Day 11: Inference for difference in means from two independent samples and Power (Sections 5.3, 5.4)

BSTA 511/611

Week 7

Author

Affiliation

Meike Niederhausen, PhD

OHSU-PSU School of Public Health

Published

November 11, 2024

MoRitz’s tip of the day

Add tabbed sections to your html file using tabset.

You can make subsections appear as different tabs in your html file.
This is the first tab.
It was created by adding ::: panel-tabset right above the subsection ### First tab (see the code file).
Look up to the right of where it says “First tab,” and you will see a second tab with the creative name “Second tab.”
If you are viewing the html output of this file, you can click on the different tabs to see what’s in them.
To stop new tabs from being created, close off the original ::: panel-tabset command with ::: at the end.
- In the source code file, you will see the ::: at the end of the ### Read up on tabsets tab.

Welcome to the second tab!

You can read up more about creating tabs at
- https://quarto.org/docs/interactive/layout.html#tabset-panel

If you are reading the source code file, the next line contains :::, which closes the tabsets.

Where are we?

Where are we? Continuous outcome zoomed in

Where are we?

CI’s and hypothesis tests for different scenarios:

$point estimate \pm z^{*} (o r t^{*}) \cdot S E, test stat = \frac{point estimate - null value}{S E}$

Day	Book	Population parameter	Symbol	Point estimate	Symbol	SE
10	5.1	Pop mean	$μ$	Sample mean	$\bar{x}$	$\frac{s}{\sqrt{n}}$
10	5.2	Pop mean of paired diff	$μ_{d}$ or $δ$	Sample mean of paired diff	${\bar{x}}_{d}$	$\frac{s_{d}}{\sqrt{n}}$
11	5.3	Diff in pop means	$μ_{1} - μ_{2}$	Diff in sample means	${\bar{x}}_{1} - {\bar{x}}_{2}$	???
12	8.1	Pop proportion	$p$	Sample prop	$\hat{p}$
12	8.2	Diff in pop proportions	$p_{1} - p_{2}$	Diff in sample proportions	${\hat{p}}_{1} - {\hat{p}}_{2}$

Goals for today

2-sample t-test (Section 5.3)

Statistical inference for difference in means from 2 independent samples
1. What are $H_{0}$ and $H_{a}$ ?
2. What is the SE for ${\bar{x}}_{1} - {\bar{x}}_{2}$ ?
3. Hypothesis test
4. Confidence Interval
5. Run test in R - using long vs. wide data
6. Satterthwaite’s df
7. Pooled SD

Power and sample size (4.3.4, 5.4, plus notes)

Critical values & rejection region
Type I & II errors
Power
How to calculate sample size needed for a study?

Examples of designs with two independent samples

Any study where participants are randomized to a control and treatment group
Study where create two groups based on whether they were exposed or not to some condition (can be observational)
Book: “Does treatment using embryonic stem cells (ESCs) help improve heart function following a heart attack?”
Book: “Is there evidence that newborns from mothers who smoke have a different average birth weight than newborns from mothers who do not smoke?”
The key is that the data from the two groups are independent of each other.

Steps in a Hypothesis Test

Set the level of significance $α$
Specify the null ( $H_{0}$ ) and alternative ( $H_{A}$ ) hypotheses
1. In symbols
2. In words
3. Alternative: one- or two-sided?
Calculate the test statistic.
Calculate the p-value based on the observed test statistic and its sampling distribution
Write a conclusion to the hypothesis test
1. Do we reject or fail to reject $H_{0}$ ?
2. Write a conclusion in the context of the problem

Does caffeine increase finger taps/min (on average)?

Study Design:

20 male college students students were trained to tap their fingers at a rapid rate.
Each then drank 2 cups of coffee (double-blind)
- Control group: decaf
- Caffeine group: ~ 200 mg caffeine
After 2 hours, students were tested.
Taps/minute recorded

Hand, David J.; Daly, Fergus; McConway, K.; Lunn, D. and Ostrowski, E. (1993). A handbook of small data sets. London, U.K.: Chapman and Hall.

