Lecture Notes on t Tests
Chapter 7: t Tests
Outline
- Estimating Population Variance from a Sample
- How to do it
- Degrees of Freedom
- The t Distribution
- t Scores
- Miscellaneous Related Points
- Effect Size with Estimated Variance
- t Tests in Psychological Research
- Types of t Tests
- Single Sample Tests
- Paired-Samples Tests
- Assumptions
Hypothesis Testing without Population Parameters
In psychology, relevant population parameters are often unknown.
- For example, when measuring scores on a memory or attention test, previous studies might not exist for comparison.
- t Test: A hypothesis testing method where the population variance is not known. It compares sample t scores to the t distribution.
- Single Sample t Test: A procedure applied when the sample population mean is known.
Estimating Population Variance from Sample Scores
- Populations with high variability tend to yield highly variable samples, while populations with low variability yield samples with low variability.
- Sample variance is generally less than true population variance, leading to biased estimates.
- Biased Estimate: An estimate that consistently over or underestimates a parameter.
Finding an Unbiased Estimate
- An unbiased estimate means it is equally likely to overestimate or underestimate a parameter.
- The formula for estimating population variance:
S² = Σ(X-M)² / (N-1) - Population standard deviation is estimated as: S = √S².
Degrees of Freedom
- Degrees of Freedom (df): the number of scores that can vary freely (in this case, N-1).
- If the mean is known, and we have N-1 scores, knowing all the other scores keeps the last score fixed.
- Example: For three scores averaging to 10, if the scores are [5, 15, x], knowing the mean helps us compute x.
The t Distribution
- The estimation of population variance assumes a different distribution than the true variance, often resulting in fatter tails compared to normal distributions.
- The shape of the t distribution is affected by degrees of freedom; more degrees of freedom yield a shape closer to the normal distribution.
- At around 30 df, the t distribution resembles the normal distribution closely.
Understanding t Scores
- A t Score indicates how many standard deviations a sample mean is from the population mean on the t distribution, similar to a Z score but adjusted for estimation.
- Critical Values are determined using a t table based on degrees of freedom and the significance level (alpha).
- t Scores are calculated using the formula:
t = (M - μM) / SM, where M is sample mean, μM is population mean, and SM is the unbiased estimator of standard deviation.
Conducting a Single Sample t Test
- Similar to hypothesis tests in Chapter 5, but with adjustments:
- Standard deviation is estimated using N-1.
- Compute Standard Error of Mean (SEM) using sample standard deviation.
- Retrieve critical value from t distribution chart and convert raw score to t score.
- Example context: Comparing ratings for new song quality against a known average rating.
Paired-Samples t Test
- Referred to as Dependent Means t Test when repeated measures for each participant are involved.
- Tests for differences between two related conditions using difference scores (M₁ - M₂) with comparison distribution mean set to zero.
- Each score pair must be linked meaningfully, such as in repeated measures or matched subjects.
Assumptions of Normality
- For hypothesis testing, scores must be normally distributed; violating this can lead to incorrect conclusions.
- However, t tests are robust to moderate violations of normal distribution premises, as distributions of means normalize with larger samples.
- Robustness reflects the accuracy of results despite assumption violations.
Effect Size with Estimated Variance
- Calculated similar to previous chapters, but using the unbiased estimator of standard deviation.
- Cohen's d Formula:
d = (μ - μ₂) / S (where μ₂ equals 0 in paired-samples cases).
Example of Cohen's d with Estimated Variance
- Based on a previous example of athletes' performance, the difference in weight lifted qualifies d calculations.
t Tests in Psychological Research
- Consideration of power in repeated measures to focus on difference score variance rather than overall levels.
- Example reporting in research: t(28) = 7.36, Bayes factor = 450,000, suggesting strong evidence against the null hypothesis.