Chapter 5 Univariate Data Exploration

“Exploratory data analysis is graphical detective work.” - John Tukey

Exploratory Data Analysis (also known as EDA) is largely a visual technique based on the human trait of pattern recognition. The purpose of EDA is simple: learn about your data by visualizing it.

Why is Exploratory Data Analysis (EDA) useful? Because getting to know a dataset is a key step towards making sense of it, and EDA is a great way to familiarize oneself with a data set. EDA is fast, efficient, and intuitive.

Think of the word “Exploratory” in terms of “Hypothesis-Building”. In other words, modeling and statistical inference testing should come after EDA. You can use EDA to jump-start ideas on “what do I do with my data?”!

“Perfect data are boring. Flawed data, on the other hand, are interesting and mysterious. Perfect data don’t get asked out on a second date.” - John Volckens

5.1 Ch. 5 Objectives

This chapter is designed around the following learning objectives. Upon completing this Chapter, you should be able to:

• Define and interpret the location, dispersion, and shape of univariate data distributions
• Calculate quantiles and basic descriptive statistics for univariate data
• Create a cumulative distribution plot and extract quantile information from it
• Create and interpret data from a histogram
• Create and interpret data from a boxplot
• Describe the differences between and the strengths/weaknesess of summary and enumerative plots
• Define skewness and its effect on measures of spread and central tendency in qualitative terms
• Create a time-series plot and identify shifts in location and dispersion
• Define autocorrelation, quantitatively and qualitatively
• Identify time lags where autocorrelation is significant, using autocorrelation and partial autocorrelation plots

5.2 Univariate Data

Univariate means that “only one variable” is being considered. Just like medical doctors, statisticians enjoy the use fancy words to describe basic things.

Univariate data analyses describe ways to discover features about a single variable or quantity. While simple, univariate analyses are a great starting point for EDA because they allow us to isolate a variable for inspection. The variation in diameter of a mass-produced component, pollutant concentration in the atmosphere, or the rate of a beating heart are all examples of things we might examine in a univariate sense. This is, of course, a potentially risky procedure because, as engineers, we are taught about mechanisms and dependencies that imply that variable A is inextricably linked to variable B through some physical process. That’s OK to admit. Univariate analyses are still useful. Trust me for now, or skip ahead to multivariate analyses or modeling—your choice.

5.2.1 Location, Dispersion, and Shape

Univariate EDA often begins with an attempt to discover three important properties about an observed variable: its location, dispersion, and shape. Once you define these properties (the “what”), you can begin to probe the underlying causes for these properties (the “how” and “why”). Discovering “how” and “why” of a variable’s location, dispersion, and shape might sound simple, and yet, answering such questions often represents the pinnacle of scientific discovery; they give out Nobel Prizes for that stuff! Let’s start with “what” (location, dispersion, and shape) and build from there.

5.2.1.1 Location

The location of univariate data means: where do the values or observations fall? Do the values tend to be large or small, based on what you know about the variable? Do the observations have a central tendency - a region where values are more likely to show up?

5.2.1.2 Dispersion

The dispersion of the data refers to its variability. Are the values tightly bound in a small range, or do they vary widely from one observation to the next? Note that the phrase “varies widely” is contextual. The variation in the cost of an ice cream cone from one location to the next might look small to you, but could mean the difference between joy and sorrow for a 10-year-old with only $1.50 in their pocket.

5.2.1.3 Shape

The shape of the distribution is actually a combination of location and dispersion but with some mathematical nuance. Knowing the shape of your data distribution means that you have insight into its probability density function. Once you know a distribution’s shape, you can model it. And if you can model the distribution, you can begin to make inferences about it (e.g., extrapolations, predictions). More on this later in Modeling.

The shape of a distribution of data is often categorized based on whether it follows that of a reference distribution, of which there are many types. You can think of reference distributions like species of living organisms; there are lots out there, but once you categorize one, you can likely predict its behavior. In other words, if the shape of your data matches a reference distribution, most of your modeling work is already done! Examples of reference distributions include the uniform distribution, normal distribution, and the lognormal distribution. More on different types of reference distributions here.

