a few situations:

1. interactive or animated 3D plots
2. multiple plots showing multiple perspectives
3. only general pattern matters: reading individual points is not important
4. data actually represents three-dimensional location data (topography)

<div class="figure" style="text-align: center">
<img src="10-multivariate-data_files/figure-html/unnamed-chunk-4-1.png" alt="The topography of the Maunga Whau Volcano through different perspectives." width="720" />
<p class="caption">The topography of the Maunga Whau Volcano through different perspectives.</p>
</div>

---

## Color

Color is often a good choice, especially with categorical variables.

```r
mtcars <- 
  rownames_to_column(mtcars, "name") %>%
  mutate_at(vars(cyl, gear), as.factor)

*ggplot(mtcars, aes(disp, hp, color = cyl)) +
  geom_point()
```

---

### Color and Continuous Variables

sometimes fine to map color to continuous variables, particularly when data is in
a grid format

```r
# see source code for dataset
ggplot(volcano_long, aes(longitude, latitude, fill = height)) + 
* geom_tile() +
  coord_fixed() # makes sense for latitude, longitude
```

---

### Scatter Plots

Mapping a continuous variable to color in a scatter plot is less appealing.

```r
ggplot(mtcars, aes(disp, hp, color = mpg)) +
  geom_point()
```

---

## Size

Mapping to size only makes sense for continuous variables!

.pull-left-40[
### Area

most common use case: bubble plots

caveat: comparing areas is **hard**
]

.pull-right-60[

```r
ggplot(mtcars, 
       aes(disp, hp, size = mpg)) +
  geom_point()
```

<img src="10-multivariate-data_files/figure-html/unnamed-chunk-9-1.png" width="396" style="display: block; margin: auto;" />
]

---

### Width

mapping to width can be useful but has few use cases

```r
# see source for dataset
ggplot(troops, 
       aes(long, lat, group = group, 
           color = direction, size = survivors)) +
  geom_path(lineend = "round")
```

<div class="figure" style="text-align: center">
<img src="10-multivariate-data_files/figure-html/unnamed-chunk-11-1.png" alt="Basic reproduction of Minard's Napoleon chart." width="720" />
<p class="caption">Basic reproduction of Minard's Napoleon chart.</p>
</div>

---

## Shape

works quite well for categorical variables, but usually not preferable over color

mapping to line shapes is generally better than point shapes

```r
*ggplot(mtcars, aes(disp, hp, shape = cyl)) +
  geom_point()
```

---

## Text

often makes sense when each observation has a meaningful identity

[ggrepel](https://CRAN.R-project.org/package=ggrepel) is a useful package when mapping
to text (`geom_text_repel()`)

.pull-left[

```r
library(ggrepel)

head(msleep, 15) %>%
  ggplot(aes(brainwt, 
             sleep_total, 
             label = name)) +
    geom_point() +
*   geom_text_repel() +
    scale_x_log10()
```
]

.pull-right[
<img src="10-multivariate-data_files/figure-html/unnamed-chunk-13-1.png" width="345.6" style="display: block; margin: auto;" />

]

---

## Facets

splits and plots datasets in small multiples

### Wrap

use `facet_wrap()` when you have **one** variable to facet on.

```r
ggplot(mtcars, aes(disp, hp)) +
  geom_point() +
* facet_wrap(vars(cyl))
```

---

### Grid

Use `facet_grid()` when you have **two** variables to facet on.

```r
ggplot(mtcars, aes(disp, hp)) +
  geom_point() +
* facet_grid(vars(cyl), vars(gear))
```

---

### Facets and Continuous Variables

If sacrificing detail is OK, you can facet by transforming a continuous
variable into an ordinal one.

```r
mtcars %>%
* mutate(hp_cat = cut_interval(hp, 3)) %>%
  ggplot(aes(mpg, disp)) +
    geom_point() +
    facet_wrap(vars(hp_cat))
```

---

## Combining Our Tools

Putting together these building blocks allows us to create interesting and complex
visualizations.

```r
tail(mtcars, 15) %>%
  ggplot(aes(hp, wt, color = cyl, label = name)) +
    geom_point() +
    geom_text_repel() +
    facet_wrap(vars(gear))
```