Parameters of probability distributions and the binomial distribution

Outline

• The following topics will be covered in this lecture:

  • Review of mean, standard deviation and variance of probability distribution
  • Binomial distribution
  • Parameters of the binomial distribution

Motivation

• Our goal in this course is to use statistics from a small, representative sample to say something general about the larger, unobservable population or phenomena.

• Recall, the measures of the population are what we referred to as parameters.

• Parameters are generally unknown and unknowable.

• If we have a representative sample we can compute the sample mean.
• The sample mean will almost surely not equal population mean, due to the natural variation (sampling error) that occurs in any given sample.

• Random variables and probability distributions give us the model for estimating population parameters.

• Generally, we will have to be satisfied with estimates of the parameters that are uncertain, but also include measures of “how uncertain”.

Characteristics of data

Courtesy of M. W. Toews CC via Wikimedia Commons.

Chebyshev's Theorem

• A very similar rule is known as Chebyshev’s theorem:

  The proportion (or fraction) of any set of data lying within \( K \) standard deviations of of the mean is always at least \( 1-\frac{1}{K^2} \) where \( K>1 \).

• Q: suppose \( K=2 \), what does this statement tell us?
• For \( K=2 \), we say at least \[ 1 - \frac{1}{2^2} = 1 - \frac{1}{4} = \frac{3}{4} \] of data lies within \( K=2 \) standard deviations of the mean.
• Note, this holds for any distribution whereas the empirical rule only holds for normal data.
• If we know the population is in fact normal, then \( 95\% > 75\% =1 - \frac{1}{2^2} \) of the population lies within \( K=2 \) standard deviations of the mean.
• We note thus there are two major differences between Chebyshev’s theorem and the empirical rule:
1. The empirical rule only holds for normal data, while Chebyshev’s theorem holds for any type of data.
2. However, Chebyshev’s theorem is only a lower bound on at least how much data lies within standard deviations and is a much weaker statement on how much.
• In either case, these parameters tell us a lot about the center and spread of data, with some qualifications.
• We will now recall how we compute the mean, variance and standard deviation from a probability distribution.

The mean of probability distributions

• Let \( X \) be a random variable that can attain the possible-to-observe values \( \{x_\alpha \in\mathbf{R}\} \).
• We say the the probability density function \( f(x_\alpha) = P(X=x_\alpha) \).
• The mean of the probability distribution \[ \mu = \sum_{x_\alpha \in \mathbf{R}} x_\alpha P(X=x_\alpha) = \sum_{x_\alpha \in \mathbf{R}} x_\alpha f(x_\alpha) \] is computed like the formula for the mean of a frequency distribution.
• However, because of the difference in the interpretation, the mean of a probability distribution has a special name:
• For a random variable \( x \) with probability distribution defined by the pairs of values \( \{x_\alpha\} \) and \( P(X=x_\alpha) \), the expected value of \( x \) is defined, \[ \mu = \mathbb{E}\left[X\right] = \sum_{x_\alpha \in \mathbf{R}} x_\alpha P(X=x_\alpha). \]
• We call the mean of the probability distribution the expected value, because it can be thought of as the theoretical mean if we repeated an experiment infinitely many times or sampled the entire population;
• we would expect this value on average, relative to infinitely many experiments.

Standard deviation of a sample

• Suppose we are considering finite population of size \( N \).
• Usually, we will not have access to the entire population \( x_1, \cdots, x_N \).
• Instead, we will only have some smaller subset of values in a sample \( x_1, \cdots, x_n \) for \( n< N \).
• Our sample standard deviation can be computed as \[ s = \sqrt{\frac{\sum_{i=1}^n\left(x_i - \overline{x}\right)^2}{n-1}} \]
• or \[ s = \sqrt{\frac{n \left(\sum_{i=1}^n x_i^2 \right) - \left(\sum_{i=1}^n x_i\right)^2}{n\left(n-1\right)}}. \]
• However, this differs from computing the population standard deviation as \[ \sigma = \sqrt{\frac{\sum_{i=1}^N \left(x_i - \mu\right)^2}{N}}. \]
• One key difference to remember is that for the sample standard deviation, we have a very different denominator than with the population standard deviation.

Variance

• The word variance also has a specific meaning in statistics and is another tool for describing the variation / dispersion / spread of the data.

• Suppose that the data has a population standard deviation of \( \sigma \) and a sample standard deviation of \( s \).

• Then, the data has a population variance of \( \sigma^2 \).

• Likewise, the data has a sample variance of \( s^2 \).