Load the data from the csv file CaffeineTaps.csv
The code below is for when the data file is in a folder called data that is in your R project folder (your working directory)

CaffTaps <- read_csv(here::here("data", "CaffeineTaps.csv"))

Rows: 20 Columns: 2
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr (1): Group
dbl (1): Taps

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

glimpse(CaffTaps)

Rows: 20
Columns: 2
$ Taps  <dbl> 246, 248, 250, 252, 248, 250, 246, 248, 245, 250, 242, 245, 244,…
$ Group <chr> "Caffeine", "Caffeine", "Caffeine", "Caffeine", "Caffeine", "Caf…

EDA: Explore the finger taps data

Dotplot of taps/minute stratified by group

ggplot(CaffTaps, aes(x=Taps)) +
  geom_dotplot() +
  facet_wrap(vars(Group), ncol=1)

Bin width defaults to 1/30 of the range of the data. Pick better value with
`binwidth`.

Summary statistics stratified by group

# get_summary_stats() from rstatix package
sumstats <- CaffTaps %>% 
  group_by(Group) %>% 
  get_summary_stats(type = "mean_sd") 
sumstats %>% gt()

Group	variable	n	mean	sd
Caffeine	Taps	10	248.3	2.214
NoCaffeine	Taps	10	244.8	2.394

diff(sumstats$mean)

[1] -3.5

Step 2: Null & Alternative Hypotheses

Question: Is there evidence to support that drinking caffeine increases the number of finger taps/min?

Null and alternative hypotheses in words

Include as much context as possible

$H_{0}$ : The population difference in mean finger taps/min between the caffeine and control groups is …
$H_{A}$ : The population difference in mean finger taps/min between the caffeine and control groups is …

Null and alternative hypotheses in symbols

$\begin{aligned} H_{0} : & μ_{c a f f} - μ_{c t r l} = \\ H_{A} : & μ_{c a f f} - μ_{c t r l} \end{aligned}$

Step 3: Test statistic (part 1)

Recall that in general the test statistic has the form:

$test stat = \frac{point estimate - null value}{S E}$ Thus, for a two sample independent means test, we have:

$test statistic = \frac{{\bar{x}}_{1} - {\bar{x}}_{2} - 0}{S E_{{\bar{x}}_{1} - {\bar{x}}_{2}}}$

What is the formula for $S E_{{\bar{x}}_{1} - {\bar{x}}_{2}}$ ?
What is the probability distribution of the test statistic?
What assumptions need to be satisfied?

What distribution does ${\bar{X}}_{1} - {\bar{X}}_{2}$ have?

Let ${\bar{X}}_{1}$ and ${\bar{X}}_{2}$ be the means of random samples from two independent groups, with parameters shown in table:

	Group 1	Group 2
sample size	$n_{1}$	$n_{2}$
pop mean	$μ_{1}$	$μ_{2}$
pop sd	$σ_{1}$	$σ_{2}$

Some theoretical statistics:

If ${\bar{X}}_{1}$ and ${\bar{X}}_{2}$ are independent normal r.v.’s, then ${\bar{X}}_{1} - {\bar{X}}_{2}$ is also normal
What is the mean of ${\bar{X}}_{1} - {\bar{X}}_{2}$ ?

$E [{\bar{X}}_{1} - {\bar{X}}_{2}] = E [{\bar{X}}_{1}] - E [{\bar{X}}_{2}] = μ_{1} - μ_{2}$

What is the standard deviation of ${\bar{X}}_{1} - {\bar{X}}_{2}$ ?

