Your company has created a new drug that may cure arthritis. How would you conduct a test to confirm the drug’s effectiveness?
The latest sales data have just come in, and your boss wants you to prepare a report for management on places where the company could improve its business. What should you look for? What should you not look for?
You and a friend are at a baseball game, and out of the blue he offers you a bet that neither team will hit a home run in that game. Should you take the bet?
You want to conduct a poll on whether your school should use its funding to build a new athletic complex or a new library. How many people do you have to poll? How do you ensure that your poll is free of bias? How do you interpret your results?
A widget maker in your factory that normally breaks 4 widgets for every 100 it produces has recently started breaking 5 widgets for every 100. When is it time to buy a new widget maker? (And just what is a widget, anyway?)
These are some of the many realworld examples that require the use of statistics. How would you approach the problem statements? There are some stepwise human algorithms, but is there a general problem statement?
 “Find possible solutions, decide on a solution, plan the solution, implement the solution, learn from the results for future solutions (or resolution).”
 “SOAP – subjective – the problem as given, objective – the problem after examination, assessment – the better defined problem, plan – decide if guidelines to management already exist, and blueprint the solution for this case, or generate a riskminimizing, new solution path”.
 “HAMRC – hypothesis, aim, methodology, results, conclusion” – the concept that there is no real difference is the null hypothesis.
Then there is the joke that compares the different ways of thinking:
“A physicist, a chemist and a statistician were working collaboratively on a problem, when the wastepaper basket spontaneously combusted (they all swore they had stopped smoking). The chemist said, ‘quick, we must reduce the concentration of the reactant which is oxygen, by increasing the relative concentration of nonreactive gases, such as carbon dioxide and carbon monoxide. Place a fire blanket over the flames.’ The physicist, interjected, ‘no, no, we must reduce the heat energy available for activating combustion ; get some water to douse the flame’. Meanwhile, the statistician was running around lighting more fires. The others asked with alarm, ‘what are you doing?’. ‘Trying to get an adequate sample size’.”
General Definition[edit]
Statistics, in short, is the study of data. It includes descriptive statistics (the study of methods and tools for collecting data, and mathematical models to describe and interpret data) and inferential statistics (the systems and techniques for making probabilitybased decisions and accurate predictions.
Etymology[edit]
As its name implies, statistics has its roots in the idea of “the state of things”. The word itself comes from the ancient Latin term statisticum collegium, meaning “a lecture on the state of affairs”. Eventually, this evolved into the Italian word statista, meaning “statesman”, and the German word Statistik, meaning “collection of data involving the State”. Gradually, the term came to be used to describe the collection of any sort of data.
Statistics as a subset of mathematics[edit]
As one would expect, statistics is largely grounded in mathematics, and the study of statistics has lent itself to many major concepts in mathematics: probability, distributions, samples and populations, the bell curve, estimation, and data analysis.
Up ahead[edit]
Up ahead, we will learn about subjects in modern statistics and some practical applications of statistics. We will also lay out some of the background mathematical concepts required to begin studying statistics.
A remarkable amount of today’s modern statistics comes from the original work of R.A. Fisher in the early 20th Century. Although there are a dizzying number of minor disciplines in the field, there are some basic, fundamental studies.
The beginning student of statistics will be more interested in one topic or another depending on his or her outside interest. The following is a list of some of the primary branches of statistics.
Probability Theory and Mathematical Statistics[edit]
Those of us who are purists and philosophers may be interested in the intersection between pure mathematics and the messy realities of the world. A rigorous study of probability — especially the probability distributions and the distribution of errors — can provide an understanding of where all these statistical procedures and equations come from. Although this sort of rigor is likely to get in the way of a psychologist (for example) learning and using statistics effectively, it is important if one wants to do serious (i.e. graduatelevel) work in the field.
That being said, there is good reason for all students to have a fundamental understanding of where all these “statistical techniques and equations” are coming from! We’re always more adept at using a tool if we can understand why we’re using that tool. The challenge is getting these important ideas to the nonmathematician without the student’s eyes glazing over. One can take this argument a step further to claim that a vast number of students will never actually use a ttest—he or she will never plug those numbers into a calculator and churn through some esoteric equations—but by having a fundamental understanding of such a test, he or she will be able to understand (and question) the results of someone else’s findings.
Design of Experiments[edit]
One of the most neglected aspects of statistics—and maybe the single greatest reason that Statisticians drink—is Experimental Design. So often a scientist will bring the results of an important experiment to a statistician and ask for help analyzing results only to find that a flaw in the experimental design rendered the results useless. So often we statisticians have researchers come to us hoping that we will somehow magically “rescue” their experiments.
A friend provided me with a classic example of this. In his psychology class he was required to conduct an experiment and summarize its results. He decided to study whether music had an impact on problem solving. He had a large number of subjects (myself included) solve a puzzle first in silence, then while listening to classical music and finally listening to rock and roll, and finally in silence. He measured how long it would take to complete each of the tasks and then summarized the results.
What my friend failed to consider was that the results were highly impacted by a learning effect he hadn’t considered. The first puzzle always took longer because the subjects were first learning how to work the puzzle. By the third try (when subjected to rock and roll) the subjects were much more adept at solving the puzzle, thus the results of the experiment would seem to suggest that people were much better at solving problems while listening to rock and roll!
The simple act of randomizing the order of the tests would have isolated the “learning effect” and in fact, a welldesigned experiment would have allowed him to measure both the effects of each type of music and the effect of learning. Instead, his results were meaningless. A careful experimental design can help preserve the results of an experiment, and in fact some designs can save huge amounts of time and money, maximize the results of an experiment, and sometimes yield additional information the researcher had never even considered!
Sampling[edit]
Similar to the Design of Experiments, the study of sampling allows us to find a most effective statistical design that will optimize the amount of information we can collect while minimizing the level of effort. Sampling is very different from experimental design however. In a laboratory we can design an experiment and control it from start to finish. But often we want to study something outside of the laboratory, over which we have much less control.
If we wanted to measure the population of some harmful beetle and its effect on trees, we would be forced to travel into some forest land and make observations, for example: measuring the population of the beetles in different locations, noting which trees they were infesting, measuring the health and size of these trees, etc.
Sampling design gets involved in questions like “How many measurements do I have to take?” or “How do I select the locations from which I take my measurements?” Without planning for these issues, researchers might spend a lot of time and money only to discover that they really have to sample ten times as many points to get meaningful results or that some of their sample points were in some landscape (like a marsh) where the beetles thrived more or the trees grew better.
Modern Regression[edit]
Regression models relate variables to each other in a linear fashion. For example, if you recorded the heights and weights of several people and plotted them against each other, you would find that as height increases, weight tends to increase too. You would probably also see that a straight line through the data is about as good a way of approximating the relationship as you will be able to find, though there will be some variability about the line. Such linear models are possibly the most important tool available to statisticians. They have a long history and many of the more detailed theoretical aspects were discovered in the 1970s. The usual method for fitting such models is by “least squares” estimation, though other methods are available and are often more appropriate, especially when the data are not normally distributed.
What happens, though, if the relationship is not a straight line? How can a curve be fit to the data? There are many answers to this question. One simple solution is to fit a quadratic relationship, but in practice such a curve is often not flexible enough. Also, what if you have many variables and relationships between them are dissimilar and complicated?
Modern regression methods aim at addressing these problems. Methods such as generalized additive models, projection pursuit regression, neural networks and boosting allow for very general relationships between explanatory variables and response variables, and modern computing power makes these methods a practical
option for many applications
Classification[edit]
Some things are different from others. How? That is, how are objects classified into their respective groups? Consider a bank that is hoping to lend money to customers. Some customers who borrow money will be unable or unwilling to pay it back, though most will
pay it back as regular repayments. How is the bank to classify customers into these two groups when deciding which ones to lend money to?
The answer to this question no doubt is influenced by many things, including a customer’s income, credit history, assets, already existing debt, age and profession. There may be other influential, measurable characteristics that can be used to predict what kind of customer a particular individual is. How should the bank decide which characteristics are important, and how should it combine this information into a rule that tells it whether or not to lend the money?
This is an example of a classification problem, and statistical classification is a large field containing methods such as linear discriminant analysis, classification trees, neural networks and other methods.
Time Series[edit]
Many types of research look at data that are gathered over time, where an observation taken today may have some correlation with the observation taken tomorrow. Two prominent examples of this are the fields of finance (the stock market) and atmospheric science.
We’ve all seen those line graphs of stock prices as they meander up and down over time. Investors are interested in predicting which stocks are likely to keep climbing (i.e. when to buy) and when a stock in their portfolio is falling. It is easy to be misled by a sudden jolt of good news or a simple “market correction” into inferring—incorrectly—that one or the other is taking place!
In meteorology scientists are concerned with the venerable science of predicting the weather. Whether trying to predict if tomorrow will be sunny or determining whether we are experiencing true climate changes (i.e. global warming) it is important to analyze weather data over time.
Survival Analysis[edit]
Suppose that a pharmaceutical company is studying a new drug which it is hoped will cause people to live longer (whether by curing them of cancer, reducing their blood pressure or cholesterol and thereby their risk of heart disease, or by some other mechanism). The company will recruit patients into a clinical trial, give some patients the drug and others a placebo, and follow them until they have amassed enough data to answer the question of whether, and by how long, the new drug extends life expectancy.
Such data present problems for analysis. Some patients will have died earlier than others, and often some patients will not have died before the clinical trial completes. Clearly, patients who live longer contribute informative data about the ability (or not) of the drug to extend life expectancy. So how should such data be analyzed?
Survival analysis provides answers to this question and gives statisticians the tools necessary to make full use of the available data to correctly interpret the treatment effect.
Categorical Analysis[edit]
In laboratories we can measure the weight of fruit that a plant bears, or the temperature of a chemical reaction. These data points are easily measured with a yardstick or a thermometer, but what about the color of a person’s eyes or her attitudes regarding the taste of broccoli? Psychologists can’t measure someone’s anger with a measuring stick, but they can ask their patients if they feel “very angry” or “a little angry” or “indifferent”. Entirely different methodologies must be used in statistical analysis from these sorts of experiments. Categorical Analysis is used in a myriad of places, from political polls to analysis of census data to genetics and medicine.
Clinical Trials[edit]
In the United States, the FDA requires that pharmaceutical companies undergo rigorous procedures called Clinical Trials and statistical analyses to assure public safety before allowing the sale or use of new drugs. In fact, the pharmaceutical industry employs more statisticians than any other business!
Further reading[edit]
Imagine reading a book for the first few chapters and then becoming able to get a sense of what the ending will be like – this is one of the great reasons to learn statistics. With the appropriate tools and solid grounding in statistics, one can use a limited sample (e.g. read the first five chapters of Pride & Prejudice) to make intelligent and accurate statements about the population (e.g. predict the ending of Pride & Prejudice). This is what knowing statistics and statistical tools can do for you.
In today’s informationoverloaded age, statistics is one of the most useful subjects anyone can learn. Newspapers are filled with statistical data, and anyone who is ignorant of statistics is at risk of being seriously misled about important reallife decisions such as what to eat, who is leading the polls, how dangerous smoking is, etc. Knowing a little about statistics will help one to make more informed decisions about these and other important questions. Furthermore, statistics are often used by politicians, advertisers, and others to twist the truth for their own gain. For example, a company selling the cat food brand “Cato” (a fictitious name here), may claim quite truthfully in their advertisements that eight out of ten cat owners said that their cats preferred Cato brand cat food to “the other leading brand” cat food. What they may not mention is that the cat owners questioned were those they found in a supermarket buying Cato.
“The best thing about being a statistician is that you get to play in everyone else’s backyard.”
