FROM VARIATION

TO VERDICT
Total SS · Regression SS · Residual SS · R² · F-Statistic · p-value
A Complete, Crystal-Clear Guide for Students
SkillArbitrage | Financial & Data Analytics

YOUR COMPLETE LEARNING ROADMAP

PART 1 — The Variation Story: What is variation? Why does it need to be explained?
PART 2 — The Three SS: Total SS, Regression SS, Residual SS — with the OLS cross-product insight.
PART 3 — R² Derived: Why R² = 1 − (Residual SS / Total SS). Built from logic, not memorisation.
PART 4 — The F-Statistic: What it is, how it is calculated, and what it tells you.
PART 5 — The p-value: Where it comes from, how to read it, and how to arrive at conclusions.
PART 6 — Three Fully Worked Examples from scratch — hand-calculated so nothing is a black box.
PART 7 — The Complete Decision Framework: How to look at an output and know exactly what to
do.
 PART 1 — WHAT IS VARIATION AND WHY DOES IT MATTER?

1. Variation — The Root of Everything

1.1 The Simplest Definition
Variation is how much the values of Y differ from each other. If everyone scored exactly 70 on a test, there is zero
variation. If scores ranged from 20 to 95, there is a lot of variation.

🎯 The Fundamental Question Regression Answers

Y varies. Some days sales are high. Some days they are low.
WHY does Y vary? What factor (X) is responsible for causing Y to move?
HOW MUCH of Y's variation can X explain?

Answering 'how much' precisely requires us to MEASURE and SPLIT variation into parts.
That is exactly what the Sum of Squares framework does.

1.2 Measuring Variation — The Mean as Baseline

Before we have a model, our best single guess for any Y value is the mean (Ȳ). The variation around this mean is
our starting point — the total 'mystery' to be solved.

Example: 5 days of coffee shop sales (₹ thousand):

Day Sales Y Mean Ȳ Deviation (Y−Ȳ) Squared Dev (Y−Ȳ)²

Mon 20 30 −10 100

Tue 25 30 −5 25
Wed 30 30 0 0
Thu 38 30 +8 64
Fri 37 30 +7 49
Σ 150 — 0 238

Total SS = Σ(Y − Ȳ)² = 238

This 238 represents ALL the variation in sales. Our regression model's job is to explain as much of this 238 as
possible using an X variable.
 PART 2 — THE THREE SUMS OF SQUARES

2. Splitting Variation into Three Pieces

2.1 The Key Identity for Any Single Data Point
For any one observation, the deviation from the mean can be perfectly split into two non-overlapping parts:

(Y − Ȳ) = (Ŷ − Ȳ) + (Y − Ŷ)

Component What It Represents

(Y − Ȳ) Total deviation The full gap between the actual point and the mean baseline. The total
'distance' that needs explaining.
(Ŷ − Ȳ) Explained The gap between the regression line's prediction and the mean. The distance
deviation the model 'moved' the prediction away from the baseline. This is the
EXPLAINED part.
(Y − Ŷ) Residual / The gap between the actual point and the regression line's prediction. This is
Unexplained what the model GOT WRONG. The leftover error.

2.2 From One Point to All Points — Squaring and Summing

When we square the identity and sum across ALL n observations, something special happens under Ordinary Least
Squares (OLS):

⚡ THE OLS CROSS-PRODUCT MIRACLE

Mathematically, squaring (Y−Ȳ) = (Ŷ−Ȳ) + (Y−Ŷ) produces three terms:

Σ(Y−Ȳ)² = Σ(Ŷ−Ȳ)² + Σ(Y−Ŷ)² + 2·Σ(Ŷ−Ȳ)(Y−Ŷ) ← cross-product

↑
In OLS regression, this cross-product term is ALWAYS exactly zero.
Why? Because OLS minimises errors such that: Σ(Y−Ŷ) = 0 and errors are uncorrelated with Ŷ.

So the cross-product vanishes and we get the clean, exact identity:

 SST = SSR + SSE (always true for OLS regression)

NOTE: This only works for OLS. If you fit a non-OLS line, SSR + SSE ≠ SST (as shown in the PDF you
uploaded).