5.2.1.3.1 Example: Location and Dispersion

Let’s plan a camping trip. Our trip is purely hypothetical, so let’s not worry about costs, logistics, or other important factors. For this exercise, we only care about comfort while outdoors. When I think of being comfortable outdoors, the first thing that comes to mind is temperature: Did I bring the proper clothing?

We will consider going camping in two lovely spots: the forest preserves along the Na Pali coast in Kauai, Hawaii or the sunny hiking and climbing region around Jack’s Canyon in southwest Colorado. Let’s examine the location and dispersion of hourly temperatures in these two regions for the month of July, 2010. We can access this information from the NOAA Climate Data Center that provides weather data for these two regions. Once we know the location and dispersion of these data, we can decide what clothes to bring.

In Figure 5.1, I use geom_jitter() to plot the hourly temperatures measured at these two regions for July 2010. For this plot, we define three aesthetics:

• y = location
• x = temperature
• color = location (not necessary for the plot, but the color between variables can be helpful)

Figure 5.1: Hourly temperature levels in Colorado and Hawaii for the month of July, 2010.

First, if we are picking spots to go camping, both locations have daily average temperatures that seem very pleasant. The average is about 78.9 \(^\circ\)F in Colorado and 78.6 \(^\circ\)F in Hawaii. When we calculate an average value, we are creating an indicator of a variable’s location (in this case the central tendency). But if we look at the dispersion of temperature observations (i.e., how the temperatures vary across the month), we can make an important distinction: the range of observed temperature values in Hawaii is fairly narrow, whereas the range in Colorado spans from 64.4 to 91.8 \(^\circ\)F!

The conclusion to be drawn here is that the central tendencies of hourly temperature are nearly identical between Hawaii and Colorado, but the dispersion of the temperature data suggests that you might want to pack more layers to be comfortable in Colorado. See Figure 5.2.

You will discover that there are many ways to communicate location (e.g., mode(), mean(), median()) and dispersion (e.g., range(), interquartile range IQR(), and standard deviation sd()). We will discuss several such descriptors in this course.

Figure 5.2: The temperature data from Hawaii and Colorado have similar locations (central tendencies) but differing dispersion (spread).

5.3 Quantiles

Let’s assume you have a sample of univariate data. A good starting point for exploring these data is to break them into quantiles and extract some basic information. Quantiles allow you to see things like the start, middle, and end rather quickly because the first step of quantile calculation is to sort your data from smallest to largest value.

A “quantile” represents a fractional portion of an ordered, univariate distribution.

Quantiles break a set of observations into similarly sized “chunks”, wherein each chunk represents an equal fraction of the total distribution, from start to finish. If you line your data up from smallest to largest and then slice that list into sections, you have created a set of quantiles. For example, when a distribution is broken into 10 equal chunks, or deciles, the first quantile (the 10^th% or the 0.1 fraction) represents the value that bounds the lower 10% of the observed data; the second quantile (0.2 fraction) represents the 20^th% value for the observed data, and so on.

There are two important aspects about quantiles to remember: (1) each quantile is defined only by its upper-end value; and (2) quantiles are defined after the data have been rank-ordered from lowest to highest value.

Quantiles are often used to communicate descriptive statistics for univariate data.

The two most extreme quantiles define the range:

• Minimum,min(): the 0% or lowest value; the zeroth quantile

• Maximum,max(): the 100^th percentile (or 1.0 in fractional terms); the highest value observed; the n^th quantile

If you break a distribution into quarters, you have created quartiles. We calculate quantiles using stats::quantile(), which comes loaded with base R. This function takes a numeric vector x = as a formal argument and, by default, returns quartiles (including min and max) for the data. If you want to extract specific quantile values, you can specify them with the optional probs = argument (default: probs = seq(from = 0, to = 1, by = 0.25)).