John Tukey, Princeton University
More seriously, those proceeding to higher education will learn that statistics is the most powerful tool available for assessing the significance of experimental data, and for drawing the right conclusions from the vast amounts of data faced by engineers, scientists, sociologists, and other professionals in most spheres of learning. There is no study with scientific, clinical, social, health, environmental or political goals that does not rely on statistical methodologies. The basic reason for that is that variation is ubiquitous in nature and probability and statistics are the fields that allow us to study, understand, model, embrace and interpret variation.
See Also
UCLA Brochure on Why Study Probability & Statistics
Statistics is a diverse subject and thus the mathematics that are required depend on the kind of statistics we are studying. A strong background in linear algebra is needed for most multivariate statistics, but is not necessary for introductory statistics. A background in Calculus is useful no matter what branch of statistics is being studied.
At a bare minimum the student should have a grasp of basic concepts taught in Algebra and be comfortable with “moving things around” and solving for an unknown.
Most of the statistics here will derive from a few basic things that the reader should become acquainted with.
Absolute Value[edit]
If the number is zero or positive, then the absolute value of the number is simply the same number. If the number is negative, then take away the negative sign to get the absolute value.
Examples[edit]
 42 = 42
 5 = 5
 2.21 = 2.21
Factorials[edit]
A factorial is a calculation that gets used a lot in probability. It is defined only for integers greaterthanorequalto zero as:
${displaystyle n!equiv {begin{cases}ncdot (n1)!,&ngeq 1\1,&n=0end{cases}}}$
Examples[edit]
In short, this means that:
0! =  1  = 1 
1! =  1 · 1  = 1 
2! =  2 · 1  = 2 
3! =  3 · 2 · 1  = 6 
4! =  4 · 3 · 2 · 1  = 24 
5! =  5 · 4 · 3 · 2 · 1  = 120 
6! =  6 · 5 · 4 · 3 · 2 · 1  = 720 
Summation[edit]
The summation (also known as a series) is used more than almost any other technique in statistics. It is a method of representing addition over lots of values without putting + after +. We represent summation using a big uppercase sigma: ∑.
Examples[edit]
Very often in statistics we will sum a list of related variables:
Here we are adding all the x variables (which will hopefully all have values by the time we calculate this). The expression below the ∑ (i=0, in this case) represents the index variable and what its starting value is (i with a starting value of 0) while the number above the ∑ represents the number that the variable will increment to (stepping by 1, so i = 0, 1, 2, 3, and then 4). Another example:
Notice that we would get the same value by moving the 2 outside of the summation (perform the summation and then multiply by 2, rather than multiplying each component of the summation by 2).
Infinite series[edit]
There is no reason, of course, that a series has to count on any determined, or even finite value—it can keep going without end. These series are called “infinite series” and sometimes they can even converge to a finite value, eventually becoming equal to that value as the number of items in your series approaches infinity (∞).
Examples[edit]
This example is the famous geometric series. Note both that the series goes to ∞ (infinity, that means it does not stop) and that it is only valid for certain values of the variable r. This means that if r is between the values of 1 and 1 (1 < r < 1) then the summation will get closer to (i.e., converge on) ^{1} / _{1r} the further you take the series out.
Linear Approximation[edit]
v / α  0.20  0.10  0.05  0.025  0.01  0.005 

40  0.85070  1.30308  1.68385  2.02108  2.42326  2.70446 
50  0.84887  1.29871  1.67591  2.00856  2.40327  2.67779 
60  0.84765  1.29582  1.67065  2.00030  2.39012  2.66028 
70  0.84679  1.29376  1.66691  1.99444  2.38081  2.64790 
80  0.84614  1.29222  1.66412  1.99006  2.37387  2.63869 
90  0.84563  1.29103  1.66196  1.98667  2.36850  2.63157 
100  0.84523  1.29007  1.66023  1.98397  2.36422  2.62589 
Let us say that you are looking at a table of values, such as the one above. You want to approximate (get a good estimate of) the values at 63, but you do not have those values on your table. A good solution here is use a linear approximation to get a value which is probably close to the one that you really want, without having to go through all of the trouble of calculating the extra step in the table.
This is just the equation for a line applied to the table of data. x_{i} represents the data point you want to know about,
${displaystyle x_{lfloor irfloor }}$is the known data point beneath the one you want to know about, and
${displaystyle x_{lceil irceil }}$is the known data point above the one you want to know about.
Examples[edit]
Find the value at 63 for the 0.05 column, using the values on the table above.
First we confirm on the above table that we need to approximate the value. If we know it exactly, then there really is no need to approximate it. As it stands this is going to rest on the table somewhere between 60 and 70. Everything else we can get from the table:
Using software, we calculate the actual value of f(63) to be 1.669402, a difference of around 0.00013. Close enough for our purposes.
Data are assignments of values onto observations of events and objects. They can be classified by their coding properties and the characteristics of their domains and their ranges.
Identifying data type[edit]
When a given data set is numerical in nature, it is necessary to carefully distinguish the actual nature of the variable being quantified. Statistical tests are generally specific for the kind of data being handled.
Data on a nominal (or categorical) scale[edit]
Identifying the true nature of numerals applied to attributes that are not “measures” is usually straightforward and apparent. Examples in everyday use include road, car, house, book and telephone numbers. A simple test would be to ask if reassigning the numbers among the set would alter the nature of the collection. If the plates on a car are changed, for example, it still remains the same car in reality.
Data on an Ordinal Scale[edit]
An ordinal scale is a scale with ranks. Those ranks only have sense in that they are ordered, that is what makes it ordinal scale. The distance [rank n] minus [rank n1] is not guaranteed to be equal to [rank n1] minus [rank n2], but [rank n] will be greater than [rank n1] in the same way [rank n1] is greater than [rank n2] for all n where [rank n], [rank n1], and [rank n2] exist. Ranks of an ordinal scale may be represented by a system with numbers or names and an agreed order.
We can illustrate this with a common example: the Likert scale. Consider five possible responses to a question, perhaps Our president is a great man, with answers on this scale
Response:  Strongly Disagree  Disagree  Neither Agree nor Disagree  Agree  Strongly Agree 

Code:  1  2  3  4  5 
Here the answers are a ranked scale reflected in the choice of numeric code. There is however no sense in which the distance between Strongly agree and Agree is the same as between Strongly disagree and Disagree.
Numerical ranked data should be distinguished from measurement data.
Measurement data[edit]
Numerical measurements exist in two forms, Meristic and continuous, and may present themselves in three kinds of scale: interval, ratio and circular.
Meristic or discrete variables are generally counts and can take on only discrete values. Normally they are represented by natural numbers. The number of plants found in a botanist’s quadrant would be an example. (Note that if the edge of the quadrant falls partially over one or more plants, the investigator may choose to include these as halves, but the data will still be meristic as doubling the total will remove any fraction).
Continuous variables are those whose measurement precision is limited only by the investigator and his equipment. The length of a leaf measured by a botanist with a ruler will be less precise than the same measurement taken by micrometer. (Notionally, at least, the leaf could be measured even more precisely using a microscope with a graticule.)
Interval Scale
Variables measured on an interval scale have values in which differences are uniform and meaningful but ratios will not be so. An oft quoted example is that of the Celsius scale of temperature. A difference between 5° and 10° is equivalent to a difference between 10° and 15°, but the ratio between 15° and 5° does not imply that the former is three times as warm as the latter.
Ratio Scale
Variables on a ratio scale have a meaningful zero point. In keeping with the above example one might cite the Kelvin temperature scale. Because there is an absolute zero, it is true to say that 400°K is twice as warm as 200°K, though one should do so with tongue in cheek. A better daytoday example would be to say that a 180 kg Sumo wrestler is three times heavier than his 60 kg wife.
Circular Scale
When one measures annual dates, clock times and a few other forms of data, a circular scale is in use. It can happen that neither differences nor ratios of such variables are sensible derivatives, and special methods have to be employed for such data.
Statistical data can be obtained from two sources primary or secondary.
Primary Data[edit]
Primary data means original data that has been collected specially for the purpose in mind. It means someone collected the data from the original source first hand. Data collected this way is called primary data.
The people who gather primary data may be an authorized organization, investigator, enumerator or they may be just someone with a clipboard. Those who gather primary data may have knowledge of the study and may be motivated to make the study a success. These people are acting as a witness so primary data is only considered as reliable as the people who gathered it.
Research where one gathers this kind of data is referred to as field research.
For example: your own questionnaire.
Secondary Data[edit]
Secondary data is data that has been collected for another purpose. When we use Statistical Method with Primary Data from another purpose for our purpose we refer to it as Secondary Data. It means that one purpose’s Primary Data is another purpose’s Secondary Data. Secondary data is data that is being reused. Usually in a different context.
Research where one gathers this kind of data is referred to as desk research.
For example: data from a book.
Why Classify Data This Way?[edit]
Knowing how the data was collected allows critics of a study to search for bias in how it was conducted. A good study will welcome such scrutiny. Each type has its own weaknesses and strengths. Primary Data is gathered by people who can focus directly on the purpose in mind. This helps ensure that questions are meaningful to the purpose but can introduce bias in those same questions. Secondary Data doesn’t have the privilege of this focus but is only susceptible to bias introduced in the choice of what data to reuse. Stated another way, those who gather Secondary Data get to pick the questions. Those who gather Primary Data get to write the questions.
<< Sources of Data  Statistics  >> Qualitative and Quantitative
Qualitative data[edit]
Qualitative data is a categorical measurement expressed not in terms of numbers, but rather by means of a natural language description. In statistics, it is often used interchangeably with “categorical” data.
For example: favorite color = "blue" height = "tall" i hated the most = "zen"
Although we may have categories, the categories may have a structure to them. When there is not a natural ordering of the categories, we call these nominal categories. Examples might be gender, race, religion, or sport.
When the categories may be ordered, these are called ordinal variables. Categorical variables that judge size (small, medium, large, etc.) are ordinal variables. Attitudes (strongly disagree, disagree, neutral, agree, strongly agree) are also ordinal variables, however we may not know which value is the best or worst of these issues. Note that the distance between these categories is not something we can measure.
Quantitative data[edit]
Quantitative data is a numerical measurement expressed not by means of a natural language description, but rather in terms of numbers. However, not all numbers are continuous and measurable. For example, the social security number is a number, but not something that one can add or subtract.
For example: molecule length = "450 nm" height = "1.8 m"
Quantitative data always are associated with a scale measure.
Probably the most common scale type is the ratioscale. Observations of this type are on a scale that has a meaningful zero value but also have an equidistant measure (i.e., the difference between 10 and 20 is the same as the difference between 100 and 110). For example, a 10 yearold girl is twice as old as a 5 yearold girl. Since you can measure zero years, time is a ratioscale variable. Money is another common ratioscale quantitative measure. Observations that you count are usually ratioscale (e.g., number of widgets).
The main portion of Statistics is the display of summarized data. Data is initially collected from a given source, whether they are experiments, surveys, or observation, and is presented in one of four methods:
 Textual Method
 The reader acquires information through reading the gathered data.
 Tabular Method
 Provides a more precise, systematic and orderly presentation of data in rows or columns.
 Semitabular Method
 Uses both textual and tabular methods.
 Graphical Method
 The utilization of graphs is most effective method of visually presenting statistical results or findings.
Experiments[edit]
In a experiment the experimenter applies ‘treatments’ to groups of subjects. For example the experimenter may give one drug to group 1 and a different drug or a placebo to group 2, to determine the effectiveness of the drug. This is what differentiates an ‘experiment’ from an ‘observational study’.
Scientists try to identify causeandeffect relationships because this kind of knowledge is especially powerful, for example, drug A cures disease B. Various methods exist for detecting causeandeffect relationships. An experiment is a method that most clearly shows causeandeffect because it isolates and manipulates a single variable, in order to clearly show its effect. Experiments almost always have two distinct variables: First, an independent variable (IV) is manipulated by an experimenter to exist in at least two levels (usually “none” and “some”). Then the experimenter measures the second variable, the dependent variable (DV).
