Business Statistics Notes (UGC NET Unit 5)

Business Statistics Syllabus Map11 topic branches · 12 sections · 107 PYQs mapped to their topic

AMeasures of Central Tendency

Averages — mean, median, mode and their empirical relationship — the starting point for every other statistic in this unit.

BMeasures of Dispersion

How spread out the data is: range, quartile deviation, mean deviation, standard deviation and the coefficient of variation.

CMeasures of Skewness

Whether a distribution leans left or right, and the Karl Pearson, Bowley and Kelly coefficients used to measure it.

DCorrelation and Regression of Two Variables

How two variables move together (correlation) and how one predicts the other (regression lines and coefficients).

EProbability: Approaches and Bayes' Theorem

The three approaches to probability, the addition and multiplication rules, and revising probabilities with Bayes' theorem.

FProbability Distributions

Binomial, Poisson and Normal distributions — their conditions, formulas and how they relate to one another.

GResearch: Concept, Types and Research Designs

What research is, how it is classified, and the exploratory, descriptive and experimental designs used to carry it out.

HData: Collection and Classification of Data

Primary vs secondary data, the methods used to collect it, and the bases and scales used to classify it.

KReport Writing

Types of research reports, the steps in writing one, its layout, and the traits of a good report.

PYQ chips link to each exam cycle’s Unit 5 question paper on webcomm.in (the question number shown is its number in the full Paper 2, not this unit-only post).

Business Statistics and Research Methods Notes: UGC NET Commerce Unit 5

Business Statistics and Research Methods is Unit 5 of the UGC NET Commerce syllabus. These exam-oriented notes are written in simple English and cover every topic of the unit: measures of central tendency, measures of dispersion, measures of skewness, correlation and regression of two variables, probability and the Bayes theorem, probability distributions (binomial, Poisson and normal), research concept and types, data collection and classification, sampling and estimation, hypothesis testing (z-test, t-test, ANOVA, chi-square, Mann-Whitney, Kruskal-Wallis, rank correlation) and report writing.

Tap a heading below to jump to that topic, or use the buttons to open or close every section at once. All diagrams and worked examples use illustrative figures. After revising, practise with the Unit 5 previous year questions.

1. Exam-Strategy / Topic-Priority Map

Based on 221 previous-year questions classified by topic across many exam cycles, here is how the eleven syllabus topics of Unit 5 rank by how often they are actually asked:

PriorityTopicPYQs classified
1Measures of central tendency, dispersion and skewness45
2Correlation and regression of two variables31
2Hypothesis testing (z, t, ANOVA, chi-square, Mann-Whitney, Kruskal-Wallis, rank correlation)31
4Sampling and estimation27
5Probability distributions: binomial, Poisson, normal26
6Research: concept, types and research designs22
7Data: collection and classification14
8Probability: approaches and the Bayes theorem13
9Report writing5

How to use this post

  • Start revision with the Priority 1-2 groups above — together they account for more than half of all Unit 5 questions asked across the audited exam cycles.
  • Each section below ends with a colour-coded mind map and dedicated diagrams for its key relationships and formulas — read the paragraph first, then use the diagram to recall the whole concept quickly before the exam.
  • After revising a topic, practise with the Unit 5 previous year questions.
2. Measures of Central Tendency

Central tendency means finding a single representative value that describes the whole of a set of data. Instead of listing every sale, every score or every wage separately, a single average tells the reader where the data is centred, which makes comparison across groups or across time possible. The three most used central tendency measures in the UGC NET syllabus are the mean, the median and the mode, and each belongs to one of two broad families.

Two families of averages

Mathematical averages are calculated using every value in the data through a fixed arithmetic or algebraic rule — the arithmetic mean, the geometric mean and the harmonic mean all belong here. Positional averages do not use every value in a formula; instead they locate the value that sits at a particular position once the data is arranged in order — the median and the mode belong here. This distinction is a favourite source of assertion-reason and matching questions in the exam, so it is worth remembering on its own before the individual formulas.

Arithmetic Mean

The arithmetic mean (AM) is the sum of all values divided by the number of values. For ungrouped data with individual observations X, the formula is simple.

Arithmetic Mean (ungrouped): AM = ΣX / N

For grouped (discrete or continuous) data, each value X (or each class midpoint m, for continuous data) carries a frequency f, so the mean is a frequency-weighted average.

Arithmetic Mean (grouped, direct method): AM = ΣfX / Σf  (or Σfm / Σf using class midpoints m)

When the class midpoints are large numbers, the direct method becomes tedious, so examiners often expect the step-deviation (short-cut) method, which shifts the data around an assumed mean A and scales by the class width i to keep the arithmetic simple.

Arithmetic Mean (step-deviation method): AM = A + (Σfd′ / Σf) × i,  where d′ = (m − A) / i

The arithmetic mean uses every observation, is capable of further algebraic treatment (which is why variance and standard deviation are both built on it) and is the most common average asked about in PYQs. Its main weakness is that it is pulled strongly by extreme values — a single very large salary in a small team can inflate the mean salary well above what most employees actually earn.

Worked Example

The monthly sales (in ₹ lakh) of 5 branches of a company are 12, 15, 18, 20 and 25. Find the arithmetic mean.

AM = (12 + 15 + 18 + 20 + 25) / 5 = 90 / 5 = ₹18 lakh

Geometric Mean

The geometric mean (GM) is used whenever values multiply together rather than add — growth rates, index numbers and compound interest are the classic business examples, because an average growth rate found by simple addition would be misleading. For n ungrouped values, the GM is the nth root of their product, which in practice is calculated through logarithms.

Geometric Mean (ungrouped): GM = Antilog [ (Σlog X) / N ]
Geometric Mean (grouped): GM = Antilog [ (Σf log X) / Σf ]

The GM is always less than or equal to the arithmetic mean for the same data (equal only when every value is identical), and unlike the AM it is not distorted by one very large value, though it cannot be used at all if any value is zero or negative.

Harmonic Mean

The harmonic mean (HM) is the reciprocal of the average of the reciprocals of the values, and it is the correct average whenever the data is expressed as a rate (speed, price per unit, output per hour) and the question asks for an average rate over equal amounts of the other quantity (equal distance, equal money spent, equal time).

Harmonic Mean (ungrouped): HM = N / Σ(1/X)
Harmonic Mean (grouped): HM = Σf / Σ(f/X)

For any set of positive values that are not all identical, the three mathematical averages always sit in the same order, which is itself a frequently tested fact.

AM ≥ GM ≥ HM  (equality holds only when all values are equal)

Median

The median is the value that lies exactly in the middle once the data is arranged in ascending or descending order — half the observations lie below it and half lie above it. For grouped continuous data, the median class is the class in which the cumulative frequency first reaches or exceeds N/2, and the exact value is interpolated within that class.

Median = L + [ (N/2 − cf) / f ] × i

Here L is the lower boundary of the median class, cf is the cumulative frequency of the class before the median class, f is the frequency of the median class itself, and i is the median class width. The median can also be read directly off a graph by plotting a cumulative frequency curve (an ogive) and reading the value on the X-axis that corresponds to N/2 on the Y-axis — a method the exam sometimes asks about by name without expecting a calculation.

Because the median depends only on the position of values and not on their exact size, it is not affected by extreme values, which makes it the preferred average for skewed data such as income or property prices.

Worked Example

Daily wages (₹) of 7 workers, arranged in order: 300, 320, 350, 400, 420, 450, 500. Find the median.

N = 7 (odd), so the median is the value at position (N+1)/2 = 4th value = ₹400

Mode

The mode is the value that occurs most frequently in the data — the single most “popular” value. For grouped continuous data, the modal class is the class with the highest frequency, and the exact modal value is interpolated using the frequencies of the classes just before and after it.

Mode = L + [ (f1 − f0) / (2f1 − f0 − f2) ] × i

Here L is the lower boundary of the modal class, f1 is the frequency of the modal class, f0 is the frequency of the class preceding it, f2 is the frequency of the class following it, and i is the class width. When a distribution has two classes with similarly high frequencies and the simple formula gives an unclear answer, examiners expect the grouping method, which systematically totals frequencies in overlapping groups of two and three classes to confirm which class is genuinely the mode.

A distribution can have one mode (unimodal), two modes (bimodal) or more, and unlike the mean and median it can even be undefined if every value occurs with equal frequency.

Empirical relationship between Mean, Median and Mode

In a perfectly symmetric distribution, the mean, median and mode all coincide at the same point. Most real business and economic data is not perfectly symmetric, so Karl Pearson’s empirical formula is used to estimate the mode from the mean and median without recalculating it directly, and the same formula is frequently inverted in exam questions to find any one of the three from the other two.

Mode = 3 × Median − 2 × Mean  (equivalently: Mean − Mode = 3 × (Mean − Median))

In a positively skewed distribution (a long tail toward higher values, such as income data), the order from left to right is Mode, then Median, then Mean. In a negatively skewed distribution, the order reverses to Mean, then Median, then Mode.

Two families of averages: mathematical and positional Central tendency splits into Mathematical Averages (Arithmetic Mean, Geometric Mean, Harmonic Mean, calculated from every value) and Positional Averages (Median, Mode, located by position in ordered data). Central Tendency Mathematical Averages Arithmetic Mean AM = ΣX / N Geometric Mean GM = Antilog(ΣlogX/N) Harmonic Mean HM = N / Σ(1/X) AM ≥ GM ≥ HM (calculated from every value) Positional Averages Median middle value Mode most frequent value not affected by extreme values Exam tip: a positional average needs the data ordered first; a mathematical average uses every value directly in a formula.
When to use Arithmetic Mean, Geometric Mean and Harmonic Mean Arithmetic Mean is used for simple averages where values add. Geometric Mean is used for growth rates and ratios where values multiply. Harmonic Mean is used for rates such as speed or price per unit, averaged over equal amounts of the other quantity. Arithmetic Mean AM = ΣX / N Use when values simply add up: sales, marks, wages, heights. Geometric Mean GM = Antilog(ΣlogX/N) Use when values multiply: growth rates, index numbers, CAGR. Harmonic Mean HM = N / Σ(1/X) Use for rates: speed, price per unit, output per hour (equal base). AM ≥ GM ≥ HM (equality only when every value in the data is identical) Exam tip: speed/average-rate questions where distance is equal across trips almost always test the Harmonic Mean.
Direct method vs step-deviation method for the arithmetic mean Direct method sums f times midpoint and divides by total frequency. Step-deviation method shifts data around an assumed mean A, scales by class width i, which is faster for large midpoint values. Grouped data: classes + frequencies Direct Method small / round midpoints AM = Σfm / Σf multiply each midpoint by its frequency, sum, then divide by total frequency Step-Deviation Method large / uneven midpoints d′ = (m − A) / i AM = A + (Σfd′/Σf) × i A = assumed mean, i = class width Both methods give exactly the same final answer — step-deviation is just faster arithmetic, never a different mean.
Reading the median and mode formulas off a histogram Median formula uses L the lower boundary of the median class, cf the cumulative frequency before it, f its own frequency, and i the class width. Mode formula uses the modal class frequency f1 and its neighbours f0 and f2. Median class — cumulative frequency reaches N/2 cf = freq. before median class f = median class freq. L L+i Median = L + [(N/2 − cf)/f] × i N/2 falls inside this shaded class, so interpolate within it Modal class — the tallest bar, f1 f0 f1 f2 f1 is the modal class Mode = L + [(f1−f0)/(2f1−f0−f2)] × i L = lower boundary of the modal class i = class width
Where mean, median and mode sit under symmetric and skewed distributions In a symmetric distribution mean, median and mode coincide. In a positively skewed distribution with a long right tail, the order left to right is mode, median, mean. In a negatively skewed distribution the order reverses to mean, median, mode. Symmetric Mean = Median = Mode Mean = Median = Mode Positively Skewed Mode Median Mean Mode < Median < Mean Negatively Skewed Mean Mode Mode Mean < Median < Mode Positive skew = long tail to the right (e.g. income data, most earn little, a few earn a lot) — this pulls the Mean further right than the Mode.
Mind map: Measures of Central Tendency Central hub Measures of Central Tendency with five colour-coded branches: Arithmetic Mean, Geometric and Harmonic Mean, Median, Mode, and the empirical relationship between mean, median and mode. Measures of Central Tendency Arithmetic Mean AM = ΣX/N most common, affected by outliers GM & HM growth rates (GM), rates like speed (HM) AM ≥ GM ≥ HM Mode most frequent value can be bimodal Mode=L+[(f1-f0)/ (2f1-f0-f2)]×i Relationship Mode = 3Median − 2Mean positive skew: Mode<Median<Mean Median middle value, not affected by extreme values
3. Measures of Dispersion

While central tendency locates the centre of the data, dispersion measures how spread out or scattered the values are around that centre. Two datasets can have exactly the same mean and still behave very differently in practice — one tightly clustered, one wildly spread — and dispersion is what tells them apart. This is why exam questions often combine a central tendency calculation with a dispersion calculation on the same dataset.