Let’s generate a sample of random numbers between 0 and 100 and break the resulting data into quartiles. A random number generator, using the R function runif(), will create a uniform distribution across the sample range provided because all values within that range have equal probability of being chosen. With that in mind, you can probably guess what the quartiles will look like, given a sufficiently large sample…

# "set seed" to generate the same results each time
set.seed(1)
# create uniform distribution
univar <- runif(n = 1000, 
                min = 0, 
                max = 100) 
# split uniform distribution into quantiles
univar_quart <- quantile(univar, 
                         probs = seq(0, 1, 0.25)) %>%
  round(1) 

Figure 5.3: Quartiles estimated from n=1000 samples of a uniform distribution between 0 and 100.

As expected, the 1^st, 2^nd, 3^rd, and 4^th quartiles from a uniform distribution between 0 and 100 fall into predictable chunks of approximately 25, 50, 75, and 100.

• The first, or lower, quartile contains the lower 25% of the distribution. In quantile terms, we define this quartile by its upper value, which occurs at the 25^th percentile (or the 0.25 quantile).

• The second quartile contains the 25^th to the 50^th percentile values of the data. But, remember, that we define this quartile at the upper end of that range: the 0.5 quantile. Note the 0.5 quantile also represents the median of the distribution.

• The third quartile contains data for the 50^th to the 75^th percentile values (0.5 to 0.75 quantiles).

• The fourth quartile contains the upper 25% of the distribution: 0.75 to 1.0 quantiles.

5.3.1 Descriptive Statistics

Quantiles allow us to calculate several important descriptive statistics for univariate data. For example, quartile calculations for highway fuel effeciencies from the mpg dataset allows us to report the following descriptive statistics:

Table 5.1: Table 5.2: Quantiles and descriptive statistics for hwy fuel efficiency from ggplot::mpg$hwy

Quantile	Descriptor	Example Values
0	minimum	12
1	maximum	44
0.5	median	24
0.25	25th%	18
0.75	75th%	27
0.75 - 0.25	IQR	9

A couple notes: You might wonder why I didn’t include summary statistics like the mean() or standard deviation, sd(), in this list. I like using these descriptors, but their use implies that one knows the type or “shape” of the distribution being described (e.g., normal, log-normal, bi-modal). Use of descriptors like mean and standard deviation are very useful when applied properly but can be misleading when the distribution is skewed (more on that later).

Quantile descriptors, on the other hand, are agnostic to the type of distribution they describe. For example, the median is ALWAYS at the 0.5 quantile, or 50^th percentile, regardless of the shape of the distribution.

The last descriptor in Table 5.1 is the interquartile range, or IQR for short. The IQR describes the spread of the data and communicates the range of values needed to go from the 25^th% to the 75^th% of the distribution. The IQR spans the “middle part” of the distribution. The IQR is similar in concept to a standard deviation but makes no assumptions about the type or shape of the distribution in question. You can calculate an IQR by subtracting the 0.75 quantile from the 0.25 quantile, or by using stats::IQR().

Now let’s create quantiles from a normal distribution of data with a mean of 50 and standard deviation of 15. We’ll start by randomly sampling 1000 values, using rnorm(), and then arranging them with the quanitle() function. Note that the second half of the code chunk below shows how to calculate quantiles manually.

# "set seed" to generate the same results each time
set.seed(2) 
# create normal distribution
norm_dist <- rnorm(1000, mean = 50, sd = 15)
# split normal distribution into quantiles and output tabular result
quantile(norm_dist, probs = seq(0, 1, 0.1)) %>% round(0)

##   0%  10%  20%  30%  40%  50%  60%  70%  80%  90% 100% 
##    9   32   38   43   46   51   54   59   64   71   95

# manual method for calculating quantiles; start with norms_dist values
norm_manual <- tibble::tibble(
  sample_data = norm_dist) %>%
  # sort the data using
  dplyr::arrange(sample_data) %>%  
  # calculate cumulative fraction for each entry
  dplyr::mutate(cum_frac = seq.int(from = 1 / length(norm_dist), 
                     to = 1,
                     by = 1 / length(norm_dist)))

# create a quantile sequence corresponding to deciles
deciles <- seq.int(from = 0, to = 1, by = 0.1)

# extract quantiles that match the decile values created above 
normal_data_deciles <- norm_manual %>% 
  dplyr::filter(cum_frac %in% deciles) 
# %in% means: return whatever "matches" between these two vectors

Normal distributions are more complicated than uniform distributions, and it’s hard to get a lot out of the decile summary when shown in tabular format. This brings us to my favorite part of exploratory data analysis: data visualization.

