# Statistics - (dependent|paired sample) t-test

> (Statistics|Probability|Machine Learning|Data Mining|Data and Knowledge Discovery|Pattern Recognition|Data Science|Data Analysis)

### Table of Contents

## 1 - About

A dependent t-test is appropriate when:

- we have the same people measured twice.
- the same subject are been compared (ex: Pre/Post Design)
- or two samples are matched at the level of individual subjects (allowing for a difference score to be calculated)

The idea is that one measure is dependent on the other. That they're related.

Is the difference between means a significant difference or is this difference just due to chance because of sampling error ?

If the mean of this different scores is significantly different from zero, we have a significant change.

## 2 - Articles Related

## 3 - Assumption

The distribution is normal

## 4 - Calculation

### 4.1 - Analysis

A thorough analysis will include:

- A measure of effect size: Cohen's d (because NHST is biased by sample size)
- Confidence Interval. An interval estimates because sample means are just point estimates.

### 4.2 - Different score

The same subjects or cases are measured twice. We can calculate a different score for each individual subject.

<MATH> \begin{array}{rrl} \text{Different Score} & = & \href{raw_score}{X}_1 - \href{raw_score}{X}_2 \\ \end{array} </MATH>where:

#### 4.2.1 - t-value

See t-value for mean

<MATH> \begin{array}{rrl} \href{t-value#mean}{\text{t-value}} & = & \frac{\href{mean}{\text{Mean of the Different Scores}}}{\href{Standard_Error#mean}{\text{Standard Error of the Different Scores}}} & \\ \end{array} </MATH>#### 4.2.2 - p-value

The p-value will be based on:

- the above t-value and which t-distribution we're in
- whether we're doing a non-directional or directional test.

#### 4.2.3 - Effect size

The most appropriate and the most common estimate of effect size is **Cohen's d**.

Because NHST is biased by sample size, we should supplement the analysis with an estimate of effect size: Cohen's d

And the effect size is calculated differently than in regression.

Cohen's d is a intuitive measure that tells us how much in terms of **standard deviation units**:

- one measurement differ from another (in a dependent t-test)
- one mean differ from another (in a independent t-test)

As you can remark:

- For the t-value, the denominator is standard error,
- For d, the denominator is the standard deviation.

Why ? Because:

- Standard error is biased by N
- whereas standard deviation is not.

A Cohen's d of 1 means that:

- score's went up a whole standard deviation.
- it's a strong effect.

0.8 is also a strong effect.

#### 4.2.4 - Confidence Interval

We can also get interval estimates around these means rather than just point estimates.

We get the mean of the difference scores and put an upper bound and a lower bound. It's the same method than for sample means or regression coefficients.

<MATH> \begin{array}{rrl} \text{Upper bound} & = & \href{Mean}{\text{Mean of the difference scores}} & + & \href{#t-value}{\text{t-value}}.\href{Standard_Error}{\text{Standard Error}} \\ \text{Lower bound} & = & \href{Mean}{\text{Mean of the difference scores}} & - & \href{#t-value}{\text{t-value}}.\href{Standard_Error}{\text{Standard Error}} \end{array} </MATH>That exact t-value value depends on:

- how confident we want to be so like a 95% confidence interval Versus an 90% confidence interval.
- which sampling distribution of t we're going to to use (because we have that family of t distribution). So it depends on the number of subjects in the sample.

When the interval does not include zero, it's significant in terms of null hypothesis significance testing.

### 4.3 - Simulation

Build a sampling distribution of the differences

Pseudo Code: Loop until you get a beautiful normal distribution

- Take the two samples
- Shuffle the observations between the two samples
- Calculate and plot the mean

After getting the normal distribution, calculate the probability of the differences.