Absolute and relative measures

Like averages, dispersion measures split into two groups. Absolute measures (range, quartile deviation, mean deviation, standard deviation) are expressed in the same unit as the original data — rupees, kilograms, marks — which makes them unsuitable for comparing two datasets measured in different units or with very different means. Relative measures (coefficient of range, coefficient of quartile deviation, coefficient of mean deviation, coefficient of variation) are pure ratios with no unit, built by dividing an absolute measure by an average, which makes cross-comparison possible.

Range

Range is the simplest measure of dispersion: the difference between the largest and the smallest value in the data.

Range = L − S  (L = largest value, S = smallest value)
Coefficient of Range = (L − S) / (L + S)

Range uses only the two extreme values and ignores everything in between, so it is easy to calculate but highly sensitive to outliers and tells nothing about how the middle of the data behaves.

Quartile Deviation

Quartile deviation (QD), also called the semi-interquartile range, fixes range’s biggest weakness by ignoring the extreme 25% at each end and looking only at the middle 50% of the data, between the first quartile (Q1) and the third quartile (Q3).

Quartile Deviation = (Q3 − Q1) / 2
Coefficient of QD = (Q3 − Q1) / (Q3 + Q1)

Because it excludes the extremes, QD is not affected by outliers, but like the median it uses only positional values and ignores the exact distances of individual observations from the centre.

Mean Deviation

Mean deviation (MD) measures the average of the absolute distances of every value from a chosen central point, usually the mean or the median. Taking the absolute value is essential — without it, deviations on either side of the average would cancel out to zero.

Mean Deviation (about A) = Σ|X − A| / N  (ungrouped);  = Σf|X − A| / Σf  (grouped)
Coefficient of MD = Mean Deviation / A

MD uses every value in the data, which range and QD do not, but the use of absolute value makes it algebraically awkward for further statistical work — this is exactly the gap that standard deviation fills.

Variance and Standard Deviation

Standard deviation (SD) is the most important and most frequently tested measure of dispersion. Instead of taking the absolute value of each deviation, it squares the deviations (which also removes the sign problem, since a square is always positive), averages them, and then takes a square root to bring the answer back to the original unit. The squared-but-not-rooted version is called the variance.

Variance (σ²) = Σ(X − Mean)² / N    Standard Deviation (σ) = √Variance

For grouped data, the direct formula is frequency-weighted, and just as with the arithmetic mean, examiners often expect the faster step-deviation shortcut for large or uneven midpoints.

SD (grouped, direct) = √[ Σf(X − Mean)² / Σf ]
SD (step-deviation): σ = i × √[ Σfd′²/Σf − (Σfd′/Σf)² ],  where d′ = (m − A)/i

Standard deviation uses every observation, is the most algebraically useful measure (it underlies correlation, regression and most of hypothesis testing later in this unit), but like the mean it is pulled by extreme values.

Worked Example

Daily sales (₹ thousand) of a shop on 5 days: 10, 12, 14, 16, 18. Mean = 14. Find the variance.

Deviations from mean: −4, −2, 0, 2, 4. Squared: 16, 4, 0, 4, 16. Sum = 40.

Variance = 40 / 5 = 8  (so SD = √8 ≈ 2.83)

Coefficient of Variation

Coefficient of variation (CV) is the relative measure built from the standard deviation, and it is the single most common way examiners ask you to compare the consistency of two different datasets.

CV = (Standard Deviation / Mean) × 100

A lower CV means the data is more consistent (less spread relative to its own average); a higher CV means more variability. CV is essential whenever the two datasets being compared have different units (for example, one set of exam marks out of 100 and another out of 50) or very different means (comparing the variability of salaries in two companies with very different average pay) — a raw standard deviation comparison would be misleading in both cases, but CV puts them on the same relative scale.

Worked Example

Branch A: mean sales ₹50,000, SD ₹5,000. Branch B: mean sales ₹80,000, SD ₹6,000. Which branch is more consistent?

CV(A) = (5,000/50,000) × 100 = 10%.  CV(B) = (6,000/80,000) × 100 = 7.5%.

Branch B has the lower CV, so Branch B is more consistent, even though its SD in rupees is larger.

Absolute and relative measures of dispersion Absolute measures (Range, Quartile Deviation, Mean Deviation, Standard Deviation) are in the data’s own unit. Relative measures (their coefficients, plus Coefficient of Variation) are unit-free ratios used to compare two different datasets. Dispersion Absolute Measures (in original unit) Range L − S Quartile Deviation Mean Deviation Standard Deviation Relative Measures (unit-free ratios) Coefficients of Range, QD, MD (absolute measure ÷ its average) used to compare two datasets Coefficient of Variation CV = (SD / Mean) × 100 Exam tip: a relative measure is unit-free, so it is the correct choice whenever a question compares two datasets with different units or means.
Range and quartile deviation on a number line Range uses the full span from the smallest value S to the largest value L. Quartile deviation uses only the middle 50 percent of the data, between the first quartile Q1 and the third quartile Q3, so it ignores outliers at either end. S Q1 Median Q3 L Range = L − S (uses the two extremes) Quartile Deviation = (Q3 − Q1) / 2 (middle 50% only — ignores outliers beyond Q1/Q3) Range is quick but one extreme value can distort it; quartile deviation is more stable because it drops the extremes.
How standard deviation is built, step by step Standard deviation is built in four steps: find deviations from the mean, square them to remove negative signs, average the squares to get variance, then take the square root to return to the original unit. 1. Deviations X − Mean can be + or − 2. Square them (X − Mean)² removes the sign 3. Average Σ(X−Mean)² / N this is the Variance 4. Square root √Variance back to original unit Standard Deviation (σ) = √[ Σ(X − Mean)² / N ] the result is called Variance before the square root, SD after it For large or uneven class midpoints, use the step-deviation shortcut: σ = i × √[ Σfd′²/Σf − (Σfd′/Σf)² ], d′ = (m−A)/i
Why coefficient of variation beats comparing standard deviation directly Branch A has a smaller standard deviation in rupees than Branch B, but once divided by its own much smaller mean, Branch A’s coefficient of variation is higher, showing it is actually less consistent. Branch A Mean = ₹50,000 SD = ₹5,000 CV = (5,000/50,000)×100 = 10% Branch B Mean = ₹80,000 SD = ₹6,000 CV = (6,000/80,000)×100 = 7.5% Comparing SD alone: ₹5,000 < ₹6,000 looks like A is more consistent — but that ignores how much bigger Branch B is already. Comparing CV instead: 10% (A) > 7.5% (B) so Branch B is actually the more consistent branch. Exam tip: whenever two datasets have different means or units, compare their coefficient of variation, never their raw SD.
Mind map: Measures of Dispersion Central hub Measures of Dispersion with five colour-coded branches: Range, Quartile Deviation, Mean Deviation, Standard Deviation and Variance, and Coefficient of Variation. Measures of Dispersion Range L − S simplest, sensitive to outliers Quartile Deviation (Q3−Q1)/2 middle 50% only, ignores outliers Mean Deviation Σ|X−A|/N uses every value, absolute distance SD & Variance √[Σ(X−Mean)²/N] most used, basis of correlation Coefficient of Variation CV=(SD/Mean)×100 compares datasets
4. Measures of Skewness

What Skewness Measures

Measures of central tendency locate the centre of a distribution, and measures of dispersion show how spread out it is. Neither tells you whether the distribution is symmetric or lopsided. Skewness measures exactly that — the degree and direction in which a distribution departs from symmetry. A perfectly symmetric distribution has its longer tail nowhere; a skewed one has a longer tail stretching out to one side, dragging the Mean away from the Median and Mode.

Positive and Negative Skewness

The sign of skewness tells you which side the long tail is on, and that in turn fixes the order of Mean, Median and Mode.

Symmetric distribution: Mean = Median = Mode, skewness = 0
Positively skewed (right-skewed): long tail to the right ⇒ Mean > Median > Mode
Negatively skewed (left-skewed): long tail to the left ⇒ Mean < Median < Mode

A quick way to remember the direction: the Mean is the average that is most pulled by extreme values, so it always sits closest to the long tail. The Mode, being the most “typical” value, sits furthest from the tail, and the Median falls in between.

Karl Pearson’s Coefficient of Skewness

This is the most commonly asked measure. It expresses skewness relative to the standard deviation, so it is a relative measure and can be used to compare two different distributions.

SkP = (Mean − Mode) / Standard Deviation

When the mode is ill-defined (bimodal data, or a class with no clear single peak), the empirical relationship Mode = 3Median − 2Mean is substituted in, giving the equivalent and more commonly used form:

SkP = 3(Mean − Median) / Standard Deviation

The value of Karl Pearson’s coefficient usually lies between −3 and +3, though in practice it is almost always a small fraction such as 0.2 or −0.4. A positive value confirms positive skew, a negative value confirms negative skew, and a value at or near zero confirms a nearly symmetric distribution.

Worked example — Karl Pearson’s coefficient

A firm’s daily sales (in ₹’000) have Mean = 62, Median = 58, and Standard Deviation = 12. Since the mode is not clearly defined in this data, use the median-based formula.

SkP = 3(62 − 58) / 12 = 3(4)/12 = 12/12 = +1
A positive coefficient of 1 confirms the distribution is positively skewed — a few unusually high-sales days are pulling the Mean above the Median.

Bowley’s Coefficient of Skewness

Karl Pearson’s measure uses every value through the Mean and SD, so a few extreme outliers can distort it. Bowley’s coefficient is a quartile-based alternative that looks only at the middle 50% of the data, making it more robust to outliers — the same robustness advantage that Quartile Deviation has over Range.

SkB = (Q3 + Q1 − 2×Median) / (Q3 − Q1)

The logic: in a symmetric distribution, Q3 and Q1 are equidistant from the Median, so Q3 + Q1 = 2×Median and the numerator is zero. When the upper quartile is pulled further from the Median than the lower quartile is, the distribution is positively skewed, and the numerator turns positive.

Bowley’s coefficient always lies between −1 and +1, which makes it easy to judge the strength of skew directly from its magnitude.

Worked example — Bowley’s coefficient

For a wage distribution, Q1 = ₹180, Median = ₹240, and Q3 = ₹360.

SkB = (360 + 180 − 2×240) / (360 − 180) = (540 − 480) / 180 = 60/180 = +0.33
A positive Bowley coefficient of 0.33 again points to positive skew: the gap between the Median and Q3 (₹120) is wider than the gap between Q1 and the Median (₹60), showing the upper half of wage-earners is more spread out.