Example: Experimentation Suppose the experimental hypothesis that concerns the scientist is that reading a wiki will enhance knowledge. Notice that the hypothesis is really an attempt to state a causal relationship like, “if you read a qiki, then you will have enhanced knowledge.” The antecedent condition (reading a wiki) causes the consequent condition (enhanced knowledge). Antecedent conditions are always IVs and consequent conditions are always DVs in experiments. So the experimenter would produce two levels of Wiki reading (none and some, for example) and record knowledge. If the subjects who got no Wiki exposure had less knowledge than those who were exposed to wikis, it follows that the difference is caused by the IV. 
So, the reason scientists utilize experiments is that it is the only way to determine causal relationships between variables. Experiments tend to be artificial because they try to make both groups identical with the single exception of the levels of the independent variable.
Sample surveys involve the selection and study of a sample of items from a population. A sample is just a set of members chosen from a population, but not the whole population. A survey of a whole population is called a census.
A sample from a population may not give accurate results but it helps in decision making.
Examples[edit]
Examples of sample surveys:
 Phoning the fifth person on every page of the local phonebook and asking them how long they have lived in the area. (Systematic Sample)
 Dropping a quad. in five different places on a field and counting the number of wild flowers inside the quad. (Cluster Sample)
 Selecting subpopulations in proportion to their incidence in the overall population. For instance, a researcher may have reason to select a sample consisting 30% females and 70% males in a population with those same gender proportions. (Stratified Sample)
 Selecting several cities in a country, several neighbourhoods in those cities and several streets in those neighbourhoods to recruit participants for a survey. (Multistage sample)
The term random sample is used for a sample in which every item in the population is equally likely to be selected.
While sampling is a more cost effective method of determining a result, small samples or samples that depend on a certain selection method will result in a bias within the results.
The following are common sources of bias:
 Sampling bias or statistical bias, where some individuals are more likely to be selected than others (such as if you give equal chance of cities being selected rather than weighting them by size)
 Systemic bias, where external influences try to affect the outcome (e.g. funding organizations wanting to have a specific result)
The most primitive method of understanding the laws of nature utilizes observational studies. Basically, a researcher goes out into the world and looks for variables that are associated with one another. Notice that, unlike experiments, in an observational study the Independent Variables are not manipulated by the experimenter. The independent variable could be something like “smoking”. It would be unethical and probably impossible to do an experiment where one randomly selected group is assigned to smoke and another group is assigned to the nonsmoking group. Therefore, in order do determine the health effects of smoking on humans, an observational study is more appropriate than an experiment. The health of smokers and nonsmokers would be compared without the experimenter assigning treatment.
Some of the foundations of modern scientific thought are based on observational research. Charles Darwin, for example, based his explanation of evolution entirely on observations he made. Case studies, where individuals are observed and questioned to determine possible causes of problems, are a form of observational research that continues to be popular today. In fact, every time you see a physician he or she is performing observational science.
There is a problem in observational science though — it cannot ever identify causal relationships because even though two variables are related both might be caused by a third, unseen, variable. Since the underlying laws of nature are assumed to be causal laws, observational findings are generally regarded as less compelling than experimental findings.
The key way to identify experimental studies is that they involve an intervention such as the administration of a drug to one group of patients and a placebo to another group. Observational studies only collect data and make comparisons.
Medicine is an intensively studied discipline, and not all phenomenon can be studied by experimentation due to obvious ethical or logistical restrictions.
 Case series: These are purely observational, consisting of reports of a series of similar medical cases. For example, a series of patients might be reported to suffer from bone abnormalities as well as immunodeficiencies. This association may not be significant, occurring purely by chance. On the other hand, the association may point to a mutation in common pathway affecting both the skeletal system and the immune system.
 CaseControl: This involves an observation of a disease state, compared to normal healthy controls. For example, patients with lung cancer could be compared with their otherwise healthy neighbors. Using neighbors limits bias introduced by demographic variation. The cancer patients and their neighbors (the control) might be asked about their exposure history (did they work in an industrial setting), or other risk factors such as smoking. Another example of a casecontrol study is the testing of a diagnostic procedure against the gold standard. The gold standard represents the control, while the new diagnostic procedure is the “case.” This might seem to qualify as an “intervention” and thus an experiment.
 Crosssectional: Involves many variables collected all at the same time. Used in epidemiology to estimate prevalence, or conduct other surveys.
 Cohort: A group of subjects followed over time, prospectively. Framingham study is classic example. By observing exposure and then tracking outcomes, cause and effect can be better isolated. However this type of study cannot conclusively isolate a cause and effect relationship.
 Historic Cohort: This is the same as a cohort except that researchers use an historic medical record to track patients and outcomes.
Data analysis is one of the more important stages in our research. Without performing exploratory analyses of our data, we set ourselves up for mistakes and loss of time.
Generally speaking, our goal here is to be able to “visualize” the data and get a sense of their values. We plot histograms and compute summary statistics to observe the trends and the distribution of our data.
Data Cleaning[edit]
‘Cleaning’ refers to the process of removing invalid data points from a dataset.
Many statistical analyses try to find a pattern in a data series, based on a hypothesis or assumption about the nature of the data. ‘Cleaning’ is the process of removing those data points which are either (a) Obviously disconnected with the effect or assumption which we are trying to isolate, due to some other factor which applies only to those particular data points.
(b) Obviously erroneous, i.e. some external error is reflected in that particular data point, either due to a mistake during data collection, reporting etc.
In the process we ignore these particular data points, and conduct our analysis on the remaining data.
‘Cleaning’ frequently involves human judgement to decide which points are valid and which are not, and there is a chance of valid data points caused by some effect not sufficiently accounted for in the hypothesis/assumption behind the analytical method applied.
The points to be cleaned are generally extreme outliers. ‘Outliers’ are those points which stand out for not following a pattern which is generally visible in the data. One way of detecting outliers is to plot the data points (if possible) and visually inspect the resultant plot for points which lie far outside the general distribution. Another way is to run the analysis on the entire dataset, and then eliminating those points which do not meet mathematical ‘control limits’ for variability from a trend, and then repeating the analysis on the remaining data.
Cleaning may also be done judgementally, for example in a sales forecast by ignoring historical data from an area/unit which has a tendency to misreport sales figures. To take another example, in a double blind medical test a doctor may disregard the results of a volunteer whom the doctor happens to know in a nonprofessional context.
‘Cleaning’ may also sometimes be used to refer to various other judgemental/mathematical methods of validating data and removing suspect data.
The importance of having clean and reliable data in any statistical analysis cannot be stressed enough. Often, in realworld applications the analyst may get mesmerised by the complexity or beauty of the method being applied, while the data itself may be unreliable and lead to results which suggest courses of action without a sound basis. A good statistician/researcher (personal opinion) spends 90% of his/her time on collecting and cleaning data, and developing hypothesis which cover as many external explainable factors as possible, and only 10% on the actual mathematical manipulation of the data and deriving results.
Summary Statistics[edit]
The most simple example of statistics “in practice” is in the generation of summary statistics. Let us consider the example where we are interested in the weight of eighth graders in a school. (Maybe we’re looking at the growing epidemic of child obesity in America!) Our school has 200 eighth graders, so we gather all their weights. What we have are 200 positive real numbers.
If an administrator asked you what the weight was of this eighth grade class, you wouldn’t grab your list and start reading off all the individual weights; it’s just too much information. That same administrator wouldn’t learn anything except that she shouldn’t ask you any questions in the future! What you want to do is to distill the information — these 200 numbers — into something concise.
What might we express about these 200 numbers that would be of interest? The most obvious thing to do is to calculate the average or mean value so we know how much the “typical eighth grader” in the school weighs. It would also be useful to express how much this number varies; after all, eighth graders come in a wide variety of shapes and sizes! In reality, we can probably reduce this set of 200 weights into at most four or five numbers that give us a firm comprehension of the data set.
Averages[edit]
An average is simply a number that is representative of data. More particularly, it is a measure of central tendency. There are several types of average. Averages are useful for comparing data, especially when sets of different size are being compared. It acts as a representative figure of the whole set of data.
Perhaps the simplest and commonly used average the arithmetic mean or more simply mean which is explained in the next section.
Other common types of average are the median, the mode, the geometric mean, and the harmonic mean, each of which may be the most appropriate one to use under different circumstances.
Statistics  Summary Statistics  >> Mean, Median and Mode
Mean, Median and Mode[edit]
Mean[edit]
The mean, or more precisely the arithmetic mean, is simply the arithmetic average of a group of numbers (or data set) and is shown using bar symbol
${displaystyle {bar {}}}$. So the mean of the variable
${displaystyle x}$is
${displaystyle {bar {x}}}$, pronounced “xbar”. It is calculated by adding up all of the values in a data set and dividing by the number of values in that data set
:
.For example, take the following set of data: {1,2,3,4,5}. The mean of this data would be:
Here is a more complicated data set: {10,14,86,2,68,99,1}. The mean would be calculated like this:
Median[edit]
The median is the “middle value” in a set. That is, the median is the number in the center of a data set that has been ordered sequentially.
For example, let’s look at the data in our second data set from above: {10,14,86,2,68,99,1}. What is its median?
 First, we sort our data set sequentially: {1,2,10,14,68,85,99}
 Next, we determine the total number of points in our data set (in this case, 7.)
 Finally, we determine the central position of or data set (in this case, the 4th position), and the number in the central position is our median – {1,2,10,14,68,85,99}, making 14 our median.
Helpful Hint!  

An easy way to determine the central position or positions for any ordered set is to take the total number of points, add 1, and then divide by 2. If the number you get is a whole number, then that is the central position. If the number you get is a fraction, take the two whole numbers on either side. 
Because our data set had an odd number of points, determining the central position was easy – it will have the same number of points before it as after it. But what if our data set has an even number of points?
Let’s take the same data set, but add a new number to it: {1,2,10,14,68,85,99,100} What is the median of this set?
When you have an even number of points, you must determine the two central positions of the data set. (See side box for instructions.) So for a set of 8 numbers, we get (8 + 1) / 2 = 9 / 2 = 4 1/2, which has 4 and 5 on either side.
Looking at our dataset, we see that the 4th and 5th numbers are 14 and 68. From there, we return to our trusty friend the mean to determine the median. (14 + 68) / 2 = 82 / 2 = 41.
find the median of 2, 4, 6, 8 => firstly we must count the numbers to determine its odd or even
as we see it is even so we can write: M=(4+6)/2=10/2=5
5 is the median of above sequential numbers.
Mode[edit]
The mode is the most common or “most frequent” value in a data set. Example: the mode of the following data set (1, 2, 5, 5, 6, 3) is 5 since it appears twice. This is the most common value of the data set.
Data sets having one mode are said to be unimodal, with two are said to be bimodal and with more than two are said to be multimodal . An example of a unimodal dataset is {1, 2, 3, 4, 4, 4, 5, 6, 7, 8, 8, 9}. The mode for this data set is 4. An example of a bimodal data set is {1, 2, 2, 3, 3}. This is because both 2 and 3 are modes.
Please note: If all points in a data set occur with equal frequency, it is equally accurate to describe the data set as having many modes or no mode.
Midrange[edit]
The midrange is the arithmetic mean strictly between the minimum and the maximum value in a data set.
Relationship of the Mean, Median, and Mode[edit]
The relationship of the mean, median, and mode to each other can provide some information about the relative shape of the data distribution. If the mean, median, and mode are approximately equal to each other, the distribution can be assumed to be approximately symmetrical. If the mean > median > mode, the distribution will be skewed to the right. If the mean < median < mode, the distribution will be skewed to the left.
