Category Exploratory Data Analysis

Exploring the Bivariate Landscape: Navigating Relationships in EDA

If you are still with me, we are now entering the phase of analysis where the rubber begins to meet the road to eventually get to the answer to the original research question: Is there a connection between a quarterback’s race and the number of times they get a Roughing the Passer (RtP) call in their favor? So far, we have completed the following steps in our EDA journey:

Now, we are diving into Bivariate Analysis. In this step, we will examine relationships between pairs of variables. Utilizing scatter plots, correlation matrices, and other tools, we will investigate how variables interact with each other, aiming to identify correlations, dependencies, or trends of interest.
For this phase, I will once again employ R to gain additional practice. The initial step involves importing the necessary libraries for this part of the analysis, as well as the QBs dataset we’ve been working with.

R
install.packages("ggcorrplot")
library(ggcorrplot)
library(ggplot2)

url <- "https://raw.githubusercontent.com/lareynolds/QB-EDA/main/cleaned_qbs.csv"

qbs = read.csv(file = url)

Now it may seem pretty easy and straight forward to follow the steps and just do what I just did. But what you don’t see is the extra tabs I have open to help through the process. For example, At first I tried to import the dataset as an Excel file, but that made the system throw some errors. So then, I thought it would be easier to import the file as a .csv file, and it sort of was, but I still got a bunch of errors. Apparently Google Colab doesn’t like to import files straight from GitHub as is, so off to Stack Overflow or Toward Data Science I go. In case you haven’t read anything on TDS, you should. This is a gem of quick tips and tricks for everything in data science.

Thanks to A Apte’s TDS article “Get Started: 3 Ways to Load CSV files into Colab”, the answer is:

“Click on the dataset in your repository, then click on View Raw. Copy the link to the raw dataset and store it as a string variable called url in Colab.”

“Get Started: 3 Ways to Load CSV files into Colab”
R
# Double Checking to make sure the data is loaded into the notebook
str(qbs)

#Output
'data.frame':	66 obs. of  26 variables:
 $ Player      : chr  "A.Dalton" "A.Luck" "A.Rodgers" "A.Smith" ...
 $ Total       : int  25 17 41 23 12 9 6 18 2 30 ...
 $ X2009       : int  0 0 3 1 0 0 1 0 0 3 ...
 $ X2010       : int  0 0 3 1 0 0 1 0 0 0 ...
 $ X2011       : int  4 0 3 1 0 2 0 0 0 3 ...
 $ X2012       : int  1 5 5 1 0 1 0 0 0 1 ...
 $ X2013       : int  1 3 2 1 0 0 0 0 0 1 ...
 $ X2014       : int  5 4 2 6 1 0 1 0 0 4 ...
 $ X2015       : int  0 1 5 4 2 2 1 0 1 2 ...
 $ X2016       : int  1 1 4 1 1 1 0 0 0 1 ...
 $ X2017       : int  1 0 0 5 7 1 1 0 0 3 ...
 $ X2018       : int  1 3 1 2 1 0 0 3 1 4 ...
 $ X2019       : int  6 0 3 0 0 0 1 4 0 0 ...
 $ X2020       : int  0 0 2 0 0 1 0 4 0 2 ...
 $ X2021       : int  3 0 5 0 0 1 0 5 0 6 ...
 $ X2022       : int  2 0 3 0 0 0 0 2 0 0 ...
 $ X2023       : int  0 0 0 0 0 0 0 0 0 0 ...
 $ Games       : int  170 94 228 149 79 62 60 74 44 190 ...
 $ Per.Game    : num  0.15 0.18 0.18 0.15 0.15 0.15 0.1 0.24 0.05 0.16 ...
 $ Attempts    : int  5557 3620 7840 4648 2719 1534 1554 2330 1230 7035 ...
 $ Per.100.Att : num  0.45 0.47 0.52 0.49 0.44 0.59 0.39 0.77 0.16 0.43 ...
 $ Sacked      : int  361 186 542 367 201 139 92 171 84 395 ...
 $ Per.Sack    : num  0.069 0.091 0.076 0.063 0.06 0.065 0.065 0.105 0.024 0.076 ...
 $ Sack.Per.Att: num  0.065 0.051 0.069 0.079 0.074 0.091 0.059 0.073 0.068 0.056 ...
 $ Third.Down..: num  40 29.4 39 26.1 33.3 ...
 $ qboc        : int  0 0 0 0 0 0 0 0 0 0 ...
 