Kelly’s Coefficient of Skewness

A less frequently tested but occasionally asked variant replaces quartiles with percentiles, using the 10th and 90th percentiles instead of Q1 and Q3. This captures skewness in the tails rather than just the middle 50%.

SkK = (P90 + P10 − 2×P50) / (P90 − P10)

P50 here is simply the Median. The interpretation is identical to Bowley’s measure — a positive value means positive skew, a negative value means negative skew.

Choosing the Right Measure

Karl Pearson’s coefficient is the default choice in most numerical problems because it is based on the Mean and SD, which are already calculated for most data. Bowley’s coefficient becomes the better choice specifically when the data has open-ended classes (no defined upper or lower limit) or extreme outliers, since Mean and SD cannot be reliably computed in those cases but quartiles still can.

Shapes of symmetric, positively skewed and negatively skewed distributions Three curve shapes side by side: a symmetric bell curve with Mean, Median and Mode together at the centre; a positively skewed curve with a long right tail where Mode is leftmost, Median in the middle and Mean furthest right; a negatively skewed curve with a long left tail where Mean is furthest left, Median in the middle and Mode furthest right. Symmetric Mean = Median = Mode skewness = 0 Positively Skewed Mode < Median < Mean long tail to the right Negatively Skewed Mean < Median < Mode long tail to the left Mean=Med=Mode Mode Median Mean Mean Median Mode Mnemonic: the Mean always sits closest to the long tail, the Mode furthest from it.
Karl Pearson’s coefficient of skewness, two forms Karl Pearson’s coefficient of skewness is Mean minus Mode divided by standard deviation. When the mode is not clearly defined, the median-based form is used instead: three times Mean minus Median, divided by standard deviation. Karl Pearson’s Coefficient of Skewness a relative measure, usually between −3 and +3 When Mode is well-defined Sk = (Mean − Mode) / SD used directly on the raw data When Mode is ill-defined Sk = 3(Mean − Median) / SD uses Mode = 3Median − 2Mean Sk > 0: positive skew  •  Sk < 0: negative skew  •  Sk ≈ 0: symmetric Both forms give nearly the same value, since Mode ≈ 3Median − 2Mean
Bowley’s quartile-based coefficient of skewness Bowley’s coefficient compares the distance from Q1 to the Median against the distance from the Median to Q3. When the upper gap is bigger, skewness is positive; when the lower gap is bigger, skewness is negative; equal gaps mean symmetric. Example: upper gap (Median→Q3) bigger than lower gap (Q1→Median) Q1 Median Q3 lower gap: Q1 → Median = 60 upper gap: Median → Q3 = 120 Sk_B = (Q3 + Q1 − 2×Median) / (Q3 − Q1) upper gap > lower gap ⇒ Sk_B > 0 ⇒ positively skewed Always between −1 and +1, and robust to outliers since only quartiles are used.
Comparing the three coefficients of skewness Karl Pearson’s coefficient uses Mean, Mode or Median and standard deviation and ranges from minus three to plus three. Bowley’s coefficient uses quartiles and ranges from minus one to plus one. Kelly’s coefficient uses the tenth and ninetieth percentiles and also ranges from minus one to plus one. Karl Pearson Sk = (Mean−Mode)/SD or 3(Mean−Median)/SD Uses: Mean, Mode/Median, SD Range: −3 to +3 Best when: no outliers, mode/mean reliable Bowley Sk = (Q3+Q1−2Med) / (Q3−Q1) Uses: Q1, Median, Q3 Range: −1 to +1 Best when: outliers or open-end classes present Kelly Sk = (P90+P10−2P50) / (P90−P10) Uses: P10, Median, P90 Range: −1 to +1 Best when: tail behaviour matters more than middle All three share the same sign rule: positive ⇒ right tail, negative ⇒ left tail, zero ⇒ symmetric.
Mind map: Measures of Skewness Central hub Measures of Skewness with four colour-coded branches: Meaning and sign rule, Karl Pearson’s coefficient, Bowley’s coefficient, and Kelly’s coefficient. Measures of Skewness Meaning & Sign Mean>Med>Mode: positive Mean<Med<Mode: negative all equal: symmetric Karl Pearson (Mean−Mode)/SD or 3(Mean−Median)/SD range −3 to +3 Bowley (Q3+Q1−2Med)/(Q3−Q1) quartile-based, robust to outliers Kelly (P90+P10−2P50) /(P90−P10) percentile-based
5. Correlation and Regression of Two Variables

What Correlation Measures

Correlation measures the degree and direction of the linear relationship between two variables — for example, advertising spend and sales, or study hours and exam marks. It tells you whether the two variables move together, move in opposite directions, or show no consistent pattern at all. Correlation does not by itself prove that one variable causes the other; it only measures how strongly they move together.

Types of Correlation

Positive correlation: both variables increase or decrease together (e.g. height and weight)
Negative correlation: one variable increases as the other decreases (e.g. price and demand)
Zero/no correlation: no consistent relationship between the two variables

Correlation can also be classified by the number of variables involved (simple — two variables; multiple — one variable against two or more others; partial — the relationship between two variables holding a third constant) and by the shape of the relationship (linear — a straight-line pattern; non-linear/curvilinear — a curved pattern). A scatter diagram, which simply plots one variable against the other as points, is the simplest way to visually judge the type and strength of correlation before calculating any coefficient.

Covariance: The Building Block of r

Before computing r itself, it helps to understand covariance — a measure of how two variables change together. Unlike variance, which describes how a single variable spreads around its own mean, covariance describes whether two variables tend to rise and fall together, move in opposite directions, or show no joint pattern.

Cov(X,Y) = Σ(X−X̄)(Y−Ȳ) / N

Covariance is positive when both variables tend to move in the same direction, negative when they move in opposite directions, and near zero when there is no linear tendency between them. Its one drawback is that it carries units (the product of X’s and Y’s units), which makes it hard to judge how strong the relationship is. Dividing covariance by the product of the two standard deviations removes the units and rescales it to the −1-to-+1 range — which is exactly how Karl Pearson’s r is defined:

r = Cov(X,Y) / (σxσy)

Karl Pearson’s Coefficient of Correlation

This is the most widely used and most frequently tested measure. It gives a precise numerical value for both the direction and the strength of a linear relationship.

r = Σ(X−X̄)(Y−Ȳ) / √[Σ(X−X̄)² × Σ(Y−Ȳ)²]

For raw data, the equivalent direct-value formula avoids calculating deviations separately:

r = [NΣXY − (ΣX)(ΣY)] / √{[NΣX² − (ΣX)²][NΣY² − (ΣY)²]}

Karl Pearson’s r always lies between −1 and +1. A value of +1 means perfect positive correlation (every point lies exactly on a rising straight line), −1 means perfect negative correlation, and 0 means no linear correlation at all. r is also a pure number with no unit, and it is symmetric: the correlation of X with Y is identical to the correlation of Y with X.

Worked example — Karl Pearson’s r

For 5 pairs of (advertising, sales) data, the calculations give: Σ(X−X̄)(Y−Ȳ) = 60, Σ(X−X̄)² = 40, Σ(Y−Ȳ)² = 160.

r = 60 / √(40 × 160) = 60 / √6400 = 60/80 = 0.75
A coefficient of 0.75 shows a strong positive correlation — as advertising spend rises, sales tend to rise along with it.

Spearman’s Rank Correlation Coefficient

When the data is already in the form of ranks (such as positions in a competition, or when the actual values are hard to measure numerically, like beauty or honesty ratings), Spearman’s coefficient is used instead of Karl Pearson’s.

R = 1 − [6Σd² / N(N²−1)]

Here d is the difference between the ranks given to the same item by the two judges/criteria, and N is the number of items ranked. When two or more items tie for the same rank, a correction term m(m²−1)/12 is added to Σd² for each group of m tied ranks before applying the formula. Like Karl Pearson’s r, Spearman’s R also ranges from −1 to +1 with the same sign interpretation.

Worked example — Spearman’s rank correlation

Two judges rank 5 contestants, and the squared rank differences work out to Σd² = 4.

R = 1 − [6(4) / 5(25−1)] = 1 − [24/120] = 1 − 0.2 = 0.8
An R of 0.8 shows the two judges largely agree on the ranking.

Regression: Meaning and the Two Lines

Regression goes a step beyond correlation: instead of just measuring how strongly two variables are related, it estimates the actual value of one variable (the dependent variable) from the value of the other (the independent variable), using a line of best fit. Because X can be used to predict Y, or Y can be used to predict X, there are always two different regression lines for the same data (except when r = ±1, when both lines coincide).

Regression line of Y on X: Y = a + byxX   (used to estimate Y from a given X)
Regression line of X on Y: X = a + bxyY   (used to estimate X from a given Y)

Both lines pass through the point (X̄, Ȳ), the means of the two series, which is why that point is where the two lines always intersect.

Regression Coefficients and Their Link to r

The two regression coefficients are calculated as:

byx = r × (σy / σx)    and    bxy = r × (σx / σy)

Three results connecting the regression coefficients and Karl Pearson’s r come up repeatedly in exams:

1. byx and bxy always carry the same sign as r.
2. r is the geometric mean of the two regression coefficients: r² = byx × bxy, so r = ±√(byx×bxy).
3. The product byx × bxy can never exceed 1, since r² itself can never exceed 1.
Worked example — regression equation of Y on X

Given r = 0.6, X̄ = 20, Ȳ = 30, σx = 4, σy = 6.

byx = r(σy/σx) = 0.6 × (6/4) = 0.9
Regression equation: Y − Ȳ = byx(X − X̄) ⇒ Y − 30 = 0.9(X − 20) ⇒ Y = 0.9X + 12

Coefficient of Determination (R²)

In the two-variable case, squaring Karl Pearson’s r gives the coefficient of determination, R², which has a very practical meaning: it is the proportion of the total variation in the dependent variable that is explained by the independent variable through the regression line, rather than left unexplained.

R² = r²   (for two variables)   ⇒   R² = byx × bxy

For instance, r = 0.8 gives R² = 0.64, meaning 64% of the variation in Y is explained by X through the regression line, while the remaining 36% is due to other factors. A higher R² means the regression line is a better fit to the data; R² = 1 means the line explains the data perfectly, and R² = 0 means X explains none of the variation in Y.

Correlation vs Regression — the Key Distinction

Correlation is symmetric and only measures the degree of association (rxy = ryx); it does not distinguish which variable depends on which. Regression is asymmetric and is specifically built for prediction — it requires you to label one variable as dependent and the other as independent, and the two regression lines (Y on X, and X on Y) are generally different lines, not the same line viewed two ways.

