Statistical Methods

Statistical Methods — Session 1: Course Overview, Data Types, and Measures of Central Tendency

📅 Published: 2026-06-27

🎓 Level: postgraduate

👥 Audience: Postgraduate students in business analytics, data science, and related fields

Statistical Methods — Session 1: Course Overview, Data Types, and Measures of Central Tendency

1. Data Types and Variables — The Foundation of All Analysis

Hook. You have a spreadsheet with 10,000 rows and 50 columns. Before you run a line of code, ask one question: what kind of stuff is in each column? If you get this wrong, every statistic you compute afterward is suspect. An algorithm that treats hair color as a number will silently produce garbage. Garbage in means garbage out.

The Two Families of Data

Every variable in your dataset belongs to exactly one of two families:

Numerical (quantitative): The values are numbers you can do arithmetic on. Age (23 years), weight (65.3 kg), voltage (3.7 V).
Categorical (qualitative): The values name groups or categories. Marital status (married/unmarried), feedback rating (excellent/good/fair/poor), hair color (black/brown/blonde).

This split is not academic — it decides which statistics you can compute and which plots you can draw.

Intuition. Think of numerical data as answers to "how much?" or "how many?" Think of categorical data as answers to "which kind?" or "which group?" If you can average the values and the result is meaningful, you have numerical data. If averaging the values would be nonsense (what is the average of "black" and "brown"?), you have categorical data.

Sub-types of Numerical Data

Numerical data splits further into two sub-types:

Discrete: The variable can only take specific, separate values — usually whole numbers you arrive at by counting. The set of possible values is {0, 1, 2, 3, …}. Example: number of children in a family. You cannot have 2.5 children.

Continuous: The variable can take any value within an interval on the number line. You arrive at these values by measuring. Example: weight (65.3 kg, 65.35 kg, 65.351 kg — there is always a finer measurement possible).

Formally: a variable $X$ with possible values ${1, 2, 3, 4}$ is discrete. A variable $X$ that can take any value in the interval $[- 1, 1]$ is continuous.

Worked classification. Classify each of these variables:

Variable	Type	Why
Number of cars per household	Discrete numerical	Counted: 0, 1, 2, … — no fractions
Temperature in Celsius	Continuous numerical	Measured: 23.7°C is meaningful
Blood type (A, B, AB, O)	Categorical — nominal	Groups with no natural order
Education level (HS, BS, MS, PhD)	Categorical — ordinal	Groups with a clear order
Number of students in a class	Discrete numerical	Counted whole numbers

Sense-check: If you find yourself saying "on average, 2.3 cars per household," that is fine — the mean of a discrete variable can be fractional. The individual values of the variable are still whole numbers.

Sub-types of Categorical Data

Categorical data splits into two sub-types based on whether order matters:

Nominal: Order does not matter. Hair color, marital status, gender, blood type. You can assign numbers (e.g., married = 1, unmarried = 0), but these are just labels. 1 is not "more" than 0. It is just different.

Ordinal: Order matters. Feedback rating (excellent > good > fair > poor), education level (PhD > MS > BS > HS), satisfaction (very satisfied > satisfied > neutral > dissatisfied). When you assign numbers, you must preserve the order: excellent = 5, good = 4, fair = 3, poor = 2, fail = 1.

Pitfall — confusing nominal and ordinal. A student survey codes "Math" as 1, "Physics" as 2, "Chemistry" as 3. A naive analyst computes the "average subject" as 2.1 and concludes students prefer Physics. This is nonsense — the numbers are nominal labels, and their average has no meaning. Always ask: does the order of these categories carry information? If not, the variable is nominal, and the mean is undefined for it.

The clustering analogy for the syllabus. The professor opened the lecture by noting that the course syllabus itself is an example of clustering — the machine learning technique that groups similar items together. The six modules of this course (Basic Probability → Bayes Theorem → Distributions → Hypothesis Testing → Regression → Time Series) are not a flat list. Each module collects topics that share a theme. You complete Module 1 before Module 2 because each builds on the previous one. This parallels how clustering groups similar data before processing each cluster. This is your first glimpse of how statistical thinking and machine learning thinking are the same thinking.

Data types determine which statistics are valid. Numerical data supports mean, median, and mode. Categorical data supports only mode. Within categorical, ordinal data preserves order; nominal data does not. Know your data type before you compute.

Domain connection. In a production machine learning pipeline, the very first step after loading data is checking the dtype of every column. A column labeled "rating" that pandas reads as object (string) instead of int64 will silently fail in a regression model. Python libraries like scikit-learn require explicit encoding of categorical variables — OneHotEncoder for nominal, OrdinalEncoder for ordinal. Getting this wrong at the start corrupts every downstream result.

2. Mean — The Arithmetic Center

Hook. A company reports that the "average salary" is ₹12 lakhs. You join and discover that 90% of employees earn under ₹6 lakhs, but the CEO earns ₹2 crores. Were you lied to? No — the mean was computed correctly. But the mean alone can be deeply misleading when the data has extreme values.