 # Checking the summary statistics of each of the variables in the dataset
summary(qbs)

#Output
    Player              Total           X2009           X2010       
 Length:66          Min.   : 2.00   Min.   :0.000   Min.   :0.0000  
 Class :character   1st Qu.: 8.25   1st Qu.:0.000   1st Qu.:0.0000  
 Mode  :character   Median :15.50   Median :0.000   Median :0.0000  
                    Mean   :18.09   Mean   :0.697   Mean   :0.8333  
                    3rd Qu.:24.75   3rd Qu.:1.000   3rd Qu.:1.0000  
                    Max.   :57.00   Max.   :5.000   Max.   :6.0000  
     X2011           X2012           X2013           X2014      
 Min.   :0.000   Min.   :0.000   Min.   :0.000   Min.   :0.000  
 1st Qu.:0.000   1st Qu.:0.000   1st Qu.:0.000   1st Qu.:0.000  
 Median :0.000   Median :0.000   Median :0.000   Median :0.500  
 Mean   :1.121   Mean   :1.258   Mean   :1.076   Mean   :1.333  
 3rd Qu.:2.000   3rd Qu.:2.000   3rd Qu.:1.000   3rd Qu.:2.000  
 Max.   :8.000   Max.   :6.000   Max.   :8.000   Max.   :7.000  
     X2015           X2016           X2017           X2018      
 Min.   :0.000   Min.   :0.000   Min.   :0.000   Min.   :0.000  
 1st Qu.:0.000   1st Qu.:0.000   1st Qu.:0.000   1st Qu.:0.000  
 Median :1.000   Median :1.000   Median :0.000   Median :1.000  
 Mean   :1.455   Mean   :1.288   Mean   :1.303   Mean   :1.561  
 3rd Qu.:2.000   3rd Qu.:2.000   3rd Qu.:2.000   3rd Qu.:3.000  
 Max.   :7.000   Max.   :6.000   Max.   :7.000   Max.   :7.000  
     X2019            X2020            X2021            X2022      
 Min.   : 0.000   Min.   : 0.000   Min.   : 0.000   Min.   :0.000  
 1st Qu.: 0.000   1st Qu.: 0.000   1st Qu.: 0.000   1st Qu.:0.000  
 Median : 0.000   Median : 0.000   Median : 0.000   Median :0.000  
 Mean   : 1.773   Mean   : 1.545   Mean   : 1.742   Mean   :1.076  
 3rd Qu.: 3.000   3rd Qu.: 2.000   3rd Qu.: 3.000   3rd Qu.:2.000  
 Max.   :10.000   Max.   :11.000   Max.   :10.000   Max.   :6.000  
     X2023            Games           Per.Game         Attempts   
 Min.   :0.0000   Min.   : 42.00   Min.   :0.0500   Min.   : 289  
 1st Qu.:0.0000   1st Qu.: 61.25   1st Qu.:0.1125   1st Qu.:1794  
 Median :0.0000   Median : 77.50   Median :0.1650   Median :2316  
 Mean   :0.0303   Mean   : 99.53   Mean   :0.1768   Mean   :3250  
 3rd Qu.:0.0000   3rd Qu.:138.00   3rd Qu.:0.2200   3rd Qu.:4476  
 Max.   :1.0000   Max.   :253.00   Max.   :0.3500   Max.   :9725  
  Per.100.Att         Sacked         Per.Sack        Sack.Per.Att    
 Min.   :0.1300   Min.   : 26.0   Min.   :0.02400   Min.   :0.03000  
 1st Qu.:0.3900   1st Qu.:131.5   1st Qu.:0.05850   1st Qu.:0.05800  
 Median :0.5500   Median :170.0   Median :0.07800   Median :0.06900  
 Mean   :0.5736   Mean   :211.2   Mean   :0.08342   Mean   :0.07009  
 3rd Qu.:0.7400   3rd Qu.:264.5   3rd Qu.:0.10725   3rd Qu.:0.08350  
 Max.   :1.2400   Max.   :542.0   Max.   :0.20200   Max.   :0.10300  
  Third.Down..        qboc       
 Min.   : 0.00   Min.   :0.0000  
 1st Qu.:25.00   1st Qu.:0.0000  
 Median :30.77   Median :0.0000  
 Mean   :31.14   Mean   :0.2727  
 3rd Qu.:39.76   3rd Qu.:1.0000  
 Max.   :77.78   Max.   :1.0000  
 
 # Creates a subset dataframe with the numeric variables
df = subset(qbs, select = c(Per.Game, Per.100.Att, Per.Sack, Sack.Per.Att, qboc)) # To drop columns use a '-' before the c and list the columns to drop

# Create the scatter plot
x <- qbs$Per.Game # x is the column with the RtP values per game for each quarterback
y <- qbs$qboc # y is either the 1 or 0 value to indicate if the quarterback is black or not

plot(x, y, main = "Scatter Plot with Trend Line", xlab = "RtP Calls per Game", ylab = "QB of Color", col = "blue", pch = 19)

# Add the trend line
abline(lm(y ~ x), col = "red") # the 'lm' in this line is calling a liner model to fit a line that best fits the data by reducing the error

Notice that the qboc is either a 1 or a 0. This is a binary variable. This might cause some implications later on. But for right now, it does appear that there is a negative association between the quaterbacks’ race and the number of RtP calls made in the quarterback’s favor. This means that the number of RtP calls decrease for quarterbacks of color.

In the Univariate Analysis post, I decided that I would examine the Per.Game and qboc variables in this analysis, but after thinking about it, I decided it might be worth taking a look to see if maybe there was another reason for the number of RtP calls. Could a quarterback that gets sacked a lot also draw more RtP calls? Could this be an indication that the offensive line is not as good and can’t protect the quarterback as well? So I’m also going to do a scatter plot for the percentage of sacks by the number of RtP calls per game.

R
# Create the scatter plot
x <- qbs$Per.Game
y <- qbs$Per.Sack
plot(x, y, main = "Scatter Plot with Trend Line", xlab = "RtP Calls per Game", ylab = "Sacks per Game", col = "blue", pch = 19)

# Add the trend line
abline(lm(y ~ x), col = "red")
# This is the same code as above, but I switched out the variable for y.

There does indeed seem to be a strong positive linear connection between the number of sacks and the number of RtP calls. One other thing to notice is that there may be some outliers for these variables. I would like to see if this is true so I am going to use the Interquartile Range (IQR) Method to determine if this is the case. This method involves calculating the IQR for your data and identifying data points outside the “whiskers” of the box plot (usually 1.5 times the IQR).

R
# Calculate the IQR for 'y' variable
q <- quantile(y, c(0.25, 0.75))
iqr <- q[2] - q[1]

# Set a threshold for outliers
threshold <- 1.5

# Identify outliers
outliers <- y < (q[1] - threshold * iqr) | y > (q[2] + threshold * iqr)

# Print the indices of outlier data points
which(outliers)

# Output
56

value = qbs[56, 1]

value

# Output
'R.Fitzpatrick'

Well that is interesting, apparently Ryan Fitzpatrick took enough sacks to literally send his stats off the charts. In this case identifying the outlier was a “good to know” exercise. However, in other datasets outliers may indicate errors in the data and should be examined further. Now let’s get back to our task at hand.

Let’s see if there is a correlation between our numeric variables. It does look like there is some level of negative correlation between the number of RtP calls and the quarterback’s race. And the scatter plot of the Sacks and RtP calls seem to be very correlated. We can create a correlation matrix to visualize the correlation coefficient between each of the variables.