Scatter diagrams for positive, negative and zero correlation Three scatter diagrams side by side: points rising left to right show positive correlation, points falling left to right show negative correlation, and scattered points with no pattern show zero correlation. Positive Correlation Negative Correlation Zero / No Correlation A scatter diagram is the first visual check before computing any correlation coefficient.
Karl Pearson’s coefficient of correlation and its strength scale Karl Pearson’s r formula using deviations, and a strength scale from minus one through zero to plus one showing perfect negative, strong negative, weak, strong positive and perfect positive correlation zones. Karl Pearson’s Coefficient of Correlation r = Σ(X−X̄)(Y−Ȳ) / √[Σ(X−X̄)² × Σ(Y−Ȳ)²] Strength scale (same shape for correlation in either direction): −1 −0.7 0 +0.7 +1 strong negative weak negative negligible weak positive strong positive r is a pure number (no unit) and symmetric: r of X with Y = r of Y with X Correlation measures association only — it never by itself proves causation.
Spearman’s rank correlation coefficient Spearman’s rank correlation compares the ranks given by two judges to the same five items, computes the squared rank differences, and applies the formula one minus six sigma d squared over n times n squared minus one. R = 1 − [6Σd² / N(N²−1)] Item Judge 1 rank Judge 2 rank d d² A 1 2 -1 1 B 2 1 1 1 C 3 3 0 0 D 4 5 -1 1 E 5 4 1 1 Σd² = 4 R = 1 − [6(4) / 5(25−1)] = 1 − [24/120] = 0.8 Tied ranks: add m(m²−1)/12 to Σd² for each group of m tied ranks
The two regression lines, Y on X and X on Y Two straight lines crossing at the point X-bar, Y-bar, both sloping upward for positively correlated data but at different steepness: the steeper line is the regression of Y on X, used to estimate Y from X, and the shallower line is the regression of X on Y, used to estimate X from Y. When r equals plus or minus one the two lines coincide; when r equals zero they are at right angles. X Y (X̄, Ȳ) Y on X X on Y Y = a + b_yx X   (estimate Y from a given X) X = a + b_xy Y   (estimate X from a given Y) Both lines pass through (X̄,Ȳ); they coincide when r = ±1, and are at right angles when r = 0.
Relationship between the two regression coefficients and r The regression coefficient of Y on X equals r times the ratio of the standard deviations, and the regression coefficient of X on Y equals r times the inverse ratio. Three rules follow: both coefficients share the sign of r, r is the geometric mean of the two coefficients, and their product can never exceed one. b_yx = r × (σy/σx) regression of Y on X b_xy = r × (σx/σy) regression of X on Y 1. Same sign b_yx and b_xy always carry the same sign as r. 2. Geometric mean r² = b_yx × b_xy, so r = ±√(b_yx×b_xy). 3. Product ≤ 1 b_yx × b_xy can never exceed 1, since r² ≤ 1. Example: if b_yx = 0.9 and b_xy = 0.4, then r = √(0.9×0.4) = √0.36 = 0.6
Mind map: Correlation and Regression of Two Variables Central hub Correlation and Regression with five colour-coded branches: Types of correlation, Karl Pearson’s r, Spearman’s rank correlation, Regression lines, and the b-yx b-xy r relationship. Correlation & Regression Types positive / negative / zero simple / partial / multiple linear / non-linear Karl Pearson r Σ(X−X̄)(Y−Ȳ)/√[Σx²Σy²] range −1 to +1 unit-free, symmetric Spearman R 1−6Σd²/N(N²−1) used for ranks tie correction m(m²−1)/12 Regression Lines Y on X & X on Y meet at (X̄,Ȳ) coincide when r=±1 b_yx, b_xy each = r×SD ratio r²=b_yx×b_xy same sign as r
6. Probability: Approaches and Bayes Theorem

Basic Terms: Experiment, Sample Space, Event

A random experiment is any well-defined process whose result cannot be predicted with certainty in advance, but all of whose possible results are known — tossing a coin, rolling a die, or drawing a card. The sample space (S) is the set of all possible outcomes of the experiment, and an event is any subset of the sample space — one outcome or a combination of outcomes that shares some characteristic of interest.

Example: Rolling one die → S = {1, 2, 3, 4, 5, 6}; the event “odd number” = {1, 3, 5}

Two events are mutually exclusive if they cannot both occur in the same trial (their intersection is empty, A∩B = ∅) — such as rolling a 2 and rolling a 5 on a single die. They are exhaustive if, together, they cover every possible outcome of the experiment, so that at least one of them must occur.

Three Approaches to Probability

There are three distinct ways of assigning a probability value to an event, and exam questions often test which approach is being used in a given scenario.

1. Classical (a priori): P(A) = (favourable outcomes) / (total equally likely outcomes) — used when all outcomes are equally likely, decided before any trial (e.g. a fair coin or an unbiased die)
2. Empirical / Frequency (a posteriori): P(A) = (number of times A occurred) / (total trials) — based on actually repeating the experiment many times and observing relative frequency
3. Subjective: a probability based on personal judgement, experience, or available information, used when neither equally-likely outcomes nor repeatable trials exist (e.g. “there is a 70% chance this project finishes on time”)

The classical and empirical approaches are both considered objective (anyone repeating the same procedure should reach the same value), while the subjective approach is personal and can vary from one person to another.

The Addition Rule

The addition rule gives the probability that at least one of two events occurs. Its exact form depends on whether the two events can happen together.

General rule: P(A or B) = P(A) + P(B) − P(A and B)
Mutually exclusive events only: P(A or B) = P(A) + P(B) — since P(A and B) = 0

The subtraction of P(A and B) in the general rule exists purely to avoid double-counting the outcomes that belong to both events; when the events cannot overlap at all, there is nothing to subtract.

Worked example — addition rule

A single card is drawn from a standard 52-card deck. Find the probability that the card is a King or a Heart.

P(King) = 4/52, P(Heart) = 13/52, P(King and Heart) = 1/52 (the King of Hearts)
P(King or Heart) = 4/52 + 13/52 − 1/52 = 16/52 = 4/13

Conditional Probability and the Multiplication Rule

Conditional probability is the probability of an event occurring given that another event has already occurred. It is written P(A|B), read as “the probability of A given B”.

P(A|B) = P(A and B) / P(B)   (provided P(B) ≠ 0)

Rearranging this definition gives the multiplication rule for the joint probability of two events:

P(A and B) = P(B) × P(A|B)

Two events are independent when the occurrence of one has no effect on the probability of the other — that is, P(A|B) = P(A). In that special case, the multiplication rule simplifies to a direct product:

Independent events: P(A and B) = P(A) × P(B)

Bayes’ Theorem

Bayes’ theorem reverses a conditional probability: it lets you start from how likely some evidence is under a hypothesis, and work out how likely the hypothesis is given that the evidence was actually observed. This updates a starting belief (the prior probability) in light of new data (the likelihood) into an updated belief (the posterior probability).

P(H|E) = [P(H) × P(E|H)] / P(E)

Here P(H) is the prior probability of the hypothesis H before seeing the evidence, P(E|H) is the likelihood — how probable the evidence E is if H is true — P(E) is the overall (unconditional) probability of observing the evidence at all, and P(H|E), the result, is the posterior probability of H after accounting for the evidence. When the sample space is partitioned into several mutually exclusive and exhaustive hypotheses, P(E) is expanded as the sum of P(Hi)×P(E|Hi) over every hypothesis Hi, so that all the ways the evidence could have arisen are accounted for.

Worked example — Bayes’ theorem

A factory has two machines. Machine 1 produces 60% of the output and has a 2% defect rate; Machine 2 produces 40% of the output and has a 5% defect rate. A randomly chosen item is found to be defective. What is the probability it came from Machine 2?

P(M2) = 0.40, P(Defective|M2) = 0.05 ⇒ P(M2 and Defective) = 0.40 × 0.05 = 0.020
P(M1) = 0.60, P(Defective|M1) = 0.02 ⇒ P(M1 and Defective) = 0.60 × 0.02 = 0.012
P(Defective) = 0.020 + 0.012 = 0.032
P(M2|Defective) = 0.020 / 0.032 = 0.625
So although Machine 2 makes fewer items, a defective item is more likely (62.5%) to have come from it, because its defect rate is so much higher.
Three approaches to probability A tree showing probability split into objective and subjective approaches. Objective splits further into classical (equally likely outcomes) and empirical (based on repeated trials). Subjective is based on personal judgement. Probability Objective Subjective personal judgement Classical (a priori) equally likely outcomes, decided before any trial P(A)=favourable/total Empirical (a posteriori) based on repeated trials, relative frequency observed No equally-likely outcomes or repeated trials needed Classical and Empirical are both objective — anyone repeating the procedure gets the same value
The addition rule: mutually exclusive versus overlapping events Two Venn diagrams. Left: two non-overlapping circles for mutually exclusive events A and B, with the simple addition rule. Right: two overlapping circles showing the shared region that must be subtracted once in the general addition rule. A B Mutually Exclusive (no overlap, A∩B=∅) A B A∩B Overlapping Events (share some outcomes) P(A or B) = P(A) + P(B) (simple sum — nothing to subtract) P(A or B) = P(A)+P(B)−P(A and B) (subtract the shared region once)
Conditional probability and the multiplication rule Formula boxes for conditional probability, the general multiplication rule, and the simplified multiplication rule for independent events, with a note on what independence means. Conditional Probability P(A|B) = P(A and B) / P(B) Multiplication Rule (general) P(A and B) = P(B) × P(A|B) Multiplication Rule (A and B independent) P(A and B) = P(A) × P(B) Independence means P(A|B) = P(A): knowing B occurred does not change the probability of A at all (e.g. two separate coin tosses are independent)
Bayes’ theorem with its components labelled The Bayes theorem formula, P of H given E equals P of H times P of E given H, divided by P of E, with callout labels identifying the prior probability, the likelihood, the overall probability of the evidence, and the posterior probability. Bayes’ Theorem P(H|E) = [P(H) × P(E|H)] / P(E) Posterior Probability of H, updated after seeing evidence E Prior Probability of H, before seeing any evidence Likelihood how likely E is, assuming H is true P(E) — total probability of the evidence, summed across all hypotheses P(E) = Σ over all H_i: Σ P(H_i) × P(E|H_i) (sum across every mutually exclusive, exhaustive hypothesis) Bayes updates a starting belief (prior) using new data (likelihood) into an updated belief (posterior)
Mind map: Probability, approaches and Bayes’ theorem Central hub Probability with four colour-coded branches: three approaches, the addition rule, conditional probability and independence, and Bayes theorem. Probability & Bayes’ Theorem Three Approaches Classical: favourable/total Empirical: relative frequency Subjective: personal judgement Addition Rule P(A or B)=P(A)+P(B)−P(AB) mutually exclusive: just add exhaustive: covers all outcomes Conditional & Independence P(A|B)=P(AB)/P(B) P(AB)=P(B)×P(A|B) independent: P(AB)=P(A)P(B) Bayes’ Theorem P(H|E)=P(H)P(E|H)/P(E) prior → likelihood → posterior P(E)=ΣP(H_i)P(E|H_i)
7. Probability Distributions: Binomial, Poisson and Normal

Discrete vs Continuous Probability Distributions

A probability distribution lists every possible value a random variable can take, together with the probability of each value. Distributions fall into two broad families depending on what kind of values the variable can take.

Discrete distributions: the variable takes only countable, separate values (0, 1, 2, 3 …) — Binomial and Poisson belong here
Continuous distributions: the variable can take any value within a range (height, weight, time) — the Normal distribution belongs here

This section covers one worked example of each family: the Binomial and Poisson distributions (discrete), and the Normal distribution (continuous).

Binomial Distribution

The Binomial distribution models the number of “successes” in a fixed number of independent trials, where every trial has only two possible outcomes. The prefix “Bi” itself signals two outcomes. Four conditions must all hold:

1. A fixed number of trials, n
2. Each trial has exactly two possible outcomes — “success” or “failure” (a Bernoulli trial)
3. The probability of success, p, stays constant across every trial (and q = 1−p is the probability of failure)
4. The trials are independent of one another
P(X = r) = nCr × pr × q(n−r)   — probability of exactly r successes in n trials
Mean (μ) = np   |   Variance (σ²) = npq   |   Standard Deviation (σ) = √(npq)

Since q is always less than 1, variance (npq) is always smaller than the mean (np) — a quick check that a claimed mean/variance pair is actually consistent with a Binomial model. The distribution is symmetric only in the special case p = q = 0.5; otherwise it is skewed toward whichever of success or failure is rarer.

Worked example — Binomial distribution

A machine produces items with a 10% defect rate. In a random sample of 5 items, find the mean and variance of the number of defective items.

n = 5, p = 0.10, q = 0.90
Mean = np = 5 × 0.10 = 0.5
Variance = npq = 5 × 0.10 × 0.90 = 0.045

Poisson Distribution

The Poisson distribution models the number of times a rare, random event occurs within a fixed interval of time, distance, area, or volume — network failures per week, customer arrivals per hour, typographical errors per page. It applies when events happen independently of one another and the average rate of occurrence is known and constant.