Questions[edit]
1. There is an old joke that states: “Using median size as a reference it’s perfectly possible to fit four pingpong balls and two blue whales in a rowboat.” Explain why this statement is true.
Statistics  Mean
Geometric Mean[edit]
The Geometric Mean is calculated by taking the nth root of the product of a set of data.
For example, if the set of data was:
1,2,3,4,5
The geometric mean would be calculated:
Of course, with large n this can be difficult to calculate. Taking advantage of two properties of the logarithm:
${displaystyle log(a^{n})=ncdot log(a)}$
We find that by taking the logarithmic transformation of the geometric mean, we get:
Which leads us to the equation for the geometric mean:
When to use the geometric mean[edit]
The arithmetic mean is relevant at any time several quantities add together to produce a total. The arithmetic mean answers the question, “if all the quantities had the same value, what would that value have to be in order to achieve the same total?”
In the same way, the geometric mean is relevant any time several quantities multiply together to produce a product. The geometric mean answers the question, “if all the quantities had the same value, what would that value have to be in order to achieve the same product?”
For example, suppose you have an investment which returns 10% the first year, 50% the second year, and 30% the third year. What is its average rate of return? It is not the arithmetic mean, because what these numbers mean is that on the first year your investment was multiplied (not added to) by 1.10, on the second year it was multiplied by 1.50, and the third year it was multiplied by 1.30. The relevant quantity is the geometric mean of these three numbers.
Written by Hafiz G m
It is known that the geometric mean is always less than or equal to the arithmetic mean (equality holding only when A=B). The proof of this is quite short and follows from the fact that
${displaystyle ({sqrt {A}}{sqrt {B}})^{2}}$is always a nonnegative number. This inequality can be surprisingly powerful though and comes up from time to time in the proofs of theorems in calculus. Source.
Harmonic Mean[edit]
The arithmetic mean cannot be used when we want to average quantities such as speed.
Consider the example below:
Example 1: The distance from my house to town is 40 km. I drove to town at a speed of 40 km per hour and returned home at a speed of 80 km per hour. What was my average speed for the whole trip?.
Solution: If we just took the arithmetic mean of the two speeds I drove at, we would get 60 km per hour. This isn’t the correct average speed, however: it ignores the fact that I drove at 40 km per hour for twice as long as I drove at 80 km per hour. To find the correct average speed, we must instead calculate the harmonic mean.
For two quantities A and B, the harmonic mean is given by:
${displaystyle {frac {2}{{frac {1}{A}}+{frac {1}{B}}}}}$
This can be simplified by adding in the denominator and multiplying by the reciprocal:
For N quantities: A, B, C……
Harmonic mean =
${displaystyle {frac {N}{{frac {1}{A}}+{frac {1}{B}}+{frac {1}{C}}+ldots }}}$
Let us try out the formula above on our example:
Harmonic mean =
${displaystyle {frac {2AB}{A+B}}}$
Our values are A = 40, B = 80. Therefore, harmonic mean
${displaystyle ={frac {2times 40times 80}{40+80}}={frac {6400}{120}}approx 53.333}$
Is this result correct? We can verify it. In the example above, the distance between the two towns is 40 km. So the trip from A to B at a speed of 40 km will take 1 hour. The trip from B to A at a speed to 80 km will take 0.5 hours. The total time taken for the round distance (80 km) will be 1.5 hours. The average speed will then be
${displaystyle {frac {80}{1.5}}approx }$53.33 km/hour.
The harmonic mean also has physical significance.
Relationships among Arithmetic, Geometric and Harmonic Mean[edit]
The Means mentioned above are realizations of the generalized mean
and ordered this way:
Measures of dispersion[edit]
Range of Data[edit]
The range of a sample (set of data) is simply the maximum possible difference in the data, i.e. the difference between the maximum and the minimum values. A more exact term for it is “range width” and is usually denoted by the letter R or w. The two individual values (the max. and min.) are called the “range limits”. Often these terms are confused and students should be careful to use the correct terminology.
For example, in a sample with values 2 3 5 7 8 11 12, the range is 11 (122+1=11) and the range limits are 2 and 12.
The range is the simplest and most easily understood measure of the dispersion (spread) of a set of data, and though it is very widely used in everyday life, it is too rough for serious statistical work. It is not a “robust” measure, because clearly the chance of finding the maximum and minimum values in a population depends greatly on the size of the sample we choose to take from it and so its value is likely to vary widely from one sample to another. Furthermore, it is not a satisfactory descriptor of the data because it depends on only two items in the sample and overlooks all the rest. A far better measure of dispersion is the standard deviation (s), which takes into account all the data. It is not only more robust and “efficient” than the range, but is also amenable to far greater statistical manipulation. Nevertheless the range is still much used in simple descriptions of data and also in quality control charts.
The mean range of a set of data is however a quite efficient measure (statistic) and can be used as an easy way to calculate s. What we do in such cases is to subdivide the data into groups of a few members, calculate their average range,
${displaystyle {bar {R}}}$and divide it by a factor (from tables), which depends on n. In chemical laboratories for example, it is very common to analyse samples in duplicate, and so they have a large source of ready data to calculate s.
(The factor k to use is given under standard deviation.)
For example:
If we have a sample of size 40, we can divide it into 10 subsamples of n=4 each.
If we then find their mean range to be, say, 3.1, the standard deviation of the parent sample of 40 items is appoximately 3.1/2.059 = 1.506.
With simple electronic calculators now available, which can calculate s directly at the touch of a key, there is no longer much need for such expedients, though students of statistics should be familiar with them.
{tBookCat}}
Quartiles[edit]
The quartiles of a data set are formed by the two boundaries on either side of the median, which divide the set into four equal sections. The lowest 25% of the data being found below the first quartile value, also called the lower quartile (Q1). The median, or second quartile divides the set into two equal sections. The lowest 75% of the data set should be found below the third quartile, also called the upper quartile (Q3). These three numbers are measures of the dispersion of the data, while the mean, median and mode are measures of central tendency.
Examples[edit]
Given the set {1,3,5,8,9,12,24,25,28,30,41,50} we would find the first and third quartiles as follows:
There are 12 elements in the set, so 12/4 gives us three elements in each quarter of the set.
So the first or lowest quartile is: 5, the second quartile is the median 12, and the third or upper quartile is 28.
However some people when faced with a set with an even number of elements (values) still want the true median (or middle value), with an equal number of data values on each side of the median (rather than 12 which has 5 values less than and 6 values greater than. This value is then the average of 12 and 24 resulting in 18 as the true median (which is closer to the mean of 19 2/3. The same process is then applied to the lower and upper quartiles, giving 6.5, 18, and 29. This is only an issue if the data contains an even number of elements with an even number of equally divided sections, or an odd number of elements with an odd number of equally divided sections.
InterQuartile Range[edit]
The inter quartile range is a statistic which provides information about the spread of a data set, and is calculated by subtracting the first quartile from the third quartile), giving the range of the middle half of the data set, trimming off the lowest and highest quarters. Since the IQR is not affected at all by outliers in the data, it is a more robust measure of dispersion than the range
IQR = Q3 – Q1
Another useful quantile is the quintiles which subdivide the data into five equal sections. The advantage of quintiles is that there is a central one with boundaries on either side of the median which can serve as an average group. In a Normal distribution the boundaries of the quintiles have boundaries ±0.253*s and ±0.842*s on either side of the mean (or median),where s is the sample standard deviation. Note that in a Normal distribution the mean, median and mode coincide.
Other frequently used quantiles are the deciles (10 equal sections) and the percentiles (100 equal sections)
Variance and Standard Deviation[edit]
Measure of Scale[edit]
When describing data it is helpful (and in some cases necessary) to determine the spread of a distribution. One way of measuring this spread is by calculating the variance or the standard deviation of the data.
In describing a complete population, the data represents all the elements of the population. As a measure of the “spread” in the population one wants to know a measure of the possible distances between the data and the population mean. There are several options to do so. One is to measure the average absolute value of the deviations. Another, called the variance, measures the average square of these deviations.
A clear distinction should be made between dealing with the population or with a sample from it. When dealing with the complete population the (population) variance is a constant, a parameter which helps to describe the population. When dealing with a sample from the population the (sample) variance is actually a random variable, whose value differs from sample to sample. Its value is only of interest as an estimate for the population variance.
Population variance and standard deviation[edit]
Let the population consist of the N elements x_{1},…,x_{N}. The (population) mean is:
The (population) variance σ^{2} is the average of the squared deviations from the mean or (x_{i} – μ)^{2} – the square of the value’s distance from the distribution’s mean.
Because of the squaring the variance is not directly comparable with the mean and the data themselves. The square root of the variance is called the Standard Deviation σ. Note that σ is the root mean squared of differences between the data points and the average.
Sample variance and standard deviation[edit]
Let the sample consist of the n elements x_{1},…,x_{n}, taken from the population. The (sample) mean is:
The sample mean serves as an estimate for the population mean μ.
The (sample) variance s^{2} is a kind of average of the squared deviations from the (sample) mean:
Also for the sample we take the square root to obtain the (sample) standard deviation s
A common question at this point is “why do we square the numerator?” One answer is: to get rid of the negative signs. Numbers are going to fall above and below the mean and, since the variance is looking for distance, it would be counterproductive if those distances factored each other out.
Example[edit]
When rolling a fair die, the population consists of the 6 possible outcomes 1 to 6. A sample may consist instead of the outcomes of 1000 rolls of the die.
The population mean is:
and the population variance:
The population standard deviation is:
Notice how this standard deviation is somewhere in between the possible deviations.
So if we were working with one sixsided die: X = {1, 2, 3, 4, 5, 6}, then σ^{2} = 2.917. We will talk more about why this is different later on, but for the moment assume that you should use the equation for the sample variance unless you see something that would indicate otherwise.
Note that none of the above formulae are ideal when calculating the estimate and they all introduce rounding errors. Specialized statistical software packages use more complicated logarithms that take a second pass of the data in order to correct for these errors. Therefore, if it matters that your estimate of standard deviation is accurate, specialized software should be used. If you are using nonspecialized software, such as some popular spreadsheet packages, you should find out how the software does the calculations and not just assume that a sophisticated algorithm has been implemented.
For Normal Distributions[edit]
The empirical rule states that approximately 68 percent of the data in a normally distributed dataset is contained within one standard deviation of the mean, approximately 95 percent of the data is contained within 2 standard deviations, and approximately 99.7 percent of the data falls within 3 standard deviations.
As an example, the verbal or math portion of the SAT has a mean of 500 and a standard deviation of 100. This means that 68% of testtakers scored between 400 and 600, 95% of test takers scored between 300 and 700, and 99.7% of testtakers scored between 200 and 800 assuming a completely normal distribution (which isn’t quite the case, but it makes a good approximation).
Robust Estimators[edit]
For a normal distribution the relationship between the standard deviation and the interquartile range is roughly: SD = IQR/1.35.
For data that are nonnormal, the standard deviation can be a terrible estimator of scale. For example, in the presence of a single outlier, the standard deviation can grossly overestimate the variability of the data. The result is that confidence intervals are too wide and hypothesis tests lack power. In some (or most) fields, it is uncommon for data to be normally distributed and outliers are common.
One robust estimator of scale is the “average absolute deviation”, or aad. As the name implies, the mean of the absolute deviations about some estimate of location is used. This method of estimation of scale has the advantage that the contribution of outliers is not squared, as it is in the standard deviation, and therefore outliers contribute less to the estimate. This method has the disadvantage that a single large outlier can completely overwhelm the estimate of scale and give a misleading description of the spread of the data.