• Therefore, for either a population parameter or a sample statistic, the variance is the square of the standard deviation.

  • Because of this, the variance has units which are the square of the original units.

• For example, measuring the heights of students in inches, the standard deviation is in the units inches.

  • However, the variance is in the unit \( \text{inches}^2 \).

The standard deviation and variance of probability distributions

• Let’s recall the formula now for the standard deviation of a population with members \( \{x_i\}_{i=1}^N \) \[ \sigma = \sqrt{\frac{\sum_{i=1}^N \left(x_i - \mu\right)^2}{N}}, \] where we take a denominator of \( N \) for the population instead of \( N-1 \) as in samples.
• Let’s suppose that the population members \( x_i \) equal values \( x_\alpha \) in the range \( \mathbf{R} \) with frequencies \( n_\alpha \).
• If we re-write the above formula in terms of \( x_\alpha \) and \( n_\alpha \) we can say, \[ \begin{align} \sigma = \sqrt{\frac{\sum_{x_\alpha\in \mathbf{R}} n_\alpha\left(x_\alpha - \mu\right)^2}{N}} = \sqrt{\sum_{x_\alpha \in \mathbf{R}} \frac{n_\alpha}{N} \left(x_\alpha - \mu\right)^2} = \sqrt{\sum_{x_\alpha\in \mathbf{R}} P(X=x_\alpha) \left(x_\alpha - \mu\right)^2 }. \end{align} \]
• We will denote, \[ \sigma = \sqrt{\sum_{x_\alpha\in \mathbf{R}} P(X=x_\alpha) \left(x_\alpha - \mu\right)^2 } \] the standard deviation of the probability distribution associated to the random variable \( X \).
• That is to say that the population standard deviation \( \sigma \) is exactly the standard deviation of the probability distribution.
• For infinite populations and ranges, we can use the same argument (with calculus) to show this holds in general.

The standard deviation and variance of probability distributions continued

• Using the derivation from the last slide, \[ \sigma = \sqrt{\sum_{x_\alpha\in \mathbf{R}} P(X=x_\alpha) \left(x_\alpha - \mu\right)^2 }, \]
• we can show directly that the variance of a probability distribution is given as, \[ \sigma^2 = \sum_{x_\alpha \in \mathbf{R}} P(X=x_\alpha) \left(x_\alpha - \mu\right)^2 . \]
• We can also derive the alternative forms for the population standard deviation and variance in terms of the probability distribution as \[ \begin{align} \sigma &= \sqrt{\sum_{x_\alpha \in \mathbf{R}} x_\alpha^2 P(X=x_\alpha) - \mu^2 }, \\ \\ \sigma^2 &= \sum_{x_\alpha\in\mathbf{R}} x_\alpha^2 P(X=x_\alpha) - \mu^2 . \end{align} \]
• This again just amounts to some algebraic manipulation and these forms are totally equivalent to the other forms.

Example of the variance and standard deviation

Binomial distribution

• A coin flipping experiment is actually a simple example of a broad category of experiments.
• For example, we can consider an experiment in which we describe two possible outcomes:
1. (success / \( S \) / \( 1 \) / \( H \)); or
2. (failure / \( F \) / \( 0 \) / \( T \)).
• We can encode the outcomes any way we like;
• it is common to encode the outcomes as \( S \) or \( F \), where the choice of “success” is arbitrary.
• More generally than coin flipping, we might consider the case where the probabilities, \[ \begin{align} P(\text{success}) \neq P(\text{failure}) \end{align} \]
• Recall, if the experiment has only two possible outcomes, if \( A= \)"success" then \( \overline{A}= \)"failure".
• Therefore, \[ \begin{align} P(\text{success}) + P(\text{failure}) = 1. \end{align} \]
• Suppose we run the experiment a total of \( n \) trials.
• Because there are only two possible outcomes for each trial, and a finite number of trials, we can create a list of all possible outcomes for \( n \) trials.
• For example, if there are \( x_\alpha \) total successes, there must be exactly \( n - x_\alpha \) failures in \( n \) trials.
• More importantly, we can also make a list of all possible ways we can have \( x_\alpha \) successes and \( n-x_\alpha \) failures.
• Let \( X \) be the random variable equal to the number of successful trials – we can therefore calculate the probability, \[ P(X = x_\alpha); \] however, the classical model for probability (equal probability of all outcomes) will no longer apply.