$\begin{aligned} V a r ({\bar{X}}_{1} - {\bar{X}}_{2}) & = V a r ({\bar{X}}_{1}) + V a r ({\bar{X}}_{2}) = \frac{σ_{1}^{2}}{n_{1}} + \frac{σ_{2}^{2}}{n_{2}} \\ S D ({\bar{X}}_{1} - {\bar{X}}_{2}) & = \sqrt{\frac{σ_{1}^{2}}{n_{1}} + \frac{σ_{2}^{2}}{n_{2}}} \end{aligned}$

Step 3: Test statistic (part 2)

$t_{{\bar{x}}_{1} - {\bar{x}}_{2}} = \frac{{\bar{x}}_{1} - {\bar{x}}_{2} - 0}{\sqrt{\frac{s_{1}^{2}}{n_{1}} + \frac{s_{2}^{2}}{n_{2}}}}$

${\bar{x}}_{1}, {\bar{x}}_{2}$ are the sample means
$μ_{0} = 0$ is the mean value specified in $H_{0}$
$s_{1}, s_{2}$ are the sample SD’s
$n_{1}, n_{2}$ are the sample sizes

Statistical theory tells us that $t_{{\bar{x}}_{1} - {\bar{x}}_{2}}$ follows a student’s t-distribution with
- $d f \approx$ smaller of $n_{1} - 1$ and $n_{2} - 1$
- this is a conservative estimate (smaller than actual $d f$ )

Assumptions:

Independent observations & samples
- The observations were collected independently.
- In particular, the observations from the two groups were not paired in any meaningful way.
Approximately normal samples or big n’s
- The distributions of the samples should be approximately normal
- or both their sample sizes should be at least 30.

Step 3: Test statistic (part 3)

Group	variable	n	mean	sd
Caffeine	Taps	10	248.3	2.214
NoCaffeine	Taps	10	244.8	2.394

$test statistic = t_{{\bar{x}}_{1} - {\bar{x}}_{2}} = \frac{{\bar{x}}_{1} - {\bar{x}}_{2} - 0}{\sqrt{\frac{s_{1}^{2}}{n_{1}} + \frac{s_{2}^{2}}{n_{2}}}}$

Based on the value of the test statistic, do you think we are going to reject or fail to reject $H_{0}$ ?

Step “3b”: Assumptions satisfied?

Assumptions:

Independent observations & samples
- The observations were collected independently.
- In particular, the observations from the two groups were not paired in any meaningful way.
Approximately normal samples or big n’s
- The distributions of the samples should be approximately normal
- or both their sample sizes should be at least 30.

Bin width defaults to 1/30 of the range of the data. Pick better value with
`binwidth`.

Step 4: p-value

The p-value is the probability of obtaining a test statistic just as extreme or more extreme than the observed test statistic assuming the null hypothesis $H_{0}$ is true.

Calculate the p-value:

Step 5: Conclusion to hypothesis test

$\begin{aligned} H_{0} : & μ_{c a f f} - μ_{c t r l} = 0 \\ H_{A} : & μ_{c a f f} - μ_{c t r l} > 0 \end{aligned}$

Recall the $p$ -value = 0.00397
Use $α$ = 0.05.
Do we reject or fail to reject $H_{0}$ ?

Conclusion statement:

Stats class conclusion
- There is sufficient evidence that the (population) difference in mean finger taps/min with vs. without caffeine is greater than 0 ( $p$ -value = 0.004).
More realistic manuscript conclusion:
- The mean finger taps/min were 244.8 (SD = 2.4) and 248.3 (SD = 2.2) for the control and caffeine groups, and the increase of 3.5 taps/min was statistically discrenible ( $p$ -value = 0.004).

95% CI for the mean difference in cholesterol levels

Group	variable	n	mean	sd
Caffeine	Taps	10	248.3	2.214
NoCaffeine	Taps	10	244.8	2.394

CI for $μ_{c a f f} - μ_{c t r l}$ :

${\bar{x}}_{c a f f} - {\bar{x}}_{c t r l} \pm t^{*} \cdot \sqrt{\frac{s_{c a f f}^{2}}{n_{c a f f}} + \frac{s_{c t r l}^{2}}{n_{c t r l}}}$

Interpretation:
We are 95% confident that the (population) difference in mean finger taps/min between the caffeine and control groups is between 1.167 mg/dL and 5.833 mg/dL.

Based on the CI, is there evidence that drinking caffeine made a difference in finger taps/min? Why or why not?