Another robust estimator of scale is the “median absolute deviation”, or mad. As the name implies, the estimate is calculated as the median of the absolute deviation from an estimate of location. Often, the median of the data is used as the estimate of location, but it is not necessary that this be so. Note that if the data are nonnormal, the mean is unlikely to be a good estimate of location.
It is necessary to scale both of these estimators in order for them to be comparable with the standard deviation when the data are normally distributed. It is typical for the terms aad and mad to be used to refer to the scaled version. The unscaled versions are rarely used.
External links[edit]
Other summaries[edit]
Moving Average[edit]
A moving average is used when you want to get a general picture of the trends contained in a data set. The data set of concern is typically a socalled “time series”, i.e a set of observations ordered in time. Given such a data set X, with individual data points
${displaystyle x_{i}}$, a 2n+1 point moving average is defined as
${displaystyle {bar {x_{i}}}={frac {1}{2n+1}}sum _{k=in}^{i+n}x_{k}}$, and is thus given by taking the average of the 2n points around
${displaystyle x_{i}}$. Doing this on all data points in the set (except the points too close to the edges) generates a new time series that is somewhat smoothed, revealing only the general tendencies of the first time series.
The moving average for many timebased observations is often lagged. That is, we take the 10 day moving average by looking at the average of the last 10 days. We can make this more exciting (who knew statistics was exciting?) by considering different weights on the 10 days. Perhaps the most recent day should be the most important in our estimate and the value from 10 days ago would be the least important. As long as we have a set of weights that sums to 1, this is an acceptable movingaverage. Sometimes the weights are chosen along an exponential curve to make the exponential movingaverage.
A single statistic tells only part of a dataset’s story. The mean is one perspective; the median yet another. And when we explore relationships between multiple variables, even more statistics arise. The coefficient estimates in a regression model, the CochranMaentelHaenszel test statistic in partial contingency tables; a multitude of statistics are available to summarize and test data.
But our ultimate goal in statistics is not to summarize the data, it is to fully understand their complex relationships. A well designed statistical graphic helps us explore, and perhaps understand, these relationships.
This section will help you let the data speak, so that the world may know its story.
Types of Data Display
 Bar_Charts
 Histograms
 Scatter_Graphs
 Box_Plots
 Pie_Charts
 Comparative_Pie_Charts
 Pictograms
 Line_Graphs
 Frequency_Polygon
External Links
 [1] “Show me the Numbers” by Stephen Few has a less technical approach to creating graphics. You might want to scan through this book if you are building a library on making graphs.
Statistics  >> Bar Charts
Bar Charts[edit]
The Bar Chart (or Bar Graph) is one of the most common ways of displaying catagorical/qualitative data. Bar Graphs consist of 2 variables, one response (sometimes called “dependent”) and one predictor (sometimes called “independent”), arranged on the horizontal and vertical axis of a graph. The relationship of the predictor and response variables is shown by a mark of some sort (usually a rectangular box) from one variable’s value to the other’s.
To demonstrate we will use the following data(tbl. 3.1.1) representing a hypothetical relationship between a qualitative predictor variable, “Graph Type”, and a quantitative response variable, “Votes”.
tbl. 3.1.1 – Favourite Graphs
Graph Type  Votes 

Bar Charts  5 
Pie Graphs  2 
Histograms  3 
Pictograms  8 
Comp. Pie Graphs  4 
Line Graphs  9 
Frequency Polygon  1 
Scatter Graphs  5 
From this data we can now construct an appropriate graphical representation which, in this case will be a Bar Chart. The graph may be orientated in several ways, of which the vertical chart (fig. 3.1.1) is most common, with the horizontal chart(fig. 3.1.2) also being used often
fig. 3.1.1 – vertical chart
fig. 3.1.2 – horizontal chart
Take note that the height and width of the bars, in the vertical and horizontal Charts, respectfully, are equal to the response variable’s corresponding value – “Bar Chart” bar equals the number of votes that the Bar Chart type received in tbl. 3.1.1
Also take note that there is a pronounced amount of space between the individual bars in each of the graphs, this is important in that it help differentiate the Bar Chart graph type from the Histogram graph type discussed in a later section.
External Links[edit]
Histograms[edit]
It is often useful to look at the distribution of the data, or the frequency with which certain values fall between preset bins of specified sizes. The selection of these bins is up to you, but remember that they should be selected in order to illuminate your data, not obfuscate it.
A histogram is similar to a bar chart. However histograms are used for continuous (as opposed to discrete or qualitative) data. The defining property of a histogram is:
 The area of each bar is proportional to the frequency.
If each bin has an equal width, then this can be easily done by plotting frequency on the vertical axis. However histograms can also be drawn with unequal bin sizes, for which one can plot frequency density.
To produce a histogram with equal bin sizes:
 Select a minimum, a maximum, and a bin size. All three of these are up to you. In the histogram data used above the minimum is 1, the maximum is 110, and the bin size is 10.
 Calculate your bins and how many values fall into each of them. For the histogram data the bins are:
 1 ≤ x < 10, 16 values.
 10 ≤ x < 20, 4 values.
 20 ≤ x < 30, 4 values.
 30 ≤ x < 40, 2 values.
 40 ≤ x < 50, 2 values.
 50 ≤ x < 60, 1 values.
 60 ≤ x < 70, 0 values.
 70 ≤ x < 80, 0 values.
 80 ≤ x < 90, 0 values.
 90 ≤ x < 100, 0 value.
 100 ≤ x < 110, 0 value.
 110 ≤ x < 120, 1 value.
 Plot the counts you figured out above. Do this using a standard bar plot.
Worked Problem[edit]
Let’s say you are an avid roleplayer who loves to play Mechwarrior, a d6 (6 sided die) based game. You have just purchased a new 6 sided die and would like to see whether it is biased (in combination with you when you roll it).
What We Expect[edit]
So before we look at what we get from rolling the die, let’s look at what we would expect. First, if a die is unbiased it means that the odds of rolling a six are exactly the same as the odds of rolling a 1–there wouldn’t be any favoritism towards certain values. Using the standard equation for the arithmetic mean find that μ = 3.5. We would also expect the histogram to be roughly even all of the way across–though it will almost never be perfect simply because we are dealing with an element of random chance.
What We Get[edit]
Here are the numbers that you collect:
1  5  6  4  1  3  5  5  6  4  1  5  6  6  4  5  1  4  3  6 
1  3  6  4  2  4  1  6  4  2  2  4  3  4  1  1  6  3  5  5 
4  3  5  3  4  2  2  5  6  5  4  3  5  3  3  1  5  4  4  5 
1  2  5  1  6  5  4  3  2  4  2  1  3  3  3  4  6  1  1  3 
6  6  1  4  6  6  6  5  3  1  5  6  3  4  5  5  5  2  4  4 
Analysis[edit]
Referring back to what we would expect for an unbiased die, this is pretty close to what we would expect. So let’s create a histogram to see if there is any significant difference in the distribution.
The only logical way to divide up dice rolls into bins is by what’s showing on the die face:
1  2  3  4  5  6 
16  9  17  21  20  17 
If we are good at visualizing information, we can simple use a table, such as in the one above, to see what might be happening. Often, however, it is useful to have a visual representation. As the amount of variety of data we want to display increases, the need for graphs instead of a simple table increases.
Looking at the above figure, we clearly see that sides 1, 3, and 6 are almost exactly what we would expect by chance. Sides 4 and 5 are slightly greater, but not too much so, and side 2 is a lot less. This could be the result of chance, or it could represent an actual anomaly in the data and it is something to take note of keep in mind. We’ll address this issue again in later chapters.
Frequency Density[edit]
Another way of drawing a histogram is to work out the Frequency Density.
 Frequency Density
 The Frequency Density is the frequency divided by the class width.
The advantage of using frequency density in a histogram is that doesn’t matter if there isn’t an obvious standard width to use. For all the groups, you would work out the frequency divided by the class width for all of the groups.
External Links
Return to Statistics.
Scatter Plots[edit]
Scatter Plot is used to show the relationship between 2 numeric variables. It is not useful when comparing discrete variables versus numeric variables. A scatter plot matrix is a collection of pairwise scatter plots of numeric variables.
External Links
Box Plots[edit]
A box plot (also called a box and whisker diagram) is a simple visual representation of key features of a univariate sample.
The box lies on a vertical axis in the range of the sample. Typically, a top to the box is placed at the 1st quartile, the bottom at the third quartile. The width of the box is arbitrary, as there is no xaxis (though see Violin Plots, below).
In between the top and bottom of the box is some representation of central tendency. A common version is to place a horizontal line at the median, dividing the box into two. Additionally, a star or asterisk is placed at the mean value, centered in the box in the horizontal direction.
Another common extension is to the ‘boxandwhisker’ plot. This adds vertical lines extending from the top and bottom of the plot to for example, the maximum and minimum values, The farthest value within 2 standard deviations above and below the mean. Alternatively, the whiskers could extend to the 2.5 and 97.5 percentiles. Finally, it is common in the boxandwhisker plot to show outliers (however defined) with asterisks at the individual values beyond the ends of the whiskers.
Violin Plots are an extension to box plots using the horizontal information to present more data. They show some estimate of the CDF instead of a box, though the quantiles of the distribution are still shown.
Pie Charts[edit]
A PieChart/Diagram is a graphical device – a circular shape broken into subdivisions. The subdivisions are called “sectors“, whose areas are proportional to the various parts into which the whole quantity is divided. The sectors may be coloured differently to show the relationship of parts to the whole. A pie diagram is an alternative of the subdivided bar diagram.
To construct a piechart, first we draw a circle of any suitable radius then the whole quantity which is to be divided is equated to 360 degrees. The different parts of the circle in terms of angles are calculated by the following formula.
Component Value / Whole Quantity * 360
The component parts i.e. sectors have been cut beginning from top in clockwise order.
Note that the percentages in a list may not add up to exactly 100% due to rounding. For example if a person spends a third of their time on each of three activities: 33%, 33% and 33% sums to 99%.
Warning: Pie charts are a poor way of communicating information. The eye is good at judging linear measures and bad at judging relative areas. A bar chart or dot chart is a preferable way of displaying this type of data.
Cleveland (1985), page 264: “Data that can be shown by pie charts always can be shown by a dot chart. This means that judgements of position along a common scale can be made instead of the less accurate angle judgments.” This statement is based on the empirical investigations of Cleveland and McGill as well as investigations by perceptual psychologists.
Threedimensional (3d) pie charts compound perceptual misinterpretation of statistical information by altering the relative angle of pie slices to create the impression of depth into a vanishing point. Angles and areas at the bottom of the chart must be exaggerated and the angles and areas at the top of the chart reduced in order to create the dimensional effect; a specifically false depiction of the data.
External Links
Comparative Pie Charts[edit]
The comparative pie charts are very difficult to read and compare if the ratio of the pie chart is not given.
Examine our example of color preference for two different groups. How much work does it take to see that it is quite challenging to work out who ate the pie? First, we have to find Fingerprints on either pie, and then remember how many sensirivity vectors it has. If we did not include the share for blue in the label, then we would probably be approximating the comparison. So, if we use multiple pie charts, we have to expect that comparisions between charts would only be approximate.
What is the most popular color in the left graph? Red. But note, that you have to look at all of the colors and read the label to see which it might be. Also, this author was kind when creating these two graphs because he used the same color for the same object. Imagine the confusion if one had made the most important color get Red in the righthand chart?
If two shares of data should not be compared via the comparative pie chart, what kind of graph would be preferred? The stacked bar chart is probably the most appropriate for sharing of the total comparisons. Again, exact comparisons cannot be done with graphs and therefore a table may supplement the graph with detailed information.
Pictograms[edit]
A pictogram is simply a picture that conveys some statistical information. A very common example is the thermometer graph so common in fund drives. The entire thermometer is the goal (number of dollars that the fund raisers wish to collect. The red stripe (the “mercury”) represents the proportion of the goal that has already been collected.