R
corr_matrix <- cor(df)
corr_plot<- ggcorrplot(corr_matrix,
           type = "lower",  # "lower" displays the lower triangular part of the matrix
           hc.order = TRUE, # reorder variables based on hierarchical clustering
           lab = TRUE,      # display variable names
           title = "Correlation Matrix Plot",
           ggtheme = theme_minimal()) # choose a theme for the plot (optional)


# Change the size of the display window
options(repr.plot.width = 10, repr.plot.height = 10)

# Display the correlation matrix plot
print(corr_plot)
R
RtP_correlation <- cor(qbs$Per.Game, qbs$qboc)
RtP_correlation

# Output
-0.0523628105182807

RtP_Sack_correlation <- cor(qbs$Per.Game, qbs$Per.Sack)
RtP_Sack_correlation

# Output
0.89240005509169

We can draw some conclusions from the correlation matrix, such as it appears as though quarterbacks of color get sacked more, but benefit less from RtP calls than their white counterparts. But don’t forget, the qboc variable is binary, so we need to make sure that we aren’t drawing some erroneous conclusions by using statistical methods meant to be used on continuous data.

For this reason, I am also going to also look at the point-biserial correlation coefficient. This is a correlation coefficient used to measure the strength and direction of the linear relationship between a binary (dichotomous) variable (0 or 1) and a continuous variable.. It is an adapted form of the Pearson correlation coefficient, which is used for two continuous variables. The point-biserial correlation is appropriate when one of the variables is binary, and the other is continuous.

Here are the key points about the point-biserial correlation coefficient:

Purpose: The point-biserial correlation is used to determine if there is a significant linear relationship between a binary predictor variable (e.g., yes/no, pass/fail) and a continuous outcome variable (e.g., test scores, income).

Range: The point-biserial correlation coefficient ranges from -1 to 1, similar to the Pearson correlation coefficient. A positive value indicates a positive relationship (as the binary variable increases, the continuous variable tends to increase), while a negative value indicates a negative relationship (as the binary variable increases, the continuous variable tends to decrease).

Interpretation:

  • A coefficient close to 1 or -1 suggests a strong linear relationship.
  • A coefficient close to 0 suggests a weak or no linear relationship.
  • The sign (+ or -) indicates the direction of the relationship.

Assumptions: Like the Pearson correlation, the point-biserial correlation assumes linearity and normality. It is also sensitive to outliers.

Use Cases:

  • In research, it is used to examine associations between a binary predictor variable (e.g., gender, treatment group) and a continuous outcome variable (e.g., test scores, response time).
  • Commonly used in educational research (e.g., comparing the performance of two groups).
  • Often used in psychology to assess the relationship between a binary variable (e.g., presence/absence of a condition) and a continuous measure (e.g., anxiety levels).
R
# Sample data: binary variable (0s and 1s) and continuous variable
binary_variable <- qbs$qboc
continuous_variable <- qbs$Per.Game

# Calculate the mean and standard deviation for both groups
mean_0 <- mean(continuous_variable[binary_variable == 0])
mean_1 <- mean(continuous_variable[binary_variable == 1])

sd_0 <- sd(continuous_variable[binary_variable == 0])
sd_1 <- sd(continuous_variable[binary_variable == 1])

# Calculate the difference in means
mean_diff <- mean_1 - mean_0

# Calculate the proportion of 1s in the binary variable
p <- mean(binary_variable)

# Calculate the pooled standard deviation
pooled_sd <- sqrt(((sd_0^2 + sd_1^2) / 2))

# Calculate the biserial correlation
biserial_correlation <- mean_diff / (pooled_sd * sqrt(p * (1 - p)))

# Print the result
print(biserial_correlation)

# Output
-0.2702969

The biserial_correlation does indicate that there is an association between the number of RtP called in favor of the quarterback is inversely related to their race. However, the strength of the correlation may be a little week given how far -0.27 is from -1. Biserial correlations are sensitive to outliers, so let’s see what happens if we remove Ryan Fitzpatrick from the dataset… just for giggles… and re-run the correlation before moving on to testing to see if the correlation is statistically significant.

R
row_index_to_remove <- 56
new_qbs <- qbs[-row_index_to_remove, ]

# Sample data: binary variable (0s and 1s) and continuous variable
binary_variable <- new_qbs$qboc
continuous_variable <- new_qbs$Per.Game

# Calculate the mean and standard deviation for both groups
mean_0 <- mean(continuous_variable[binary_variable == 0])
mean_1 <- mean(continuous_variable[binary_variable == 1])

sd_0 <- sd(continuous_variable[binary_variable == 0])
sd_1 <- sd(continuous_variable[binary_variable == 1])

# Calculate the difference in means
mean_diff <- mean_1 - mean_0

# Calculate the proportion of 1s in the binary variable
p <- mean(binary_variable)

# Calculate the pooled standard deviation
pooled_sd <- sqrt(((sd_0^2 + sd_1^2) / 2))

# Calculate the biserial correlation
biserial_correlation <- mean_diff / (pooled_sd * sqrt(p * (1 - p)))

# Print the result
print(biserial_correlation)

# Output
-0.1595576

So that is interesting that the strength of the correlation decreased by removing the outlier. Since I want as much data as possible, and there was no error in the data for the outlier, I am going to use the entire dataset when testing to see if the correlation is statistically significant.

The next step in this process is hypothesis testing. In this study:

  • Ho = There is no correlation between a quarterback’s race and the number of RtP calls made in his favor
  • Ha = There is a correlation between race and RtP calls

The alpha for this test will be 0.05.

R
categorical_variable <- qbs$qboc
continuous_variable <- qbs$Per.Game

# Compute the Point-Biserial Correlation and perform the test
correlation_test <- cor.test(categorical_variable, continuous_variable)

# Check the p-value
p_value <- correlation_test$p.value

# Set your significance level (e.g., 0.05)
alpha <- 0.05

# Determine if the correlation is statistically significant
if (p_value < alpha) {
  cat("The correlation is statistically significant (p-value:", p_value, ")\n")
} else {
  cat("The correlation is not statistically significant (p-value:", p_value, ")\n")
}

# Output
The correlation is not statistically significant (p-value: 0.6762716)

There you have it, folks. According to the findings of our analysis, the NFL is doing a good job of minimizing bias in penalty calls.