2.3 The Three SS — Formally Defined

Total SS (SST) = Σ(Y − Ȳ)²

The TOTAL variation in Y around its mean. This is the total 'size of the mystery'. Calculated from raw data alone —
no model needed.

Regression SS (SSR) = Σ(Ŷ − Ȳ)²

The variation EXPLAINED by X. How much the regression line's predictions differ from the flat mean baseline. The
'work the model did'. Larger SSR = better model.

Residual SS (SSE) = Σ(Y − Ŷ)²

The variation NOT explained by X. The sum of squared prediction errors. The 'work still undone'. Smaller SSE =
better model.

SST = SSR + SSE (The Golden Identity)

2.4 A Visual Metaphor That Never Fails

🍕 THE PIZZA SLICE ANALOGY

Imagine SST is an entire pizza — 8 slices representing total variation in Y.

SSR = the slices your regression model ATE. The variation it explained and consumed.
SSE = the slices LEFT ON THE PLATE. The variation the model couldn't touch.
R² = SSR/SST = slices eaten / total slices = proportion the model consumed.

A great model eats most of the pizza (SSR ≈ SST, SSE ≈ 0, R² ≈ 1).
A poor model barely touches it (SSR ≈ 0, SSE ≈ SST, R² ≈ 0).

And no matter what: slices eaten + slices left = total slices. (SSR + SSE = SST always)
 PART 3 — R² DERIVED FROM FIRST PRINCIPLES

3. Why R² = 1 − (SSE / SST) — The Logical Derivation

3.1 What R² Should Mean — Starting Point
R² answers: 'What proportion of Y's total variation did our model explain?' It must be a number between 0 and 1.

Step 1 — Write what we want:

R² = Explained Variation / Total Variation = SSR / SST

Step 2 — Substitute using the golden identity (SSR = SST − SSE):

R² = (SST − SSE) / SST

Step 3 — Split the fraction:

R² = SST/SST − SSE/SST = 1 − SSE/SST

✅ RESULT — Never memorise, always derive

R² = 1 − ( SSE / SST )

This is not magic. It is plain algebra from one identity: SST = SSR + SSE.
SSE/SST = the fraction of variation LEFT UNEXPLAINED.
1 − SSE/SST = the fraction of variation EXPLAINED. That is R².

When SSE = 0 (perfect model): R² = 1 − 0 = 1.0 (explains 100%)

When SSE = SST (useless model): R² = 1 − 1 = 0.0 (explains 0%)
3.2 R² Interpretation Guide
R² Range Interpretation
R² = 0.95 to 1.00 Excellent. Model explains 95-100% of variation. Highly reliable for
prediction.
R² = 0.80 to 0.94 Good. Strong explanatory power. Suitable for most business decisions.
R² = 0.60 to 0.79 Moderate. Useful but predictions carry noticeable uncertainty.
R² = 0.40 to 0.59 Weak. Large portion unexplained. Use with significant caution.
R² < 0.40 Poor. Model misses most of the variation. Other predictors likely needed.

Important: R² tells you how well the model FITS the data. It does NOT tell you if the relationship is statistically real
(that requires the p-value, covered in Part 5).
 PART 4 — THE F-STATISTIC: THE BRIDGE BETWEEN SS AND p-value

4. The F-Statistic — Making SS Comparable

4.1 The Problem: SS Values Are Not Comparable As-Is
Here is something students miss: you cannot directly compare SSR and SSE because they are based on different
numbers of 'free calculations' (degrees of freedom). A large SSR does not automatically mean a good model — it
depends on how many parameters the model used.

The solution: convert each SS into a mean square (average SS per degree of freedom). Then compare the two
means.

4.2 Degrees of Freedom — A Simple Explanation

🎓 What Are Degrees of Freedom?

Think of degrees of freedom as 'how many independent pieces of information went into this
calculation?'

For Regression SS (SSR):

In simple linear regression (one X), we estimated ONE slope. So df_Regression = 1.
(More X variables = more df. For k predictors, df_Regression = k.)

For Residual SS (SSE):

We have n observations but used 2 parameters (slope and intercept).
So df_Residual = n − 2.
(For k predictors: df_Residual = n − k − 1.)