Another example is a picture that represents the gender constitution of a group. Each small picture of a male figure might represent 1,000 men and each small picture of a female figure would, then, represent 1,000 women. A picture consisting of 3 male figures and 4 female figures would indicate that the group is made up of 3,000 men and 4,000 women.
An interesting pictograph is the Chernoff Faces. It is useful for displaying information on cases for which several variables have been recorded. In this kind of plot, each case is represented by a separate picture of a face. The sizes of the various features of each face are used to present the value of each variable. For instance, if blood pressure, high density cholesterol, low density cholesterol, body temperature, height, and weight are recorded for 25 individuals, 25 faces would be displayed. The size of the nose on each face would represent the level of that person’s blood pressure. The size of the left eye may represent the level of low density cholesterol while the size of the right eye might represent the level of high density cholesterol. The length of the mouth could represent the person’s temperature. The length of the left ear might indicate the person’s height and that of the right ear might represent their weight. Of course, a legend would be provided to help the viewer determine what feature relates to which variable. Where it would be difficult to represent the relationship of all 6 variables on a single (6dimensional) graph, the Chernoff Faces would give a relatively easy to interpret 6dimensional representation.
Line Graphs[edit]
Basically, a line graph can be, for example, a picture of what happened by/to something (a variable) during a specific time period (also a variable).
On the left side of such a graph usually is as an indication of that “something” in the form of a scale, and at the bottom is an indication of the specific time involved.
Usually a line graph is plotted after a table has been provided showing the relationship between the two variables in the form of pairs. Just as in (x,y) graphs, each of the pairs results in a specific point on the graph, and being a LINE graph these points are connected to one another by a LINE.
Many other line graphs exist; they all CONNECT the points by LINEs, not necessarily straight lines. Sometimes polynomials, for example, are used to describe approximately the basic relationship between the given pairs of variables, and between these points. The higher the degree of the polynomial, the more accurate is the “picture” of that relationship, but the degree of that polynomial must never be higher than n1, where n is the number of the given points.
See also
Graph theory
[[../../Curve fittingCurve fitting]]
From Wikipedia: Line graph and Curve fitting
External Links
Frequency Polygon[edit]
Midpoints of the interval of corresponding rectangle in a histogram are joined together by straight lines. It gives a polygon i.e. a figure with many angles.
It is used when two or more sets of data are to be illustrated on the same diagram such as death rates in smokers and nonsmokers, birth and death rates of a population etc.
One way to form a frequency polygon is to connect the midpoints at the top of the bars of a histogram with line segments (or a smooth curve). Of course the midpoints themselves could easily be plotted without the histogram and be joined by line segments. Sometimes it is beneficial to show the histogram and frequency polygon together.But sometimes, the frequency polygon is much more accurate than the histogram because you can evaluate which is the low point and the high point.
Unlike histograms, frequency polygons can be superimposed so as to compare several frequency distributions.
Frequency polygons were created in the 19th Century^{[1]} as a way of not only storing data^{[citation needed]}, but making it easily accessible for people who are illiterate^{[citation needed]}.
Probability is connected with some unpredictability. We know what outcomes may occur, but not exactly which one. The set of possible outcomes plays a basic role. We call it the sample space and indicate it by S. Elements of S are called outcomes. In rolling a dice the sample space is S = {1,2,3,4,5,6}. Not only do we speak of the outcomes, but also about events, sets of outcomes (or subsets of the sample space). E.g. in rolling a dice we can ask whether the outcome was an even number, which means asking after the event “even” = E = {2,4,6}. In simple situations with a finite number of outcomes, we assign to each outcome s (∈ S) its probability (of occurrence) p(s) (written with a small p), a number between 0 and 1. It is a quite simple function, called the probability function, with the only further property that the total of all the probabilities sum up to 1. Also for events A do we speak of their probability P(A) (written with a capital P), which is simply the total of the probabilities of the outcomes in A. For a fair dice p(s) = 1/6 for each outcome s and P(“even”) = P(E) = 1/6+1/6+1/6 = 1/2.
The general concept of probability for nonfinite sample spaces is a little more complex, although it rests on the same ideas.
Introduction[edit]
Why have probability in a statistics textbook?[edit]
Very little in mathematics is truly self contained. Many branches of mathematics touch and interact with one another, and the fields of probability and statistics are no different. A basic understanding of probability is vital in grasping basic statistics, and probability is largely abstract without statistics to determine the “real world” probabilities.
This section is not meant to give a comprehensive lecture in probability, but rather simply touch on the basics that are needed for this class, covering the basics of Bayesian Analysis for those students who are looking for something a little more interesting. This knowledge will be invaluable in attempting to understand the mathematics involved in various Distributions that come later.
Set notion[edit]
A set is a collection of objects. We usually use capital letters to denote sets, e.g. A is the set of females in this room.
 The members of a set A are called the elements of A, e.g. Patricia is an element of A (Patricia ∈ A); Patrick is not an element of A (Patrick ∉ A).
 The universal set, U, is the set of all objects under consideration, e.g., U is the set of all people in this room.
 The null set or empty set, ∅, has no elements, e.g., the set of males above 2.8m tall in this room is an empty set.
 The complement A^{c} of a set A is the set of elements in U outside A, i.e. x ∈ A^{c} iff x ∉ A.
 Let A and B be 2 sets. A is a subset of B if each element of A is also an element of B. Write A ⊂ B, e.g. the set of females wearing metal frame glasses in this room ⊂ the set of females wearing glasses in this room ⊂ the set of females in this room.
• The intersection A ∩ B of two sets A and B is the set of the common elements. I.e. x ∈ A ∩ B iff x ∈ A and x ∈ B.
• The union A ∪ B of two sets A and B is the set of all elements from A or B. I.e. x ∈ A ∪ B iff x ∈ A or x ∈ B.
Venn diagrams and notation[edit]
A Venn diagram visually models defined events. Each event is expressed with a circle. Events that have outcomes in common will overlap with what is known as the intersection of the events.
Probability Axioms[edit]
Calculating Probability[edit]
Negation[edit]
Negation is a way of saying “not A“, hence saying that the complement of A has occurred. Note: The complement of an event A can be expressed as A’ or A^{c}
For example: “What is the probability that a sixsided die will not land on a one?” (five out of six, or p = 0.833)
Or, more colloquially, “the probability of ‘not X’ together with the probability of ‘X’ equals one or 100%.”
Relative frequency describes the number of successes over the total number of outcomes. For example if a coin is flipped and out of 50 flips 29 are heads then the relative frequency is
${displaystyle {frac {29}{50}}}$
The Union of two events is when you want to know Event A OR Event B.
This is different from “And.” “And” is the intersection, whereas “OR” is the union of the events (both events put together).
In the above example of events you will notice that…
Event A is a STAR and a DIAMOND.
Event B is a TRIANGLE and a PENTAGON and a STAR
(A ∩ B) = (A and B) = A intersect B is only the STAR
But (A ∪ B) = (A or B) = A Union B is EVERYTHING. The TRIANGLE, PENTAGON, STAR, and DIAMOND
Notice that both event A and Event B have the STAR in common. However, when you list the Union of the events you only list the STAR one time!
Event A = STAR, DIAMOND EVENT B = TRIANGLE, PENTAGON, STAR
When you combine them together you get (STAR + DIAMOND) + (TRIANGLE + PENTAGON + STAR) BUT WAIT!!! STAR is listed two times, so one will need to SUBTRACT the extra STAR from the list.
You should notice that it is the INTERSECTION that is listed TWICE, so you have to subtract the duplicate intersection.
Conjunction[edit]
Formula for the Union of Events: P(A ∪ B) = P(A) + P(B) – P(A ∩ B)
Example:
Let P(A) = 0.3 and P(B) = 0.2 and P(A ∩ B) = 0.15. Find P(A ∪ B).
P(A ∪ B) = (0.3) + (0.2) – (0.15) = 0.35
Example:
Let P(A) = 0.3 and P(B) = 0.2 and P(A ∩ B) = 0. Find P(A ∪ B).
Note: Since the intersection of the events is the null set, then you know the events are DISJOINT or MUTUALLY EXCLUSIVE.
P(A ∪ B) = (0.3) + (0.2) – (0) = 0.5
Disjunction[edit]
Law of total probability[edit]
Conditional Probability[edit]
What is the probability of one event given that another event occurs? For example, what is the probability of a mouse finding the end of the maze, given that it finds the room before the end of the maze?
This is represented as:
or “the probability of A given B.”
${displaystyle P(AB)={frac {P(Acap B)}{P(B)}}}$
If A and B are independent of one another, such as with coin tosses or child births, then:
Thus, “what is the probability that the next child a family bears will be a boy, given that the last child is a boy.”
This can also be stacked where the probability of A with several “givens.”
or “the probability of A given that B_{1}, B_{2}, and B_{3} are true?”
–>
Conclusion: putting it all together[edit]
Bernoulli Trials[edit]
A lot of experiments just have two possible outcomes, generally referred to as “success” and “failure”. If such an experiment is independently repeated we call them (a series of) Bernoulli trials. Usually the probability of success is called p. The repetition may be done in several ways:
 a fixed number of times (n); as a consequence the observed number of successes is stochastic;
 until a fixed number of successes (m) is observed; as a consequence the number of experiments is stochastic;
In the first case the number of successes is Binomial distributed with parameters n and p. For n=1 the distribution is also called the Bernoulli distribution.
In the second case the number of experiments is Negative Binomial distributed with parameters m and p. For m=1 the distribution is also called the Geometric distribution.
Bayesian analysis is the branch of statistics based on the idea that we have some knowledge in advance about the probabilities that we are interested in, so called a priori probabilities. This might be your degree of belief in a particular event, the results from previous studies, or a general agreedupon starting value for a probability. The terminology “Bayesian” comes from the Bayesian rule or law, a law about conditional probabilities. The opposite of “Bayesian” is sometimes referred to as “Classical Statistics.”
Example[edit]
Consider a box with 3 coins, with probabilities of showing heads respectively 1/4, 1/2 and 3/4. We choose arbitrarily one of the coins. Hence we take 1/3 as the a priori probability
${displaystyle P(C_{1})}$of having chosen coin number 1. After 5 throws, in which X=4 times heads came up, it seems less likely that the coin is coin number 1. We calculate the a posteriori probability that the coin is coin number 1, as:
In words:
 The probability that the Coin is the first Coin, given that we know heads came up 4 times… Is equal to the probability that heads came up 4 times given we know it’s the first coin, times the probability that the coin is the first coin. All divided by the probability that heads comes up 4 times (ignoring which of the three Coins is chosen). The binomial coefficients cancel out as well as all denominators when expanding 1/2 to 2/4. This results in
In the same way we find:
and
This shows us that after examining the outcome of the five throws, it is most likely we did choose coin number 3.
Actually for a given result the denominator does not matter, only the relative Probabilities
${displaystyle p(C_{i})=P(C_{i}X=4)/P(X=4)}$
When the result is 3 times heads the Probabilities change in favor of Coin 2 and further as the following table shows:
Heads  

5  1  32  243 
4  3  32  81 
3  9  32  27 
2  27  32  9 
1  81  32  3 
0  243  32  1 
How are the results of the latest SAT test? What is the average height of females under 21 in Zambia? How does beer consumption among college students at engineering college compare to college students in liberal arts colleges?
To answer these questions, we would collect data and put them in a form that is easy to summarize, visualize, and discuss. Loosely speaking, the collection and aggregation of data result in a distribution. Distributions are most often in the form of a histogram or a table. That way, we can “see” the data immediately and begin our scientific inquiry.
