Standard Error Slides Standard Error
Dynamics
Why variances add and how we measure difference
Unit: Statistical Inference for Means
The Scoring Paradox
Imagine two basketball teams:
Team A
Avg score: 100 | Var: 25
Team B
Avg score: 90 | Var: 16
If we look at the difference (A - B)...
Does the variation in the final point spread get smaller because we are subtracting, or larger because both teams contribute uncertainty?
The Golden Rule
When combining independent random variables...
Variances ALWAYS Add
\[ \sigma_{X \pm Y}^2 = \sigma_X^2 + \sigma_Y^2 \]
Whether we add or subtract the means, the uncertainty (variance) grows. You can't subtract uncertainty!
Deriving the Formula
1. Start with the individual SE
\[ SE_{\bar{x}} = \frac{s}{\sqrt{n}} \]
2. Convert to Variance
\[ Var(\bar{x}) = (SE_{\bar{x}})^2 = \frac{s^2}{n} \]
3. Sum the Variances & Root it!
\[ SE_{\bar{x}_1 - \bar{x}_2} = \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}} \]
This is the Standard Error for the difference between two independent sample means.
The Big Picture
Independence
This formula ONLY works if the two groups are independent. No matched pairs allowed here!
The t-Score
We use this SE to calculate our t-statistic: \[ t = \frac{(\bar{x}_1 - \bar{x}_2) - 0}{SE_{diff}} \]
Precision
As sample sizes ($n_1, n_2$) increase, our Standard Error decreases, giving us more precise estimates.
Variance Addition Worksheet Variance Addition
Independent Means & Standard Error
Name:
Date:
1
The Core Concept: You Can't Subtract Uncertainty
In statistics, when we combine two independent random variables $X$ and $Y$, the variation in the result always increases, regardless of whether we are adding or subtracting the variables.
\[ \sigma^2_{X+Y} = \sigma^2_X + \sigma^2_Y \] \[ \sigma^2_{X-Y} = \sigma^2_X + \sigma^2_Y \]
Scenario: Commute Times
Your morning commute time ($M$) has a mean of 25 minutes and a variance of 4 minutes. Your evening commute ($E$) has a mean of 35 minutes and a variance of 9 minutes. Assume morning and evening times are independent.
A) Calculate the mean and variance of your total daily commute ($M + E$).
B) Calculate the mean and variance of the difference between your evening and morning commute ($E - M$).
C) Explain in one sentence why the variance for the difference is not $9 - 4 = 5$.
2
Calculating Standard Error for Means
The Two-Sample SE Formula:
\[ SE_{\bar{x}_1 - \bar{x}_2} = \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}} \]
Case Study A: Battery Life Showdown
A consumer group tests two brands of AA batteries to see which lasts longer in a digital camera. They collect the following data (in hours):
Brand X
$n = 30$
$\bar{x} = 4.5$
$s = 0.8$
Brand Y
$n = 45$
$\bar{x} = 5.2$
$s = 1.1$
Calculate the standard error for the difference in sample means ($\bar{x}_Y - \bar{x}_X$). Show your calculation.
Case Study B: Sleep and Test Scores
An education researcher compares test scores between students who slept >8 hours (Group 1) and <8 hours (Group 2).
Group 1 (>8 hrs)
$n = 50$
$s = 12$
Group 2 (<8 hrs)
$n = 50$
$s = 15$
Find the standard error for the difference in means. If the sample sizes were both doubled to $n=100$, what would happen to the Standard Error? (Explain or show calculation).
3
The Independence Requirement
For each scenario below, determine if the two-sample standard error formula is appropriate. Explain why or why not.
Scenario 1:
A researcher measures the heart rate of 20 participants before exercise and then measures the heart rate of the same 20 participants after exercise.
Scenario 2:
A gardener applies Fertilizer A to 15 tomato plants in her backyard and Fertilizer B to 15 different tomato plants in her front yard.
Standard Error Teacher Guide Teacher Resource Lesson 1: Standard Error Dynamics
Standard Error Facilitation Guide
Guidance for deriving the variance addition rule and the two-sample t-statistic foundation.
Learning Objectives
Conceptualize why variances add for independent variables.
Derive the standard error formula for the difference in sample means.
Identify conditions under which two-sample procedures are valid.
