Probability distributions: matching a model to the question
Relate possible outcomes, probability models and sampling distributions without assuming that all datasets are normally distributed.
Possible outcomes and their probabilities
A probability distribution describes how probability is spread over a random variable’s possible values. Discrete random variables have distinct possible values, such as counts. Continuous models describe values over a range, such as duration.
The sum of all probabilities equals 1 when dealing with discrete distributions. However, when using continuous densities we consider the area under the curve within a specified interval. The height at any given location along a density function does NOT relate to the probability of that exact occurrence.
Match the model to the mechanism
Binomial models use a fixed number of independent trials, each with a success-or-failure outcome and the same success probability. A Poisson model can describe event counts over an interval when its rate assumptions are appropriate. Neither model fits every kind of count data.
A Normal Model is symmetric and is completely defined by a Mean and a Positive Standard Deviation. Because it is symmetric, it may fit strongly right-skewed positive durations poorly. A Z-Score measures how many Standard Deviations a value lies from the Mean: (Value − Mean)/SD. Calculating a Z-Score does not make non-Normal Data Normal.
Individual values and sample means differ
Under independent, identically distributed sampling with finite variance, standardised sample means approach a normal distribution as sample size increases. This refers to the sampling distribution of means, not a distribution of sampling distributions or a transformation of individual observations.
The Sample Size necessary to obtain a Good Approximation to the Sampling Distribution varies depending upon both the Distribution from which the Sample was drawn and what level of Accuracy is desired. There is no Universal Rule about Thirty (30). Any Skewness, Heavy Tails or Dependence present will require Caution and could render a Simple Argument Invalid. Instead of Relying Upon Numerical Rules; Evaluate the Assumptions of Your Models Before Proceeding.
Worked example: a probability under a normal model
Consider the case of an idealized study measuring something. The study assumes that the measure follows a normal distribution with a mean = 100 and a standard deviation = 15.
For a value of 130, z = (130 − 100)/15 = 2. The normal-model probability of a value at or below 130 is approximately 0.97725.
The existence of this probability would depend on the use of a Normal Distribution Model. The fact that a mean and a standard deviation of 100 and 15 have been calculated from your sample data does not automatically make it true that a normal distribution model can be used to describe your data.
Regardless of the calculations performed on a model, an assumption will still exist. Simply performing the various calculations associated with a model does not ensure that the model accurately represents your data.