Mini-homework 7 – Who busts the Mythbusters?

Instructions

Obtain the GitHub repository you will use to complete the mini-homework, which contains a starter file named mythbusters.Rmd. This mini-homework will guide you through examples demonstrating how to conduct hypothesis tests and calculate confidence intervals using the infer package. Read the About the dataset section to get some background information on the dataset that you’ll be working with. Each of the below exercises are to be completed in the provided spaces within your starter file mythbusters.Rmd. Then, when you’re ready to submit, follow the directions in the How to submit section below.

The dataset we are using in this mini-homework comes from an experiment shown on an episode of the Mythbusters television show1, a science entertainment TV program that aired on the Discovery Channel. The show’s build team, Kari Byron, Scottie Chapman, and Tory Belleci, conducted the experiment in order to test the following hypothesis,

It becomes more likely for a person to yawn if another person near them yawns.

To conduct the experiment, the team drove 50 volunteers, one at a time, to a flea market in a van, and one of the team members, would either yawn (treatment) or not yawn (no treatment) during the ride. It was randomly determined ahead of time whether or not Kari yawned. Therefore, the 50 people were split into two randomly assigned groups:

• Treatment group: 34 people where a person near them yawned

• Control group: 16 people where there wasn’t a person yawning near them

The results of the experiment are presented in the following table,

group yawns total fraction_yawned
Control 4 16 0.2500
Treatment 10 34 0.2941

They concluded that the myth was confirmed because there was a 4% increase in yawns between the control and treatment groups. However, their analysis did not appear to include a formal hypothesis test, nor did the team discuss how likely it was that random chance alone could generate their results. Without this kind of analysis, we can not meaningfully determine whether the data supports the conclusion that yawning is contagious.

Our main activity in this mini-homework will be fact-checking the Mythbusters by carrying out a formal hypothesis test using the infer package. This will allow us to assess how likely it is that a random chance model could generate the observed experimental results and whether or not the collected data allow us to infer the same conclusion as the Mythbusters build team.

The dataset yawn contains 3 variables and 50 rows,

Variable Description
subject_id Integer uniquely identifying an experiment volunteer
yawn Whether or not a volunteer yawned, yes if he/she did, no if he/she did not
group Specifies if the volunteer was part of the Control group or the Treatment group

Exercises

1. Whenever we conduct an experiment, we need to know what variable we will change between the treatment and control groups, and what the kinds of outcomes we will be measuring. Look again as the three variables for this experiment. Which one is the explanatory variable and which one is the response variable? What value in the response value should be classified as a success?

2. To conduct the hypothesis test, we need to simulate the null distribution. The null distribution represents the range of outcomes for this experiment if yawning has no effect on the treatment group and is due to random chance alone. To generate a null distribution we need to know how the observed experimental result was calculated, as we will need to repeat this same calculation during our simulations. Review the results table in the About the dataset section again. What quantity should we use or compute in order to build the null distribution? Choose the correct answer below.

1. The average number of yawns in the treatment group

2. The average value of fraction_yawned in the treatment group

3. The average difference in yawns between the treatment and control groups

4. The average difference in fraction_yawned between the treatment and control groups

3. Running a hypothesis test for this dataset requires us to fill in the following code template that contains four functions from the infer package,

yawn_null <- yawn %>%
specify(<variable1> ~ <variable2>, success = "...") %>%
hypothesize(null = "...") %>%
generate(reps = ...,  type = "...") %>%
calculate(stat = "...",  order = combine("...", "..."))

We will step through filling in this template over the next few questions.

We start with the function specify(), which we see has the following structure,

specify(<variable1> ~ <variable2>, success = "...")

<variable1>, <variable2>, and “…” are placeholders. Using your answers from Exercise 1, figure out what to fill in for these three placeholders. You only need to write out the specify() function.