The Hook: The Basketball Point Spread
Start by asking students: "If Team A is very inconsistent (wide spread of scores) and Team B is also very inconsistent, will the final point spread ($A - B$) be more or less consistent than the individual teams?"
Misconception Alert:
Students often think subtraction of means implies subtraction of variation. Use the physical analogy of "stacking errors": if you make an error in one direction, and another independent error in the other, you are further from "zero" certainty, not closer.
Mathematical Foundations
1. Variance of Mean
Remind students that $Var(X) = \sigma^2$, so $Var(\bar{x}) = \left(\frac{\sigma}{\sqrt{n}}\right)^2 = \frac{\sigma^2}{n}$. We use sample standard deviation $s$ to estimate this.
2. Summing Variances
Since $X$ and $Y$ are independent, $Var(\bar{x}_1 - \bar{x}_2) = Var(\bar{x}_1) + Var(\bar{x}_2)$. This is the most critical conceptual hurdle.
3. Square Root Back to SE
We must take the square root of the combined variance to return to the scale of the original data (Standard Error).
Pacing
10 min: Basketball Hook & Discussion
15 min: Slide presentation & Derivation
20 min: Worksheet guided practice
5 min: Wrap-up: Checking Independence
Differentiation
Support: Use color-coding for the two groups ($n_1$ in red, $n_2$ in blue) throughout the formula.
Extension: Ask students to prove why $Var(X-Y) = Var(X) + Var(Y)$ using the definition of variance if they are in an honors/advanced track.
Closing Discussion Prompts
"If Group A has a much larger sample size than Group B, which group's variation contributes more to the overall standard error of the difference?"
"Why do we use $s$ instead of $\sigma$? How does this change our distribution from $z$ to $t$?"
Degrees of Freedom Slides Degrees of
Freedom
Navigating the complexity of two-sample t-distributions
The Calculator Mystery
You run a 2-sample t-test on your calculator and see this:
df = 37.4218...
"Wait... how can you have 0.42 of a 'degree of freedom'? And why isn't it just (n - 1)?"
The Challenge
In two-sample inference, the math behind the "exact" t-distribution is incredibly messy. We have two different sample sizes and two different standard deviations trying to share one distribution.
Today, we explore two ways to solve this.
Method 1: The Conservative Approach
The "By Hand" Rule
\[ df = min(n_1 - 1, n_2 - 1) \]
Why use it?
Safe, easy, and yields a higher p-value (conservative). You are less likely to make a Type I error.
The Downside
It lacks power. It underestimates the "true" degrees of freedom, making it harder to reject the null.
Method 2: Technology Approximation
Welch-Satterthwaite Equation
\[ df = \frac{(\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2})^2}{\frac{1}{n_1-1}(\frac{s_1^2}{n_1})^2 + \frac{1}{n_2-1}(\frac{s_2^2}{n_2})^2} \]
More Accurate
Greater Statistical Power
Impossible to do by hand
Does It Really Matter?
Large Samples
When $n_1$ and $n_2$ are large, the difference between the two methods is negligible. Both distributions look like the Normal curve.
Small Samples
When samples are small, technology gives a significantly more precise t-distribution, resulting in smaller p-values.
Exam Strategy
If using a calculator: Always report the tech df (e.g., 37.4).
If using a table: Use the conservative df (round down if your exact df isn't on the table).
DF Investigation Worksheet Degrees of Freedom
Conservative vs. Technology
STUDENT:
DATE:
When conducting a 2-sample t-test, we must determine which t-distribution to use. Today, you will compare the "Simple Conservative" method used with tables and the "Welch-Satterthwaite" method used by calculators.
The Data Set
A researcher is comparing the mean height of two species of sunflowers grown in different soil types.
Species A
Sample Size ($n_1$): 12
Sample Mean ($\bar{x}_1$): 74.2 cm
Sample Std Dev ($s_1$): 4.5 cm
Species B
Sample Size ($n_2$): 25
Sample Mean ($\bar{x}_2$): 68.5 cm
Sample Std Dev ($s_2$): 8.2 cm
1
The Conservative Way (By Hand)
Use the formula: $df = min(n_1 - 1, n_2 - 1)$
$n_1 - 1$
$n_2 - 1$
Conservative $df$
Using a t-table, find the critical value $t^*$ for a 95% confidence level with this $df$:
2
The Technology Way (Calculator)
Enter the data above into your calculator's 2-Sample t-Test function. (Select "Inpt: Stats"). Do NOT pool variances.