R: 2-sample t-test (with long data)

The CaffTaps data are in a long format, meaning that
- all of the outcome values are in one column and
- another column indicates which group the values are from
This is a common format for data from multiple samples, especially if the sample sizes are different.

(Taps_2ttest <- t.test(formula = Taps ~ Group, 
                       alternative = "greater", 
                       data = CaffTaps))


    Welch Two Sample t-test

data:  Taps by Group
t = 3.3942, df = 17.89, p-value = 0.001628
alternative hypothesis: true difference in means between group Caffeine and group NoCaffeine is greater than 0
95 percent confidence interval:
 1.711272      Inf
sample estimates:
  mean in group Caffeine mean in group NoCaffeine 
                   248.3                    244.8

`tidy` the `t.test` output

# use tidy command from broom package for briefer output that's a tibble
tidy(Taps_2ttest) %>% gt()

estimate	estimate1	estimate2	statistic	p.value	parameter	conf.low	conf.high	method	alternative
3.5	248.3	244.8	3.394168	0.001627703	17.89012	1.711272	Inf	Welch Two Sample t-test	greater

Pull the p-value:

tidy(Taps_2ttest)$p.value  # we can pull specific values from the tidy output

[1] 0.001627703

R: 2-sample t-test (with wide data)

# make CaffTaps data wide: pivot_wider needs an ID column so that it 
# knows how to "match" values from the Caffeine and NoCaffeine groups
CaffTaps_wide <- CaffTaps %>% 
  mutate(id = rep(1:10, 2)) %>% #  "fake" IDs for pivot_wider step
  pivot_wider(names_from = "Group",
              values_from = "Taps")

glimpse(CaffTaps_wide)

Rows: 10
Columns: 3
$ id         <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
$ Caffeine   <dbl> 246, 248, 250, 252, 248, 250, 246, 248, 245, 250
$ NoCaffeine <dbl> 242, 245, 244, 248, 247, 248, 242, 244, 246, 242

t.test(x = CaffTaps_wide$Caffeine, y = CaffTaps_wide$NoCaffeine, alternative = "greater") %>% 
  tidy() %>% gt()

estimate	estimate1	estimate2	statistic	p.value	parameter	conf.low	conf.high	method	alternative
3.5	248.3	244.8	3.394168	0.001627703	17.89012	1.711272	Inf	Welch Two Sample t-test	greater

Why are the df’s in the R output different?

From many slides ago:

Statistical theory tells us that $t_{{\bar{x}}_{1} - {\bar{x}}_{2}}$ follows a student’s t-distribution with
- $d f \approx$ smaller of $n_{1} - 1$ and $n_{2} - 1$
- this is a conservative estimate (smaller than actual $d f$ )

The actual degrees of freedom are calculated using Satterthwaite’s method:

$ν = \frac{[(s_{1}^{2} / n_{1}) + (s_{2}^{2} / n_{2})]^{2}}{(s_{1}^{2} / n_{1})^{2} / (n_{1} - 1) + (s_{2}^{2} / n_{2})^{2} / (n_{2} - 1)} = \frac{[S E_{1}^{2} + S E_{2}^{2}]^{2}}{S E_{1}^{4} / d f_{1} + S E_{2}^{4} / d f_{2}}$

Verify the p-value in the R output using $ν$ = 17.89012:

pt(3.3942, df = 17.89012, lower.tail = FALSE)

[1] 0.001627588

Pooled standard deviation estimate

Sometimes we have reasons to believe that the population SD’s from the two groups are equal, such as when randomizing participants to two groups

In this case we can use a pooled SD:

$s_{p o o l e d}^{2} = \frac{s_{1}^{2} (n_{1} - 1) + s_{2}^{2} (n_{2} - 1)}{n_{1} + n_{2} - 2}$

$n_{1}$ , $n_{2}$ are the sample sizes, and
$s_{1}$ , $s_{2}$ are the sample standard deviations
of the two groups

We use the pooled SD instead of $s_{1}^{2}$ and $s_{2}^{2}$ when calculating the standard error