Hint

The relationship between the response and explanatory variables is written using a special formula syntax, <variable1> ~ <variable2>. The word on the left side of the tilde, <variable1>, is the response variable. The word on the right side of the tilde, <variable2>, is the explanatory variable.

4. The next step in our code template for generating a null hypothesis is the hypothesize() function,

hypothesize(null = "...")

This function only takes one input, and you have two choices for what to use. What input should we use?

Values for null

• “independence”: used when the explanatory and response variables both refer to columns in the dataset.

• “point”: used when we want to compare a column against a reported point estimate (i.e. mean or median).
5. The third step in our code template is the generate() function,

generate(reps = ...,  type = "...")

We use the reps input to set how many simulation trials we want to run to generate the null distribution. This input should be set equal to an integer value. The more trials we run, the more accurate our simulated range of outcomes will be, although it will also mean it will take our code longer to run. The type keyword is used to choose the type of simulation that we want to run.

Generating a null distribution for this dataset means we want to see what our outcomes will look like if the response variable and explanatory variable have no relationship to each other. If we want to run 10,000 simulations to create the null distribution for this dataset, what should I replace the placeholders with?

Values for type

There are three choices for type,

• “bootstrap”: Run the simulation by randomly sampling rows with replacement from the dataset. The relationship between different columns in each row is preserved.

• “permute”: Run the simulation by randomly shuffling the order of the cells in the explanatory and response variables. This random shuffling procedure is done separately in each column, so the relationship between different columns in each row is not preserved.

• “simulate”: Run the simulation by flipping a weighted coin, meaning that heads or tails may not be equally probable (i.e. the coin may be unfair). This coin, being virtual, may have more than two sides, meaning there are several discrete outcomes that each have certain probabilities of happening. This type is typically needed if you are using null = “point” in hypothesize() and your response variable is categorical.
6. The last step in our code template is the calculate() function,

calculate(stat = "...",  order = combine("...", "..."))

The first input stat is where we specify what exactly it is that we’re calculating, which must mirror how the observed experimental result was calculated. The second input is order and specifies the order of subtraction between the two groups listed under the explanatory variable.

Based on your response to question 2 and the above information, what should I replace the placeholders with?

Values for stat

The available choices for stat are as follows,

• “mean”: For simulations based on a single numerical variable. Compute the mean of each simulation trial.

• “median”: For simulations based on a single numerical variable. Find the median of each simulation trial.

• “sum”: For simulations based on a single numerical variable. Compute the sum of each simulation trial.

• “sd”: For simulations based on a single numerical variable. Compute the standard deviation of each simulation trial.

• “prop”: For simulations based on a single categorical variable. Count the number of successes and convert this to a fraction (proportion).

• “count”: For simulations based on a single categorical variable. Count the number of successes.

• “diff in means”: Calculate the mean of the numerical response variable for the two groups defined in the categorical explanatory variable. Subtract these two means.

• “diff in medians”: Find the median of the numerical response variable for the two groups defined in the categorical explanatory variable. Subtract these two medians.

• “diff in props”: Calculate the number of successes in the categorical response variable and convert this to a fraction (proportion) in each of the two groups defined in the categorical explanatory variable. Subtract these two fractions.

• “Chisq”: Calculate the chi-squared statistic for a response and explanatory variable that are both categorical and contain two or more groups each.

• “F”: Calculate the F statistic for a numerical response variable and categorical explanatory variable with two or more groups.

• “t”: Calculate the two-sample t-test statistic for a numerical response variable and categorical explanatory variable with two groups. This is similar to the “diff in means” calculation.

• “z”: Calculate the two-sample z statistic for a numerical response variable and categorical explanatory variable with two groups. It is assumed that the number of samples in your dataset is large, otherwise the “t” test is preferred. This is similar to the “diff in means” calculation.

• “slope”: Calculate the slope using linear regression for a response and explanatory variable that are both numerical.