Summary

In this phase of the EDA, I examined the relationship between a quarterback’s race and the number of RtP calls made in favor of the QB. I utilized scatter plots with trend lines to visualize potential relationships, employed a correlation matrix to identify the strength and direction of these relationships, checked for outliers to assess their impact on the relationship’s strength, ensured the use of the appropriate correlation technique to account for a categorical independent variable, and conducted hypothesis testing to determine the statistical significance of the correlation.

Voyage Through Univariate Analysis: Charting the Solo Attributes of Roughing the Passer Penalties in the NFL

Univariate analysis is a crucial step in data analysis that focuses on examining and summarizing the characteristics of a single variable or attribute from the dataset. Univariate analysis provides a foundation for understanding the characteristics of individual variables, which is essential for more advanced multivariate analyses and modeling. It helps identify patterns, outliers, and potential data quality issues, making it a crucial step in the data analysis pipeline.

Python
import pandas as pd
# import and store the dataset
qbs = pd.read_excel("https://myordinaryjourney.com/wp-content/uploads/2023/09/cleaned_qbs.xlsx")
print(qbs.head())

#Output
      Player  Total  2009  2010  2011  2012  2013  2014  2015  2016  ...  \
0   A.Dalton     25     0     0     4     1     1     5     0     1  ...   
1     A.Luck     17     0     0     0     5     3     4     1     1  ...   
2  A.Rodgers     41     3     3     3     5     2     2     5     4  ...   
3    A.Smith     23     1     1     1     1     1     6     4     1  ...   
4  B.Bortles     12     0     0     0     0     0     1     2     1  ...   

   2023  Games  Per Game  Attempts  Per 100 Att  Sacked  Per Sack  \
0     0    170      0.15      5557         0.45     361     0.069   
1     0     94      0.18      3620         0.47     186     0.091   
2     0    228      0.18      7840         0.52     542     0.076   
3     0    149      0.15      4648         0.49     367     0.063   
4     0     79      0.15      2719         0.44     201     0.060   

   Sack Per Att  Third Down %  qboc  
0         0.065         40.00     0  
1         0.051         29.41     0  
2         0.069         39.02     0  
3         0.079         26.09     0  
4         0.074         33.33     0  

First, we should look to see what the variable datatypes are and if there are any null or missing values. 

Python
qbs.info(verbose=True)

#Output
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 66 entries, 0 to 65
Data columns (total 26 columns):
 #   Column        Non-Null Count  Dtype  
---  ------        --------------  -----  
 0   Player        66 non-null     object 
 1   Total         66 non-null     int64  
 2   2009          66 non-null     int64  
 3   2010          66 non-null     int64  
 4   2011          66 non-null     int64  
 5   2012          66 non-null     int64  
 6   2013          66 non-null     int64  
 7   2014          66 non-null     int64  
 8   2015          66 non-null     int64  
 9   2016          66 non-null     int64  
 10  2017          66 non-null     int64  
 11  2018          66 non-null     int64  
 12  2019          66 non-null     int64  
 13  2020          66 non-null     int64  
 14  2021          66 non-null     int64  
 15  2022          66 non-null     int64  
 16  2023          66 non-null     int64  
 17  Games         66 non-null     int64  
 18  Per Game      66 non-null     float64
 19  Attempts      66 non-null     int64  
 20  Per 100 Att   66 non-null     float64
 21  Sacked        66 non-null     int64  
 22  Per Sack      66 non-null     float64
 23  Sack Per Att  66 non-null     float64
 24  Third Down %  66 non-null     float64
 25  qboc          66 non-null     int64  
dtypes: float64(5), int64(20), object(1)

The variables I will examine more closely are the total number of Roughing the Passer (RTP) calls per quarterback, the number of RTP calls per year for each quarterback, and the number of RTP calls per games played for their distribution, central tendency, and variance. I will be using Python and Jupyter Notebooks.

Python
table_stats = qbs.describe()
print(table_stats)

#Output
           Total       2009       2010       2011       2012       2013  \
count  66.000000  66.000000  66.000000  66.000000  66.000000  66.000000   
mean   18.090909   0.696970   0.833333   1.121212   1.257576   1.075758   
std    12.505663   1.176301   1.452672   1.767355   1.774266   1.825550   
min     2.000000   0.000000   0.000000   0.000000   0.000000   0.000000   
25%     8.250000   0.000000   0.000000   0.000000   0.000000   0.000000   
50%    15.500000   0.000000   0.000000   0.000000   0.000000   0.000000   
75%    24.750000   1.000000   1.000000   2.000000   2.000000   1.000000   
max    57.000000   5.000000   6.000000   8.000000   6.000000   8.000000   

            2014       2015       2016      2017  ...       2023       Games  \
count  66.000000  66.000000  66.000000  66.00000  ...  66.000000   66.000000   
mean    1.333333   1.454545   1.287879   1.30303  ...   0.030303   99.530303   
std     1.842518   1.832878   1.698553   1.75385  ...   0.172733   53.915952   
min     0.000000   0.000000   0.000000   0.00000  ...   0.000000   42.000000   
25%     0.000000   0.000000   0.000000   0.00000  ...   0.000000   61.250000   
50%     0.500000   1.000000   1.000000   0.00000  ...   0.000000   77.500000   
75%     2.000000   2.000000   2.000000   2.00000  ...   0.000000  138.000000   
max     7.000000   7.000000   6.000000   7.00000  ...   1.000000  253.000000   

        Per Game     Attempts  Per 100 Att      Sacked   Per Sack  \
count  66.000000    66.000000    66.000000   66.000000  66.000000   
mean    0.176818  3250.303030     0.573636  211.212121   0.083424   
std     0.073801  2085.250348     0.241951  117.910594   0.034212   
min     0.050000   289.000000     0.130000   26.000000   0.024000   
25%     0.112500  1794.500000     0.390000  131.500000   0.058500   
50%     0.165000  2315.500000     0.550000  170.000000   0.078000   
75%     0.220000  4476.000000     0.740000  264.500000   0.107250   
max     0.350000  9725.000000     1.240000  542.000000   0.202000   