REPORTED BY CALCULATOR:
t =
p =
df =
3
The Great Comparison
A) Compare the $df$ from Part 1 and Part 2. Which one is larger?
B) Look at your t-distribution reference sheet. As $df$ increases, what happens to the area in the tails of the distribution for a fixed t-score?
C) Based on your answer to B, why is the calculator's p-value usually smaller than the one you'd get using the conservative method?
D) "Conservative" means we err on the side of caution. In the context of hypothesis testing, how is the $min(n_1-1, n_2-1)$ method more "cautious"? (Think about Type I errors).
Feature Conservative Method Technology Method Value of $df$ Smaller (integer) Larger (decimal) p-value Larger Smaller Risk of Type I Error Lower Higher (but accurate)
T-Distribution Reference Sheet t-Distribution Guide
Inference for Means of Independent Groups
When to use 't' vs 'z'
Z
Normal Distribution
Used for Proportions OR when the population standard deviation ($\sigma$) is known (rare!).
t
t-Distribution
Used for Means when $\sigma$ is unknown and we estimate it using sample standard deviation ($s$).
Properties of t
Symmetric and bell-shaped, centered at 0.
Has "fresher/fatter" tails than the Z-curve (more area far from the mean).
As $df$ increases, the t-distribution approaches the Standard Normal distribution.
The "Safe" Rule
If you are ever unsure of the technology approximation for $df$, use the conservative value: the smaller of $(n_1 - 1)$ and $(n_2 - 1)$. This ensures your results are robust and not over-stated.
Required Conditions
1. Randomness
Data must come from two independent random samples or a randomized experiment.
2. Independence
Samples must be independent of each other. If sampling without replacement, $n < 10\%$ of population.
3. Normality
Both populations are Normal, OR both sample sizes are large ($n \ge 30$). If small, check for outliers/skewness.
Formula Cheat Sheet
Standard Error
\[ SE = \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}} \]
t-statistic
\[ t = \frac{(\bar{x}_1 - \bar{x}_2) - 0}{SE} \]
© Statistical Reference Material • AP Statistics Standards Compatible
Clinical Trial Slides Clinical Trial
Evidence
Testing Significance in Modern Medicine
The $500 Million Decision
A pharmaceutical company develops "Cephala-Gone" for migraine relief.
"Our data shows that Cephala-Gone reduces pain significantly faster than the leading competitor."
Before the FDA approves it, they need PROOF.
The Key Questions
Is the difference just due to random chance?
Were the groups truly independent?
Is the sample size large enough to trust?
The 2-Sample t-Test Workflow
01
State
Define $H_0: \mu_1 - \mu_2 = 0$ and $H_a$. Identify the parameter and significance level ($\alpha$).
02
Plan
Check conditions:
03
Do
Calculate the t-statistic, degrees of freedom, and the p-value.
04
Conclude
Compare p to $\alpha$. Reject or Fail to Reject $H_0$ in context.
The "Normal" Hurdle
For 2-sample tests, we must check Normality for BOTH groups independently.
If $n_1, n_2 \ge 30$, we are safe (CLT).
If small, graph the data! Look for extreme skew or outliers.
Two Samples, Two Checks
Group 1 Plot
Group 2 Plot
The Final Verdict
"Since our p-value ($0.031$) is less than $\alpha = 0.05$..."
"...we REJECT the null hypothesis. There is convincing evidence that the mean relief time for Cephala-Gone is significantly lower than for the placebo."
Statistical Tip:
Never say we 'proved' the medicine works. We say we have 'evidence' against the null hypothesis.
Clinical Trial Case Study Packet Case Study: Medical Statistics
The Cephala-Gone Trial
STUDENT ID: ________________
Background
Pharma-Global is testing a new fast-acting headache relief pill called Cephala-Gone. They randomly assigned 80 volunteers with chronic tension headaches into two groups. One group received Cephala-Gone, and the other received a placebo (sugar pill).
The primary outcome measured was time to relief (in minutes). The results were as follows:
Group 1: Cephala-Gone
$n_1 = 40$
$\bar{x}_1 = 28.5 \text{ min}$
$s_1 = 8.2 \text{ min}$
Group 2: Placebo
$n_2 = 40$
$\bar{x}_2 = 35.1 \text{ min}$
$s_2 = 12.4 \text{ min}$
Research Question
Do patients taking Cephala-Gone experience a significantly faster mean relief time than those taking a placebo?