Binomial distribution continued

• Recall our random variable \( X \) equal to the number of successful trials in \( n \) total trials.
• Unlike with coin flipping, we suppose that it is possible for \[ \begin{align} P(\text{success}) \neq P(\text{failure}) \end{align} \]
• However, there are finite trials, finite possible outcomes and, for each possible number of successes \( x_\alpha \), there are a finite number of ways \( X=x_\alpha \).
• Provided all trials are independent (like coin flipping) and the probability of success is constant, we can still make a counting argument using
1. The rule of complementary probability \[ \begin{align} P(\text{success}) + P(\text{failure}) = 1; \end{align} \]
2. independence;
3. the list of all possible ways we can make \( x_\alpha \) successes;
4. the list of all possible ways we can make \( n-x_\alpha \) failures; and
5. a total of \( n \) trials exactly;
• to compute the probability exactly for each \( x_\alpha \) where \( x_\alpha \) ranges from \( 0, 1, \cdots, n \).
• The list of all possible number of successes \( x_\alpha = 0, 1, \cdots, n \) and the associated probabilities \( P(X= x_\alpha) \) for \( x_\alpha = 0, 1, \cdots, n \) is called the binomial distribution.
• The argument itself is somewhat long, but it really only uses tools we already know.
• Therefore, if you can understand the principles of the points 1 - 5 above, we don’t need to belabor the details.

Binomial distribution continued

• Formally, we will now describe the binomial distribution.
• Suppose we run an experiment with two possible outcomes \( S= \)"success" and \( F= \)"failure", where \[ \begin{align} P(S) = p && P(F) = 1 - P(S) = q. \end{align} \]
• Suppose we run exactly \( n \) total trials of the above experiment and suppose that:
1. each trial is independent; and
2. \( P(S)=p \) for every trial.
• Let \( X \) be the random variable equal to the total number of successful trials.
• Let \( x_\alpha \) be one of the possible number of successful trials in the range \( 0, 1 ,\cdots , n \).
• Then the probability of exactly \( x_\alpha \) successful trials (the event \( X= x_\alpha \)) is given by \[ \begin{align} P(X=x_\alpha) = \frac{n!}{\left( n - x_\alpha\right)! x_\alpha !} p^{x_\alpha} q^{(n - x_\alpha)}, \end{align} \]
• where:
1. The total number of ways that we can have exactly \( x_\alpha \) successes in \( n \) trials is given by \[ \frac{n!}{\left( n - x_\alpha\right)! x_\alpha !}. \]
2. The probability of \( x_\alpha \) independent successes (or \( n-x_\alpha \) independent failures) is \( p^{x_\alpha} \) \( \big( \) or \( q^{(n-x_\alpha)}\big) \) respectively.

Binomial distribution example

• Recall our notation:
  1. \( n \) - the number of trials;
  2. \( X \) - the random variable;
  3. \( x_\alpha \) - a specific number of successes that \( X \) could possibly attain;
  4. \( P(S)= p \) - the probability of an independent trial’s success;
  5. \( P(F)=q \) - the probability of an independent trial’s failure.
• Suppose that when an adult is randomly selected with replacement, there is a \( 0.85 \) probability that this person knows what Twitter is (based on results from a Pew Research Center survey from several years ago…).
• Suppose that we want to find the probability that exactly three of five random adults know what Twitter is.
• Q: can you identify what \( n \), \( X \), \( x_\alpha \), \( p \) and \( q \) are in the above word problem?
• Here we consider the random selection to be a “trial” so that the number of trials is \( n=5 \)
• If we consider a “successful” trial to be “select an adult who knows what Twitter is”, then \( X \) is “number of adults who know what Twitter is out of five”.
• \( x_\alpha \) is the specific number of successful trials we are interested in, i.e., \( x_\alpha = 3 \).
• \( p \) is the probability of an independent trial’s successs, i.e, \( p=0.85 \)
• \( q \) is the probability of an independent trial’s failure, i.e., \( q=1-p = 0.15 \).