• “correlation”: Calculate the correlation coefficient for a response and explanatory variable that are both numerical.
7. Take your answers to questions 3 through 6 and combine them to fill in the template,

yawn_null <- yawn %>%
specify(formula = <variable1> ~ <variable2>, success = "...") %>%
hypothesize(null = "...") %>%
generate(reps = ...,  type = "...") %>%
calculate(stat = "...",  order = combine("...", "..."))

This will simulate the null distribution for the yawn experiment. This may take a few seconds to run in RStudio.

8. Now that we have a simulated null distribution, let’s quantify how likely it is that a random chance model would generate the Mythbuster’s experimental result. We will do this by computing the p-value of a one-sided hypothesis test, which is just the probability that the result of a simulation trial is the same or more extreme than the experimental result. We will also assume a significance level of $$\alpha = 0.05$$. If our p-value is less than $$\alpha = 0.05$$, then we will conclude that we have sufficient evidence to reject the null hypothesis in favor of the alternative hypothesis.

First, let’s use infer to compute the experimental result for comparison with the null distribution. Fill in the template below with the values you determined for specify() and calculate(),

yawn_obs_stat <- yawn %>%
specify(formula = <variable1> ~ <variable2>, success = "...") %>%
calculate(stat = "...",  order = combine("...", "..."))

Notice that we just took the template from Exercise 7 and removed the hypothesize() and generate() lines.

Next, if we wanted to compute the p-value manually, we would need to count how many simulations gave us a result that was the same or more extreme than the experimental result and then divide by 10,000, which is the total number of simulations we ran. Lucky for us, the infer package has a convenience function, get_p_value(), for calculating p-values so that we don’t have to do it manually. To compute the p-value of a one-sided hypothesis test, we run,

yawn_null %>%
get_p_value(obs_stat = yawn_obs_stat, direction = "right")

To visualize the range of outcomes in the null distribution that are the same or more extreme than yawn_obs_stat, we run,

yawn_null %>%
visualize() +
shade_p_value(obs_stat = yawn_obs_stat, direction = "right")

Based on these results and a significance level of $$\alpha = 0.05$$, conclude whether or not we have sufficient evidence to reject the null hypothesis.

9. It’s often prudent to run a two-sided hypothesis test instead of the one-sided version, as it is more strict and hence stronger evidence is needed before the $$\alpha = 0.05$$ significance threshold is crossed. Switching from a one-sided test to a two-sided test simply requires that you replace direction = “right” with direction = “both” in get_p_value() and shade_p_value(). Do this and compute the p-value and visualization for the two-sided hypothesis test. Assuming a significance level of $$\alpha = 0.05$$, do we have sufficient evidence to reject the null hypothesis?

10. Our final piece of analysis will be to compute a 95% confidence interval for the experimental result. The 95% confidence interval represents a plausible range of values for the experimental result based on the variability of the dataset. Computing this interval requires us to run a bootstrap simulation. To do this, copy and paste your code from Exercise 7, delete the hypothesize() line, and change the type input in generate() to type = “bootstrap”. A template is shown below,

yawn_bootstrap <- yawn %>%
specify(formula = <variable1> ~ <variable2>, success = "...") %>%
generate(reps = ..., type = "bootstrap") %>%
calculate(stat = "...", order = combine("...", "..."))
11. We obtain the 95% confidence interval by running the following code,

yawn_ci <- yawn_bootstrap %>%
get_confidence_interval()

To visualize the 95% confidence interval, we run the following code,

yawn_bootstrap %>%
visualize() +
shade_confidence_interval(yawn_ci)

After computing and visualizing the confidence interval, choose the statement that correctly interprets what it means,

1. People in this sample, on average, yawn approximately 23% less to 30% more of the time when someone near them yawns.

2. People will, on average, yawn 23% less to 30% more when someone near them yawns.

3. A randomly chosen person yawns 23% less to 30% more when someone near them yawns.

4. 95% of people yawn 23% less to 30% more when someone near them yawns.