For example, if we want to know more about students’ latest performance on the SAT, we would collect SAT scores from ETS, compile them in a way that is pertinent to us, and then form a distribution of these scores. The result may be a data table or it may be a plot. Regardless, once we “see” the data, we can begin asking more interesting research questions about our data.
The distributions we create often parallel distributions that are mathematically generated. For example, if we obtain the heights of all high school students and plot this data, the graph may resemble a normal distribution, which is generated mathematically. Then, instead of painstakingly collecting heights of all high school students, we could simply use a normal distribution to approximate the heights without sacrificing too much accuracy.
In the study of statistics, we focus on mathematical distributions for the sake of simplicity and relevance to the realworld. Understanding these distributions will enable us to visualize the data easier and build models quicker. However, they cannot and do not replace the work of manual data collection and generating the actual data distribution.
What percentage lie within a certain range? Distributions show what percentage of the data lies within a certain range. So, given a distribution, and a set of values, we can determine the probability that the data will lie within a certain range.
The same data may lead to different conclusions if it is interposed on different distributions. So, it is vital in all statistical analysis for data to be put onto the correct distribution.
Distributions
 Probability Distributions
 Discrete Distributions
 Uniform Distribution
 Bernoulli Distribution
 Binomial Distribution
 Poisson Distribution
 Geometric Distribution
 Negative Binomial Distribution
 Hypergeometric Distribution
 Continuous Distributions
 Uniform Distribution
 Exponential Distribution
 Gamma Distribution
 Normal Distribution
 ChiSquare Distribution
 Studentt Distribution
 F Distribution
 Beta Distribution
 Weibull Distribution
 Gumbel Distribution
Comparison of Some Distributions[edit]
Name  Notation  Formula  Symbols  Use  Continuous/ discrete  Notes 

Bernoulli  f(x)=  p x 
2 outcomes  Discrete  1 trial  
Binomial  b(x;n, p)=  n trials k successes p probability 
number of times success specific probabilities not random 
Discrete  
Poisson  P(x)=  outcome/time outcome/region 
Discrete  
Hypergeometric  h(x;N,n,k) =  n samples from N items k of N items are successes, Nk are failures 
X times success occurs irregardless of location is random 
Discrete  Without Replacement  
Multivariate Hypergeometric 
n sample size N items k cells 
Discrete  Without Replacement  
Normal  x 
:  Continuous  Z is a random variable with 

ChiSquare  variances of random sample related to the pop  Continuous  
Studentt  T=  If don’t know  Continuous  v=n1  
F  F=  Continuous 
Discrete Distributions[edit]
‘Discrete’ data are data that assume certain discrete and quantized values. For example, truefalse answers are discrete, because there are only two possible choices. Valve settings such as ‘high/medium/low’ can be considered as discrete values. As a general rule, if data can be counted in a practical manner, then they can be considered to be discrete.
To demonstrate this, let us consider the population of the world. It is a discrete number because the number of civilians is theoretically countable. But since this is not practicable, statisticians often treat this data as continuous. That is, we think of population as within a range of numbers rather than a single point.
For the curious, the world population is 6,533,596,139 as of August 9, 2006. Please note that statisticians did not arrive at this figure by counting individual residents. They used much smaller samples of the population to estimate the whole. Going back to Chapter 1, this is a great reason to learn statistics – we need only a smaller sample of data to make intelligent descriptions of the entire population!
Discrete distributions result from plotting the frequency distribution of data which is discrete in nature.
General Properties[edit]
Cumulative Distribution Function[edit]
A discrete random variable has a cumulative distribution function that describes the probability that the random variable is below the point. The cumulative distribution must increase towards 1. Depending on the random variable, it may reach one at a finite number, or it may not. The cdf is represented by a capital F.
ń====Probability Mass Function====
A discrete random variable has a probability mass function that describes how likely the random variable is to be at a certain point. The probability mass function must have a total of 1, and sums to the cdf. The pmf is represented by the lowercase f.
Special Values[edit]
The expected value of a discrete variable is
The expected value of any function of a discrete variable g(
${displaystyle X}$) is
The variance is equal to
External Links
Simulating binomial, hypergeometric, and the Poisson distribution: Discrete Distributions
Discrete Uniform Distribution[edit]
The discrete uniform distribution (not to be confused with the continuous uniform distribution) is where the probability of equally spaced possible values is equal. Mathematically this means that the probability density function is identical for a finite set of evenly spaced points. An example of would be rolling a fair 6sided die. In this case there are six, equally like probabilities.
One common normalization is to restrict the possible values to be integers and the spacing between possibilities to be 1. In this setup, the only two parameters of the function are the minimum value (a), the maximum value (b). (Some even normalize it more, setting a=1.) Let n=ba+1 be the number of possibilities. The probability density function is then
${displaystyle fcolon {a,a+1,ldots ,b1,b}to mathbb {R} }$
${displaystyle fleft(xright)={frac {1}{n}}}$
Mean[edit]
Let
${displaystyle S={a,a+1,ldots ,b1,b}}$. The mean (notated as
${displaystyle operatorname {E} [X]}$) can then be derived as follows:
Remember that in
${displaystyle sum _{i=0}^{m}i=(m^{2}+m)/2}$
Variance[edit]
The variance (
${displaystyle E[(XEX)^{2}]}$) can be derived:
Remember that in
${displaystyle sum _{i=0}^{m}i^{2}=m(m+1)(2m+1)/6}$
Bernoulli Distribution: The coin toss[edit]
Parameters  

Support  
PMF  
CDF  
Mean  
Median  
Mode  
Variance  
Skewness  
Ex. kurtosis  
Entropy  
MGF  
CF  
PGF  
Fisher information 
There is no more basic random event than the flipping of a coin. Heads or tails. It’s as simple as you can get! The “Bernoulli Trial” refers to a single event which can have one of two possible outcomes with a fixed probability of each occurring. You can describe these events as “yes or no” questions. For example:
 Will the coin land heads?
 Will the newborn child be a girl?
 Are a random person’s eyes green?
 Will a mosquito die after the area was sprayed with insecticide?
 Will a potential customer decide to buy my product?
 Will a citizen vote for a specific candidate?
 Is an employee going to vote prounion?
 Will this person be abducted by aliens in their lifetime?
The Bernoulli Distribution has one controlling parameter: the probability of success. A “fair coin” or an experiment where success and failure are equally likely will have a probability of 0.5 (50%). Typically the variable p is used to represent this parameter.
If a random variable X is distributed with a Bernoulli Distribution with a parameter p we write its probability mass function as:
Where the event X=1 represents the “yes.”
This distribution may seem trivial, but it is still a very important building block in probability. The Binomial distribution extends the Bernoulli distribution to encompass multiple “yes” or “no” cases with a fixed probability. Take a close look at the examples cited above. Some similar questions will be presented in the next section which might give an understanding of how these distributions are related.
Mean[edit]
The mean (E[X]) can be derived:
Variance[edit]
External links[edit]
Where the Bernoulli Distribution asks the question of “Will this single event succeed?” the Binomial is associated with the question “Out of a given number of trials, how many will succeed?” Some example questions that are modeled with a Binomial distribution are:
 Out of ten tosses, how many times will this coin land heads?
 From the children born in a given hospital on a given day, how many of them will be girls?
 How many students in a given classroom will have green eyes?
 How many mosquitos, out of a swarm, will die when sprayed with insecticide?
The relation between the Bernoulli and binomial distributions is intuitive: The binomial distribution is composed of multiple Bernoulli trials. We conduct
${displaystyle n}$repeated experiments where the probability of success is given by the parameter
${displaystyle p}$and add up the number of successes. This number of successes is represented by the random variable X. The value of X is then between 0 and
${displaystyle n}$.
When a random variable X has a binomial distribution with parameters
${displaystyle p}$and
${displaystyle n}$we write it as X ~ Bin(n,p) or X ~ B(n,p) and the probability mass function is given by the equation:
where
${displaystyle {n choose k}={n! over k!(nk)!}}$
For a refresher on factorials (n!), go back to the Refresher Course earlier in this wiki book.
An example[edit]
Let’s walk through a simple example of the binomial distribution. We’re going to use some pretty small numbers because factorials can be hard to compute. We are going to ask five random people if they believe there is life on other planets. We are going to assume in this example that we know 30% of people believe this to be true. We want to ask the question: “How many people will say they believe in extraterrestrial life?” Actually, we want to be more specific than that: “What is the probability that exactly 2 people will say they believe in extraterrestrial life?”
We know all the values that we need to plug into the equation. The number of people asked, n=5. The probability of any given person answering “yes”, p=0.3. (Remember, I said that 30% of people believe in life on other planets!) Finally, we’re asking for the probability that exactly 2 people answer “yes” so k=2. This yields the equation:
Here are the probabilities for all the possible values of X. You can get these values by replacing the k=2 in the above equation with all values from 0 to 5.
Value for k  Probability f(k) 
0  0.16807 
1  0.36015 
2  0.30870 
3  0.13230 
4  0.02835 
5  0.00243 
What can we learn from these results? Well, first of all we’ll see that it’s just a little more likely that only one person will confess to believing in life on other planets. There’s a distinct chance (about 17%) that nobody will believe it, and there’s only a 0.24% (a little over 2 in 1000) that all five people will be believers.
Explanation of the equation[edit]
Take the above example. Let’s consider each of the five people one by one.
The probability that any one person believes in extraterrestrial life is 30%, or 0.3. So the probability that any two people both believe in extraterrestrial life is 0.3 squared. Similarly, the probability that any one person does not believe in extraterrestrial life is 70%, or 0.7, so the probability that any three people do not believe in extraterrestrial life is 0.7 cubed.
Now, for two out of five people to believe in extraterrestrial life, two conditions must be satisfied: two people believe in extraterrestrial life, and three do not. The probability of two out of five people believing in extraterrestrial life would thus appear to be 0.3 squared (two believers) times 0.7 cubed (three nonbelievers), or 0.03087.
However, in doing this, we are only considering the case whereby the first two selected people are believers. How do we consider cases such as that in which the third and fifth people are believers, which would also mean a total of two believers out of five?
The answer lies in combinatorics. Bearing in mind that the probability that the first two out of five people believe in extraterrestrial life is 0.03087, we note that there are C(5,2), or 10, ways of selecting a set of two people from out of a set of five, i.e. there are ten ways of considering two people out of the five to be the “first two”. This is why we multiply by C(n,k). The probability of having any two of the five people be believers is ten times 0.03087, or 0.3087.
feruzi ^{[2]}[edit]
The mean can be derived as follow.
Now let w=x1 and m=n1. We see that mw=nx. We can now rewrite the summation as
We now see that the summation is the sum over the complete pmf of a binomial random variable distributed Bin(m, p). This is equal to 1 (and can be easily verified using the Binomial theorem). Therefore, we have
Variance[edit]
We derive the variance using the following formula:
We have already calculated E[X] above, so now we will calculate E[X^{2}] and then return to this variance formula:
We can use our experience gained above in deriving the mean. We use the same definitions of m and w.
The first sum is identical in form to the one we calculated in the Mean (above). It sums to mp.
The second sum is 1.
Using this result in the expression for the variance, along with the Mean (E(X) = np), we get
External links[edit]
Poisson Distribution[edit]
Probability mass function The horizontal axis is the index k, the number of occurrences. The function is only defined at integer values of k. The connecting lines are only guides for the eye. 

Cumulative distribution function The horizontal axis is the index k, the number of occurrences. The CDF is discontinuous at the integers of k and flat everywhere else because a variable that is Poisson distributed only takes on integer values. 

Notation  

Parameters  λ > 0 (real) 
Support  k ∈ { 0, 1, 2, 3, … } 
PMF  
CDF  
Mean  
Median  
Mode  
Variance  
Skewness  
Ex. kurtosis  
Entropy  
MGF  
CF  
PGF 
Any French speaker will notice that “Poisson” means “fish”, but really there’s nothing fishy about this distribution. It’s actually pretty straightforward. The name comes from the mathematician SiméonDenis Poisson (17811840).
