Describing Scale Variables, Central Tendency, and Variability

Descriptive Statistics and Scale Variable Distributions

• Context of Variable Description:

  • Categorical variables (nominal and ordinal) group individuals or objects into distinct categories, which are visually displayed using frequency bar graphs where the bars do not touch.

  • Scale variables (interval and ratio measurements) represent quantitative values along a continuous scale.

  • Describing scale variables requires analyzing distribution models to understand the shape, clustering, and dispersion of data scores.

• Normal Distribution Overview:

  • Known as the bell curve, Gaussian curve, or velocity curve.

  • Highly symmetrical model where data points cluster heavily in the middle and taper off equally toward both extremes.

  • Serves as the baseline comparative model for identifying distribution deviations.

Skewness and Distribution Shapes

• Definition of Skewness:

  • Skewness refers to the asymmetry or deviation in a distribution where the central mass of data shifts to either the left or the right, resulting in an extended tail on one side.

• Naming Conventions for Skewed Distributions:

  • Skewness is named exclusively after the direction in which the long, thin tail extends, regardless of where the majority of data points cluster.

  • Positive Skew:

  • The long, thin tail extends toward the right (positive side) of the horizontal axis.

  • The majority/bulk of data scores are clustered on the left (lower end) of the scale.

  • Negative Skew:

  • The long, thin tail extends toward the left (negative side) of the horizontal axis.

  • The majority/bulk of data scores are clustered on the right (higher end) of the scale.

• Directional Skew Metaphor:

  • Envision a miniature skier positioned at the peak of the curve with ski poles.

  • The direction in which the skier would slide down toward the extended tail determines the name of the slope (e.g., sliding down toward the positive side denotes positive skew).

• Statistical Terminology vs. Value Judgments:

  • Statistical terms like "positive" and "negative" are strictly mathematical and spatial indicators.

  • They must be completely divorced from qualitative value judgments such as "good" or "bad."

Visualizing Scale Data: Histograms, Line Graphs, and Scale Distortions

• Histograms:

  • The standard graphic representation for scale variables.

  • Differs from a categorical bar graph because the vertical bars in a histogram physically touch one another, representing continuous numerical ranges or bins.

  • Binning Example:

  • Money amounts spanning from $0$ to $\$5\,$ \\$10\, \\$20\, and \\$30\,.

  • Bars are erected over interval bins (e.g., $\\$5\text{--}\\$10\,$ $\\$10\text{--}\\$20\,$ $\\$20\text{--}\\$30$) showing the exact frequency of individuals falling into each monetary range.

  • Transition from Histogram to Distribution Curve:

  • Increasing the dataset size and adding substantially higher numbers of narrower bin columns smooths out the surface of the histogram bars.

  • A continuous mathematical distribution curve can be superimposed over histogram bars in statistical software to assess distribution shape.

• Frequency Graphs vs. Line Graphs:

  • Frequency Graph: The vertical axis (Y-axis) represents frequency (count of occurrences) while the horizontal axis (X-axis) represents the score values of the variable (e.g., how many years exhibited a specific average August temperature).

  • Time-Series / Line Graph: The X-axis represents ordered time increments (e.g., years $1900\,$ $1920\,$ $1930$) and the Y-axis represents the variable value itself (e.g., average temperature recorded in each year).

• Visual Distortion and Software Auto-Scaling Warnings:

  • Visual graphs are powerful tools because visual memory typically persists longer than raw numeric memory.

  • Statistical software often automatically rescales graph axes to fit data within a display frame rather than maintaining standard uniform increments.

  • Scale Magnification Effects:

  • Adjusting a vertical scale from increments of $5\,$ $10\,$ $15$ to increments of $10\,$ $20\,$ $30$ visually compresses data height.

  • Zooming in on a narrow vertical axis makes minor fluctuations appear deceptively large.

  • Zooming out on a broad vertical axis makes substantial fluctuations appear deceptively small.

Measures of Central Tendency

• Definition and Purpose:

  • Measures of central tendency summarize an entire large spreadsheet of numbers into a single representative score.

  • Identifies where the center mass or highest concentration of data resides.

• The Three Primary Measures:

  • Mode:

  • The score or value that appears most frequently in a dataset.

  • Applicable to scale data as well as categorical data (nominal and ordinal levels).

  • Median ($MDN$):

  • The exact middle score when all data points are arranged sequentially from lowest to highest.

  • If dataset count $N$ is odd, it is the exact middle score.