For Total SS (SST):

df_Total = n − 1. Note: df_Total = df_Regression + df_Residual (always).
4.3 Mean Squares — Dividing SS by df

Mean Square Regression (MSR) = SSR / df_Regression = SSR / 1

Mean Square Error / Residual (MSE) = SSE / df_Residual = SSE /

(n−2)

4.4 The F-Statistic — Signal to Noise Ratio

The F-statistic compares the variation EXPLAINED by the model (signal) to the variation NOT explained (noise):

F = MSR / MSE = ( SSR / 1 ) ÷ ( SSE / (n−2) )

💡 Intuition for F
Think of F as a Signal-to-Noise ratio:

MSR = the 'signal' — how much variation per degree of freedom did X explain?
MSE = the 'noise' — how much variation per degree of freedom is still random error?

Large F → signal is much stronger than noise → X is genuinely explaining Y.

Small F (close to 1) → signal ≈ noise → X is not doing better than random chance.

RULE OF THUMB:
F > 4 typically starts to be meaningful (for small samples)
The larger F is, the more confident we are the model is real.
But we need the exact p-value to be precise — which comes from the F-distribution.

4.5 The ANOVA Table — How Excel Presents All This

The Data Analysis ToolPak presents everything in an ANOVA (Analysis of Variance) table. Here is how to read every
single cell:
 Source df SS MS F p-value (Significance F)

Regression 1 (= k) SSR = Σ(Ŷ−Ȳ)² MSR = SSR/1 F= ← This is the p-value you

MSR/MSE use
Residual n−2 SSE = Σ(Y−Ŷ)² MSE = SSE/(n−2) — —
Total n−1 SST = SSR+SSE — — —

The 'Significance F' column is the p-value for the overall regression. It answers: 'What is the probability of
observing this F-statistic by random chance if X actually has no relationship with Y?'
 PART 5 — THE p-VALUE: FROM F-STATISTIC TO CONCLUSION

5. The p-value — Making the Final Call

5.1 What the p-value Actually Is
This is the most misunderstood concept in statistics. Let us be precise and clear:

📖 THE EXACT DEFINITION

The p-value is the probability of observing an F-statistic as large as (or larger than)
the one you calculated, IF the null hypothesis were true — i.e., if X had NO real
relationship with Y whatsoever.

In plain language:
p-value = 'How likely is it that I got these results purely by luck?'

A SMALL p-value means: it would be very unlikely to see this result by chance.
→ We conclude: the relationship is real. Reject H₀.

A LARGE p-value means: these results could easily appear by random chance.
→ We cannot conclude the relationship is real. Fail to reject H₀.

5.2 The Hypothesis Test Framework

Element Explanation
H₀ (Null Hypothesis) X has NO effect on Y. The slope b = 0. The model is useless. Any pattern
you see is random noise.
H₁ (Alternative Hypothesis) X DOES have an effect on Y. The slope b ≠ 0. The pattern is real.
Test Statistic F = MSR / MSE. The larger this is, the more evidence against H₀.
Decision Rule If p-value < α (significance level, usually 0.05): Reject H₀. Conclude X
matters. If p-value ≥ α: Fail to reject H₀. Cannot conclude X matters.
What 0.05 means You accept a 5% risk of wrongly concluding the relationship is real when
it isn't. This is the standard threshold used in business and academia.
5.3 How the p-value Is Calculated From F
You do not calculate the p-value manually — you look it up from the F-distribution. But here is the logic:

📐 The F-Distribution Logic

Under H₀ (no real relationship), F-statistics follow a known statistical distribution
called the F-distribution, which depends on two numbers:
df1 = df_Regression = 1 (for simple linear regression)
df2 = df_Residual = n − 2

The F-distribution is a bell-shaped-ish curve (skewed right). Most of its area is near
small F values (when the model is no better than chance).

The p-value = the AREA IN THE RIGHT TAIL beyond your calculated F.

Large F → very little tail area → very small p-value → strong evidence against H₀.
Small F → lots of tail area → large p-value → weak evidence against H₀.

In Excel: p-value = [Link](F_statistic, df1, df2)

e.g. =[Link](85.3, 1, 8) gives you the p-value directly.