Definition and Intuition

Think of your data as weights placed at different positions on a long wooden plank. The mean is the exact spot where you would place a fulcrum (pivot) to make the plank balance perfectly horizontal. Each data point pushes down with equal weight; the mean is the balance point that equalizes all the pushes.

In the reference text (Anderson et al., Ch. 3), this is called the center of balance for the dot plot. If you shift one data point far to the right, the balance point must also shift right to compensate — this is why the mean chases outliers.

Where the analogy breaks: The fulcrum analogy assumes equal weights. When observations have different importance (e.g., a GPA with varying credit hours), you need the weighted mean. The fulcrum shifts toward heavier weights.

Formal Definition

The sample mean, written $x ˉ$ (pronounced "x-bar"), is the arithmetic average of $n$ observations:

$x ˉ = \frac{x _{1} + x _{2} + \dots + x _{n}}{n} = \frac{1}{n} i = 1 \sum n x_{i}$

where:

$x_{i}$ is the $i$ -th observation (i = 1, 2, …, n).
$n$ is the total number of observations.
$\sum$ (capital Greek sigma) means "sum over all i from 1 to n".

The population mean is written $μ$ (Greek letter mu) and uses the same formula applied to the entire population rather than a sample.

Notation note: The professor uses $x ˉ$ throughout this course. Textbook treatments (Anderson et al., Ch. 3) use the same convention: $x ˉ$ for sample, $μ$ for population.

Weighted Mean (Extension)

When observations have different importance, use the weighted mean:

$x ˉ_{w} = \frac{\sum _{i = 1}^{n} w _{i} x _{i}}{\sum _{i = 1}^{n} w _{i}}$

where $w_{i}$ is the weight assigned to observation $i$ . The GPA is a weighted mean: $x_{i}$ = grade points (A=4, B=3, …), $w_{i}$ = credit hours for each course. A 4-credit course pulls the mean twice as hard as a 2-credit course.

Worked example — computing the mean. Five college classes have these enrollments: 46, 54, 42, 46, 32. Compute the mean.

$x ˉ = \frac{46 + 54 + 42 + 46 + 32}{5} = \frac{220}{5} = 44$

The mean class size is 44 students.

Now suppose we discover an error: the class with 54 actually has 114 students (an outlier). Recompute:

$x ˉ_{new} = \frac{46 + 114 + 42 + 46 + 32}{5} = \frac{280}{5} = 56$

The mean jumped from 44 to 56 — a 12-student shift — because of one changed value. This is the sensitivity to outliers in action. The other four values stayed exactly the same, yet the mean moved by 27%.

Sense-check: Is 56 a representative class size? No — four of five classes have 46 or fewer students. The mean of 56 describes none of them well. This dataset calls for the median.

Pitfall 1 — using mean for skewed data. Income data, house prices, and reaction times are almost always right-skewed. Reporting only the mean for skewed data gives a distorted picture. Always report the median alongside the mean when skewness is suspected.

Pitfall 2 — mean of categorical codes. If you encode "Math" = 1, "Physics" = 2, "Chemistry" = 3, the mean of these codes is mathematically computable but statistically meaningless. The mean requires interval-level measurement — the distance between values must be meaningful.

Pitfall 3 — mean imputation with outliers. Filling missing values with the mean is common in data preprocessing. But if your data has outliers, the mean you compute is already distorted — and you are injecting that distortion into every missing slot. Use the median for imputation when outliers are present. The professor explicitly flagged this trap.

Visual intuition. Picture a dot plot with values 32, 42, 46, 46, 54 on a number line from 30 to 60. The mean (44) sits slightly left of center because the low value 32 pulls the balance point leftward — but the high value 54 counterbalances it. Now replace 54 with 114. The dot at 114 is far to the right, like a heavy weight at the end of a seesaw. The fulcrum must slide right to 56 to keep the plank level. The mean follows the outlier.

The mean is the balance point of the data. The mean is the most widely used measure of center, but it is fragile. A single extreme value can pull it far from where most data sits.

Domain connection. In A/B testing, the mean difference between control and treatment groups is the primary metric. If your metric is "revenue per user," a few whales in one group can create a significant mean difference. This may not reflect typical user experience. Top tech companies routinely winsorize (cap extreme values) before computing means, or they use the median as a complementary metric.

3. Median — The Positional Center

Hook. You are at a party with 9 other people. Bill Gates walks in. The mean net worth in the room just skyrocketed. But the median — the net worth of the person exactly in the middle of the sorted list — barely moved. Which number better describes the "typical" person in that room?

Definition and Intuition

The median is the value that splits the sorted data exactly in half. Half the observations are below it, half are above it. It does not care how far the extreme values are — only their position in the sorted order matters.

Think of lining up all your friends by height, shortest to tallest. The person standing in the exact middle is the median height. If you replace the tallest person with a professional basketball player (7'2"), the middle person does not change — same position, same height. The median ignores magnitude and respects only rank.