5.4 Univariate Data Visualization

Data visualization is a powerful technique for exploring data. Visualizing univariate data is the first step in getting to know your data, once they have been imported and wrangled.

Below, I introduce formal approaches to visualizing the location, dispersion, and shape of univariate data. Later, we will explore relationships between two or more variables in multivariate data analyses. The techniques shown below are simple, powerful, and often underutilized.

5.4.1 Summary vs. Enumerative Plots

Univariate data visualizations can be broken down into two categories: summary plots and enumerative plots. A summary plot shows you the data summarized in some way; a data analysis or manipulation happens prior generating a summary plot. Enumerative is a fancy word meaning “to name or show” all of the data without modification. Enumerative plots visualize all the data in a single plot. There are strengths and weaknesses to both:

• Summary plots
  • Strengths: Easy to read; easy to see the “main points”.
  • Weaknesses: Don’t show all of the data; only show you a summary. You are more likely to miss nuances or special features when looking at a summary plot.
• Enumerative plots
  • Strengths: Show all of the data without manipulation. This is considered a strength because many analysts like to see raw data; “the data don’t lie!”
  • Weaknesses: Can take more time and energy to understand. Sometimes more data mean more work for your brain.

5.4.2 Cumulative Distribution Plot

A cumulative distribution plot is an enumerative plot that lays out the raw quantiles, typically on the y-axis, against the observed data on the x-axis. Because these plots show all the data, each observation gets assigned a unit quantile; if there are 100 data points in a sample, each datum represents 1% of the total, or a quantile of 0.01 . Thus, quantiles are shown on the y-axis, and the x-axis shows the observed data, ordered from smallest to largest value.

The cumulative distribution plot shows the rolling fraction (or percent) of data on the y-axis that are “less than or equal to” the x-axis values of that data.

An annotated cumulative distribution plot is shown below for a sample of temperature data that follow a normal distribution. For this plot, I calculated the unit_quantiles manually to show you how this can be done:

# randomly sample two variables from normal distributions with `rnorm()`
normal_data <- tibble::tibble(a = rnorm(n=1000, mean = 15, sd = 5),
                              b = rnorm(n=1000, mean = 10, sd = 3))
# create unit quantile values for the `normal_data`
normal_data_cdf <- normal_data %>%
  dplyr::select(a) %>%
  dplyr::arrange(a) %>%
  dplyr::mutate(unit_quantiles = seq.int(from = 1/length(a), 
                                         to = 1,
                                         by = 1/length(a)))

Figure 5.4: Annotated Cumulative Distribution Plot for Normally-Distributed Temperature Data

As you can see, there is a lot of information contained in a cumulative distribution plot! We can see the range (minimum to maximum), median (0.5 quantile), and any other percentile values for the distribution. We can quickly pick out any value and find out what fraction of the data are below it. For example, picking out the quantile of 0.9 on the y-axis shows that 90% of the observed temperature values are below 21 \(^\circ\)C. In fact, most of the basic descriptive statistics can be accessed on a cumulative distribution plot:

• range: minimum and maximum values at quantile 0 and 1.0, respectively
  
• median: x-axis value at quantile 0.5
  
• 25^th and 75^th percentiles (or quartiles): x-axis values at quantile 0.25 and 0.75, respectively
  
• interquartile range (IQR): x-axis distance between quantiles 0.25 and 0.75

Cumulative distribution plots also reveal details about the dispersion of the observed data, and, more subtly, about the nature or shape of underlying data distribution. For example, in the temperature example above (Figure 5.4), the curve has geometric symmetry about the median; this symmetry indicates that the dispersion of the data is equal as you move away from the median in either direction.