       Sack Per Att  Third Down %       qboc  
count     66.000000     66.000000  66.000000  
mean       0.070091     31.143788   0.272727  
std        0.016449     13.872813   0.448775  
min        0.030000      0.000000   0.000000  
25%        0.058000     25.000000   0.000000  
50%        0.069000     30.770000   0.000000  
75%        0.083500     39.755000   1.000000  
max        0.103000     77.780000   1.000000  

[8 rows x 25 columns]

Well, that was easy. Using the .describe() function allows for the examination of the data set’s values for central tendency (mean in this case) and the spread (standard deviation) and dispersion (technically). Although all the information is present to observe the dispersion of the information, it may hard to conceptualize the shape without using a visualization help. Histograms can aid in our observations.

Python
qbs.hist(figsize=(10, 15))

Adjust the figure size to be able to view all the histogram outputs more clearly. The distributions for Per Game and Per 100 Attempts look nearly normal, so that will allow us to use some parametric tests for analysis.

Another visual that is helpful to see the distribution is a boxplot (box-and whisker plot). This time let us just look at just the Per Game and Per 100 Attempts.

Python
qbs.boxplot(column=['Per Game', 'Per 100 Att'], figsize=(10,10))

The wonderful thing about boxplots is that they make it easy to identify outliers – observations that are outside the whiskers (either top or bottom). You can also easily see the centrality and spread of the data.  

One additional variable that we should look at is the ‘qboc’. Did you notice that there was an interesting distribution when the histograms plotted above? Let’s take a closer look.  

Python
qbs['qboc'].hist(figsize=(10, 15))

As you can see there are only two values for this variable. And this makes sense since we are categorizing the quarterbacks based on if they are (1) or are not (0) a quarterback of color. I am going to do a bit of foreshadowing, but this means that if we wanted to do any sort of predictive analysis, we need to think about some additional models beyond regular linear regression, logistic regression to be specific… but that is a post for another day. For now, I have identified the variables we can use for some multivariate analysis:

  • Per Game
  • Per 100 Attempts
  • qboc

In this data exploration process, univariate analysis was applied to understand and summarize the characteristics of individual variables, specifically focusing on those related to Roughing the Passer (RTP) calls. The dataset was examined using Python and Jupyter Notebooks. The summary statistics and visualizations were generated to gain insights into the central tendency, dispersion, and distribution of the data.

The analysis revealed that variables such as “Per Game” and “Per 100 Attempts” exhibited nearly normal distributions, making them suitable for parametric tests. Additionally, the “qboc” variable, which categorizes quarterbacks based on their ethnicity, showed a binary distribution, indicating potential utility in predictive modeling.

This initial exploration sets the stage for further multivariate analysis and modeling, offering a foundation for more in-depth investigations into the relationships between these variables and RTP calls in NFL quarterbacks.

Cruising the Data Landscape: Exploring the Fundamentals of Data Exploration

Recap

Well, well, well! Looks like we’ve embarked on quite the adventure here! Our first task was to get to the bottom of the burning question: “Do quarterbacks of color get the short end of the stick when it comes to roughing the passer penalties compared to their fair-skinned counterparts?” Exciting stuff, huh?

The next step in our journey involved gathering the all-important data. We needed to make sure we were dealing with top-notch, premium-quality data. None of that shady, questionable stuff for us! We demanded data that was reliable, traceable, and oh-so-accurate. Because in the world of data science, it’s all about that golden rule: “garbage in, garbage out!” Can’t have our analysis going awry now, can we?

For this particular project, we scoured the depths of NFL Penalties and even tapped into the vast knowledge reserves of the mighty Wikipedia. We left no stone unturned, my friend!

Now, onto the thrilling next stage of our expedition – data cleaning and reshaping! Brace yourself, because this is where things get spicy. We had to create a nifty variable to indicate the race of those quarterbacks. Easy-peasy, right? Well, almost. Turns out, we stumbled upon a little naming convention conundrum that resulted in a sneaky duplicate value. But fear not! With our eagle-eyed attention to detail, we swiftly detected and corrected that pesky error. Crisis averted!

And now, dear explorer, we venture forth to the exhilarating realm of Data Exploration. Excited? You should be! There’s so much more to discover and uncover on this grand data-driven expedition! So, buckle up and get ready for the thrill ride ahead! Let’s dive into the vast ocean of information and see what fascinating insights await us!

Data exploration is an essential step in the data analysis process. It allows us to gain an initial understanding of the dataset by examining its general properties and characteristics. By delving into the data, we can uncover valuable insights.

One of the key aspects of data exploration involves assessing the size of the dataset. This includes understanding the number of observations or records and the number of variables or features present in the dataset. Knowing the scope of the dataset helps us gauge its comprehensiveness and potential usefulness for our analysis.

Furthermore, examining the data types is crucial in data exploration. By identifying the types of data contained in the dataset, such as numerical, categorical, or textual, we can determine the appropriate statistical techniques or visualization methods to apply during analysis. This enables us to make accurate interpretations and draw meaningful conclusions from the data.

Another crucial aspect is the identification of initial patterns or trends within the dataset. Exploring the data can reveal inherent relationships, correlations, or anomalies that may exist. By uncovering these patterns, we can develop hypotheses, generate insights, and pose relevant research questions for further investigation.

You may have noticed that I talk about using Python a lot in my work. I actually prefer to use in my day job since it seems to get the most use and support from other data professionals. However, for this portion of my project, I’m going to be using R instead. It has been some time since I’ve used R (maybe before I even finished my master’s program) so full discloser, I’m pretty sure I had to look up most of that I’m doing to refresh my memory. But since it is a language, if you don’t use it, you lose it, so I’m going to shake things up a bit and revisit R.

Data Exploration

The first things I need to do is load in any packages I need and my dataset. I am loading the ‘cleaned_qbs.xlsx’ Excel file that I created during the data cleaning process.