Use a significance level of $\alpha = 0.05$ for all calculations.
1
STATE
Define the parameters, hypotheses, and significance level.
Parameter of Interest ($\mu_1 - \mu_2$):
Null Hypothesis ($H_0$):
Alternative Hypothesis ($H_a$):
2
PLAN
Check conditions for a two-sample t-test.
Random:
Independence:
Normal / Large Sample:
3
DO
Calculate the test statistic and p-value.
A) Calculate the Standard Error ($SE$):
B) Calculate the t-statistic ($t$):
C) Degrees of Freedom ($df$):
(Choose one: Conservative or Technology approximation)
D) P-Value:
4
CONCLUDE
Make your final decision in the context of the study.
Final Verdict (Compare p-value to $\alpha$ and state conclusion):
Clinical Trial Answer Key Answer Key Teacher Reference Only
The Cephala-Gone Trial: Key
1. State
Parameters: $\mu_1 = \text{true mean relief time for Cephala-Gone}$, $\mu_2 = \text{true mean relief time for Placebo}$.
Hypotheses: $H_0: \mu_1 - \mu_2 = 0$ (or $\mu_1 = \mu_2$) | $H_a: \mu_1 - \mu_2 < 0$ (or $\mu_1 < \mu_2$).
Significance Level: $\alpha = 0.05$.
2. Plan
Random: Yes, patients were "randomly assigned" to groups.
Independence: 40 is likely < 10% of all potential migraine sufferers. Groups are independent of each other.
Normal / Large Sample: $n_1 = 40 \ge 30$ and $n_2 = 40 \ge 30$. The Central Limit Theorem applies for both groups.
3. Do
Standard Error:
\[ SE = \sqrt{\frac{8.2^2}{40} + \frac{12.4^2}{40}} \approx 2.351 \]
t-statistic:
\[ t = \frac{28.5 - 35.1}{2.351} = \frac{-6.6}{2.351} \approx -2.807 \]
Degrees of Freedom:
Conservative: $df = 39$
Technology: $df \approx 67.8$
P-Value:
Using tech ($df=67.8$): $p \approx 0.0033$
Using table ($df=39$): $0.0025 < p < 0.005$
4. Conclude
"Since the p-value ($0.0033$) is less than $\alpha = 0.05$, we reject the null hypothesis. There is convincing evidence that the true mean relief time for Cephala-Gone is significantly faster than the true mean relief time for the placebo."
Facilitation Tip
Encourage students to use the calculator's 2-SampTTest for the "Do" step. If they get a different p-value, check if they used "Not Pooled" (correct) vs "Pooled" (incorrect). In this course, we almost never pool variances for means.
Mean Difference Slides Estimating the
Difference
From Hypothesis Tests to Confidence Intervals
2-Sample t-Intervals Inference for Means
The Better Battery?
We rejected the null hypothesis: Brand Y lasts longer than Brand X.
Great news! But wait...
"If Brand Y is \$2 more expensive, is the difference in life span worth the extra cost? How much longer does it actually last?"
The Limit of p-values
A p-value tells us if there is a difference. A Confidence Interval tells us how big that difference is.
Constructing the Estimate
\[ (\bar{x}_1 - \bar{x}_2) \pm t^* \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}} \]
1
Statistic
The difference in sample means.
2
Critical Value
$t^*$ based on your confidence level and $df$.
3
Standard Error
The same formula we used for the test.
The Zero Connection
When interpreting a confidence interval for $\mu_1 - \mu_2$, look for Zero.
If 0 is in the interval:
The difference could be zero! We do NOT have evidence of a significant difference. (Corresponds to Fail to Reject $H_0$)
If 0 is NOT in the interval:
Zero is not a plausible value. We have evidence of a significant difference. (Corresponds to Reject $H_0$)
Visualizing the Interval
Lower Bound
Upper Bound
Zero is here
"If both bounds are positive, Group 1 is significantly larger than Group 2."
Quick Check
A 95% Confidence Interval for $\mu_{\text{New}} - \mu_{\text{Old}}$ is $(3.2, 8.5)$.
Is it significant?