$S E = \sqrt{\frac{s_{p o o l e d}^{2}}{n_{1}} + \frac{s_{p o o l e d}^{2}}{n_{2}}} = s_{p o o l e d} \sqrt{\frac{1}{n_{1}} + \frac{1}{n_{2}}}$

Test statistic with pooled SD:

$t_{{\bar{x}}_{1} - {\bar{x}}_{2}} = \frac{{\bar{x}}_{1} - {\bar{x}}_{2} - 0}{s_{p o o l e d} \sqrt{\frac{1}{n_{1}} + \frac{1}{n_{2}}}}$

CI with pooled SD:

$({\bar{x}}_{1} - {\bar{x}}_{2}) \pm t^{⋆} \cdot s_{p o o l e d} \sqrt{\frac{1}{n_{1}} + \frac{1}{n_{2}}}$

The $t$ distribution degrees of freedom are now:

$d f = (n_{1} - 1) + (n_{2} - 1) = n_{1} + n_{2} - 2.$

R: 2-sample t-test with pooled SD

# t-test with pooled SD
t.test(formula = Taps ~ Group, 
       alternative = "greater", 
       var.equal = TRUE,  # pooled SD 
       data = CaffTaps) %>% 
  tidy() %>% 
  gt()

estimate	estimate1	estimate2	statistic	p.value	parameter	conf.low	conf.high	method	alternative
3.5	248.3	244.8	3.394168	0.001616497	18	1.711867	Inf	Two Sample t-test	greater

# t-test without pooled SD
t.test(formula = Taps ~ Group, 
       alternative = "greater", 
       var.equal = FALSE,  # default, NOT pooled SD 
       data = CaffTaps) %>% 
  tidy() %>% 
  gt()

estimate	estimate1	estimate2	statistic	p.value	parameter	conf.low	conf.high	method	alternative
3.5	248.3	244.8	3.394168	0.001627703	17.89012	1.711272	Inf	Welch Two Sample t-test	greater

Similar output in this case - why??

What’s next?

CI’s and hypothesis tests for different scenarios:

$point estimate \pm z^{*} (o r t^{*}) \cdot S E, test stat = \frac{point estimate - null value}{S E}$

Day	Book	Population parameter	Symbol	Point estimate	Symbol	SE
10	5.1	Pop mean	$μ$	Sample mean	$\bar{x}$	$\frac{s}{\sqrt{n}}$
10	5.2	Pop mean of paired diff	$μ_{d}$ or $δ$	Sample mean of paired diff	${\bar{x}}_{d}$	$\frac{s_{d}}{\sqrt{n}}$
11	5.3	Diff in pop means	$μ_{1} - μ_{2}$	Diff in sample means	${\bar{x}}_{1} - {\bar{x}}_{2}$	$\sqrt{\frac{s_{1}^{2}}{n_{1}} + \frac{s_{2}^{2}}{n_{2}}}$ or pooled
12	8.1	Pop proportion	$p$	Sample prop	$\hat{p}$	???
12	8.2	Diff in pop proportions	$p_{1} - p_{2}$	Diff in sample proportions	${\hat{p}}_{1} - {\hat{p}}_{2}$	???

Power and sample size calculations

Critical values & rejection region
Type I & II errors
Power
How to calculate sample size needed for a study?

Materials are from
- Section 4.3.4 Decision errors
- Section 5.4 Power calculations for a difference of means
- plus notes

Critical values

Critical values are the cutoff values that determine whether a test statistic is statistically significant or not.
If a test statistic is greater in absolute value than the critical value, we reject $H_{0}$

Critical values are determined by
- the significance level $α$ ,
- whether a test is 1- or 2-sided, &
- the probability distribution being used to calculate the p-value (such as normal or t-distribution).
The critical values in the figure should look very familiar!
- Where have we used these before?

How can we calculate the critical values using R?

Rejection region

If the absolute value of the test statistic is greater than the critical value, we reject $H_{0}$
- In this case the test statistic is in the rejection region.
- Otherwise it’s in the nonrejection region.

What do rejection regions look like for 1-sided tests?