R
# import and store the dataset
library(openxlsx)

qbs = read.xlsx("https://myordinaryjourney.com/wp-content/uploads/2023/09/cleaned_qbs.xlsx")

# Check to make sure the data loaded correctly
head(qbs, n=5)

#Output
A data.frame: 5 × 26
Player	Total	2009	2010	2011	2012	2013	2014	2015	2016	...	2023	Games	Per.Game	Attempts	Per.100.Att	Sacked	Per.Sack	Sack.Per.Att	Third.Down.%	qboc
<chr>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	...	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>
1	A.Dalton	25	0	0	4	1	1	5	0	1	...	0	170	0.15	5557	0.45	361	0.069	0.065	40.00	0
2	A.Luck	17	0	0	0	5	3	4	1	1	...	0	94	0.18	3620	0.47	186	0.091	0.051	29.41	0
3	A.Rodgers	41	3	3	3	5	2	2	5	4	...	0	228	0.18	7840	0.52	542	0.076	0.069	39.02	0
4	A.Smith	23	1	1	1	1	1	6	4	1	...	0	149	0.15	4648	0.49	367	0.063	0.079	26.09	0
5	B.Bortles	12	0	0	0	0	0	1	2	1	...	0	79	0.15	2719	0.44	201	0.060	0.074	33.33	0

Looking over the raw data, it is hard to see any particular patterns in the data. However, a few data points do jump out as they are either larger or smaller than the data around the point. For example, the number of RTP penalties called each year numbers in the single digits with the exception of 2020 for the quarterback Josh Allen. That year he received 11 RTP called in his favor. Another example is the number of sacks recorded against Taysom Hill. Most of the QBs have a sack count in the triple digits, while only a few have double digit sack counts and Hill has the lowest of any QB in the data set. The next thing I want to do is check how many records are in my table and the total number of variables. It is helpful to know what your sample size is so you can select the correct analytic approach later on.

R
nrow(qbs)

# Output
66

length(qbs)

#Output
26

The next step is to understand the data types of each of the variables. Luckily, when the head of the table was printed, it also included that data types for each of the variables. In our cleaned_qbs table, we have one variable with a character data type (Player) and 25 variables with the data type of ‘dbl’, otherwise known float.

R
#Output
A data.frame: 5 × 26
Player	Total	2009	2010	2011	2012	2013	2014	2015	2016	...	2023	Games	Per.Game	Attempts	Per.100.Att	Sacked	Per.Sack	Sack.Per.Att	Third.Down.%	qboc
<chr>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	...	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>	<dbl>

Next Steps

Now that we understand how much and what kind of data we have, we can develop a game plan on what analytic techniques we can take in the next step. With Univariate Analysis, Bivariate Analysis, and Multivariate Analysis we will begin to explore the data and any relationships between the variables that might exist. This will get us closer to a possible answer to our original question.

First the Idea, then the Data: Navigating the Depths of Information

When working on a project that requires data, it is essential to consider the various sources and formats in which the information may be available. Often, the desired data cannot be found in one single location, requiring careful compilation from multiple sources. This process can be time-consuming and challenging, as each source may present its own set of complications and intricacies.

While some data may be readily accessible through online databases, APIs, or official government reports, it may not always be presented in the most convenient or usable form. For instance, a dataset could be in a raw, unstructured format, making it difficult to analyze and extract meaningful insights. In such cases, data manipulation techniques are necessary to transform the data into a suitable format.

Data manipulation involves various tasks such as cleaning, merging, and reshaping the data. Cleaning the data involves removing any inconsistencies, errors, or missing values. This step ensures that the data is accurate and reliable for analysis. Merging data from multiple sources requires aligning and combining different datasets based on common attributes or keys. This consolidation allows for a comprehensive and holistic view of the data.

Reshaping the data involves rearranging its structure to fit the desired analysis or visualization needs. This can include tasks such as pivoting, aggregating, or disaggregating the data. By reshaping the data, researchers and analysts can uncover patterns, trends, and relationships that might otherwise go unnoticed.

In the given scenario, the QB Penalty Stats table lacked information on the racial background of the players. While this data may not be readily available in the original source, alternative sources can be explored. In this case, a reliable and well-sourced list of black quarterbacks could be found on Wikipedia. Although caution is advisable when using information from Wikipedia, the aforementioned table demonstrates reliability in this particular case.

Working with data often goes beyond simply retrieving information from a single source. It requires diligent gathering, cleaning, merging, and reshaping of data to ensure accuracy and usability. With the right techniques and approaches, researchers can unlock the full potential of data and derive valuable insights.

Embarking on an Epic Journey

I now have a list of all QBs from 2009 until present (that’s 2023 for those of you that stumble upon my ramblings in the future) in the qb_rtp table. I also have a list of black QBs in the qbs_oc table. Sometimes I find it can be faster to do some of the data cleaning work in the native application. In this case, I needed to create a column in the black_qbs table that I could use to compare the qb_rtp[‘Player’] column in the other table. Luckily Excel’s autofill feature made this particularly easy. I inserted a blank column next to the [‘Quarterback’] called [‘short_name’]. After typing a few examples, I selected the “Autofill” option in the Data drop down menu, and voilà, I had the new column that I can use for comparison.

Starting to Use Notebooks

I haven’t used a Jupyter Notebook to code in a very long time. I usually use the Spyder IDE as I can see all my variables’ values at the same time. But I’m going to give it a shot. You can get the entire notebook here. The first step is to import the different libraries that I think I’ll need.