P(X = r) = (e−λ × λr) / r!   — where λ (lambda) is the average number of occurrences in the interval
Mean (μ) = λ   |   Variance (σ²) = λ   — the defining feature: mean and variance are always equal

The Poisson distribution can also be seen as a limiting case of the Binomial distribution: when the number of trials n becomes very large, the probability of success p becomes very small, but the product np stays fixed at a moderate value λ, the Binomial distribution converges to the Poisson distribution with that same λ. This is useful when n is large and p is small enough that computing Binomial probabilities directly becomes impractical.

Worked example — Poisson distribution

A call centre receives an average of 4 calls per minute. Find the probability that exactly 2 calls arrive in a given minute. (e−4 ≈ 0.0183)

λ = 4, r = 2
P(X=2) = (e−4 × 4²) / 2! = (0.0183 × 16) / 2 = 0.2928 / 2 = ≈ 0.1464

Normal Distribution

The Normal distribution is the most important continuous distribution in statistics: a smooth, bell-shaped, symmetric curve defined completely by two parameters — its mean (μ) and standard deviation (σ). Its key properties:

1. Bell-shaped and perfectly symmetric about the mean
2. Mean = Median = Mode, all located at the centre
3. Total area under the curve = 1 (i.e. 100% of all probability)
4. Both tails extend infinitely and approach, but never touch, the horizontal axis

Any Normal variable X can be converted to the standard normal variable Z, which always has mean 0 and standard deviation 1, making it possible to use one common table (the Z-table) for every Normal distribution regardless of its own mean and standard deviation.

Z = (X − μ) / σ

The empirical rule describes how probability is spread around the mean in any Normal distribution:

μ ± 1σ covers ≈ 68.27% of the data   |   μ ± 2σ covers ≈ 95.45%   |   μ ± 3σ covers ≈ 99.73%
Worked example — standard normal / Z-score

The heights of a population are Normally distributed with mean 165 cm and standard deviation 10 cm. Find the probability that a randomly chosen person is between 155 cm and 175 cm tall.

Z for 155 = (155−165)/10 = −1    Z for 175 = (175−165)/10 = +1
This is exactly the μ ± 1σ range, so P(155 < X < 175) ≈ 68.27% (by the empirical rule)

Choosing the Right Distribution

A quick way to decide which distribution a question is describing: if the variable counts successes out of a known, fixed number of trials, think Binomial; if it counts rare occurrences over a continuous interval of time/space with no natural upper limit, think Poisson; if it describes a continuously measurable quantity that clusters symmetrically around an average, think Normal.

Probability distributions: discrete versus continuous A tree with Probability Distribution splitting into Discrete, which holds Binomial and Poisson, and Continuous, which holds Normal distribution. Probability Distribution Discrete Continuous Binomial fixed n trials, two outcomes each mean=np, var=npq Poisson rare events per interval of time/space mean=var=λ Normal bell-shaped, symmetric mean=median=mode Z = (X−μ)/σ
Binomial distribution: conditions, formula and mean/variance Four conditions for a Binomial distribution, its probability mass formula, and its mean, variance and standard deviation formulas, with a note that variance is always less than the mean. Binomial Distribution — 4 Conditions 1. Fixed number of trials, n 2. Each trial: exactly two outcomes (success/failure) 3. Probability of success p constant every trial 4. Trials are independent Probability Mass Function P(X=r) = nCr × p^r × q^(n−r) Mean (μ) = np   |   Variance (σ²) = npq   |   SD (σ) = √(npq) since q < 1, variance is always smaller than the mean
Poisson distribution: formula, examples and relation to Binomial The Poisson probability formula, its mean equals variance property, a strip of real-life examples, and a note that Poisson is the limiting case of Binomial when n is large and p is small. Poisson Distribution P(X=r) = (e^−λ × λ^r) / r! Mean (μ) = λ   =   Variance (σ²) the defining feature: mean and variance are always equal Real-life examples of Poisson events Network failures per week Customer arrivals per hour Calls at a call centre per minute Typos per page Website visitors per hour Poisson as a limiting case of Binomial when n → very large, p → very small, but np stays fixed at λ, Binomial → Poisson(λ)
The normal distribution and the empirical rule A bell-shaped normal curve with shaded bands at one, two and three standard deviations from the mean, labelled with the empirical rule percentages 68.27, 95.45 and 99.73 percent. −3σ −2σ −1σ μ 1σ 2σ 3σ 68.27% 95.45% 99.73% Empirical Rule: μ±1σ≈68.27%  |  μ±2σ≈95.45%  |  μ±3σ≈99.73% Mean = Median = Mode at the centre; curve is symmetric; total area = 1 Standardization: Z = (X − μ) / σ  —  always mean 0, SD 1 lets any Normal variable be read off one common Z-table
Mind map: Binomial, Poisson and Normal distributions Central hub Probability Distributions with four colour-coded branches: discrete versus continuous, Binomial, Poisson, and Normal with the empirical rule. Probability Distributions Discrete vs Continuous Discrete: Binomial, Poisson Continuous: Normal countable vs measurable Binomial fixed n, 2 outcomes mean=np, var=npq P(X=r)=nCr p^r q^(n-r) Poisson rare events/interval mean=var=λ limit of Binomial, n→∞, p→0 Normal & Empirical Rule bell-shaped, symmetric Z=(X−μ)/σ μ±1,2,3σ≈68/95/99.7%
8. Research: Concept, Types and Research Designs

Meaning and Objectives of Research

Research is a systematic, objective and logical process of investigation undertaken to discover new facts, verify existing facts, or find a reliable solution to a problem.
Key features: systematic (follows a planned sequence of steps), objective (free from the researcher’s personal bias), empirical (based on observable, verifiable evidence), and replicable (another researcher following the same steps should reach similar conclusions).
Objectives of research (Exploration → Description → Diagnosis → Prediction/Hypothesis-testing): to explore a new or poorly-understood area, to describe characteristics of a group or phenomenon accurately, to diagnose the frequency with which something occurs or is associated with something else, and to test a hypothesis about a causal relationship between variables.

Types of Research

By purpose — Pure (Fundamental) Research: undertaken purely to add to the existing stock of knowledge, with no immediate practical application in mind; e.g. a theoretical study of consumer-behaviour models.
By purpose — Applied Research: undertaken to solve a specific, practical, real-world problem faced by an organisation or society; e.g. a study to find why a particular product’s sales are declining.
By approach — Quantitative Research: deals with numerical data, uses statistical and mathematical tools, and aims at measurement and generalisable conclusions (e.g. a survey measuring customer satisfaction scores).
By approach — Qualitative Research: deals with non-numerical data — opinions, motivations, behaviours — gathered through interviews, focus groups or observation; aims at depth of understanding rather than statistical generalisation.
By method: Historical research (studies past events/records), Survey research (structured questionnaires on a sample), Case-study research (in-depth study of one unit — a firm, an event), and Experimental research (controlled manipulation of variables to establish cause-and-effect).

Research Design — Meaning and Features

Research design is the overall blueprint or master plan of a research study. It specifies the methods and procedures for collecting, measuring, and analysing data — i.e. what, where, when, how much, and by what means the study will be conducted.
Features of a good research design: Flexibility (can adapt to new information), Appropriateness (fits the research problem and objectives), Efficiency (minimises time and effort), Economy (minimises cost), Reliability (yields consistent results on repetition), and Validity (actually measures what it is intended to measure).

Types of Research Design

Exploratory research design: used when the research problem is loosely defined or not yet well understood. Flexible and unstructured; typically uses literature review, expert interviews, focus groups, and case studies. Goal: to gain insight and formulate a sharper problem or hypothesis for later study — not to reach final conclusions.
Descriptive research design: used when the problem is already well-defined and the goal is to describe characteristics accurately — the “who, what, when, where, how” of a population or phenomenon. Structured and pre-planned; typically uses large-sample surveys and observational studies. Can be cross-sectional (data at one point in time) or longitudinal (data over repeated time periods on the same subjects).
Causal (experimental) research design: used to establish cause-and-effect relationships between variables. Involves a controlled experiment where an independent variable is deliberately manipulated and its effect on a dependent variable is measured, usually comparing a test group against a control group to rule out other explanations.
Worked Example

A company has no idea why a new product is underperforming and wants to generate possible explanations before committing to a full study. Which research design fits best?

Answer: Exploratory research design — the problem is loosely defined at this stage, so the goal is to gain insight and generate hypotheses (e.g. via focus groups with customers) rather than to test or describe anything precisely yet.

Research Process — Key Steps

1. Identify and formulate the research problem — state clearly what is being investigated.
2. Review of literature — study existing work to avoid duplication and sharpen the problem.
3. Formulate the hypothesis — a tentative, testable statement about the relationship between variables.
4. Prepare the research design — choose exploratory, descriptive, or causal design as appropriate, and plan the sampling and data-collection method.
5. Collect data, analyse, and interpret — gather data per the design, apply suitable statistical tools, and draw conclusions.
6. Prepare the research report — present findings, limitations, and recommendations (see Report Writing, a later section of this unit).
Worked Example

A researcher wants to measure the average monthly savings of salaried employees across different income brackets in a city, using a large structured questionnaire administered once. Which research design is this, and is it cross-sectional or longitudinal?

Answer: Descriptive research design — the problem (measuring average savings by income bracket) is already well-defined, and the goal is to describe, not explore or establish causation. Since data is collected once rather than repeatedly over time, it is cross-sectional.
Types of research by purpose, approach and method A tree showing Research splitting by purpose into Pure and Applied, by approach into Quantitative and Qualitative, and by method into Historical, Survey, Case study and Experimental. Research By Purpose By Approach By Method Pure (Fundamental) adds to knowledge Applied solves a specific practical problem Quantitative numerical data, statistical tools Qualitative opinions, depth, interviews/groups Historical past records Survey questionnaire on a sample Case Study in-depth, one unit Experimental controlled, cause-effect Purpose = why; Approach = kind of data; Method = how data is gathered (a single study can combine one choice from each axis)
Research process: six key steps A flowchart of the research process from identifying the problem through literature review, hypothesis formulation, research design, data collection and analysis, to the final report. 1. Identify the Problem state clearly what is being investigated 2. Review of Literature study existing work to sharpen the problem 3. Formulate the Hypothesis a tentative, testable statement 4. Prepare the Research Design choose exploratory/descriptive/causal; plan sampling 5. Collect, Analyse & Interpret Data gather per design, apply statistical tools 6. Prepare the Research Report present findings, limitations, recommendations
Exploratory, descriptive and causal research designs compared Three side-by-side panels comparing exploratory, descriptive and causal research design by when it is used, its structure, typical methods, and goal. Exploratory When used problem loosely defined Structure flexible, unstructured Methods lit. review, focus groups, case study Goal gain insight, sharpen the problem Descriptive When used problem well-defined Structure pre-planned, structured Methods large-sample surveys, observation Goal describe “who, what, when, where, how” Causal When used relationship suspected Structure controlled experiment Methods test group vs control group Goal establish cause-and-effect Studies often move Exploratory → Descriptive → Causal as understanding deepens
Six features of a good research design Six colour-coded badges: flexibility, appropriateness, efficiency, economy, reliability and validity, each with a one-line meaning. A Good Research Design Is… Flexibility adapts to new information Appropriateness fits the problem & objectives Efficiency minimises time & effort Economy minimises cost Reliability consistent results on repetition Validity measures what it intends to measure
Mind map: research concept, types and research designs Central hub Research with five colour-coded branches: meaning, types of research, research design meaning, types of research design, and features of a good design. Research: Concept & Designs Meaning systematic, objective, empirical, replicable investigation Types of Research Pure vs Applied Quantitative vs Qualitative Research Design the blueprint: what, where, when, how data is gathered Types of Design Exploratory Descriptive Causal/Experimental Good-Design Features Flexible, Appropriate, Efficient, Economical, Reliable, Valid
9. Data: Collection and Classification

Collection of Data — Primary and Secondary Data

Primary data is original data collected firsthand by the investigator for a specific research purpose — it does not exist before the study begins. It is more reliable and relevant to the exact problem, but collecting it is time-consuming and expensive.
Secondary data is data that was originally collected by someone else for a different purpose and is now reused for the current study — e.g. government reports, RBI bulletins, company annual reports, published research. It is quicker and cheaper to obtain, but must be checked for reliability, suitability and adequacy before use.
Sources of secondary data: Published sources — government publications, international bodies (RBI, CSO, World Bank), journals, newspapers; Unpublished sources — theses, records of research institutions, internal company records.