  • If dataset count $N$ is even, it is the arithmetic average of the two central middle scores.

  • Usable with ordinal, interval, and ratio data; cannot be used with nominal data.

  • Mean (Sample Mean: $\bar{x}$):

  • The mathematical or arithmetic average of all scores in a sample.

  • Calculated using the formula:       $\bar{x} = \frac{\sum x_i}{N}$       Where:

    • $\bar{x}$ = Sample mean (lowercase $x$ with a bar over top).

    • $\sum$ = Sigma symbol, representing the mathematical operation of summation.

    • $x_i$ = Individual score for a given observation ($i$ stands for individual).

    • $N$ = Total number of scores/observations in the sample.

• Central Tendency Alignment Across Shapes:

  • Symmetrical / Normal Distribution:

  • The Mean, Median, and Mode align perfectly at the exact center peak of the distribution, creating symmetry on both sides.

  • Positively Skewed Distribution:

  • The Mode remains under the peak.

  • The Median shifts slightly right.

  • The Mean is pulled furthest to the right into the long tail: $\text{Mode} < \text{Median} < \bar{x}$.

  • Negatively Skewed Distribution:

  • The Mode remains under the peak.

  • The Median shifts slightly left.

  • The Mean is pulled furthest to the left into the long tail: $\bar{x} < \text{Median} < \text{Mode}$.

Outliers and Ethical Handling of Extreme Data Points

• Definition of an Outlier:

  • An extreme data score that falls significantly outside the normal boundary or cluster of the rest of the dataset (excessively high or low value).

• Mechanism of Mean Distortion:

  • Because the mean calculation requires summing every single numeric value ($\sum x_i$), extreme outlier values weigh heavily on the total sum, pulling the mean away from the cluster toward the tail.

  • Metaphor: "Wrecking the curve" on an exam, where a class clustered around a C grade average has a single student score $99$ out of $100\,$, shifting the overall average upward.

• Data Auditing and Ethical Removal Protocols:

  • Extreme values must first be audited for human data-entry errors (e.g., a research assistant accidentally double-typing digits like 77 instead of 7). Physical survey archives are maintained to verify suspicious scores against original source documents.

  • Deleting legitimate outliers cleans the visual distribution, but editing datasets modifies authentic findings and can obscure unexpected insights.

  • Ethical standards require researchers to disclose any outlier removals. Modern digital publishing allows reporting both raw dataset results and cleaned dataset results side-by-side.

• Sensitivity Comparison:

  • Mean: Highly sensitive to extreme outliers because it acts like a physical balance scale (Dickensian brass balance scale) where distance and mass dictate equilibrium.

  • Median: Highly resistant/insensitive to outliers because it relies exclusively on positional rank order rather than numeric weight.

  • Example of Median Insensitivity:

  • Dataset: $9\,$ $9\,$ $10\,$ $11\,$ $12\,$ $28$

  • Changing the outlier $28$ to an extreme value of $48$ leaves the median completely unchanged ($MDN = 10$).

  • Removing $28$ entirely leaves $9\,$ $9\,$ $10\,$ $11\,$ $12$, yielding a revised median of $\frac{9 + 10}{2} = 9.5$.

Contextual Application and Reporting of Central Tendency

• Central Tendency as a "Best Guess":

  • In the absence of specific individual information, the reported measure of central tendency serves as the single best statistical prediction for an unknown individual's score from that group.

• Real-World Scenarios Favoring the Median:

  • Financial Metrics (Income, Net Worth, Real Estate):

  • Real estate values on a single street might contain several $\$200{,}000$ homes and one $\$1{,}000{,}000$ home.

  • Using the mean artificially inflates home values; reporting the median home value accurately reflects the neighborhood center without unethically discarding high-value properties.

  • Demographics (College Student Age):

  • While traditional college populations enter directly from high school ($18\text{--}22$ years old), returning adult learners create high positive outliers.

  • Reporting median age accurately characterizes the student body, whereas mean age is artificially inflated.

• Real-World Scenarios Favoring the Mode:

  • Categorical Inventory Sales: Department stores (e.g., Target) identifying which product category (e.g., clothing vs. toys) generates the highest transaction volume.

  • Discrete Scale Levels: Identifying which video game level contains the highest absolute concentration of active players (e.g., Level $1$).

Non-Unimodal Distributions: Bimodal and Multimodal Curves

• Bimodal Distributions:

  • A distribution containing two distinct high-frequency peaks (resembling two humps on a camel).