Formal Definition

To find the median of $n$ observations:

Sort the data in ascending order: $x_{(1)} \leq x_{(2)} \leq \dots \leq x_{(n)}$
If $n$ is odd, the median is the single middle value at position $\frac{n + 1}{2}$ :

$Median = x_{(\frac{n + 1}{2})}$

If $n$ is even, the median is the average of the two middle values:

$Median = \frac{x _{(\frac{n}{2})} + x _{(\frac{n}{2} + 1)}}{2}$

The median is also called the 50th percentile — 50% of observations fall below it. (Percentiles are covered formally in Module 3.)

Worked example — odd n. Class sizes: 46, 54, 42, 46, 32. Sort: 32, 42, 46, 46, 54. $n = 5$ (odd). Middle position: $\frac{5 + 1}{2} = 3$ . The 3rd value is 46.

Median = 46.

Worked example — even n. Monthly starting salaries (in $): 3710, 3755, 3850, 3880, 3880, 3890, 3920, 3940, 3950, 4050, 4130, 4325. $n = 12$ (even). Middle two positions: 6th and 7th. Values: 3890 and 3920.

$Median = \frac{3890 + 3920}{2} = \frac{7810}{2} = 3905$

Median starting salary = $3,905.

Sense-check: Count: 6 values below 3905, 6 values above. The split is exact.

Outlier robustness — a direct comparison. Take the class-size data {32, 42, 46, 46, 54}. Mean = 44, Median = 46. Now replace 54 with 500:

Mean: $\frac{32 + 42 + 46 + 46 + 500}{5} = \frac{666}{5} = 133.2$ — jumped by 89.2.
Median: sorted = {32, 42, 46, 46, 500}, middle value = 46 — unchanged.

The median stays at 46 while the mean triples. This is why the median is called a robust statistic.

Pitfall 1 — median requires sorting. Unlike the mean, you cannot compute the median from a running sum. You need the full sorted list. For a dataset with 100 million rows, sorting is expensive (O(n log n) vs. O(n) for the mean). In practice, approximate medians (using histograms or quantile sketches) are used for massive datasets.

Pitfall 2 — median for multimodal data. If your data has two distinct peaks (bimodal), the median can land in the valley between them — a value that describes almost nobody. Example: shoe sizes with modes at 8 and 11. The median might be 9.5, but very few people wear 9.5. Always visualize before summarizing.

Pitfall 3 — median of small even-n samples. When n is small and even, the median is the average of two values — and that average may not be an actual data point. This is normal and expected. Do not round it to a real observation.

Visual intuition. On a dot plot, draw a vertical line that has exactly half the dots to its left and half to its right. That line is the median. Now slide the rightmost dot from 54 to 500, to 5000. The vertical line does not move — it still has 2 dots on each side. The median line is anchored by count, not by distance.

The median is the middle-ranked value. It is immune to outliers because it uses position, not magnitude. When the data is skewed, the median is usually the better measure of "typical."

Domain connection. Government statistics agencies report median household income, not mean, precisely because income distributions are heavily right-skewed. Real estate portals report median home price for the same reason. In data preprocessing for machine learning, SimpleImputer(strategy='median') is the default choice in scikit-learn for strong missing-value imputation — a direct implementation of the professor's advice.

4. Mode — The Most Frequent Value

Hook. A shoe manufacturer surveys 10,000 customers about their shoe size. Sizes 8 and 9 each appear 3,000 times; all other sizes trail far behind. The mean shoe size is 8.7, and the median is 8.5. Which number tells the manufacturer what to produce? Neither. The mode — sizes 8 and 9 — is the only statistic that directly answers the question. "Which sizes do most people actually wear?"

Definition and Intuition

The mode is the value that appears most often. It answers "what is most common?" — a different question than "what is average?" or "what is in the middle?"

Think of a classroom vote. The option that gets the most hands raised wins — that is the mode. Nobody averages the votes or finds the median vote. Elections, popularity contests, and market demand all run on the mode, not the mean.

Formal Definition

The mode is the value (or values) with the highest frequency in the dataset. Frequency is always a whole number (a count).

Unimodal: Exactly one mode. Example: hair colors {black, brown, black, blonde, black} → mode = "black" (frequency = 3).
Bimodal: Exactly two modes (a tie for highest frequency). Example: shoe sizes {8: 300 times, 9: 300 times, 7: 200 times, 10: 200 times} → modes = 8 and 9.
Multimodal: More than two modes.

The mode is the only measure of central tendency defined for categorical (nominal) data. For "hair color," you cannot compute a mean or median — but you can report the mode.

Worked example — categorical data. A survey asks 20 people their favorite color. Results:

Blue: 8, Green: 5, Red: 4, Yellow: 3

The mode is Blue (frequency = 8).

Worked example — numerical data with ties. Quiz scores: 70, 75, 80, 80, 80, 85, 90, 90, 90, 95.

Frequencies: 80 appears 3 times, 90 appears 3 times. Both are modes. The data is bimodal with modes 80 and 90.