5.4 p-value Decision Table

p-value Range Decision What It Means Confidence Level

< 0.001 Reject H₀ (strongly) Less than 0.1% chance of random luck. Very 99.9%+
strong evidence X matters.
0.001 – 0.01 Reject H₀ Less than 1% chance. Strong evidence X 99%+
matters.
0.01 – 0.05 Reject H₀ Less than 5% chance. Standard threshold — 95%+
X likely matters.
0.05 – 0.10 Grey zone 5–10% chance. Marginal evidence — 90–95%
context-dependent decision.
> 0.10 Fail to Reject H₀ More than 10% chance this is random. < 90%
Cannot conclude X matters.
5.5 The Court Room Analogy — Never Forget It

⚖️ THE COURT ROOM

In a court case, the null hypothesis is: 'The defendant is INNOCENT.'
You need evidence to convict (reject H₀).

p-value = 'What is the probability that an innocent person would produce this evidence by
coincidence?'

p = 0.001 → Only 0.1% chance an innocent person generates this evidence → CONVICT (reject H₀).
p = 0.40 → 40% chance an innocent person generates this evidence → ACQUIT (fail to reject H₀).

Similarly in regression:
'Innocent' (H₀) = X has no real effect on Y
'Evidence' = Your F-statistic
p-value = chance of getting this F if X truly has no effect
Verdict = reject or fail to reject H₀ at your chosen significance level (usually 5%)

CRITICAL REMINDER: Failing to reject H₀ does NOT mean X has no effect.

It means you don't have enough evidence to conclude it does. Small sample? Noisy data?
These can produce a high p-value even when a real effect exists.
 PART 6 — THREE FULLY WORKED EXAMPLES

6. Examples: From Raw Data to Final Conclusion

Example 1 — Strong Model: Study Hours → Exam Marks

📖 The Setup
5 students. X = Hours studied per day, Y = Exam score (out of 100).
We want to know: Does study time significantly explain exam scores?

Step 1: Raw Data and Summary Calculations

Student X (hrs) Y (score) X² XY

Arun 2 52 4 104
Bina 4 63 16 252
Chirag 6 74 36 444
Deepa 8 85 64 680
Eshan 10 96 100 960
Σ 30 370 220 2440

n=5, X̄ =6, Ȳ=74

Step 2: Regression Equation

Slope and Intercept
b = (n·ΣXY − ΣX·ΣY) ÷ (n·ΣX² − (ΣX)²)
b = (5×2440 − 30×370) ÷ (5×220 − 900)
b = (12200 − 11100) ÷ (1100 − 900) = 1100/200 = 5.5

a = Ȳ − b·X̄ = 74 − 5.5×6 = 74 − 33 = 41

Equation: Score = 41 + 5.5 × Hours

Step 3: Predicted Values and All Three SS
Student Y Ŷ=41+5.5X (Y−Ȳ)² Total (Y−Ŷ)² Residual (Ŷ−Ȳ)² Regression

Arun 52 52 484 0 484

Bina 63 63 121 0 121
Chirag 74 74 0 0 0
Deepa 85 85 121 0 121
Eshan 96 96 484 0 484
TOTAL (Σ) — — 1210 0 1210

SS Summary — Example 1
SST (Total SS) = 1210
SSE (Residual SS) = 0 ← The line passes THROUGH every point perfectly
SSR (Regression SS) = 1210 ← All variation explained

Verification: SSR + SSE = 1210 + 0 = 1210 = SST ✓

Step 4: F-Statistic and p-value

ANOVA Table — Example 1
df_Regression = 1, df_Residual = n−2 = 5−2 = 3

MSR = SSR / df_Reg = 1210 / 1 = 1210

MSE = SSE / df_Res = 0 / 3 = 0

F = MSR / MSE = 1210 / 0 = ∞ (undefined/infinity)

Interpretation: When SSE = 0, the model is perfect and F → infinity.

The p-value in this case → 0.000.

R² = 1 − (SSE/SST) = 1 − (0/1210) = 1.00 (100% of variation explained)

Step 5: Conclusion for Example 1

✅ CONCLUSION — Example 1
R² = 1.00: Study hours explain 100% of the variation in scores. Perfect linear fit.
F → ∞, p-value → 0: The probability of this occurring by chance is essentially zero.

DECISION: Reject H₀ strongly. Study hours are a highly significant predictor of exam scores.

