stats midterm 2

1. Categorical vs. Quantitative Variables & Comparing Groups

Understanding Variable Types

You must know how to classify variables, because this determines which statistical method you use.

Categorical variables

Describe groups or categories

Examples: gender, previous quit attempt (yes/no), treatment (AZT vs placebo)

Typically summarized using: counts, proportions, bar graphs, segmented bar graphs

Quantitative variables

Numerical values where arithmetic makes sense

Examples: age, weight, number of cigarettes, number of words memorized

Summarized with: mean, median, SD, boxplots, histograms

What dictates a comparison?

Compare proportions → when variables are categorical

Compare means → when variables are quantitative

Statistical Goals in Experimental Design

When comparing background variables between treatment groups:

You hope to fail to reject null hypotheses (i.e., groups are similar)

Because well-balanced groups help guarantee fairness and reduce confounding

---

2. Randomization Tests & Two-Way Tables (Shift Study Example)

Observational Units

Know how to identify what “one row of data” represents — here it was one shift.

Explanatory vs. Response Variables

Explanatory: “Gilbert on shift?” (yes/no)

Response: “At least one death?” (yes/no)

Two-way tables

Used when both variables are categorical.

Statistics Commonly Used

Difference in proportions

Risk ratio

Odds ratio

Randomization Test Logic

A randomization test:

1. Assumes shifts are assigned randomly under the null

2. Reassigns shift labels many times

3. Measures how often the simulated statistic is as extreme as observed

Interpreting Simulated Null Distributions

If observed statistic is far in the tail → reject H₀

If it’s common → fail to reject H₀

“Lawyer interpretation” Skills

Be able to argue against causation:

Observational study?

Confounding variables?

Imbalanced shifts?

Patterns may appear by chance

Alternate Measures of Extremeness

Examples:

z-score

standardized statistic

tail proportion in permutation distribution

---

3. Observational Studies vs. Experiments & Confounding

(Positive/Negative Emotion and Colds Study)

Identifying explanatory & response variables

Explanatory: emotional state score (quantitative → categorized into thirds)

Response: did the person catch a cold? (binary categorical)

Study type matters

Experiment → researcher assigns explanatory variable

Observational study → just observes, no assignment

This example: observational study.

Implication for causal conclusions

Cannot conclude causation

Must consider confounding variables

Common Confounders

Examples:

stress levels

sleep

income

underlying health

exposure to virus outside study

You need to be able to:

1. Name a potential confounder

2. Explain how it affects both the explanatory & response variables

---

4. Two-Sample Z Tests, Segmented Bar Graphs, and Random Assignment

(AZT vs. placebo example)

Segmented Bar Graphs

Understand how they visualize:

Proportion infected within each treatment group

Validity conditions for a two-sample z test

For comparing two proportions, both groups must have:

At least 10 successes and

At least 10 failures

in both groups under the null OR observed counts.

Study Design

Important distinction:

Random assignment → supports causal inference

Random sampling → supports generalization to population

In this example: random assignment only

→ causation is justified, generalization is limited

---

5. Interpreting Two-Sample t-test Outputs & Effects on p-value

Understanding Software Output

Key pieces to interpret:

Sample means

Standard deviations

SE of difference

Test statistic (t-value)

p-value

Confidence interval

Predicting how changes affect p-values

You need conceptual understanding of how p-values behave:

1. Adding 1 to every observation

Increases both means equally → difference stays identical

p-value unchanged

2. Increasing sample standard deviations

SE increases

t-statistic decreases

p-value increases (less significant)

3. Increasing sample sizes

SE decreases

t-statistic increases

p-value decreases (more significant)

---

6. Matched Pairs Designs, Simulation Based Tests, Paired t-tests

(Jumping jacks & memory example)

Identifying Explanatory & Response Variables

Explanatory: condition (exercise vs. not) — categorical

Response: number of words memorized — quantitative

Null & Alternative Hypotheses

Know symbolic forms:

μᵈ = 0, μᵈ > 0, μᴇ - μₙ = 0, etc.

Simulation Plots (Red, Blue, Black)

You need to know:

Black plot = null distribution

Each dot = randomized difference in means (or paired differences) under null

Randomization mechanism

Understand how an applet reassigns values for matched pairs:

Randomly swap the two condition labels for each subject

Conclusion logic in randomization tests

Reject H₀ if:

Observed difference is in extreme tail of null distribution

Validity conditions for paired t-test

Need:

Differences are approximately normal

No extreme outliers

Sample size n ≥ 15 gives robustness

Paired t-test & CI Concepts

You must know formulas conceptually:

t = (mean diff) / (SD(diff)/√n)

CI = mean diff ± t* × SE(diff)

Interpretation:

A CI describes plausible values for the true mean difference.

---

7. Independent Samples vs Matched Pairs Design

(Milking methods example)

Know how to classify designs:

Independent samples → different cows in each group (Design A)

Matched pairs → same cow measured twice OR paired based on similarity (Designs B & C)

Key idea:

Matched pairs controls for cow-to-cow variability → reduces noise → increases power.

---

8. One-Sample Paired t-test: Violation of Expectation Study

(Helper vs Hinderer looking-time example)

Parameter of Interest

Always:

Mean difference in population (μᵈ)

Hypotheses

H₀: μᵈ = 0

Hₐ: μᵈ > 0 (longer looking at hinderer)

Appropriate test

One-sample paired t-test on differences

Validity conditions

Differences ~ normal shape OR

sample size ≥ 15

Here n = 16, so t-test is appropriate.

Interpretation of p-value

Always in context:

Probability of observing a difference as large or larger if infants in population truly have no preference.

Interpretation of Confidence Interval

CI gives range of plausible mean differences

If interval does not include 0 → supports significance