Hypothesis Testing “Errors”

Justice system analogy

Type I and Type II Errors - Making Mistakes in the Justice System

Type I & II Errors

$α$ = probability of making a Type I error
- This is the significance level (usually 0.05)
- Set before study starts
$β$ = probability of making a Type II error
Ideally we want
- small Type I & II errors and
- big power

Applet for visualizing Type I & II errors and power: https://rpsychologist.com/d3/NHST/

Relationship between Type I & II errors

Type I vs. Type II error
- Decreasing P(Type I error) leads to
  - increasing P(Type II error)
- We typically keep P(Type I error) = $α$ set to 0.05

From the applet at https://rpsychologist.com/d3/NHST/

Relationship between Type II errors and power

Power = P(correctly rejecting the null hypothesis)

Power is also called the
- true positive rate,
- probability of detection, or
- the sensitivity of a test

Power vs. Type II error
- Power = 1 - P(Type II error) = 1 - $β$
- Thus as $β$ = P(Type II error) decreases, the power increases
- P(Type II error) decreases as the mean of the alternative population shifts further away from the mean of the null population (effect size gets bigger).
- Typically want at least 80% power; 90% power is good

Example calculating power

Suppose the mean of the null population is 0 ( $H_{0} : μ = 0$ ) with standard error 1
Find the power of a 2-sided test if the actual $μ = 3$ , assuming the SE doesn’t change.

Power = $P ($ Reject $H_{0}$ when alternative pop is $N (3, 1))$
When $α$ = 0.05, we reject $H_{0}$ when the test statistic z is at least 1.96
Thus for $X \sim N (3, 1)$ we need to calculate $P (X \leq - 1.96) + P (X \geq 1.96)$ :

# left tail + right tail:
pnorm(-1.96, mean=3, sd=1, lower.tail=TRUE) + pnorm(1.96, mean=3, sd=1, lower.tail=FALSE)

[1] 0.8508304

The left tail probability pnorm(-1.96, mean=3, sd=1, lower.tail=TRUE) is essentially 0 in this case.

Note that this power calculation specified the value of the SE instead of the standard deviation and sample size $n$ individually.

Sample size calculation for testing one mean

Recall in our body temperature example that $μ_{0} = 98.6$ °F and $\bar{x} = 98.25$ °F.
- The p-value from the hypothesis test was highly significant (very small).
- What would the sample size $n$ need to be for 80% power?
Calculate $n$ ,
- given $α$ , power ( $1 - β$ ), “true” alternative mean $μ$ , and null $μ_{0}$ ,
- assuming the test statistic is normal (instead of t-distribution):

$n = {(s \frac{z_{1 - α / 2} + z_{1 - β}}{μ - μ_{0}})}^{2}$

mu <- 98.25
mu0 <- 98.6
sd <- 0.73
alpha <- 0.05
beta <- 0.20
n <- (sd*(qnorm(1-alpha/2) + qnorm(1-beta)) / (mu-mu0))^2
n

[1] 34.14423

ceiling(n)  # always round UP to the next highest integer

[1] 35

We would only need a sample size of 35 for 80% power!
However, this is an under-estimate since we used the normal instead of t-distribution.

See http://powerandsamplesize.com/Calculators/Test-1-Mean/1-Sample-Equality.

Power calculation for testing one mean

Conversely, we can calculate how much power we had in our body temperature one-sample test, given the sample size of 130.

Calculate power,
- given $α$ , $n$ , “true” alternative mean $μ$ , and null $μ_{0}$ ,
- assuming the test statistic is normal (instead of t-distribution)

$1 - β = Φ (z - z_{1 - α / 2}) + Φ (- z - z_{1 - α / 2}), where z = \frac{μ - μ_{0}}{s / \sqrt{n}}$

$Φ$ is the probability for a standard normal distribution

mu <- 98.25; mu0 <- 98.6; sd <- 0.73; alpha <- 0.05; n <- 130
(z <- (mu-mu0) / (sd/sqrt(n)) )

[1] -5.466595

Power <- pnorm(z-qnorm(1-alpha/2)) + pnorm(-z-qnorm(1-alpha/2))
Power

[1] 0.9997731

If the population mean is 98.2 instead of 98.6, we have a 99.98% chance of correctly rejecting $H_{0}$ when the sample size is 130.