BUSINESS/ACADEMIC STATEMENT:
'The regression analysis reveals a perfect linear relationship between daily study hours
and exam performance (R²=1.00, p<0.001). For every additional hour of daily study,
a student gains 5.5 marks. The model is statistically significant beyond any reasonable
doubt and can be used with full confidence for predictions within this dataset.'

NOTE: A perfect R²=1.0 is rare in real data. It often indicates the X and Y are
directly mathematically linked (e.g., cost = fixed rate × units) or the dataset is tiny.
Example 2 — Moderate Model: Rainfall → Crop Yield

📖 The Setup
An agricultural analyst studies whether monthly rainfall (mm) predicts wheat yield (tonnes per
hectare).
n = 7 months of data. The relationship is positive but not perfect due to soil, fertiliser, etc.

Step 1: Data
Month Rainfall X (mm) Yield Y (t/ha) X² XY

Jan 40 2.5 1600 100

Feb 55 3.1 3025 170.5
Mar 70 3.8 4900 266
Apr 85 4.0 7225 340
May 100 4.9 10000 490
Jun 115 5.2 13225 598
Jul 130 5.5 16900 715
Σ 595 29.0 56875 2679.5

n=7, X̄ =85, Ȳ=29/7=4.143

Step 2: Regression Equation

Slope and Intercept
b = (7×2679.5 − 595×29) ÷ (7×56875 − 595²)
b = (18756.5 − 17255) ÷ (398125 − 354025)
b = 1501.5 ÷ 44100 = 0.03404

a = 4.143 − 0.03404×85 = 4.143 − 2.893 = 1.25

Equation: Yield = 1.25 + 0.034 × Rainfall

Step 3: Predicted Values and SS

Month Y Ŷ=1.25+0.034X Residual (Y−Ȳ)² Total (Y−Ŷ)² Residual (Ŷ−Ȳ)² Regression
(Y−Ŷ)

Jan 2.5 2.61 -0.11 2.697 0.012 2.362

 Feb 3.1 3.12 -0.02 1.083 0.000 1.035
Mar 3.8 3.63 +0.17 0.117 0.029 0.262
Apr 4.0 4.14 -0.14 0.020 0.020 0.000
May 4.9 4.65 +0.25 0.574 0.063 0.257
Jun 5.2 5.16 +0.04 1.125 0.002 1.030
Jul 5.5 5.67 -0.17 1.839 0.029 2.327
TOTAL 7.455 0.135 7.273

SS Summary — Example 2
SST = 7.455
SSE = 0.135 (small — model fits well)
SSR = 7.273 (most variation explained)

Verification: 7.273 + 0.135 = 7.408 ≈ 7.455 ✓ (small rounding difference)

R² = 1 − (0.135 / 7.455) = 1 − 0.0181 = 0.982

Interpretation: Rainfall explains 98.2% of crop yield variation. Excellent fit.

Step 4: ANOVA Table — Full Calculation

Source df SS MS F p-value

Regression 1 7.273 7.273/1 = 7.273 7.273/0.0225 = < 0.0001

323.2
Residual n−2 = 5 0.135 0.135/5 = 0.027 — —
Total 6 7.408 — — —

MSE = 0.135/5 = 0.027. F = 7.273/0.027 = 323.2

In Excel: p-value = [Link](323.2, 1, 5) → approximately 0.000004 (essentially zero)

Step 5: Conclusion for Example 2

✅ CONCLUSION — Example 2
R² = 0.982: Rainfall explains 98.2% of crop yield variation. Excellent model.
F = 323.2, p-value < 0.0001: Extremely unlikely to occur by chance.

DECISION: Reject H₀ strongly. Monthly rainfall is a highly significant predictor of crop yield.
WRITTEN CONCLUSION:
'Simple linear regression was performed with monthly rainfall as the predictor of
wheat yield. The model was statistically significant (F(1,5) = 323.2, p < 0.0001),
with rainfall explaining 98.2% of the variance in yield (R² = 0.982). For every
1 mm increase in monthly rainfall, yield increases by approximately 0.034 tonnes
per hectare. The model is highly reliable for yield forecasting within the
observed rainfall range of 40–130 mm.'
Example 3 — Weak/Insignificant Model: Noise vs. Performance

📖 The Setup
A misguided manager claims: office noise level (decibels) predicts employee productivity score.
n = 6 employees measured over one month.
This example shows what a BAD, INSIGNIFICANT model looks like in every statistic.