5.4.2.1 Example: Cumulative Distribution Plot

The ggplot2 package can create cumulative distribution plots with different geom options:

• geom_step(aes(x = your_univariate_data), stat = "ecdf")
  
• stat_ecdf(aes(x = your_univariate_data))

In the example below, we create a cumulative distribution plot of annual salaries reported by individuals with a degree in Mechanical Engineering in the United States. These data come from the NSF Survey of College Graduates. I’ve also taken the liberty of identifying the minimum, maximum, and quartile values on the plot.

# create cumulative fractions 
ordered_salaries <- raw_salaries %>%
  dplyr::select(salary) %>%
  dplyr::filter(salary < 500000) %>% # to remove numeric identifiers 
  dplyr::arrange(salary) %>%
  dplyr::mutate(cum_pct = seq.int(from = 1/length(salary),
                                  to = 1, 
                                  by = 1/length(salary))) 

# alternate method if you don't want to calculate cumulative fractions
ordered_salaries %>% 
  ggplot2::ggplot(mapping = aes(x = salary)) +
  geom_step(stat = "ecdf") + # "empirical cumulative distribution function"
  labs(x = "Salaries of ME Graduates", y = "Cumulative Fraction") +
  scale_y_continuous(limits = c(-0.05, 1.03), 
                     expand = c(0,0),
                     breaks = seq(from = 0, 
                                  to = 1, 
                                  by = 0.1)) +
  scale_x_continuous(labels = scales::label_dollar(scale = 0.001, 
                                                   prefix = '$', 
                                                   suffix = 'k'),
                     minor_breaks = seq(from = 0,
                                        to = 450000,
                                        by = 10000))+
  geom_segment(data = data.frame(x = quantile(ordered_salaries$salary),
                                 y = rep.int(-.05, 5),
                                 xend = quantile(ordered_salaries$salary),
                                 yend = seq(from = 0, to = 1, by = 0.25)),
               aes(x = x, y = y, xend = xend, yend = yend),
               color = "red",
               linetype = "dashed") +
  theme_bw()

ggsave("./images/cdf_me_salaries.png", dpi = 150)

Figure 5.5: Annual Salaries (2017) for US Graduates (all ages) with a Mechanical Engineering Degree

The good news is that if you can survive the undergraduate grind in Mechanical Engineering, you have a reasonable chance of making a great living wage. The median salary is $ 96,000. Note this is for graduates of all ages across the country in the year 2017. Don’t expect that kind of paycheck on your first day on the job!

Look at the “tails” of the curve in Figure 5.5; the shape of the curves on either side of the median are not symmetrical. One tail looks longer than the other. This lack of symmetry means that the data are skewed. Why is that? We will explore skewed data at several points in this course and discuss some of the mechanisms that cause data to be skewed.

5.4.3 Histogram

A histogram is a summary plot of counts, or frequency of observations, as a function of magnitude, or levels. Histograms are useful because they allow you to visualize the spread and shape of your data as a distribution. A histogram can tell you several important things about a variable:

• The location or central tendency: what’s the most common value?
• The approximate range of the data: what are the maximum and minimum values?
• The dispersion: how spread out are the data?
• The nature of the distribution: do the data appear normally distributed or skewed?
• The presence or absence of outliers in the data: what are those observations doing way out there?

Below is a basic histogram that I’ve annotated to show some key features. This histogram was created by sampling a normal distribution rnorm(). I created the plot using geom_histogram(), which requires only an x = aesthetic to run.

Examination of this histogram quickly reveals a lot about the underlying data. For example, the central tendency is located around a value of ~30 and the majority of the data appear to fall between a range of 20 - 40.

Figure 5.6: A histogram is a plot of counts, or frequency of observations, as a function of magnitude, or levels, of univariate data.

5.4.3.1 What to look for in a histogram?

In addition to being able to “see” key data descriptors such as mean, range, and spread, a histogram also allows one (a) to get a feel for skewness, that is whether the distribution is symmetric about the central tendency, (b) to see outliers, and (c) to visualize other potential interesting features in the dataset.