Sense-check: The mean is $(70 + 75 + 80 + 80 + 80 + 85 + 90 + 90 + 90 + 95) /10 = 83.5$ . The median is $(80 + 85) /2 = 82.5$ . Neither 83.5 nor 82.5 describes the two distinct performance clusters (around 80 and around 90). The bimodal nature is the real story, and only the mode reveals it.

Pitfall 1 — frequency is always an integer. A student asked. "What if two values have frequencies 8 and 8.1?" This cannot happen. Frequency is a count — you count how many times each value appears. Counts are whole numbers. A frequency of 8.1 is impossible.

Pitfall 2 — mode is not unique. The mean is always a single number. The median is always a single number. The mode can be multiple numbers. If two values tie at the highest frequency, the data is bimodal and both are valid modes. Reporting only one would be misleading.

Pitfall 3 — statistical mode vs. business decision. Statistically, if value A has frequency 8 and value B has frequency 7, the mode is A only. But a business might still produce both A and B if the difference is small. The statistical definition gives you the facts; domain context drives the decision. As the professor said: the tool gives understanding, the domain context determines the action.

Visual intuition. Draw a bar chart with values on the x-axis and frequency (count) on the y-axis. The mode is the tallest bar. In a bimodal distribution, two bars rise above all others — a "camel back" shape. The mean is a single point somewhere along the x-axis; it cannot capture the two-hump structure.

The mode is the most frequent value. It is the only measure of center for nominal categorical data. It is not unique — a dataset can have multiple modes. When the goal is to serve the majority, target the mode.

Domain connection. Product designers target the mode. Door handles are built for right-handed users (the mode of the population). Clothing retailers stock more modal sizes. Streaming services recommend the modal genre for a user segment. In machine learning, the mode sets the baseline for classification. A "dummy classifier" that always predicts the majority class sets the floor that real models must beat.

5. Symmetry and Skewness — Reading the Shape of Data

Hook. You compute the mean, median, and mode for your dataset. They are 120, 85, and 70. Three different "centers" for the same data. Which one is right? All three are right — and the fact that they differ is a signal. The pattern of differences is the shape of your data.

The Three-Center Relationship

The relationship between mean, median, and mode reveals the skewness (asymmetry) of the data:

Symmetric (no skew): mean ≈ median ≈ mode. The left half of the distribution mirrors the right half. The classic bell curve (normal distribution) has this property.

Positively skewed (right-skewed): mean > median > mode. The right tail is longer. Most values cluster on the left, but a few very large values pull the mean rightward. Income data, house prices, and reaction times are typically right-skewed.

Negatively skewed (left-skewed): mean < median < mode. The left tail is longer. Most values cluster on the right, but a few very small values pull the mean leftward. Exam scores in an easy test (most score high, a few score very low) are left-skewed.

Formal Skewness

The reference text (Anderson et al., Ch. 3) gives the sample skewness formula:

$Skewness = \frac{n}{( n - 1 ) ( n - 2 )} i = 1 \sum n (\frac{x _{i} - x ˉ}{s})^{3}$

where $s$ is the sample standard deviation (covered in a later lecture). Interpretation:

Skewness = 0 → symmetric.
Skewness > 0 → right-skewed (positive skew).
Skewness < 0 → left-skewed (negative skew).

For a quick diagnostic without computing the full formula, the professor's rule is practical: compare the mean and median. If mean > median, suspect right skew. If mean < median, suspect left skew. If mean ≈ median, the data is roughly symmetric.

The Normal Distribution Preview

The normal distribution (Gaussian distribution) is the most important distribution in all of statistics. It is symmetric, bell-shaped, and has mean = median = mode. It is fully described by two numbers: the mean $μ$ (center) and the standard deviation $σ$ (spread). About 68% of data falls within $\pm 1 σ$ of the mean, about 95% within $\pm 2 σ$ , and about 99.7% within $\pm 3 σ$ .

The normal distribution appears everywhere: measurement errors, heights, IQ scores, blood pressure readings, and manufacturing tolerances. It is the foundation of hypothesis testing, confidence intervals, and Six Sigma quality control. Module 3 covers it in full detail.

Worked example — skewness diagnosis. A women's apparel store records purchase amounts (Anderson et al., Ch. 3): mean = $77.60, median = $59.70. Since mean > median by a large margin, the data is right-skewed. Most customers spend around $60, but a few big spenders inflate the mean to $77.60. The median ($59.70) is the better summary of a "typical" purchase.

Salary dataset from the lecture: If mean salary = ₹12L but median = ₹6L, the mean is pulled up by high executive salaries. The distribution is right-skewed. Report the median for a representative picture.

Pitfall 1 — mean = median = mode is an idealization. In real data, these three are rarely exactly equal — even for roughly symmetric data. "Approximately equal" is the practical standard. Small differences do not indicate skewness.

Pitfall 2 — skewness does not mean "bad data." Skewed data is not wrong data. Income is naturally right-skewed in every economy. The skewness is the information. Do not transform away skewness without understanding why it exists.

Pitfall 3 — the professor's ordinal relationship (mean > median > mode) holds for unimodal distributions. For multimodal data, the pattern can break. Always visualize before concluding.