Step 1: Data
Emp Noise X (dB) Productivity Y X² XY

1 55 72 3025 3960
2 60 68 3600 4080
3 60 80 3600 4800
4 65 70 4225 4550
5 65 75 4225 4875
6 70 71 4900 4970
Σ 375 436 23575 27235

n=6, X̄ =62.5, Ȳ=72.67

Step 2: Regression Equation

Slope and Intercept
b = (6×27235 − 375×436) ÷ (6×23575 − 375²)
b = (163410 − 163500) ÷ (141450 − 140625)
b = −90 ÷ 825 = −0.109

a = 72.67 − (−0.109)×62.5 = 72.67 + 6.82 = 79.49

Equation: Productivity = 79.49 − 0.109 × Noise

Slope is tiny and negative — barely any relationship.

Step 3: SS Calculations
Emp Y Ŷ (Y−Ȳ)² Total (Y−Ŷ)² Residual (Ŷ−Ȳ)² Regression

1 72 73.50 0.44 2.25 0.67

 2 68 72.96 21.78 24.60 0.10
3 80 72.96 53.78 50.80 0.10
4 70 72.41 7.11 5.80 0.07
5 75 72.41 5.44 6.72 0.07
6 71 71.87 2.78 0.76 0.36
TOTAL 91.33 90.93 1.37

SS Summary — Example 3
SST = 91.33
SSE = 90.93 (enormous — model barely reduced the variation!)
SSR = 1.37 (tiny — noise level explained almost nothing)

Verification: 1.37 + 90.93 = 92.30 ≈ 91.33 ✓ (small rounding difference)

R² = 1 − (90.93/91.33) = 1 − 0.9956 = 0.0044 → less than 0.5%!

Interpretation: Noise explains only 0.44% of productivity variation.

Step 4: ANOVA Table — Weak Model

Source df SS MS F p-value

Regression 1 1.37 1.37 1.37/22.73 = 0.060 ~0.815

Residual 4 90.93 90.93/4 = 22.73 — —
Total 5 91.33 — — —

F = 0.060. In Excel: =[Link](0.060, 1, 4) → p-value ≈ 0.815

Step 5: Conclusion for Example 3

❌ CONCLUSION — Example 3 (Do NOT use this model)
R² = 0.004: Noise explains only 0.4% of productivity variation. Essentially nothing.
F = 0.060, p-value = 0.815: 81.5% chance these results occurred purely by random chance.

DECISION: FAIL TO REJECT H₀. There is no statistical evidence that noise level
predicts employee productivity.

WRITTEN CONCLUSION:
 'A simple linear regression was conducted to test whether office noise level predicts
employee productivity. The model was not statistically significant (F(1,4) = 0.060,
p = 0.815), and noise level accounted for less than 1% of variance in productivity
scores (R² = 0.004). We fail to reject the null hypothesis. There is insufficient
evidence to conclude that noise level is a meaningful predictor of productivity.
Other factors — such as task complexity, experience, or work design — should
be investigated as more relevant predictors.'

NOTE TO MANAGER: High p-value does not prove noise has zero effect. It means
we cannot statistically establish a link with this small dataset.
 PART 7 — THE COMPLETE DECISION FRAMEWORK

7. Reading Any Regression Output — Your Step-by-Step

Guide
7.1 The 5-Step Protocol — Do This Every Time

📋 STEP-BY-STEP PROTOCOL FOR ANY REGRESSION OUTPUT

STEP 1 — Check p-value (Significance F) first.
• p < 0.05 → Model is statistically significant. PROCEED.
• p ≥ 0.05 → Model is NOT significant. STOP. Do not interpret R² or make predictions.
(A non-significant model tells you nothing reliable.)

STEP 2 — Check R².

• How much variation does X explain? (Higher = better fit.)
• Is the explanatory power adequate for your decision context?

STEP 3 — Check Standard Error (√MSE).

• Predictions lie within approximately ±2 × SE of the true value (95% confidence).
• Is this precision acceptable for your purpose?

STEP 4 — Check the Coefficients (slope and intercept).

• What does the slope tell you? How much does Y change per unit of X?
• Is the direction (+ or −) as expected? Does it make business sense?