Methods of Collecting Primary Data

Direct personal interview: the investigator meets respondents face-to-face and collects information directly — high accuracy, but costly and time-consuming for a large area.
Indirect oral investigation: information is gathered from third parties (witnesses, experts) who are familiar with the matter, rather than from the respondent directly — used when direct contact is not possible.
Questionnaire method: a set of questions is mailed or sent electronically to respondents who fill and return it themselves — low cost, covers a wide area, but suffers from low response rates and depends on respondents’ literacy.
Schedule method: similar to a questionnaire, but it is filled in by trained enumerators who visit respondents in person — higher response rate than a questionnaire, suited to illiterate respondents, but more expensive.
Through local correspondents/agents: local agents stationed in different areas regularly send information to a central office — used by newspapers and government departments for continuous data such as market prices.
Worked Example

A government department wants nationwide literacy data covering both literate and illiterate respondents in rural areas, and is willing to spend more for an accurate, high response rate. Which method of collecting primary data fits best?

Answer: The schedule method — trained enumerators personally fill in the schedule, so illiterate respondents can still be covered and the response rate stays high, unlike a self-administered questionnaire.

Classification of Data — Meaning and Objectives

Classification is the process of arranging data into groups or classes based on similarities and common characteristics, so that items within a class are alike and items in different classes differ.
Objectives: to condense a large, unwieldy mass of data into a form that is simple to understand; to highlight similarities and differences between items; to make comparison easier; and to prepare the data for further statistical treatment such as tabulation and analysis.

Bases of Classification

Qualitative (attribute-based) classification: data is classified by a quality or attribute that cannot be measured numerically — e.g. gender, literacy, marital status. It can be simple (one attribute, two classes — literate/illiterate) or manifold (more than one attribute studied together — literacy crossed with gender).
Quantitative classification: data is classified by a characteristic that can be measured numerically, such as age, height, income or marks, typically grouped into class-intervals.
Geographical (spatial) classification: data is classified by location or region — e.g. state-wise or country-wise sales figures.
Chronological (temporal) classification: data is classified by the time period in which it occurs — e.g. year-wise population figures, forming a time series.

Scales of Measurement

Nominal scale: classifies data into named categories with no inherent order, distance or true zero — e.g. gender, religion, blood group. Only counting and mode are meaningful.
Ordinal scale: categories can be ranked in order, but the gaps between ranks are not necessarily equal — e.g. customer satisfaction (poor/average/good/excellent), class rank. Median and mode can be used, but not meaningful arithmetic averaging.
Interval scale: ordered, with equal and meaningful intervals between values, but no true (non-arbitrary) zero point — e.g. temperature in Celsius or Fahrenheit, calendar years. Differences are meaningful, but ratios are not (20°C is not “twice as hot” as 10°C).
Ratio scale: ordered, with equal intervals AND a true zero point representing total absence of the quantity — e.g. income, weight, height, age. All arithmetic operations including ratios are meaningful (₹20,000 income is genuinely twice ₹10,000).
Memory aid: each scale adds one property to the one before it — Nominal (names only) → Ordinal (+ order) → Interval (+ equal spacing) → Ratio (+ true zero).
Worked Example

A researcher records respondents’ body temperature in Celsius and separately records their annual income in rupees. Which scale of measurement does each variable belong to?

Answer: Temperature in Celsius is an interval scale — it has equal, meaningful intervals but 0°C does not mean “no temperature”. Annual income is a ratio scale — it has a true zero (₹0 means no income at all), so ratios like “twice the income” are meaningful.
Primary versus secondary data Two side-by-side panels comparing primary data, collected firsthand for the current study, with secondary data, originally collected for another purpose and reused. Primary Data What collected firsthand For this exact research Examples interview, survey, questionnaire Trade-off reliable but costly, slow Secondary Data What already collected by someone else For a different original purpose Examples govt. reports, RBI data, journals Trade-off cheap, fast, but check reliability
Five methods of collecting primary data Five colour-coded rows: direct personal interview, indirect oral investigation, questionnaire, schedule, and local correspondents, each with a one-line description. Methods of Collecting Primary Data Direct Personal Interview face-to-face; high accuracy but costly & slow Indirect Oral Investigation via third parties familiar with the matter Questionnaire mailed/online, self-filled; cheap but low response Schedule enumerator-filled in person; covers illiterate respondents Local Correspondents/Agents regular reports from field agents, e.g. market prices
Four bases of classification of data A tree showing Classification splitting into Qualitative, Quantitative, Geographical and Chronological bases, each with a short example. Classification of Data Qualitative by attribute, not numerical e.g. gender, literacy Quantitative by numerical characteristic e.g. age, income, marks Geographical by location or region e.g. state-wise sales Chronological by time period e.g. year-wise population Qualitative classification can be simple (1 attribute) or manifold (several) (e.g. literacy alone, vs literacy crossed with gender)
Four scales of measurement: Nominal, Ordinal, Interval, Ratio A four-step ladder from Nominal through Ordinal and Interval up to Ratio, each step adding one property: naming, order, equal intervals, and a true zero. Scales of Measurement 4. Ratio — order + equal intervals + TRUE ZERO e.g. income, weight, age — ratios ARE meaningful (Rs.20k is 2x Rs.10k) 3. Interval — order + equal intervals, no true zero e.g. temperature C, calendar year — differences meaningful, ratios not 2. Ordinal — ordered categories, gaps not equal e.g. satisfaction rating, class rank 1. Nominal — named categories only, no order e.g. gender, religion, blood group Each step keeps every earlier property and adds one more (Nominal → + order → + equal spacing → + true zero)
Mind map: data collection and classification Central hub Data with five colour-coded branches: primary versus secondary data, collection methods, classification meaning, bases of classification, and scales of measurement. Data: Collection & Types Primary vs Secondary firsthand vs reused data Collection Methods interview, questionnaire, schedule, agents Classification groups data by similarity Bases of Classification Qualitative, Quantitative, Geographical, Chronological Scales of Measurement Nominal, Ordinal, Interval, Ratio
10. Sampling and Estimation

Sampling — Basic Concepts

Population is the entire mass of observations or units about which a conclusion is to be drawn — e.g. all registered voters in a state. Sample is a smaller, representative subset of the population actually selected for study, and sampling is the process/technique used to select that subset.
Why sample instead of a complete census? Sampling is more economical (lower cost), more practical for a large or spread-out population, faster (saves time), and — when done with a sound method — can still give accurate, reliable results, sometimes even more accurate than a rushed full census because more care can be given to fewer units.
Sampling error is the difference between a sample statistic (e.g. sample mean) and the true population parameter, arising purely because a sample — not the whole population — was studied; it tends to fall as sample size increases.

Probability Methods of Sampling

Probability sampling: every unit of the population has a known, non-zero (often equal) chance of being selected. These methods allow valid statistical inference about the population.
1. Simple Random Sampling: every unit has an equal chance of selection, typically via random numbers or lottery — e.g. drawing 50 names from a hat of 500 employees.
2. Stratified Random Sampling: the population is first divided into homogeneous, non-overlapping strata (e.g. by department or income group), then units are randomly sampled from each stratum — ensures every subgroup is represented.
3. Cluster Sampling: the population is divided into naturally occurring clusters (e.g. city blocks, colleges); entire clusters are randomly selected and every unit within a chosen cluster is studied — cost-effective for large, geographically spread populations, though less precise than stratified sampling.
4. Systematic Sampling: units are selected at fixed, regular intervals from an ordered list after a random start — e.g. every 10th customer entering a store.
5. Multi-stage Sampling: sampling is carried out in several successive stages, each stage narrowing the population further (e.g. randomly select states → then districts within those states → then households within those districts) — only the final-stage sample is actually studied.

Non-Probability Methods of Sampling

Non-probability sampling: units are selected using non-random, judgment-based criteria, so not every unit has a known chance of selection. Easier and cheaper, but results cannot be safely generalised to the whole population.
Convenience Sampling: includes whichever individuals are easiest for the researcher to reach — e.g. surveying shoppers at a mall entrance. Quick and low-cost, but highly prone to bias.
Judgemental / Purposive Sampling: the researcher deliberately selects units believed to be most informative or representative, based on expert judgment — e.g. interviewing only senior doctors for a medical-policy study.
Quota Sampling: the population is divided into groups (as in stratified sampling), but within each group units are picked non-randomly until a fixed quota is filled — e.g. interviewing exactly 20 men and 20 women from a mall, chosen as they come.
Snowball Sampling: existing respondents refer or recruit further respondents from among their acquaintances — used for hard-to-reach or rare populations (e.g. a study of a specific rare illness).
Worked Example

A researcher wants to study customer satisfaction and decides to interview the first 100 people who agree to talk to them outside a railway station. Is this a probability or a non-probability method of sampling, and which specific method is it?

Answer: This is a non-probability method — specifically convenience sampling, since respondents are chosen simply because they happen to be accessible at that place and time, not because every member of the population had a known chance of being selected.

Sampling Distribution and the Central Limit Theorem

Sampling distribution is the probability distribution of a sample statistic (most commonly the sample mean, x̄) obtained by repeatedly drawing all possible samples of a fixed size n from a population and computing the statistic for each sample. Its own mean equals the population mean (μ), and its own standard deviation is called the standard error.
Central Limit Theorem (CLT): as the sample size n gets larger (conventionally n ≥ 30), the sampling distribution of the sample mean approaches a normal distribution, regardless of the shape of the original population’s distribution — even if the population itself is skewed or irregular. This is what allows normal-distribution methods (Z-scores, confidence intervals) to be used for sample means drawn from almost any population.
Why the CLT matters: it is the theoretical bridge that justifies using the normal distribution for hypothesis testing and estimation about population means, even when nothing is known about the population’s own shape, as long as the sample is reasonably large.

Standard Error and Statistical Estimation

Standard Error (SE) is the standard deviation of a sampling distribution — it measures how much a sample statistic (e.g. sample mean) is expected to vary from the true population parameter purely due to sampling. For the sample mean: SE = σ / √n (σ = population standard deviation, n = sample size). Unlike the standard deviation, which measures spread among individual data values, SE measures the reliability of the sample statistic itself — SE falls as n rises.
Statistical Estimation uses sample information to infer an unknown population parameter. Point estimation gives a single value (e.g. the sample mean x̄ as an estimate of μ). Interval estimation gives a range — a confidence interval — within which the true parameter is expected to lie with a stated probability (the confidence level, commonly 95% or 99%), computed as point estimate ± (critical value × standard error).
Good estimator properties (memory aid): Unbiasedness (its expected value equals the true parameter), Consistency (gets closer to the true value as sample size grows), Efficiency (has the smallest possible variance among unbiased estimators).
Worked Example

A population has standard deviation σ = 40. A random sample of n = 64 units is drawn. What is the standard error of the sample mean, and what does increasing the sample size to 400 do to it?