Yes! 0 is not in the interval.
What does it mean?
The new product is between 3.2 and 8.5 units better than the old one.
Confidence intervals give us context, magnitude, and direction—all in one estimate.
Battery Life Interval Sheet Confidence Intervals
Estimating $\mu_1 - \mu_2$
NAME: ___________________________
DATE: ___________________________
\[ \text{Interval} = (\bar{x}_1 - \bar{x}_2) \pm t^* \sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}} \]
The Smartphone Battery Duel
A tech reviewer is comparing the battery life of the "Titan-Phone" and the "Ultra-Phone." They take independent random samples of users and record their average screen-on time (in hours) until the battery dies.
Titan-Phone
Group 1
$n_1 = 35$
$\bar{x}_1 = 9.4 \text{ hrs}$
$s_1 = 1.2 \text{ hrs}$
Ultra-Phone
Group 2
$n_2 = 35$
$\bar{x}_2 = 8.8 \text{ hrs}$
$s_2 = 1.5 \text{ hrs}$
1
Calculate the difference in sample means ($\bar{x}_1 - \bar{x}_2$):
2
Find the Standard Error for the difference ($SE$):
3
Find the critical value $t^*$ for a 99% Confidence Level.
Use the conservative degrees of freedom method ($df = n-1$).
Degrees of Freedom:
Critical Value $t^*$:
4
Construct the Confidence Interval:
(Point Estimate $\pm$ Margin of Error)
Interpretation
Write a sentence interpreting the interval in the context of the smartphone battery life.
Based on your 99% confidence interval, is there evidence of a significant difference between the two phones? Explain using the inclusion (or exclusion) of zero.
Effect Size Slides Significance
vs. Impact
When "Statistically Significant" isn't enough
The $5 Million SAT Boost
A school district spends \$5 million on a new AI tutoring program. After a year, they find:
Result:
SAT scores improved by 1.2 points (out of 1600).
p-value = 0.0001
Is it Significant?
Yes! (Statistically)
Is it Useful?
Maybe not. An improvement of 1 point is practically meaningless.
The Great Divide
Statistical
Determined by the p-value.
As $n$ gets very large, even tiny differences become "significant."
"The result is likely not due to random chance."
Practical
Determined by Subject Matter Expertise.
Does the magnitude of the difference matter in the real world?
"The result is large enough to change our behavior or policy."
Effect Size (Cohen's d)
To measure magnitude, we use the number of standard deviations the means are apart.
\[ d = \frac{\bar{x}_1 - \bar{x}_2}{s_{pooled}} \]
0.2
Small Effect
0.5
Medium Effect
0.8+
Large Effect
The Researcher's Checklist
1
Is $p < \alpha$? (Is there a real difference?)
2
How large is the interval? (What is the likely range of the difference?)
3
Is the effect size practically meaningful? (Should we care?)
Never report just the p-value. Tell the whole story.
Statistical vs Practical Worksheet Significance & Scope
Statistical vs. Practical Significance
NAME: ___________________________
Statistical Significance
"Is the observed effect likely due to random chance?"
Key Indicator:
p-value < $\alpha$
Practical Significance
"Is the magnitude of the effect meaningful in the real world?"
Key Indicator:
Confidence Interval / Effect Size
Case Analysis
Case A: The "Miracle" Weight Loss Supplement
A study of 10,000 participants found that those taking "Burn-Fast" lost an average of 0.5 pounds more than the placebo group over 6 months. The p-value was 0.002.
Statistical Significance?
Yes
No
Practical Significance?
Yes
No
Justify your answer for Practical Significance:
Case B: The Surgical Technique
A small pilot study ($n=20$) compared a new robotic surgical technique to traditional surgery. The new technique reduced recovery time by 4 days on average, but the p-value was 0.12.
Statistical Significance?
Yes
No
Practical Significance?
Yes
No
What is likely preventing this study from reaching statistical significance despite a large physical difference?
Quantifying Magnitude
Calculate Cohen's $d$ for a study where $\bar{x}_1 = 50$, $\bar{x}_2 = 45$, and the pooled standard deviation is $s_p = 4$.
\[ d = \frac{50 - 45}{4} = \text{______} \]
Interpret Effect Size:
Circle one:
Small Medium Large
Final Thought: Why is it important to report a confidence interval and a p-value in a scientific report?
