Chapter 6 Hypothesis Testing and Interval Estimation; Answering Research Questions

6.1 Computing Corner

We will learn the basisc for hypothesis testing in R.

6.1.1 Probability Distributions in R

For every probability distribution there are four commands. These command for each distribution are prepended by a letter to indicate the functionality.

  • “d” returns the height of the probability “d”ensity function
  • “p” returns the cummulative density function or the “p”robability of being being between two values of the random variable.
  • “q” returns the inverse density function or the value of the random variable (“q”uantile) given a probability.
  • “r” returns a “r”andomly generated number from the probability distribution

The distributions you are most likely to encounter in econometrics are the normal (norm), the F distribution (f), the chi-square distribution (chisq), and Student’s t-distribution (t). Others include the uniform (unif), binomial (binom), Poisson (pois), etc. Use of the help tab in the Files/Plots/Packages/Help pane or use of args will list the arguments necessary to extract value for each distribution.

6.1.2 Critical Values in R

To calculate critical values to perform a hypothesis test use the “q” version of the probability distribution. This will return the quantile for the given probability. The probability under the curve will be cummulative from \(-\infty\) to the quantile returned. The “q” version will return the critical value for a one-tail test. Suppose you’d like to test the following hypothesis about \(\mu\):

\[H_0:\mu=0\] \[H_1:\mu<0\] at the \(\alpha=.05\) level of significance. To calculate the critical t-stastic call qt(p = .05, df = n-1). You know from args(qt) the default value of the argument lower.tail is TRUE. Suppose, instead, you’d like to test the following hypothesis about \(\mu\)

\[H_0:\mu=0\] \[H_1:\mu>0\] at the \(\alpha = .10\) level of significance. You can call qt in two ways:

  1. qt(p = .10, df = n-1, lower.tail = FALSE) or
  2. qt(p = .90, df = n-1)

Finally, suppose you’d like to test the following hypothesis about \(\mu\)

\[H_0:\mu=0\] \[H_1:\mu\ne0\] at the \(\alpha=.01\) level of significance. Since the t-distribution is symmetric you can use the lower tail or upper tail value and -1 times it. You can call qt in three ways:

  1. qt(p = .005, df = n-1) or
  2. qt(p = .005, df = n-1, lower.tail = FALSE) or
  3. qt(p = .995, df = n-1)

You can find crtical values for the normal, F, and \(\chi^2\) distributions with similar function calls. p values in R

To calculate p values in R, use the “p” version of the distribution call. So suppose we test the following hypothesis:

\[H_0:\sigma_1^2=\sigma_2^2\] \[H_0:\sigma_1^2\ne\sigma_2^2\]

at the \(\alpha=.05\) level of significance. We could use an F test of the form


where \(s_x^2\) and \(s_y^2\) are the sample variances with n-1 and m-1 degrees of freedom. To calculate the p value, call pf(F, n-1, m-1) where F is the value calculated above.

6.1.3 Confidence Intervals for OLS estimates

In addition to confint(), confint_tidy() from the broom package will create a tibble of the low and high values for each estimate. The default level of confidence is 95%.

6.1.4 Power Curves

The power curve represents the probability of making Type II error under alternative null hypotheses. We can generate the power of the test with the pwr.norm.test(d = NULL, n = NULL, sig.level =.05, power = NULL, alternative = c("two-sided", "less", "greater")) call from the pwr package and plot the power with ggplot. To estimate the power we need the effect size \(d = \beta_i - \beta\) where \(\beta\) is the hypothesised paramater. We will use \[H_0: \beta = 0\] \[H_1: \beta > 0\]

The \(\beta_i\) represent alternative null hypothseses for \(\beta\). Let’s let \(0 < beta < 7\). Let the significance level be \(\alpha=.01\) and \(se_{\beta} = 1\).