Answer: SE = σ/√n = 40/√64 = 40/8 = 5. If n increases to 400, SE = 40/√400 = 40/20 = 2 — SE fell to less than half even though n rose only about 6-fold, because SE shrinks with the square root of n. This is the general rule: larger samples give more precise (less variable) estimates of the population mean, but with diminishing returns — to halve SE again, n would need to be quadrupled once more.
Probability versus non-probability methods of sampling Two side-by-side panels: five probability sampling methods (simple random, stratified, cluster, systematic, multi-stage) and four non-probability methods (convenience, judgemental, quota, snowball), each with a one-line description. Probability Methods Non-Probability Methods Simple Random every unit has an equal chance Stratified divide into strata, sample each Cluster sample whole natural clusters Systematic every nth unit, random start Multi-stage successive narrowing stages Convenience whoever is easiest to reach Judgemental researcher picks by judgment Quota fixed quota per group, non-random Snowball respondents refer respondents Probability methods allow valid statistical inference; non-probability methods do not (but are faster and cheaper for exploratory work)
Population, sample and sampling error A large outer circle representing the population containing a smaller inner circle representing the sample drawn from it, with a note on why sampling is used instead of a full census and on sampling error. Population, Sample and Sampling Error Population (entire group studied) Sample (subset studied) Why sample, not census? Economical (lower cost) Practical for large/spread populations Faster (saves time) Can still be accurate & reliable if well designed Sampling Error difference between a sample statistic and the true population parameter — shrinks as n rises Sampling studies a part to draw conclusions about the whole, within a known margin of error
The Central Limit Theorem An irregular, skewed population distribution on the left feeds into increasingly bell-shaped sampling distributions of the mean as sample size n grows from small to large, illustrating that the sampling distribution approaches normal regardless of the population’s own shape. Central Limit Theorem Population irregular / skewed shape n = 5 samples roughly bell-shaped already n = 30+ samples closely normal, narrower spread → → Central Limit Theorem (CLT) As sample size n grows (conventionally n ≥ 30), the sampling distribution of the sample mean approaches a NORMAL distribution — whatever the shape of the original population. This lets normal-based methods (Z-scores, confidence intervals) be used for the sample mean even when the population itself is skewed.
Standard error and statistical estimation Standard error formula and its meaning, a comparison of standard deviation versus standard error, and the difference between point estimation and interval estimation with the confidence interval formula. Standard Error & Estimation Standard Error of the Mean: SE = σ / √n (σ = population standard deviation, n = sample size) — falls as n rises Standard Deviation (SD) spread of INDIVIDUAL data values around the mean of one data set Standard Error (SE) spread of a SAMPLE STATISTIC (e.g. sample mean) across many possible samples Point Estimation a single value estimate e.g. sample mean x̄ for μ Interval Estimation a range likely to contain the true parameter Confidence Interval = Point Estimate ± (Critical Value × Standard Error) e.g. a 95% confidence interval means: if we repeated the sampling many times, about 95% of such intervals would contain the true population parameter.
Mind map: sampling and estimation Central hub Sampling and Estimation with five colour-coded branches: concepts, probability and non-probability methods, sampling distribution and the central limit theorem, standard error, and statistical estimation. Sampling & Estimation Concepts population, sample, sampling error Sampling Methods probability (5) vs non-probability (4) Sampling Distribution + Central Limit Theorem (n≥30→normal) Standard Error SE = σ/√n, falls as n rises Estimation point vs interval, confidence level
11. Hypothesis Testing

Hypothesis Testing — Basic Concepts

A hypothesis is a predictive, testable statement about a population, usually about a parameter such as a mean or a proportion. Hypothesis testing is the statistical procedure used to decide, from sample evidence, whether to accept or reject a claim about that population.

Null Hypothesis (H₀) — the “no difference / no effect” statement, assumed true unless strong evidence says otherwise, e.g. H₀: μ = μ₀.
Alternative Hypothesis (H₁ or Ha) — the statement we accept if H₀ is rejected, e.g. H₁: μ ≠ μ₀ (two-sided) or μ > μ₀ (one-sided).
Level of Significance (α) — the risk of rejecting H₀ when it is actually true, usually 1%, 5% or 10%. (1−α) is the confidence level.
Critical Value & Critical Region — the critical value is the boundary between the acceptance and rejection regions; the critical (rejection) region is the set of test-statistic values that lead to rejecting H₀.
p-value — the probability of getting a result at least as extreme as the one observed, assuming H₀ is true. If p-value < α, reject H₀.
Type I Error (α) — rejecting H₀ when it is actually true (a false alarm). Type II Error (β) — failing to reject H₀ when it is actually false (a missed effect). The two cannot both be minimised at once for a fixed sample size.
One-tailed vs Two-tailed test — a one-tailed test places the entire rejection region in a single tail (used when the interest is only in “greater than” or only “less than”); a two-tailed test splits the rejection region between both tails (used when either direction of difference matters).
Worked Example A quality inspector wants to know only whether a new machine produces MORE defective items than the old one (not fewer). Answer: This calls for a one-tailed test, because the inspector’s interest is on one side only (an increase in defects) — the entire 5% rejection region is placed in the right tail of the sampling distribution.

Parametric Tests — Z-Test, T-Test and F-Test (ANOVA)

Parametric tests assume the population is normally distributed and typically require knowledge of (or an estimate of) the population variance. The choice between Z, T and F depends mainly on sample size and how many groups are being compared.

Z-test — used when the population is normal, the sample is large (n ≥ 30), and the population standard deviation (σ) is known. Tests the significance of the difference between a sample mean and a population mean, or between two large-sample means. Formula: Z = (X̄ − μ) / (σ/√n).
T-test — used when the sample is small (n < 30) and the population SD is unknown (only the sample SD, s, is available). Developed by Gosset (“Student”). Formula: t = (X̄ − μ) / (s/√n), compared against the t-distribution with (n−1) degrees of freedom. As n grows large, the t-distribution converges to the normal (Z) distribution.
F-test / ANOVA — Analysis of Variance, developed by R.A. Fisher, compares the means of THREE OR MORE groups at once using the F-statistic (ratio of between-group variance to within-group variance). One-way ANOVA is effectively an extension of the t-test to more than two groups — using several separate t-tests instead would inflate the overall error rate.
Worked Example A researcher compares the average marks of a sample of 25 students against a claimed population mean, with the population SD unknown. Which test applies? Answer: T-test — the sample is small (n=25 < 30) and the population SD is unknown, so the sample SD must be used with the t-distribution (24 degrees of freedom).

Chi-Square Test

The Chi-square (χ²) test, developed by Karl Pearson, is a non-parametric test that measures how far observed (categorical) data departs from what would be expected under H₀. It requires no assumption of a normally distributed population.

Formula: χ² = Σ [(O − E)² / E], where O = observed frequency and E = expected frequency under H₀, summed over all categories/cells.
Key properties — χ² is always non-negative (zero or positively skewed); it is never symmetrical; degrees of freedom depend on the number of categories (goodness-of-fit) or (rows−1)×(columns−1) for a contingency table (test of independence).
Common uses — testing goodness of fit (does the observed distribution match an expected/theoretical one?) and testing independence between two categorical attributes (e.g. gender and product preference).
Worked Example A die is rolled 60 times; a fair die should give 10 of each face. The observed frequencies are 8,12,9,11,7,13. Is the die fair at α=5%? Answer: χ² = (8−10)²/10 + (12−10)²/10 + (9−10)²/10 + (11−10)²/10 + (7−10)²/10 + (13−10)²/10 = 0.4+0.4+0.1+0.1+0.9+0.9 = 2.8. With df=5, the 5% critical value is about 11.07; since 2.8 < 11.07, we fail to reject H₀ — the die appears fair.

Non-Parametric Tests — Mann-Whitney U and Kruskal-Wallis H

Non-parametric (“distribution-free”) tests are used when the population is not normally distributed, the sample is small, or the data is only ordinal (ranked) rather than truly numerical — they work on RANKS instead of raw values.

Mann-Whitney U test — the true non-parametric counterpart of the (independent-samples) t-test. Compares two INDEPENDENT groups when the dependent variable is ordinal or continuous but not normally distributed. Gives the most accurate estimate of significance when sample sizes are small. Test statistic is denoted U.
Kruskal-Wallis H test — the non-parametric counterpart of one-way ANOVA, sometimes called “one-way ANOVA on ranks”. Determines whether there are statistically significant differences between TWO OR MORE independent groups on a continuous or ordinal dependent variable, using ranks rather than raw data.
Rule of thumb — Mann-Whitney U replaces the t-test when comparing 2 groups; Kruskal-Wallis H replaces one-way ANOVA when comparing 3+ groups; both are used when normality cannot be assumed.
Worked Example A researcher wants to compare customer-satisfaction RANKINGS (not scores) between two competing stores, with a small sample and no assurance of normal distribution. Answer: Mann-Whitney U test — two independent groups, ordinal (ranked) data, small sample, non-normal population: exactly the Mann-Whitney scenario rather than a t-test.

Rank Correlation as a Hypothesis Test, and Choosing the Right Test

Spearman’s rank correlation coefficient (R, covered under Correlation in this unit) can itself be put through a significance test — testing H₀: there is no real association between the two rankings in the population, against Ha: there is. It is the non-parametric counterpart of testing Pearson’s correlation coefficient (r), used whenever data is ranked rather than measured on an interval/ratio scale.