STEP 5 — Check Residuals.

• Are there outliers (very large residuals)?
• Is any pattern visible? (Curved residuals → model may need to be non-linear.)
7.2 The Three Examples Compared — Side by Side
Metric Example 1 (Study/Scores) Example 2 (Rain/Yield) Example 3
(Noise/Productivity)

SST 1210 7.455 91.33

SSR 1210 7.273 1.37
SSE 0 0.135 90.93
SSR/SST 1.000 0.976 0.015
SSE/SST 0.000 0.018 0.995
R² 1.000 0.982 0.004
F-statistic ∞ 323.2 0.060
p-value ~0 <0.0001 0.815
Significant? YES YES NO
Model Useful? YES YES NO

Notice: As SSE/SST goes up (more unexplained), R² goes down, F goes down, and p-value goes up. They all move
together — they are measuring the same thing from different angles.

7.3 Writing a Professional Statistical Conclusion

📝 THE TEMPLATE — Fill in the blanks for any regression

'Simple linear regression was performed to examine the effect of [X variable] on [Y variable].
The model was [statistically significant / not statistically significant]
(F([df_Reg], [df_Res]) = [F value], p = [p-value]).
[X variable] accounted for [R² × 100]% of the variance in [Y variable] (R² = [value]).
The regression equation was [Y] = [a] + [b] × [X],
indicating that for every one-unit increase in [X], [Y] [increases/decreases] by [b] units.
[Based on these results, we conclude / We cannot conclude] that [X] is a significant predictor of [Y].'

EXAMPLE FILLED IN (for Example 2):

'Simple linear regression was performed to examine the effect of monthly rainfall on wheat yield.
The model was statistically significant (F(1,5) = 323.2, p < 0.0001).
Rainfall accounted for 98.2% of the variance in yield (R² = 0.982).
The regression equation was Yield = 1.25 + 0.034 × Rainfall,
indicating that for every 1 mm increase in rainfall, yield increases by 0.034 t/ha.
 Based on these results, we conclude that monthly rainfall is a significant predictor of wheat yield.'

7.4 The Master Summary — All Concepts Connected

🔗 HOW EVERYTHING CONNECTS — THE BIG PICTURE

DATA has variation → measured as SST = Σ(Y−Ȳ)²

REGRESSION tries to explain that variation → the line produces predictions (Ŷ)

WHAT IT EXPLAINED = SSR = Σ(Ŷ−Ȳ)² ← how far the line moved from the mean
WHAT IT MISSED = SSE = Σ(Y−Ŷ)² ← how far actual points are from the line

IDENTITY: SST = SSR + SSE (always exact under OLS — cross-product = 0)

R² = SSR/SST = 1 − SSE/SST → 'What fraction of the mystery did we solve?'

Mean Squares remove the degrees of freedom effect:

MSR = SSR/1 (for simple regression)
MSE = SSE/(n−2)

F = MSR/MSE → 'Is the signal stronger than the noise?'

F → F-distribution → p-value → 'How likely is this F by chance alone?'

p < 0.05 → Reject H₀ → X IS a significant predictor of Y → USE THE MODEL

p ≥ 0.05 → Fail to reject H₀ → INSUFFICIENT evidence → DO NOT rely on this model

7.5 Quick Reference Card — All Formulas

Formula What It Means
SST = Σ(Y−Ȳ)² Total variation in Y — the full mystery
SSR = Σ(Ŷ−Ȳ)² Variation explained by X — what the model solved
SSE = Σ(Y−Ŷ)² Variation unexplained — the leftover errors
SST = SSR + SSE The golden identity — always true under OLS
R² = SSR/SST Proportion explained — ranges 0 to 1
R² = 1 − SSE/SST Equivalent derivation — one minus unexplained fraction
MSR = SSR/1 Mean square for regression (simple model)
MSE = SSE/(n−2) Mean square for residuals — the noise floor
F = MSR/MSE Signal-to-noise ratio — larger = better model
p-value = [Link](F, 1, n−2) Excel formula for the p-value
p < 0.05 Reject H₀ — model is statistically significant
p ≥ 0.05 Fail to reject H₀ — insufficient evidence

© SkillArbitrage | Financial & Data Analytics | Variation, R², F-Statistic & p-value