Visual intuition. Draw a bell curve (symmetric): the peak, the balance point, and the halfway line all align. Now grab the right tail and stretch it outward. The peak (mode) stays put. The halfway line (median) slides right a bit. The balance point (mean) slides right the most. It chases the stretched tail. This is the visual meaning of mean > median > mode for right-skewed data.

Skewness is the asymmetry of the data. Comparing mean and median is the fastest skewness diagnostic. For skewed data, the median is the preferred measure of center.

Domain connection. Machine learning models that assume normally distributed errors (linear regression, ANOVA) produce unreliable results when applied to heavily skewed data. Data scientists routinely apply log transforms or Box-Cox transforms to reduce skewness before modeling. The skew() function in scipy.stats quantifies skewness; a value outside [-1, 1] often warrants transformation.

6. Preview — Conditional Probability and Bayes Theorem

Hook. You are in a dark room. Something touches your foot. On the 30th floor of a city apartment, you panic — maybe an intruder. In a village hut, you panic — definitely a snake. A 10-month-old baby, in either room, plays happily. Same sensation, three different reactions. The difference is past data — and this is Bayes theorem at work in your brain.

Conditional Probability

Conditional probability is the probability of an event, given that another event has already occurred. It answers questions like. "Given that a customer already has bread in their basket, what is the probability they will also buy butter?"

This thinking drives market basket analysis — the technique supermarkets use to decide shelf placement, product bundling, and promotional offers. The same logic powers website clickstream analysis. "Given that a user visited page A, what is the probability they navigate to page B?"

The BP-diabetes example. Government health authorities observe that blood pressure (BP) and diabetes often appear together. By analyzing medical records, they compute: 70% of patients developed BP first, then diabetes. 30% developed diabetes first, then BP. Conclusion: BP is the more common "root cause." Focusing public health resources on BP prevention may reduce downstream diabetes more effectively. This is conditional probability guiding public policy.

Bayes Theorem — A Thought Process

Bayes theorem is a formula, but more importantly, it is a way of thinking: update your belief based on new evidence.

Prior belief: What you think before seeing data.
Evidence: New data you observe.
Posterior belief: Your updated belief after incorporating the evidence.

The COVID-19 lockdown example. First lockdown (zero data): people stopped everything — maids, newspapers, milk, mail. Extreme caution born of complete uncertainty. Second lockdown (data accumulated): authorities confirmed newspapers do not spread the virus. Hospitals had treatment protocols. Behavior was more relaxed — not careless, but informed by data. The shift from "fear everything" to "selective caution" is Bayes theorem at societal scale. Prior: no data, uniform fear. Evidence: newspapers are safe, treatments exist. Posterior: targeted precautions.

Bayes theorem is the mathematical engine of learning from experience. It underlies the Naive Bayes classifier — one of the simplest yet most effective machine learning algorithms for text classification, spam detection, and sentiment analysis.

Notation note: The formal Bayes formula $P (A ∣ B) = \frac{P ( B ∣ A ) P ( A )}{P ( B )}$ will be derived in Module 2. This preview is about building the intuition first. The professor's approach: understand the thought process before seeing the equation.

7. Preview — Probability Distributions

Hook. Roll a die 600 times. Count how many times you get each face. The counts will be roughly equal — about 100 each. Now measure the heights of 600 random adults. The counts form a bell shape — most people near average, fewer at the extremes. Each pattern is a probability distribution, and recognizing the pattern tells you which statistical tools to use.

What is a Probability Distribution?

A probability distribution describes the pattern that data follows — the shape formed when you plot values against how often they occur. Instead of looking at individual numbers, you ask: "What process generated this shape?"

Named Distributions (Module 3 Preview)

Distribution	What it models	Example
Bernoulli	A single yes/no trial	Coin flip: heads or tails
Binomial	Number of successes in n independent trials	Number of heads in 10 coin flips
Poisson	Count of events in a fixed interval	Number of customer arrivals per hour
Normal (Gaussian)	The bell curve — symmetric, continuous	Heights, measurement errors, IQ scores

Why the Normal Distribution Dominates

The normal distribution appears so often that it has many names: bell curve, Gaussian, normal. The professor gave six reasons it matters:

Nature: Hills, waves, and natural phenomena form bell shapes.
Human traits: Emotions cluster around "normal" with occasional extreme highs and lows.
Performance appraisals: Healthy organizations have ~60% average, ~20% above, ~20% below.
Six Sigma: Manufacturing quality control is built on the normal curve. Sigma (σ) = standard deviation.
Electronics: Gaussian noise is the fundamental model in signal processing.
Psychology: You cannot jump from "normal" to "euphoric" — the path goes through the continuum. Shortcuts (drugs) create artificial spikes; the person feels at the peak but is not actually there.

A probability distribution is a pattern. The normal distribution (bell curve) is the most important pattern in statistics — symmetric, predictable, and the foundation of most inferential methods.