Python
import pandas as pd
import openpyxl

Next, I will import the data sets.

qbs = pd.read_excel('https://myordinaryjourney.com/wp-content/uploads/2023/09/QB_Table-1.xlsx', engine='openpyxl')
qbs_oc = pd.read_excel('https://myordinaryjourney.com/wp-content/uploads/2023/09/QB_Table-2.xlsx', engine='openpyxl')

Make sure to verify the data has been loaded by viewing a few rows from each dataframe. If you are wondering what the ‘\n’ means, it is the regular expression for a new line. It creates some extra space between the table headings and the table data to make them more readable. Dataquest has a great quick guide and cheat sheet available here.

print('Quarterback Stats table:\n', qbs.head(), '\n')
print('Quarterback of Color table:\n', qbs_oc.head())

#Output
Quarterback Stats table:
       Player  Total  2009  2010  2011  2012  2013  2014  2015  2016  ...  \
0   A.Dalton     25     0     0     4     1     1     5     0     1  ...   
1     A.Luck     17     0     0     0     5     3     4     1     1  ...   
2  A.Rodgers     41     3     3     3     5     2     2     5     4  ...   
3    A.Smith     23     1     1     1     1     1     6     4     1  ...   
4  B.Bortles     12     0     0     0     0     0     1     2     1  ...   

   2022  2023  Games  Per Game  Attempts  Per 100 Att  Sacked  Per Sack  \
0     2     0    170      0.15      5557         0.45     361     0.069   
1     0     0     94      0.18      3620         0.47     186     0.091   
2     3     0    228      0.18      7840         0.52     542     0.076   
3     0     0    149      0.15      4648         0.49     367     0.063   
4     0     0     79      0.15      2719         0.44     201     0.060   

   Sack Per Att  Third Down %  
0         0.065         40.00  
1         0.051         29.41  
2         0.069         39.02  
3         0.079         26.09  
4         0.074         33.33  

[5 rows x 25 columns] 

Quarterback of Color table:
          Quarterback    short_name Years active  \
0      Fritz Pollard     F.Pollard    19201926   
1        Joe Lillard     J.Lillard    19321933   
2  George Taliaferro  G.Taliaferro    19501955   
3     Marlin Briscoe     M.Briscoe        1968*   
4       James Harris      J.Harris    19691981   

                                                Team  
0  Akron Pros, Milwaukee Badgers, Hammond Pros, P...  
1                                  Chicago Cardinals  
2  New York Yanks, Dallas Texans, Baltimore Colts...  
3                                     Denver Broncos  
4  Buffalo Bills, Los Angeles Rams, San Diego Cha...  

Smushing the Data Together

Now that the two initial datasets have been read into Pandas dataframes, it’s time to smush them together in a way that makes them useful for further analysis. The first step I take is adding a new variable to the Quarterback Stats Table (qbs) called ‘qboc’. I set all the values to 0 for right now.

qbs['qboc'] = 0

Once again check to make sure the variable was added to the dataframe as expected.

print(qbs.head())

#Output
Player  Total  2009  2010  2011  2012  2013  2014  2015  2016  ...  \
0   A.Dalton     25     0     0     4     1     1     5     0     1  ...   
1     A.Luck     17     0     0     0     5     3     4     1     1  ...   
2  A.Rodgers     41     3     3     3     5     2     2     5     4  ...   
3    A.Smith     23     1     1     1     1     1     6     4     1  ...   
4  B.Bortles     12     0     0     0     0     0     1     2     1  ...   

   2023  Games  Per Game  Attempts  Per 100 Att  Sacked  Per Sack  \
0     0    170      0.15      5557         0.45     361     0.069   
1     0     94      0.18      3620         0.47     186     0.091   
2     0    228      0.18      7840         0.52     542     0.076   
3     0    149      0.15      4648         0.49     367     0.063   
4     0     79      0.15      2719         0.44     201     0.060   

   Sack Per Att  Third Down %  qboc  
0         0.065         40.00     0  
1         0.051         29.41     0  
2         0.069         39.02     0  
3         0.079         26.09     0  
4         0.074         33.33     0  

[5 rows x 26 columns]

The ‘qboc’ column was added to the end of the dataframe. The next step I take is to make the qbs[‘Player’] and the qbs_oc[‘short_name’] columns the indexes of the dataframes. This will speed up the comparison process. Although it wouldn’t be noticeable with these datasets, if you have a dataset will millions of records, it will bog down the system… ask me how I know.

qbs = qbs.set_index('Player')
qbs_oc = qbs_oc.set_index('short_name')

Now this is the really fun part. All I need to do now is compare the indexes of the two dataframes. If the indexes match, then I can assign a ‘1’ to the qbs[‘qboc’] variable to signify that that quarterback is indeed a quarterback of color.

qbs.loc[(qbs.index.isin(qbs_oc.index)), 'qboc'] = 1

This one line of code is pretty cool, I think. I want to dig into what this line actually does. I know this isn’t a Python coding post, but explaining will only take a minute and it unpacks some interesting concepts.

  1. qbs.index and qbs_oc.index specify that we are going to be doing something with the indexes of the two dataframes.
  2. The .isin is a function that returns a Boolean mask. So what (qbs.index.isin(qbs_oc.index) does is takes the index values of the qbs_oc dataframe and look to see if those values are in the qbs dataframe index. If they are, then the Boolean value = True, if not, the value = False. You can also save that part of the code to a variable and use over and over in other parts of your project, but in this instance, we don’t need to.
  3. The qbs.loc is locating all the the indexes in the qbs dataframe where the mask value = True.
  4. ‘qboc’ specifies the column we want to assign our value to. In this case a 1 to signify that the quarterback is a person of color.

The qbs.loc() designates the dataframe we are intending to modify; (qbs.index.isin(qbs_oc.index)) designates the row we want to work with; and ‘qboc’ designates the column.

print(qbs.head())

#Output
           Total  2009  2010  2011  2012  2013  2014  2015  2016  2017  ...  \
Player                                                                  ...   
A.Dalton      25     0     0     4     1     1     5     0     1     1  ...   
A.Luck        17     0     0     0     5     3     4     1     1     0  ...   
A.Rodgers     41     3     3     3     5     2     2     5     4     0  ...   
A.Smith       23     1     1     1     1     1     6     4     1     5  ...   
B.Bortles     12     0     0     0     0     0     1     2     1     7  ...   