R package `pwr` for power analyses

Use pwr.t.test for both one- and two-sample t-tests.
Specify all parameters except for the one being solved for.

pwr.t.test(n = NULL, d = NULL, sig.level = 0.05, power = NULL,
type = c("two.sample", "one.sample", "paired"),
alternative = c("two.sided", "less", "greater"))

d is Cohen’s d effect size: small = 0.2, medium = 0.5, large = 0.8

One-sample test (or paired t-test):

$d = \frac{μ - μ_{0}}{s}$

Two-sample test (independent):

$d = \frac{{\bar{x}}_{1} - {\bar{x}}_{2}}{s_{p o o l e d}}$

${\bar{x}}_{1} - {\bar{x}}_{2}$ is the difference in means between the two groups that one would want to be able to detect as being significant,
$s_{p o o l e d}$ is the pooled SD between the two groups - often assume have same sd in each group
R package pwr for basic statistical tests
- https://cran.r-project.org/web/packages/pwr/vignettes/pwr-vignette.html

`pwr`: sample size for one mean test

pwr.t.test(n = NULL, d = NULL, sig.level = 0.05, power = NULL,
type = c("two.sample", "one.sample", "paired"), alternative = c("two.sided", "less", "greater"))

d is Cohen’s d effect size: $d = \frac{μ - μ_{0}}{s}$

Specify all parameters except for the sample size:

library(pwr)
t.n <- pwr.t.test(
  d = (98.6-98.25)/0.73, 
  sig.level = 0.05, 
  power = 0.80, 
  type = "one.sample")

t.n


     One-sample t test power calculation 

              n = 36.11196
              d = 0.4794521
      sig.level = 0.05
          power = 0.8
    alternative = two.sided

plot(t.n)

`pwr`: power for one mean test

pwr.t.test(n = NULL, d = NULL, sig.level = 0.05, power = NULL,
type = c("two.sample", "one.sample", "paired"), alternative = c("two.sided", "less", "greater"))

d is Cohen’s d effect size: $d = \frac{μ - μ_{0}}{s}$

Specify all parameters except for the power:

t.power <- pwr.t.test(
  d = (98.6-98.25)/0.73, 
  sig.level = 0.05, 
  # power = 0.80, 
  n = 130,
  type = "one.sample")

t.power


     One-sample t test power calculation 

              n = 130
              d = 0.4794521
      sig.level = 0.05
          power = 0.9997354
    alternative = two.sided

plot(t.power)

`pwr`: Two-sample t-test: sample size

pwr.t.test(n = NULL, d = NULL, sig.level = 0.05, power = NULL,
type = c("two.sample", "one.sample", "paired"), alternative = c("two.sided", "less", "greater"))

d is Cohen’s d effect size: $d = \frac{{\bar{x}}_{1} - {\bar{x}}_{2}}{s_{p o o l e d}}$

Example: Suppose the data collected for the caffeine taps study were pilot day for a larger study. Investigators want to know what sample size they would need to detect a 2 point difference between the two groups. Assume the SD in both groups is 2.3.

Specify all parameters except for the sample size:

t2.n <- pwr.t.test(
  d = 2/2.3, 
  sig.level = 0.05, 
  power = 0.80, 
  type = "two.sample") 

t2.n


     Two-sample t test power calculation 

              n = 21.76365
              d = 0.8695652
      sig.level = 0.05
          power = 0.8
    alternative = two.sided

NOTE: n is number in *each* group

plot(t2.n)

`pwr`: Two-sample t-test: power

pwr.t.test(n = NULL, d = NULL, sig.level = 0.05, power = NULL,
type = c("two.sample", "one.sample", "paired"), alternative = c("two.sided", "less", "greater"))

d is Cohen’s d effect size: $d = \frac{{\bar{x}}_{1} - {\bar{x}}_{2}}{s_{p o o l e d}}$

Specify all parameters except for the power:

t2.power <- pwr.t.test(
  d = 2/2.3, 
  sig.level = 0.05, 
  # power = 0.80, 
  n = 22,
  type = "two.sample") 

t2.power


     Two-sample t test power calculation 

              n = 22
              d = 0.8695652
      sig.level = 0.05
          power = 0.8044288
    alternative = two.sided

NOTE: n is number in *each* group

plot(t2.power)

What information do we need for a power (or sample size) calculation?