  • Contains two separate statistical modes.

  • Failure of the Mean in Bimodal Data: The mean calculates to a point in the central valley between the two peaks where very few actual data points exist, rendering the mean misleading.

• Multimodal / Amodal Distributions:

  • Distributions displaying three or more distinct frequency peaks.

  • Occurs when data clusters around multiple independent anchor points or categories.

  • Real-world example: YouTube video scrub bar analytics showing multiple distinct replayed peaks where viewers repeatedly rewind to rewatch complex tutorials or poorly filmed segments.

Measures of Variability and Dispersion

• Definition and Purpose:

  • Variability measures the extent to which data scores are spread apart, dispersed, or differ from one another relative to the center.

  • While central tendency measures similarity and clustering, variability measures difference and dispersion.

• Three Primary Measures of Variability:

  • 1. Range:

  • The simplest measure of variability, defined as the difference between the highest score and the lowest score:       $\text{Range} = x_{\text{max}} - x_{\text{min}}$

  • Highly sensitive to context and measurement units (e.g., an age range of $10$ years in a classroom indicates high variability, whereas a bank balance range of $\$10$ represents minimal variability).

  • 2. Individual Deviance (Deviation):

  • The distance of any specific score from the sample mean:       $\text{Deviance} = x_i - \bar{x}$

  • Scores below the mean yield negative deviance values; scores above the mean yield positive deviance values.

  • 3. Total Deviance and the Zero-Sum Problem:

  • Summing all individual deviations across a dataset yields the total deviance formula:       $\text{Total Deviance} = \sum (x_i - \bar{x})$

  • Fundamental Mathematical Rule: The sum of all individual deviances from the mean ALWAYS equals zero ($0$):       $\sum (x_i - \bar{x}) = 0$

  • Mathematical Demonstration:

    • Dataset scores: $1\,$ $2\,$ $3$

    • Sample size $N = 3$

    • Sum of scores: $1 + 2 + 3 = 6$

    • Sample mean: $\bar{x} = \frac{6}{3} = 2$

    • Individual Deviances:

    • $1 - 2 = -1$

    • $2 - 2 = 0$

    • $3 - 2 = +1$

    • Total Sum of Deviance: $(-1) + 0 + (+1) = 0$

  • Because the mean is the exact mathematical balance point, positive and negative deviations cancel each other out completely, requiring advanced metrics (Variance and Standard Deviation) to quantify dispersion.

Questions and Audience Discussion

• Question on Positive Skew Naming Confusion:

  • Listener Prompt: In positive skew, the majority of people are clustered on the left (negative side). Why is it called positive skew?

  • Response: The naming convention is strictly defined by the location of the extended tail, not the cluster of people. Utilizing the skier metaphor, a person skiing down the long tail toward the right is sliding toward the positive direction, defining positive skew.

• Question on Temperature as a Distribution Example:

  • Listener Prompt: Is temperature an example of a specific distribution shape?

  • Response: Temperature distributions depend entirely on time scale, geographic location, and sampling parameters. August temperatures in New Jersey over a 10-year span exhibit a negative skew due to global warming trends (more hot days, tail extending left to cooler days). Over a 100-year span, August temperatures form a normal distribution centered around an average (e.g., $85\,^\circ\text{F}$).

  • Clarification: Frequency graphs tracking how many years experienced a given temperature must not be confused with time-series line graphs tracking temperature over sequential years.

• Question on Practical Applications of the Mode:

  • Listener Prompt: What is a practical example where reporting the mode is most appropriate?

  • Response: The mode is appropriate when evaluating categorical inventory data (e.g., Target identifying whether clothing or toys sold the most individual units) or discrete player level distribution in games (e.g., determining that Level $1$ contains the single most frequent player count).

• Question on Outlier Deletion Rules:

  • Listener Prompt: Is there a set rule for deleting outliers, and will students be required to identify outlier causes on exams?

  • Response: No absolute rule exists; handling outliers depends on research goals. Students will learn to spot outliers visually and mathematically, but will not be asked subjective questions regarding the underlying human cause of an outlier.

• Discussion on Bimodal YouTube Analytics:

  • Listener Prompt: How does bimodal data appear in real-world media?

  • Response: Video platforms like YouTube feature visual scrub bars showing "most replayed" sections. Multiple peaks occur where viewers repeatedly rewatch specific key moments, creating a bimodal or multimodal distribution across video timestamp intervals.