Figure 5.7: A histogram depicting a skewed distribution.

Figure 5.7 depicts just such a skewed distribution. Examination of Figure 5.7 also reveals the presence of potential outliers in the data: the apparent spike in observations around x = 200. Outliers are interesting features and should neither be ignored nor deleted outright. Outliers exist for some reason, and usually only through detective work or mechanistic knowledge can you elucidate their source. We will delve into managing outliers in Chapter 9; for now, it’s enough to know that they exist and may warrant attention.

In Figure 5.7, the central tendency is evident, but the spread of the data is not symmetrical about that central tendency. One outcome of skewed data is that traditional location measures of central tendency such as mode, median, and mean do not agree with each other. According to Figure 5.7, the central tendency calculations differ based on type of location measure.

Figure 5.8: Describing the spread of skewed data with a simple standard deviation can lead to confusion and trouble: in this case, the existence of physically impossible values.

Another, potentially more troubling, outcome of skewed data is that certain measures of the spread of the data such as the standard deviation can lead to misleading conclusions. For example, if we calculate and extend “one standard deviation” to the mean of the data in Figure 5.8, this would imply the existence of negative values (87 - 113 = -26), which is often impossible in the case of physical measurements. This is a common mistake in data analysis: the application of “normal” descriptors like mean and standard deviation to “non-normal” data.

Skewed distributions can be more challenging to handle because they are less predictable than normal distributions. That said, statisticians have figured out how to model just about any type of distribution, so don’t fret. In any case, we can always rely on quantile calculations for descriptors such as median and IQR because those descriptors do not make assumptions about the shape of the distribution.

5.4.3.2 What causes skewed histograms?

I want to spend some time thinking about mechanisms that cause variability in an observation. These mechanisms are often responsible for whether a set of observations appears normally distributed or skewed. We will focus on two types of distributions for now: normal and log-normal data. Many other types of data distributions exist in the real world.

A general “rule of thumb” is that additive variability tends to produce normally distributed data, whereas mechanisms that cause multiplicative variability tend to produce skewed data—often log-normal, as seen in Figure 5.8). By “additive”, I mean that variable x tends to vary in a “plus/minus” sense about its central tendency. Examples of additive variability include the distribution of heights measured in a sample of 3^rd graders or the variability in the mass of a 3D-printed part produced 100 times by the same machine.

Multiplicative (or log-normal) variability arises when the mechanism(s) controlling the variation of x are multipliers (or divisors) of x. Many real-world phenomena create multiplicative variability in observed data: the strength of a WiFi signal at different locations within a building, the magnitude of earthquakes measured at a given position on the Earth’s surface, the size of rocks found in a section of a riverbed, or the size of particles found in air. All of these phenomena tend to be governed by multiplicative factors. In other words, all of these observations are controlled by mechanisms that suggest \(x = a * b * c\) not \(x = a\cdot b\cdot c\). Explicit examples of multiplicative variability and how it leads to log-normal data are provided here.

5.4.3.3 Probability Density Function

The histogram is really an attempt to visualize the probability density function (pdf) for a distribution of univariate data. The pdf is a mathematical representation of the likelihood of sampling a given value from the distribution, assuming a random draw. Unlike the histogram (which bins data into groups), the pdf is a continuous function that can be solved at any datum. A pdf can be expressed empirically (as a numeric approximation) or as an exact analytic equation (in the case normal, lognormal, and other known distributional forms).

5.4.4 Boxplot

The boxplot is summary plot that allows one to view a distribution by its quartiles (Figure 5.9). The “box” in a boxplot consists of three parallel lines, usually oriented vertically, connected on their sides to form a box. The outer box lines, also called the “hinges”, represent the 0.25 and 0.75 quantiles, and the inner line represents the 0.5 quantile, or median. Thus, the height of the box depicts the inter-quartile range of the data.