There are 4 pieces of information:

Level of significance $α$
- Usually fixed to 0.05
Power
- Ideally at least 0.80
Sample size
Effect size (expected change)

Given any 3 pieces of information, we can solve for the 4th.

pwr.t.test(
  d = (98.6-98.25)/0.73,
  sig.level = 0.05, 
  # power = 0.80, 
  n=130,
  type = "one.sample")


     One-sample t test power calculation 

              n = 130
              d = 0.4794521
      sig.level = 0.05
          power = 0.9997354
    alternative = two.sided

More software for power and sample size calculations: PASS

PASS is a very powerful (& expensive) software that does power and sample size calculations for many advanced statistical modeling techniques.
- Even if you don’t have access to PASS, their documentation is very good and free online.
- Documentation includes formulas and references.
- PASS documentation for powering means
  - One mean, paired means, two independent means
One-sample t-test documentation: https://www.ncss.com/wp-content/themes/ncss/pdf/Procedures/PASS/One-Sample_T-Tests.pdf

OCTRI-BERD power & sample size presentations

Power and Sample Size 101
- Presented by Meike Niederhausen; April 13, 2023
- Slides: http://bit.ly/PSS101-BERD-April2023
- Recording
Power and Sample Size for Clinical Trials: An Introduction
- Presented by Yiyi Chen; Feb 18, 2021
- Slides: http://bit.ly/PSS-ClinicalTrials
- Recording
Planning a Study with Power and Sample Size Considerations in Mind
- Presented by David Yanez; May 29, 2019
- Slides
- Recording
Power and Sample Size Simulations in R
- Presented by Robin Baudier; Sept 21, 2023
- Slides
- Recording

MoRitz’s tip of the day

Where are we?

Where are we? Continuous outcome zoomed in

Where are we?

Goals for today

2-sample t-test (Section 5.3)

Power and sample size (4.3.4, 5.4, plus notes)

Examples of designs with two independent samples

Steps in a Hypothesis Test

Does caffeine increase finger taps/min (on average)?

EDA: Explore the finger taps data

Step 2: Null & Alternative Hypotheses

Step 3: Test statistic (part 1)

What distribution does X¯1−X¯2 have?

Step 3: Test statistic (part 2)

Step 3: Test statistic (part 3)

Step “3b”: Assumptions satisfied?

Step 4: p-value

Step 5: Conclusion to hypothesis test

95% CI for the mean difference in cholesterol levels

R: 2-sample t-test (with long data)

tidy the t.test output

R: 2-sample t-test (with wide data)

Why are the df’s in the R output different?

Pooled standard deviation estimate

R: 2-sample t-test with pooled SD

What’s next?

Power and sample size calculations

Critical values

Rejection region

Hypothesis Testing “Errors”

Justice system analogy

Type I & II Errors

Relationship between Type I & II errors

Relationship between Type II errors and power

Example calculating power

Sample size calculation for testing one mean

Power calculation for testing one mean

R package pwr for power analyses

pwr: sample size for one mean test

pwr: power for one mean test

pwr: Two-sample t-test: sample size

pwr: Two-sample t-test: power

What information do we need for a power (or sample size) calculation?

More software for power and sample size calculations: PASS

OCTRI-BERD power & sample size presentations

What distribution does ${\bar{X}}_{1} - {\bar{X}}_{2}$ have?

`tidy` the `t.test` output

R package `pwr` for power analyses

`pwr`: sample size for one mean test

`pwr`: power for one mean test

`pwr`: Two-sample t-test: sample size

`pwr`: Two-sample t-test: power