           2023  Games  Per Game  Attempts  Per 100 Att  Sacked  Per Sack  \
Player                                                                      
A.Dalton      0    170      0.15      5557         0.45     361     0.069   
A.Luck        0     94      0.18      3620         0.47     186     0.091   
A.Rodgers     0    228      0.18      7840         0.52     542     0.076   
A.Smith       0    149      0.15      4648         0.49     367     0.063   
B.Bortles     0     79      0.15      2719         0.44     201     0.060   

           Sack Per Att  Third Down %  qboc  
Player                                       
A.Dalton          0.065         40.00     0  
A.Luck            0.051         29.41     0  
A.Rodgers         0.069         39.02     0  
A.Smith           0.079         26.09     1  
B.Bortles         0.074         33.33     0 

As you can see, A.Smith now has a 1 in the qboc column. However, the A.Smith in the qbs data set is actually Alex Smith not Akili Smith who is the actual quarterback referenced in the qbs_oc dataset. This illustrates the importance of going through your data to ensure accuracy. Luckily these datasets are small enough to go through to spot any errors.

double_check = qbs.loc[(qbs['qboc'] == 1)]
print(double_check.index)

#Output
Index(['A.Smith', 'C.Kaepernick', 'C.Newton', 'D.Prescott', 'D.Watson',
       'G.Smith', 'J.Brissett', 'J.Campbell', 'J.Freeman', 'J.Hurts',
       'J.Winston', 'K.Murray', 'L.Jackson', 'M.Vick', 'P.Mahomes',
       'R.Griffin III', 'R.Wilson', 'T.Bridgewater', 'T.Taylor'],
      dtype='object', name='Player')

qbs['qboc']['A.Smith'] = 0
print(qbs)

#Output
         Total  2009  2010  2011  2012  2013  2014  2015  2016  2017  ...  \
Player                                                                ...   
A.Smith     23     1     1     1     1     1     6     4     1     5  ...   

         2023  Games  Per Game  Attempts  Per 100 Att  Sacked  Per Sack  \
Player                                                                    
A.Smith     0    149      0.15      4648         0.49     367     0.063   

         Sack Per Att  Third Down %  qboc  
Player                                     
A.Smith         0.079         26.09     0  

[1 rows x 25 columns]

Next Steps

Get ready for some quarterback madness in the upcoming post! We’ll be diving into our fancy new quarterbacks table and unleashing some wild analysis on the gridiron.

The First Detour

Once upon a time, in the realm of weather-related nerdiness, I embarked on a quest to decipher the secrets of changing weather patterns. Armed with my mighty keyboard and a burning hatred for sweltering summers, I planned to uncover the truth about my local area’s climate evolution. You see, summer and I have never been the best of friends. The scorching heat and suffocating humidity make me cringe harder than a cat stuck in a cucumber maze. So, I figured, why not dive into the delightful world of data and investigate if there’s any hope for an early arrival of autumn? I dubbed it my “Project Meteorological Marvel.”

My cunning plan involved sifting through decades of weather records, gathering juicy tidbits on how temperatures have tortured us poor mortals over the years. I wanted to spot trends, make dazzling graphs, and perhaps even predict when autumn would grace us with its blessed presence. Oh, how I yearned for a reliable sign that summer’s reign of terror would soon be over! Of course, this was no ordinary undertaking. I needed a trustworthy data source, and what better place to turn to than the National Oceanic and Atmospheric Administration (NOAA)? If you can’t trust the NOAA to provide accurate historical weather data, well, I guess we’re all doomed!

Now, I must confess, I had no intention of becoming a weather forecaster during this escapade. That’s a whole different level of sorcery reserved for the truly brave and slightly crazy souls. No, my friends, my mission was solely to unravel the mysteries of the past, not predict the future. So, off I went, armed with my web-scraping skills and a fervent desire to put an end to endless summers. And thus, my epic journey into the realm of weather data began… but did it?

Well, it seems that once people discover your supreme data-crunching powers, they start throwing project ideas at you like confetti at a parade. Take my poor, football-obsessed husband for example. He came up with the brilliant notion of analyzing if there’s any connection between a quarterback’s race and the number of times they get a sweet, sweet roughing the passer call in their favor. And as if that wasn’t enough, I thought, why not spice it up even more and explore if the defender’s race also plays a role in how many roughing the passer flags rain down upon them? Heck, let’s even toss in the officials’ race for good measure. Who knew the football field could have so many hidden layers of sociology and statistics? But hey, I’ll play along and start with quarterbacks for now. Let the mind-boggling journey begin! Just don’t blame me if we end up in a statistical black hole of absurdity.

At first, I was going to look at NCAA football statistics given that the sample size would be much larger than for the NFL. However, I didn’t really find a good source for the data to either download or extract. It just doesn’t seem like the NCAA collects that data down to the player level. As luck would have it I was able to find a source for NFL penalty data. the aptly named NFL Penalties is a sited dedicated to capturing penalty data so that users can basically settle disputes “over a player and their frequent ability to get away with murder, or not.” The site’s author does a pretty good job at articulating problems with the data and any mitigation actions taken. Ultimately the data on the site is provided by nflfastR. 1

Now that I’ve talked about the general concept and the search for a data source, here my next steps.:

  1. Collect Roughing the Passer (RTP) data for the quarterback from NFL Penalties.
  2. Collect the relevant biographical data on each of the quarterbacks.
  3. Use Python and relevant libraries such as Pandas2 and matplotlib3 to perform data cleaning, exploration, univariate and bivariate analysis, and visualizations.
  4. Publish the findings along with the documented methodology.

I’m not sure how long this will take me, could be a day, could be weeks. Either way, check back often for updates on the project’s progress.


  1. Carl S, Baldwin B (2023). nflfastR: Functions to Efficiently Access NFL Play by Play Data. https://www.nflfastr.com/, https://github.com/nflverse/nflfastR. ↩︎
  2. The pandas development team. (2023). pandas-dev/pandas: Pandas (v2.1.0). Zenodo. https://doi.org/10.5281/zenodo.8301632 ↩︎
  3. J. D. Hunter, “Matplotlib: A 2D Graphics Environment”, Computing in Science & Engineering, vol. 9, no. 3, pp. 90-95, 2007. ↩︎