8. Preview — Hypothesis Testing

Hook. You feel thirsty and suspect you are dehydrated. You don't drain all your blood to check. You prick your finger, test a drop, and conclude. That is hypothesis testing — forming a belief about a large population and testing it with a small sample.

The Core Idea

Hypothesis testing is the process of:

Form an assumption (hypothesis) about a population.
Take a representative sample.
Test the sample.
Conclude whether the sample evidence supports or contradicts the hypothesis.

The vegetable market analogy. You see a heap of beans. Hypothesis: "These beans are fresh." You cannot eat the entire heap. You take a small sample, taste it, and conclude about the whole heap. This is hypothesis testing as everyday reasoning.

The medical diagnosis analogy. You suspect diabetes (hypothesis). The doctor takes a few milliliters of blood — not all 5 liters. The sample is tested for blood sugar. Based on the sample, the doctor concludes about your health. Sampling is how you test; hypothesis testing is the entire what and why.

Key distinction — hypothesis testing vs. sample testing. Sample testing is the narrower term — it is the mechanism. Hypothesis testing is the broader framework: forming the hypothesis, choosing the test, setting decision criteria, and drawing the conclusion. The professor explicitly clarified this when a student asked.

You cannot test an entire population. Sampling lets you draw reliable conclusions about the whole from a small, representative piece. This is the inferential nature of statistics.

9. Preview — Correlation, Regression, and Time Series

Hook. Does more study time lead to higher exam scores? Can we predict tomorrow's temperature from today's data? Will the stock market go up or down next week? These three questions span correlation, regression, and time series — the topics that bridge statistics directly into machine learning.

The Three Bridges

Correlation (Module 5): Quantifies the strength and direction of a relationship between two variables. Answers: "Are X and Y related, and how strongly?" Example: correlation between study hours and exam scores.

Regression (Module 5): Builds a model to predict a continuous outcome from one or more predictors. Answers: "Given X, what is the best estimate of Y?" Example: predict salary from years of experience.

Time Series (Module 6): Analyzes data collected sequentially over time. Answers: "What will happen next, based on past patterns?" Example: forecast next month's sales from the last 24 months of data.

Real-world applications from the lecture:

Stock market: Analyze past Sensex/Nifty values to forecast tomorrow's index.
Weather: Use past temperature, humidity, and pressure to predict tomorrow's conditions.
Demand forecasting: Predict product demand from past sales patterns for inventory planning.

Correlation measures relationship. Regression predicts outcomes. Time series forecasts the future. Together, these three topics form the statistical backbone of predictive modeling in machine learning.

10. Course Logistics and Exam Guidance

Study strategy from the professor:

Attend live sessions. They are the primary resource.
Skim slides (shared 2-3 days before class) for 10-15 minutes. Identify what is new vs. familiar.
Make your own handwritten notes in your own words. They are always allowed in open-book exams.
After each session, spend 15-20 minutes summarizing. This locks in understanding.
Use the LMS discussion forum for questions — no question is too basic.
PPTs and your own notes are enough for exams. Textbooks are supplementary.

Evaluation components:

Quiz 1 + Quiz 2: 10% combined (online, LMS Takshila).
Assignment + Situated Learning: online (LMS).
Mid-semester exam: open book, traditional.
Comprehensive exam: open book, traditional.

No makeup for missed quizzes or assignments. Set reminders. Use your peer network. Open book: PPTs, textbooks, and your own handwritten notes are allowed. Solution manuals and printed guides are typically not. Grading: Relative (not absolute). Your grade depends on performance relative to the class.

Deadline discipline. Once a quiz or assignment deadline passes, there is no reopening, no redo, no makeup. With busy work schedules, it is easy to miss a deadline. Form a study group — peers reminding each other is the most reliable safety net.

Key Industry Applications

Application Area	Statistical Concept	Real-World Use
Market Basket Analysis	Conditional Probability	Supermarket shelf arrangement, product bundling, promotional offers
Public Health Policy	Conditional Probability	Identifying root-cause diseases for targeted prevention
User Journey Analysis	Conditional Probability	Website clickstream analysis, conversion funnel optimization
Performance Management	Normal Distribution	Employee appraisal rating distribution
Quality Control	Normal Distribution	Six Sigma defect reduction in manufacturing
Signal Processing	Gaussian Distribution	Noise modeling in electronics and communications
Medical Diagnosis	Hypothesis Testing	Blood sample testing to diagnose conditions
Stock Market Forecasting	Time Series Analysis	Index value prediction
Weather Prediction	Time Series Analysis	Temperature, rainfall, storm forecasting
Demand Forecasting	Time Series Analysis	Inventory planning from past sales
Customer Ticket Classification	Clustering	Grouping support tickets by issue type
Data Preprocessing	Mean vs Median	Choosing imputation strategy for missing values
Spam Detection	Bayes Theorem (Naive Bayes)	Email spam filtering, sentiment analysis
A/B Testing	Hypothesis Testing	Comparing two versions of a product or webpage
Risk Assessment	Bayes Theorem	Updating fraud probability from transaction patterns

ISM Lecture 1 notes · Session 1: Course Overview, Data Types, and Measures of Central Tendency