Boxplots are often drawn with lines extending out from each end of the box called “whiskers”. These whisker lines vary in their representation. For some plots, the whiskers represent either the maximum and minimum values, respectively, or a distance of \(1.5\cdot IQR\) if either the maximum or minimum value are larger than \(1.5\cdot IQR\). Other times the whiskers represent the 0.9 or 0.95 quantiles or \(\frac{1.58\cdot IQR}{\sqrt(2)}\). Because of these variations, it’s best to define the whiskers for the viewer either in the text or the Figure caption.

Here is the R definition from the help file when you type ?geom_boxplot into the console:

The upper whisker extends from the hinge to the largest value no further than 1.5 * IQR from the hinge (where IQR is the inter-quartile range, or distance between the first and third quartiles). The lower whisker extends from the hinge to the smallest value at most 1.5 * IQR of the hinge. Data beyond the end of the whiskers are called “outlying” points and are plotted individually.

Figure 5.9: A generic boxplot. The box lines represent the lower, middle, and upper quartiles, respectively. The whiskers represent 1.5*IQR in each direction.

Boxplots can be created for a single univariate distribution, as shown in Figure 5.9, but the strength of the boxplot really comes from drawing several together, so that one univariate distribution can be compared to another. In my opinion, single boxplots are quite boring by themselves, and the same information presented by a boxplot is just as easy to get from a cumulative distribution plot. Nonetheless, I think it’s faster to view quartiles and the IQR on the boxplot, so a plot with multiple boxplots has advantages. For example, let’s look at the data from Figure 5.1 as two boxplots.

Figure 5.10: Boxplots of hourly temperatures in CO and HI for July, 2010. Whiskers represent 1.5*IQR.

Sometimes, it’s helpful to combine enumerative and summary plots together. For example, we can add a Boxplot layer to Figure 5.1 so that we can visualize quartiles in addition to the location, dispersion, and shape of the raw data. The ability to layer geoms is one of the many strengths of the ggplot2 package.

ggplot(data = noaa_temp,
       mapping = aes(y = temp_hr_f, 
                     x = location, 
                     fill = location)) +
  geom_boxplot(width = 0.55) +
  geom_jitter(width = 0.25,
              alpha = 0.2,
              shape = "circle filled",
              color = "black") +
  labs(y = "Region",
       x = "Temperature (°F)") +
  theme_bw() 

Figure 5.11: Boxplots of hourly temperatures in CO and HI for July, 2010

5.4.5 Time-Series Plot

The time-series plot (also called a “run sequence” plot) is something you have undoubtedly seen before. This plot, typically shown as an enumerative plot, depicts the value of a variable on the y-axis plotted in the sequence that the observations were made on the x-axis, often shown as units of time. A time-series plot allows you to do several things:

• Identify shifts in “location” of the variable: when and by how much do values change over time?
• Identify shifts in “variation” of the variable: does the data get more/less noisy over time?
• Identify (potential) outliers and when they occurred
• Identify the timing of events!

Let’s plot the same temperature data from Hawaii as shown in Figure 5.1 as a time series. We will use geom_point() to add the data and geom_line() to connect the dots for this plot.

Figure 5.12: Time series of hourly temperature measurements in Kauai, Hawaii for July 2010.

What can we see in these data? Here are some quick conclusions:

1. The temperature appears to rise and fall in a predictable pattern each day—duh!

2. Over the course of the month, we can see a gradual increase in daily mean temperature.

Neither of these conclusions are groundbreaking, but they both allude to an important outcome: there are patterns to be seen in these data! None of those patterns were evident in the cumulative distribution or histogram plots. If we zoom in on just the first 5 days of the month, we see the daily temperature patterns in more detail.

Figure 5.13: Time series of hourly temperature measurements in Kauai, Hawaii for July 1-5, 2010.

5.4.6 Autocorrelation Plot

Autocorrelation, or serial correlation, means correlated with oneself across time. Autocorrelation is a concept that suggests “whatever is happening in this moment is still likely to be happening in the next moment.” Lots of things are autocorrelated: your mood right now is likely related to your mood 15 minutes from now (or 15 minutes ago). The weather outside at 08:00 on a given day is a good predictor of the weather outside an hour later at 09:00, or two hours later at 10:00, and so on, but less so in fourteen days hence.