Binomial distribution example continued

• Let’s recall our values from the last slide,
  • Here we consider the random selection to be a “trial” so that the number of trials is \( n=5 \)
  • If we consider a “successful” trial to be “select an adult who knows what Twitter is”, then \( X \) is “number of adults who know what Twitter is out of five”.
  • \( x_\alpha \) is the specific number of successful trials we are interested in, i.e., \( x_\alpha = 3 \).
  • \( p \) is the probability of an independent trial’s success, i.e, \( p=0.85 \)
  • \( q \) is the probability of an independent trial’s failure, i.e., \( q=1-p = 0.15 \).
• Suppose we wanted to compute the probability of one particular outcome,
• say, \( S_i = \)"the \( i \)-th particpant knows what Twitter is" and \( F_i= \)"the \( i \)-th participant does not know what twitter is", where \[ A = S_1 \text{ and } S_2 \text{ and } S_3 \text{ and } F_4 \text{ and } F_5. \]
• We can use independence and the multiplication rule to show \[ \begin{align} P(A) &= P(S_1)\times P(S_2)\times P(S_3)\times P(F_4)\times P(F_5) \\ &= 0.85 \times 0.85 \times 0.85 \times 0.15 \times 0.15 \\ &= 0.85^3 \times 0.15^2. \end{align} \]
• This shows how we get one part of the binomial distribution formula.
• However, there are many combinations of \( S_i \) and \( F_i \) that arise in \( X=3 \).
• Using a counting argument, we can show that the total number of ways \( X=3 \) is \[ \begin{align} \frac{n!}{(n- x_\alpha)! x_\alpha!} = \frac{5!}{(5 - 3)! 3!} = \frac{5!}{(2)!3!} = 10 \end{align} \]

Binomial distribution example continued

• Let’s recall our values from the last slide,
  • Here we consider the random selection to be a “trial” so that the number of trials is \( n=5 \)
  • If we consider a “successful” trial to be “select an adult who knows what Twitter is”, then \( X \) is “number of adults who know what Twitter is out of five”.
  • \( x_\alpha \) is the specific number of successful trials we are interested in, i.e., \( x_\alpha = 3 \).
  • \( p \) is the probability of an independent trial’s successs, i.e, \( p=0.85 \)
  • \( q \) is the probability of an independent trial’s failure, i.e., \( q=1-p = 0.15 \).
  • The total number of ways \( X=3 \) is \[ \begin{align} \frac{n!}{(n- x_\alpha)! x_\alpha!} = \frac{5!}{(5 - 3)! 3!} = \frac{5!}{(2)!3!} = 10 \end{align} \]
• The binomial distribution formula can then be read as,
• The probability of finding exactly \( 3 \) out of \( 5 \) independently, randomly selected adults who know what Twitter is, is equal to \[ \begin{align} \frac{n!}{(n- x_\alpha)! x_\alpha!} p^{x_\alpha} q^{n- x_\alpha} = 10 \times 0.85^3 \times 0.15^2 \approx 0.138, \end{align} \]
• or in plain English,

  the probability of three independent successful trials, times the probability of two independent failure trials, times all possible ways we can have exactly \( 3 \) successful trials out of five.

• Again, the counting argument follows directly from the material we studied before the midterm.

Binomial distribution example continued

• We should remark the following on the last calculation.
• Technically, we could only make use of the binomial distribution because we sampled with replacement to enforce independent trials.
• If sampled our population without replacement, we know
1. that the trials are dependent; and
2. that the probability of success changes at each trial.
• These conditions make it so the binomial distribution does not apply to the random variable \( x \) when we do not replace samples.
• However it is common to approximate sampling without replacement as independent when the sample size is less than \( 5\% \) of the population.
• In practice for polls of, e.g., all US adults, this approximation will often be used.

Parameters of the binomial distribution

Courtesy of Ania Panorska CC

Parameters of the binomial distribution continued

Public domain via Wikimedia Commons

Review of the binomial distribution

• The binomial distribution is a key distribution that gives us a way to model a wide range of experiments probabilistically.
• This applies when we run an experiment with two possible outcomes \( S= \)"success" and \( F= \)"failure", where \[ \begin{align} P(S) = p && P(F) = 1 - P(S) = q. \end{align} \]
• When we run exactly \( n \) total trials of the above experiment, assuming that:
1. each trial is independent; and
2. \( P(S)=p \) for every trial.
• We can model the probability of a particular number of successes \( x_\alpha \) like a (possibly) non-fair coin flipping experiment.
• We model the probability of exactly \( x_\alpha \) successful trials as \[ \begin{align} P(X=x_\alpha) = \underbrace{\frac{n!}{\left( n - x_\alpha\right)! x_\alpha !}}_{(1) } \times \underbrace{ p^{x_\alpha}}_{(2)} \times \underbrace{q^{(n - x_\alpha)}}_{(3)} \end{align} \] where:
1. Total number of ways to find exactly \( x_\alpha \) successful trials out of \( n \) total trials;
2. Probability of \( x_\alpha \) independent succesful trials;
3. Probability of \( n-x_\alpha \) independent failure trials;
• The special structure of this distribution also allows us to compute the mean and standard deviation directly as \[ \begin{align} \mu = n \times p & & \sigma = \sqrt{n \times p\times q} \end{align} \]