Introduction to Statistical Methods· postgraduate· 2026-06-27

Summary

Statistical Methods Session 1 covers the foundational concepts of data classification and measures of central tendency. Students learn to distinguish numerical data (discrete vs continuous) from categorical data (nominal vs ordinal) and understand how data type determines which statistical methods are valid. The three measures of central tendency — mean, median, and mode — are each defined with intuition, formal computation, worked examples, and common pitfalls. The mean is presented as the balance point sensitive to outliers; the median as the robust positional center; and the mode as the most frequent value essential for categorical data. The relationship between these three measures reveals skewness: mean > median > mode indicates right skew, while mean < median < mode indicates left skew. The lecture also previews upcoming topics including conditional probability, Bayes theorem as belief updating, probability distributions (Bernoulli, Binomial, Poisson, Normal), hypothesis testing, correlation, regression, and time series — building the statistical foundation for machine learning.

Learning Objectives

1Distinguish between numerical (discrete, continuous) and categorical (nominal, ordinal) data types

2Compute and interpret the mean, median, and mode for a given dataset

3Understand the weighted mean and its applications (e.g., GPA calculation)

4Explain why the median is robust to outliers while the mean is not

5Diagnose skewness by comparing mean and median

6Recognize the normal distribution as symmetric with mean = median = mode

7Understand the intuition behind conditional probability and Bayes theorem

8Identify appropriate measures of central tendency based on data type and distribution shape

Sections Breakdown

1Data Types and Variables

Classification of data into numerical (discrete, continuous) and categorical (nominal, ordinal) types with worked examples

2Mean — The Arithmetic Center

Definition, intuition, weighted mean, computation, and pitfalls of the arithmetic average

3Median — The Positional Center

Definition, computation for odd/even n, outlier robustness, and comparison with mean

4Mode — The Most Frequent Value

Definition, unimodal/bimodal/multimodal distributions, and use for categorical data

5Symmetry and Skewness

Relationship between mean, median, and mode; positive and negative skew; normal distribution preview

6Preview — Conditional Probability and Bayes Theorem

Intuition for conditional probability and Bayes theorem as updating beliefs with evidence

7Preview — Probability Distributions

Bernoulli, Binomial, Poisson, and Normal distributions

8Preview — Hypothesis Testing

Core idea of forming and testing hypotheses using samples

9Preview — Correlation, Regression, and Time Series

Relationship measurement, prediction modeling, and forecasting

10Course Logistics and Exam Guidance

Study strategy, evaluation components, and deadline discipline

Postgraduate students in business analytics, data science, and related fields

Exam Revision Notes

Below is the distilled, exam-ready core of this lecture. Every entry is built from the full textbook notes above. Use this section for rapid review — but if something doesn't make sense, go back to the full explanation in the main content.

Data Types and Measurement Scales

Must-know: Data type determines which statistics are valid. Numerical data (discrete or continuous) supports mean, median, and mode. Categorical data (nominal or ordinal) supports only mode.

No formula. Rules: Numerical = answers how much/how many. Categorical = answers which kind/which group.

Top pitfall: Computing the mean of nominal codes (e.g., Math=1, Physics=2 → average 1.5 is meaningless).

Self-check: Is education level (HS, BS, MS, PhD) numerical or categorical? If categorical, is it nominal or ordinal?

Connects to: Mean, Median, Mode, Skewness

Mean (Arithmetic Average)

Must-know: The mean is the balance point of the data. It is sensitive to outliers.

$x ˉ = \frac{1}{n} i = 1 \sum n x_{i}$ where $x ˉ$ is the sample mean, $n$ is the number of observations, and $x_{i}$ is the i-th observation.

Top pitfall: Using the mean for skewed data without reporting the median alongside.

Self-check: Five class sizes: 32, 42, 46, 46, 54. What is the mean?

Connects to: Weighted Mean, Median, Skewness

Weighted Mean

Must-know: When observations have different importance, use the weighted mean. GPA is the classic example.

$x ˉ_{w} = \frac{\sum _{i = 1}^{n} w _{i} x _{i}}{\sum _{i = 1}^{n} w _{i}}$ where $w_{i}$ is the weight and $x_{i}$ is the value.

Top pitfall: Using the unweighted mean when observations have different significance.

Self-check: A student scores A=4 in a 4-credit course and B=3 in a 2-credit course. What is the GPA?

Connects to: Mean

Median

Must-know: The median is the middle-ranked value, immune to outliers. For skewed distributions, the median is usually better.

Sort ascending. Odd n: value at position (n+1)/2. Even n: average of values at positions n/2 and n/2+1.

Top pitfall: Median requires sorting (expensive for massive datasets). For bimodal data, median can fall in the valley.

Self-check: Salaries: 3710, 3755, 3850, 3880, 3880, 3890, 3920, 3940, 3950, 4050, 4130, 4325. What is the median?

Connects to: Mean, Skewness, Percentiles

Mode

Must-know: The mode is the most frequent value. It is the only measure of center for nominal categorical data. Datasets can have multiple modes.