Almost all physical phenomena have a degree of autocorrelation and that’s important because if a datapoint at time \(i\) is correlated with time \(i + 1\), then those two datapoints are NOT independent. Yet, many statistical tests assume that the underlying data ARE independent from one another! Thus, autocorrelation can pose a problem when we want to make inferences using data collected over a span of time. Fortunately, we have ways of detecting the presence and relative strength of autocorrelation with a simple exploratory plot.

The autocorrelation plot is a summary plot that gives the Pearson correlation coefficient (r) between all measurements (x_i) and their lags (x_i+n), where x represents an observation and i represents a point in time. You can learn more about correlation coefficients (and how to calculate them) from the cor() function: type ?cor into the console.

An autocorrelation plot can help answer the following questions:

• Are the data correlated with each other over time?
  • Note: if the answer is yes, then your data are not necessarily independent measures.
• Are there patterns (periodicity) to discover?
  • hourly?
    
  • daily?
    
  • weekly?
    
  • seasonally?

Let’s examine the Hawaii temperature data shown in Figures 5.1 and 5.12 using an autocorrelation plot. The ggplot2 package doesn’t have explicit geoms for autocorrelation plots (sigh), but there are simple alternatives.

• The acf() function from the stats package (base R)
• The ggAcf() function in the forecast package (requires a time-series object ts)

Both of these functions calculate autocorrelation lags and plot them through time, wherein each unit of time represents the duration of a single lag. Here, the y-axis represents the Pearson correlation coefficient and the x-axis represents time lags for the basic unit of time in which the data are arranged.

# subset data to contain only measures from Hawaii
noaa_temp_hi <- noaa_temp %>% 
  dplyr::filter(location == "Hawaii")

# call the autocorrelation plot from base R `stats` package
stats::acf(noaa_temp_hi$temp_hr_f,
           main = " " ,
           xlab = "Lag (hours)",
           ylab = "Correlation coefficient",
           lwd = 2) 

The plot suggests a fairly strong level of autocorrelation in these temperature data. We should not be surprised, especially after seeing the time series shown in Figures 5.12 and 5.13. Can you explain why the Pearson correlation coefficient is negative at a 12-hr lag? Why does the coefficient trend back upwards to nearly perfect correlation at 24-hr lags?

Other notes about autocorrelation plots:

• Bars above the blue horizontal lines indicate that autocorrelation is worth attention.
• One important implication of autocorrelated data is that the data wihtin a given lag duration are not independent.
  
• This lack of independence violates the assumptions of many modeling approaches. This is potentially bad.

5.4.7 Partial Autocorrelation Plot

The partial autocorrelation plot is a summary plot that gives the correlation coefficient (r) between all timepoints (x_i) and their lags (x_i+n) after accounting for autocorrelation for all previous lags.

The partial autocorrelation plot can help you decide: how far out in time should I go in order to assume that measurement (x_i) is independent from (x_i+n)?

# create partial autocorrelation plot
stats::pacf(noaa_temp_hi$temp_hr_f,
            main = " ",
            xlab = "Lag (hours)",
            ylab = "Partial Correlation Coefficient",
            lwd = 2) 

In this case, after a 2-hr lag, the data loses most autocorrelation, except for some periodicity around 12- and 24-hr lags.

5.5 Exploring Data

The whole point of this chapter comes down to this: when you have new, unfamiliar data, the best thing to do is to explore that data graphically. Create a panel of plots from available candidates: histogram, cumulative distribution, time-series, boxplot, autocorrelation. Together, these plots allow you to see basic features and extract important descriptors such as central tendency, spread, shape, and quantiles. EDA plots also frequently lead to more questions; the answers to which will bring you even closer to understanding what the data are trying to tell you. Let your eyes do the listening at first.

For example, here is a “3-plot” for the ME Salary data originally shown in Figure 5.5:

Figure 5.14: EDA Plots of Annual Salaries among Mech Eng Graduates

5.6 Chapter 5 Homework

You will continue to work with the ozone measurement data introduced in the previous chapters. Download the R Markdown template from our Github page to get started.