Parametric test ↔ non-parametric counterpart — t-test (independent samples) ↔ Mann-Whitney U; t-test (paired samples) ↔ Wilcoxon signed-rank test; one-way ANOVA ↔ Kruskal-Wallis; Pearson’s r ↔ Spearman’s rank correlation (or Kendall’s tau); Z-test for proportions ↔ Chi-square test for independence.
Decision rule for exam questions — first ask: is the data numerical/continuous and roughly normal? If yes, go parametric (Z if n≥30 and σ known; t if n<30; F/ANOVA if 3+ groups). If the data is ranked, categorical, or clearly non-normal, go non-parametric (Chi-square for categorical association; Mann-Whitney/Kruskal-Wallis/rank correlation for ranked data).
Hypothesis testing: the six-step process A six-step flowchart from stating the null and alternative hypotheses through choosing significance level, selecting and computing the test statistic, comparing to the critical value, to the final decision to reject or not reject the null hypothesis. 1. State H₀ and Ha the null and alternative hypotheses 2. Choose level of significance (α) usually 1%, 5% or 10% 3. Select the test statistic Z, t, F or Chi-square — based on sample size, σ known?, data type 4. Compute the test statistic from the sample data 5. Compare with critical value / p-value against the chosen α 6. Decision: Reject or fail to reject H₀ state the conclusion in context
Type I and Type II error A two-by-two matrix of decision (reject or fail to reject the null hypothesis) against reality (null hypothesis true or false), labelling Type I error, Type II error and the two correct-decision cells, with a closing note on the alpha-beta trade-off. Type I vs Type II Error Decision: Reject H₀ Decision: Fail to Reject H₀ Reality: H₀ is TRUE Type I Error (α) False alarm — rejecting a true null hypothesis Probability = α Reality: H₀ is TRUE Correct decision Correctly not rejecting a true null hypothesis Probability = 1−α Reality: H₀ is FALSE Correct decision Correctly rejecting a false null hypothesis Probability = Power = 1−β Reality: H₀ is FALSE Type II Error (β) Missed effect — failing to reject a false null hypothesis Probability = β For a fixed sample size, lowering α (fewer Type I errors) raises β (more Type II errors). The only way to reduce BOTH errors together is to increase the sample size.
One-tailed versus two-tailed hypothesis tests Two bell-curve panels comparing a one-tailed test, with its entire rejection region in a single tail, against a two-tailed test, which splits the rejection region between both tails, with a closing note on critical-value placement. One-Tailed Test critical value Rejection region (α) Two-Tailed Test α/2 α/2 Rejection region in ONE tail only Used when H1 is directional (“ > ” or “ < ”) e.g. does a new method INCREASE scores? All of α placed on one side Rejection region SPLIT between tails Used when H1 is non-directional (“ ≠ ”) e.g. does a drug CHANGE blood pressure? α/2 placed on each side Same α overall, but a one-tailed test has a less extreme critical value in its single tail — making it easier to reject H₀ in that one direction.
Comparing the Z-test, T-test and F-test A three-column comparison panel showing when to use the Z-test, T-test and F-test (ANOVA): sample size, whether the population standard deviation is known, what each test compares, and which distribution it uses, with a closing note on the t-distribution converging to the normal distribution as sample size grows. Z-Test vs T-Test vs F-Test Z-Test Sample size n ≥ 30 Population σ KNOWN Compares a sample mean to a population mean, or two large samples Distribution: Normal Given by: (classical) T-Test Sample size n < 30 Population σ UNKNOWN (sample SD, s, used) Compares a sample mean to a population mean, or two small samples Given by: Gosset F-Test / ANOVA 3 or more groups Compares several group MEANS at once using group variances (ratio of variances) Distribution: F Given by: R.A. Fisher As n → ∞, the t-distribution converges to the normal (Z) distribution — small samples need the WIDER t-distribution because they carry more uncertainty.
Non-parametric tests and their parametric counterparts A three-panel comparison of the Chi-square test, Mann-Whitney U test and Kruskal-Wallis H test, with a closing table mapping each common parametric test to its non-parametric counterpart. Non-Parametric Tests Chi-Square Test Categorical (nominal) data χ² = Σ (O−E)²/E Tests goodness-of-fit OR independence of 2 attributes Always non-negative, skewed right By Karl Pearson Mann-Whitney U 2 INDEPENDENT groups Ordinal/continuous data, NOT normally distributed True non-parametric counterpart of the t-test Uses ranks, not raw data Test statistic: U Kruskal-Wallis H 3+ INDEPENDENT groups “One-way ANOVA on ranks” Non-parametric counterpart of one-way ANOVA Uses ranks, not raw data Test statistic: H Parametric Test ↔ Non-Parametric Counterpart t-test (independent samples) ↔ Mann-Whitney U test t-test (paired samples) ↔ Wilcoxon signed-rank test One-way ANOVA ↔ Kruskal-Wallis H test Pearson’s correlation (r) ↔ Spearman’s rank correlation (R) Z-test for proportions ↔ Chi-square test for independence
Mind map: hypothesis testing Central hub Hypothesis Testing with six colour-coded branches: Basic Concepts, Z/T/F Tests, Chi-Square, Mann-Whitney U, Kruskal-Wallis H, and Rank Correlation. Hypothesis Testing Basic Concepts H₀/Ha, α, p-value, Type I & II errors Z / T / F Tests parametric, by sample size & σ Chi-Square χ²=Σ(O−E)²/E, categorical data Mann-Whitney U non-parametric, 2 groups, ranks Kruskal-Wallis H non-parametric, 3+ groups, ranks Rank Correlation Spearman’s R, tests significance
12. Report Writing

Meaning and Significance of a Research Report

Report writing is the final stage of the research process: converting the findings of a study into a clear, organised written document that communicates what was investigated, how it was investigated, and what was found. A good report is what makes research usable — the most carefully designed study (see Sections 8–11) is of little value to anyone if its results cannot be understood and acted upon.
Significance: the report is the only lasting record of the research process (methodology, data, and conclusions survive beyond the researcher’s own memory); it is the basis on which readers judge the study’s quality and credibility; and it is what ultimately persuades policymakers, managers, or other researchers to act on or build upon the findings.

Types of Research Reports

By audience — Technical Report: written for the scientific/academic community (other researchers, examiners, specialists). Uses formal and technical language, and includes full methodological detail — research design, sampling procedure, statistical tests and formulae used, limitations of the study. Example: a PhD thesis or a peer-reviewed journal article.
By audience — Popular Report: written for the general public, policymakers, or non-specialist readers. Uses simple, non-technical language, avoids statistical jargon, and emphasises practical implications and recommendations over methodological detail. Example: a government white paper summary or a newspaper report on a survey’s findings.
By form — Oral Report: presented verbally (e.g. at a meeting or conference). Quick to deliver and allows immediate clarification through questions, but leaves no permanent, citable record and relies on the audience’s memory/notes.
By form — Written Report: presented as a physical or digital document. Provides a permanent, detailed, and precisely citable record that can be referred back to at any time, though it takes longer to prepare and does not allow for immediate interactive clarification.
Worked Example

A university doctoral committee requires a full account of the sampling procedure, the hypothesis tests used, and their statistical formulae, written for an audience of subject specialists. Which type of report is this?

Answer: A Technical Report — the audience is specialists/examiners, and the document must carry full methodological and statistical detail, not a simplified summary.

Steps in Writing a Research Report

1. Logical analysis of the subject matter: the researcher first divides the material into sections and sub-sections following a logical sequence, so the report builds up its argument step by step rather than presenting findings in a random order.
2. Preparation of the final outline: a tentative skeleton/outline of chapters and headings is drawn up before writing begins, keeping the eventual report internally consistent and complete.
3. Preparation of the rough draft: the researcher writes out the findings in the planned sequence, with supporting tables, charts and illustrations placed where they are first discussed.
4. Rewriting and polishing: the most time-consuming step — the draft is critically re-examined for weaknesses in logical development and presentation style, and errors, repetition, and verbosity are removed.
5. Preparing the final bibliography: a complete list of every book, journal, and source consulted during the study, usually arranged alphabetically by author’s surname, in a consistent citation format.
6. Writing the final draft: the polished version is written out in a simple, clear style, avoiding vague or ambiguous expressions, ready for submission or publication.

Layout (Structure) of a Research Report

Preliminary pages: title page, acknowledgement/preface, table of contents, and a list of tables/graphs used in the report — these orient the reader before the substantive content begins.
Main text — Introduction: statement of the research problem, objectives and hypotheses, the methodology adopted, and the scope and limitations of the study.
Main text — Statement of findings and results: a detailed, usually chapter-wise, presentation of the results using text, tables, and charts, followed by a discussion of what the findings mean.
Main text — Conclusion and recommendations: a concise restatement of the main findings and the practical recommendations or suggestions for further research that follow from them.
End matter: appendices (supporting technical material such as the questionnaire used, detailed statistical workings, or raw data tables), the bibliography, and an index if the report is long enough to need one.
Worked Example

In which part of a research report would a reader expect to find the questionnaire that was used to collect the primary data?

Answer: The end matter, as an appendix — appendices hold supporting technical material (like the instrument used for data collection) that would clutter the main text if placed there directly.

Characteristics of a Good Report

Clarity and simplicity: written in simple, direct language, avoiding vague or ambiguous expressions that could be read more than one way.
Logical sequence: ideas and sections follow one another in a planned, easy-to-follow order rather than jumping between topics.
Accuracy: all facts, figures, and statistical results presented must be checked and correctly reported — a report is only as trustworthy as the numbers inside it.
Effective use of visual aids: tables, charts, and diagrams are used wherever they communicate a pattern or comparison more clearly than plain text would.
Conciseness: the report says what needs to be said as briefly as possible without sacrificing completeness — padding and repetition weaken a report rather than strengthening it.
Proper bibliography/referencing: every source used is acknowledged in a consistent citation format, both for academic honesty and so a reader can trace a claim back to its source.
Draft and review: a good report is never written in a single pass — a rough draft is prepared, then critically reviewed and rewritten before the report is finalised.
Uniform style: headings, numbering, citation style, and terminology stay consistent throughout the whole document, rather than changing partway through.
Worked Example

A government department needs a quick finding shared at a review meeting, with no permanent document required, and expects a chance for officials to ask immediate questions. What form of report fits best?

Answer: An Oral Report — it suits a quick, interactive setting where immediate clarification matters more than producing a permanent citable record.

Report Writing in the Research Cycle

Report writing is the culmination, not a separate add-on, of the research process: the research problem and design (Section 8), the data collected and classified (Section 9), the sample drawn and estimates made (Section 10), and the hypotheses tested (Section 11) all ultimately feed into the report — it is the single document in which every earlier stage of the study is finally communicated to a reader.
Choosing the report type correctly matters for exam questions: look first at WHO the report is for (specialists → Technical; general public/policymakers → Popular) and HOW it is delivered (a permanent document → Written; a live, interactive setting → Oral) — these two questions settle almost every classification question on this topic.
Types of research report: Technical vs Popular, Oral vs Written A four-box panel comparing Technical and Popular reports by audience and detail level, and Oral and Written reports by permanence and delivery form. Types of Research Report Technical Report Audience: specialists / researchers Language: formal, technical Detail: full methodology, formulae, tests, limitations Example: a PhD thesis or peer-reviewed journal article “For the expert reader” keeps every statistical and design detail intact Popular Report Audience: general public, policymakers Language: simple, non-technical Detail: practical implications and recommendations emphasised Example: a government white paper summary “For the general reader” avoids jargon, keeps it short Classified by AUDIENCE — who the report is written for Oral Report vs Written Report Oral Report Delivered verbally; quick, allows immediate questions but NO permanent record relies on memory / notes Written Report A permanent document; detailed, precisely citable record takes longer to prepare no live interactive clarification Classified by FORM — how the report is delivered
Steps in writing a research report A six-step vertical flowchart from logical analysis of the subject matter through to writing the final draft. 1. Logical analysis of the subject matter Divide material into sections, in a logical sequence 2. Preparation of the final outline A tentative skeleton of chapters & headings, before writing 3. Preparation of the rough draft Write findings in sequence, with tables/charts placed in 4. Rewriting and polishing Most time-consuming step — remove errors, repetition 5. Preparing the final bibliography Alphabetical list of every source consulted 6. Writing the final draft Clear, simple language — ready for submission Decision: Submit the finished, polished report
Layout (structure) of a research report A three-column panel showing the Preliminary Pages, Main Text, and End Matter of a research report, with their typical contents. Layout of a Research Report Preliminary Pages Title page Acknowledgement / Preface Table of contents List of tables & graphs → Main Text Introduction (problem, objectives, methodology, scope & limitations) Findings & Results (text, tables, charts) Conclusion & recommendations → End Matter Appendices (questionnaire, raw data) Bibliography Index (if needed) Reading order: Preliminary Pages -> Main Text -> End Matter the reader is oriented first, then walked through the study, then given the supporting detail
Characteristics of a good research report An eight-badge grid listing the characteristics of a good research report: clarity, logical sequence, accuracy, visual aids, conciseness, proper referencing, drafting and review, and uniform style. A Good Report Is… Clear & Simple avoids vague, ambiguous language Logically Sequenced ideas follow a planned order Accurate facts & figures are checked Visual where useful tables/charts aid clarity Concise as brief as possible, complete Properly Referenced consistent bibliography format Drafted & Reviewed never written in one pass Uniform in Style consistent headings & terms Memory aid: Clear, Logical, Accurate, Visual, Concise, Referenced, Reviewed, Uniform
Mind map: report writing Central hub Report Writing with five colour-coded branches: Types of Report, Steps in Writing, Layout, Good-Report Traits, and Research Cycle. Report Writing Types of Report Technical/Popular, Oral/Written Steps in Writing Outline -> draft -> polish -> final Layout Preliminary, Main Text, End Matter Good-Report Traits Clear, logical, accurate, concise Research Cycle Final stage after design & analysis

Discover more from WebComm.in

Subscribe to get the latest posts sent to your email.

Yashraj Patidar, author of WebComm.in
MEET THE AUTHOR
Mr. Yashraj

Founder of WebComm.in · PhD scholar, Dept. of Commerce & Management.

B.ComM.ComNET qualified (4 times)JRF qualifiedMP SETGSETMH SETPhD (pursuing)
View All Posts →
Leave a Comment

Comments

No comments yet. Why don’t you start the discussion?

    Leave a Reply