No formula. The mode is the value(s) with the highest frequency (count).

Top pitfall: Confusing statistical mode with business decisions. Frequency is always an integer.

Self-check: Shoe sizes: 8 appears 300 times, 9 appears 300 times, 7 appears 200 times. What are the modes?

Connects to: Data Types, Mean, Median

Skewness and Symmetry

Must-know: Comparing mean and median is the fastest skewness diagnostic. Mean > median suggests right skew.

$Skewness = \frac{n}{( n - 1 ) ( n - 2 )} i = 1 \sum n (\frac{x _{i} - x ˉ}{s})^{3}$ where $s$ is sample standard deviation. > 0 = right-skewed, < 0 = left-skewed.

Top pitfall: Assuming mean = median = mode is an exact requirement. Approximately equal is the standard.

Self-check: A store reports mean = $77.60 and median = $59.70. Is the distribution left-skewed or right-skewed?

Connects to: Mean, Median, Mode, Normal Distribution

Normal Distribution

Must-know: The normal distribution is symmetric, bell-shaped, with mean = median = mode. Described by mean μ and std dev σ.

68-95-99.7 rule: ~68% within ±1σ, ~95% within ±2σ, ~99.7% within ±3σ.

Top pitfall: Assuming all data is normally distributed. Many real-world datasets are skewed.

Self-check: IQ scores are normally distributed with mean 100 and std dev 15. What range contains ~95% of scores?

Connects to: Skewness, Hypothesis Testing, Regression

Conditional Probability and Bayes Theorem

Must-know: Conditional probability is the probability of an event given another event. Bayes theorem updates beliefs with new evidence.

$P (A ∣ B) = \frac{P ( B ∣ A ) P ( A )}{P ( B )}$ where P(A|B) is posterior, P(B|A) is likelihood, P(A) is prior, P(B) is evidence.

Top pitfall: Confusing P(A|B) with P(B|A). They are not the same.

Self-check: If 70% develop BP before diabetes and 30% develop diabetes before BP, which is more likely the root cause?

Connects to: Hypothesis Testing, Naive Bayes Classifier

Practice Quiz

Test your understanding of ISM Lecture 1 notes. Select an answer for each question — results are instant.

Which measure of central tendency is the only one defined for nominal categorical data?

A dataset has mean = 120, median = 85, and mode = 70. What does this tell you about the distribution?

Why do government agencies report median household income instead of mean income?

Previous LectureLecture 20 · Microservices, Kubernetes, and Serverless Computing

Next LectureLecture 02 · Measures of Variability, Dispersion, and Introduction to Probability

Was this lecture useful?

Loading comments…

🤖

BitsNotes AI Assistant

Subject Notes Assistant

Configure AI Chat

Choose how to access the chatbot

Have your own API key?

Switch to "Bring Your Own Key" tab above for unlimited access with any OpenAI-compatible provider.

AI Provider Preset

API Endpoint URL (OpenAI-Compatible)

API Key

Model Name

🔑 Enter API key above to fetch live models from provider, or enter model name manually.

Remember key for this session (until tab closes)

OpenAI-Compatible API Support

Choose any provider preset (Gemini, DeepSeek, Kimi, GLM, MiniMax, Qwen, OpenAI, Groq, Ollama, etc.) or enter a custom endpoint URL.

Security & Privacy First

Your API key is sent directly from your browser to your specified provider. BitsNotes servers never store or see your key.

Statistical Methods — Session 1: Course Overview, Data Types, and Measures of Central Tendency

1. Data Types and Variables — The Foundation of All Analysis

The Two Families of Data

Sub-types of Numerical Data

Sub-types of Categorical Data

2. Mean — The Arithmetic Center

Definition and Intuition

Formal Definition

Weighted Mean (Extension)

3. Median — The Positional Center

Definition and Intuition

Formal Definition

4. Mode — The Most Frequent Value

Definition and Intuition

Formal Definition

5. Symmetry and Skewness — Reading the Shape of Data

The Three-Center Relationship

Formal Skewness

The Normal Distribution Preview

6. Preview — Conditional Probability and Bayes Theorem

Conditional Probability

Bayes Theorem — A Thought Process

7. Preview — Probability Distributions

What is a Probability Distribution?

Named Distributions (Module 3 Preview)

Why the Normal Distribution Dominates

8. Preview — Hypothesis Testing

The Core Idea

9. Preview — Correlation, Regression, and Time Series

The Three Bridges

10. Course Logistics and Exam Guidance

Key Industry Applications

ISM Lecture 1 notes · Session 1: Course Overview, Data Types, and Measures of Central Tendency

Summary

Learning Objectives

Sections Breakdown

Exam Revision Notes

Data Types and Measurement Scales

Mean (Arithmetic Average)

Weighted Mean

Median

Mode

Skewness and Symmetry

Normal Distribution

Conditional Probability and Bayes Theorem

Practice Quiz

0 / 3

Was this lecture useful?

BitsNotes AI Assistant

Configure AI Chat

Sign in to use AI Chat