The Poisson Distribution is very similar to the Binomial Distribution. We are examining the number of times an event happens. The difference is subtle. Whereas the Binomial Distribution looks at how many times we register a success over a fixed total number of trials, the Poisson Distribution measures how many times a discrete event occurs, over a period of continuous space or time. There isn’t a “total” value n. As with the previous sections, let’s examine a couple of experiments or questions that might have an underlying Poisson nature.
 How many pennies will I encounter on my walk home?
 How many children will be delivered at the hospital today?
 How many mosquito bites did you get today after having sprayed with insecticide?
 How many angry phone calls did I get after airing a particularly distasteful political ad?
 How many products will I sell after airing a new television commercial?
 How many people, per hour, will cross a picket line into my store?
 How many alien abduction reports will be filed this year?
 How many defects will there be per 100 metres of rope sold?
What’s a little different about this distribution is that the random variable X which counts the number of events can take on any nonnegative integer value. In other words, I could walk home and find no pennies on the street. I could also find one penny. It’s also possible (although unlikely, short of an armoredcar exploding nearby) that I would find 10 or 100 or 10,000 pennies.
Instead of having a parameter p that represents a component probability like in the Bernoulli and Binomial distributions, this time we have the parameter “lambda” or λ which represents the “average or expected” number of events to happen within our experiment. The probability mass function of the Poisson is given by
An example[edit]
We run a restaurant and our signature dish (which is very expensive) gets ordered on average 4 times per day. What is the probability of having this dish ordered exactly 3 times tomorrow? If we only have the ingredients to prepare 3 of these dishes, what is the probability that it will get sold out and we’ll have to turn some orders away?
The probability of having the dish ordered 3 times exactly is given if we set k=3 in the above equation. Remember that we’ve already determined that we sell on average 4 dishes per day, so λ=4.
Here’s a table of the probabilities for all values from k=0..6:
Value for k  Probability f(k) 
0  0.0183 
1  0.0733 
2  0.1465 
3  0.1954 
4  0.1954 
5  0.1563 
6  0.1042 
Now for the big question: Will we run out of food by the end of the day tomorrow? In other words, we’re asking if the random variable X>3. In order to compute this we would have to add the probabilities that X=4, X=5, X=6,… all the way to infinity! But wait, there’s a better way!
The probability that we run out of food P(X>3) is the same as 1 minus the probability that we don’t run out of food, or 1P(X≤3). So if we total the probability that we sell zero, one, two and three dishes and subtract that from 1, we’ll have our answer. So,
 1 – P(X≤3) = 1 – ( P(X=0) + P(X=1) + P(X=2) + P(X=3) ) = 1 – 0.4335 = 0.5665
In other words, we have a 56.65% chance of selling out of our wonderful signature dish. I guess crossing our fingers is in order!
Mean[edit]
We calculate the mean as follows:
Remember that
${displaystyle mathrm {e} ^{lambda }=sum _{x=0}^{infty }{frac {lambda ^{x}}{x!}}}$
Variance[edit]
We derive the variance using the following formula:
We have already calculated E[X] above, so now we will calculate E[X^{2}] and then return to this variance formula:
The first sum is E[X]=λ and the second we also calculated above to be 1.
Returning to the variance formula we find that
External links[edit]
Geometric Distribution[edit]
There are two similar distributions with the name “Geometric Distribution”.
 The probability distribution of the number X of Bernoulli trials needed to get one success, supported on the set { 1, 2, 3, …}
 The probability distribution of the number Y = X − 1 of failures before the first success, supported on the set { 0, 1, 2, 3, … }
These two different geometric distributions should not be confused with each other. Often, the name shifted geometric distribution is adopted for the former one. We will use X and Y to refer to distinguish the two.
Shifted[edit]
The shifted Geometric Distribution refers to the probability of the number of times needed to do something until getting a desired result. For example:
 How many times will I throw a coin until it lands on heads?
 How many children will I have until I get a girl?
 How many cards will I draw from a pack until I get a Joker?
Just like the Bernoulli Distribution, the Geometric distribution has one controlling parameter: The probability of success in any independent test.
If a random variable X is distributed with a Geometric Distribution with a parameter p we write its probability mass function as:
${displaystyle Pleft(X=iright)=pleft(1pright)^{i1}}$
With a Geometric Distribution it is also pretty easy to calculate the probability of a “more than n times” case. The probability of failing to achieve the wanted result is
${displaystyle left(1pright)^{k}}$.
Example: a student comes home from a party in the forest, in which interesting substances were consumed. The student is trying to find the key to his front door, out of a keychain with 10 different keys. What is the probability of the student succeeding in finding the right key in the 4th attempt?
${displaystyle Pleft(X=4right)={frac {1}{10}}left(1{frac {1}{10}}right)^{41}={frac {1}{10}}left({frac {9}{10}}right)^{3}=0.0729}$
Unshifted[edit]
The probability mass function is defined as:
Mean[edit]
Let q=1p
We can now interchange the derivative and the sum.
Variance[edit]
We derive the variance using the following formula:
We have already calculated E[X] above, so now we will calculate E[X^{2}] and then return to this variance formula:
Let q=1p
We now manipulate x^{2} so that we get forms that are easy to handle by the technique used when deriving the mean.
We then return to the variance formula
External links[edit]
Negative Binomial Distribution[edit]
Just as the Bernoulli and the Binomial distribution are related in counting the number of successes in 1 or more trials, the Geometric and the Negative Binomial distribution are related in the number of trials needed to get 1 or more successes.
The Negative Binomial distribution refers to the probability of the number of times needed to do something until achieving a fixed number of desired results. For example:
 How many times will I throw a coin until it lands on heads for the 10th time?
 How many children will I have when I get my third daughter?
 How many cards will I have to draw from a pack until I get the second Joker?
Just like the Binomial Distribution, the Negative Binomial distribution has two controlling parameters: the probability of success p in any independent test and the desired number of successes m. If a random variable X has Negative Binomial distribution with parameters p and m, its probability mass function is:
Example[edit]
A travelling salesman goes home if he has sold 3 encyclopedias that day. Some days he sells them quickly. Other days he’s out till late in the evening. If on the average he sells an encyclopedia at one out of ten houses he approaches, what is the probability of returning home after having visited only 10 houses?
Answer:
The number of trials X is Negative Binomial distributed with parameters p=0.1 and m=3, hence:
Mean[edit]
The mean can be derived as follows.
Now let s = r+1 and w=x1 inside the summation.
We see that the summation is the sum over a the complete pmf of a negative binomial random variable distributed NB(s,p), which is 1 (and can be verified by applying Newton’s generalized binomial theorem).
Variance[edit]
We derive the variance using the following formula:
We have already calculated E[X] above, so now we will calculate E[X^{2}] and then return to this variance formula:
Again, let let s = r+1 and w=x1.
The first summation is the mean of a negative binomial random variable distributed NB(s,p) and the second summation is the complete sum of that variable’s pmf.
We now insert values into the original variance formula.
Hypergeometric Distribution[edit]
The hypergeometric distribution describes the number of successes in a sequence of n draws without replacement from a population of N that contained m total successes.
Its probability mass function is:
Technically the support for the function is only where x∈[max(0, n+mN), min(m, n)]. In situations where this range is not [0,n], f(x)=0 since for k>0,
${displaystyle {0 choose k}=0}$.
Probability Density Function[edit]
We first check to see that f(x) is a valid pmf. This requires that it is nonnegative everywhere and that its total sum is equal to 1. The first condition is obvious. For the second condition we will start with Vandermonde’s identity
We now see that if a=m and b=Nm that the condition is satisfied.
Mean[edit]
We derive the mean as follows:
We use the identity
${displaystyle {binom {a}{b}}={frac {a}{b}}{binom {a1}{b1}}}$in the denominator.
Next we use the identity
${displaystyle b{binom {a}{b}}=a{binom {a1}{b1}}}$in the first binomial of the numerator.
Next, for the variables inside the sum we define corresponding prime variables that are one less. So N′=N−1, m′=m−1, x′=x−1, n′=n1.
Now we see that the sum is the total sum over a Hypergeometric pmf with modified parameters. This is equal to 1. Therefore
Variance[edit]
We first determine E(X^{2}).
We use the same variable substitution as when deriving the mean.
The first sum is the expected value of a hypergeometric random variable with parameteres (n’,m’,N’). The second sum is the total sum that random variable’s pmf.
We then solve for the variance
Continuous Distributions[edit]
A continuous statistic is a random variable that does not have any points at which there is any distinct probability that the variable will be the corresponding number.
General Properties[edit]
Cumulative Distribution Function[edit]
A continuous random variable, like a discrete random variable, has a cumulative distribution function. Like the one for a discrete random variable, it also increases towards 1. Depending on the random variable, it may reach one at a finite number, or it may not. The cdf is represented by a capital F.
Probability Distribution Function[edit]
Unlike a discrete random variable, a continuous random variable has a probability density function instead of a probability mass function. The difference is that the former must integrate to 1, while the latter must have a total value of 1. The two are very similar, otherwise. The pdf is represented by a lowercase f.
Special Values[edit]
Let R be the set of points of the distribution.
The expected value for a continuous variable X with probability density function f is defined as
.
More generally, the expected value of any continuously transformed variable g(X) with probability density function f is defined as
.
The mean of a continuous or discrete distribution is defined as
${displaystyle E[X]}$.
The variance of a continuous or discrete distribution is defined as
${displaystyle E[(XE[X]^{2})]}$.
Expectations can also be derived by producing the Moment Generating Function for the distribution in question. This is done by finding the expected value
${displaystyle E[exp(tX)]}$. Once the Moment Generating Function has been created, each derivative of the function gives a different piece of information about the distribution function.
${displaystyle {frac {dE[exp(tX)]}{dt}}}$
= mean
= variance
= skewness
= kurtosis
Continuous Uniform Distribution[edit]
Probability density function Using maximum convention 

Cumulative distribution function 

Notation  

Parameters  
Support  
CDF  
Mean  
Median  
Mode  any value in 
Variance  
Skewness  0 
Ex. kurtosis  
Entropy  
MGF  
CF 
The (continuous) uniform distribution, as its name suggests, is a distribution with probability densities that are the same at each point in an interval. In casual terms, the uniform distribution shapes like a rectangle.
Mathematically speaking, the probability density function of the uniform distribution is defined as
${displaystyle fcolon [a,b]to mathbb {R} }$
${displaystyle fleft(xright)={1 over {ba}}}$
And the cumulative distribution function is:
${displaystyle Fleft(xright)={begin{cases}0,&{mbox{if }}xleq a\{{xa} over {ba}},&{mbox{if }}a$
Mean[edit]
We derive the mean as follows.
As the uniform distribution is 0 everywhere but [a, b] we can restrict ourselves that interval
Variance[edit]
We use the following formula for the variance.
External links[edit]
Normal Distribution[edit]
Normal distribution is without exception the most widely used distribution. It also goes under the name Gaussian distribution. It assumes that the observations are closely clustered around the mean, μ, and this amount is decaying quickly as we go farther away from the mean. The measure of spread is quantified by the variance,
${displaystyle sigma ^{2}}$.
Some examples of applications are:
 If the average man is 175 cm tall with a variance of 6 cm, what is the probability that a man found at random will be 183 cm tall?
 If the average man is 175 cm tall with a variance of 6 cm and the average woman is 168 cm tall with a variance of 3cm, what is the probability that the average man will be shorter than the average woman?
 If cans are assumed to have a variance of 4 grams, what does the average weight need to be in order to ensure that the 99% of all cans have a weight of at least 250 grams?
The density function is:
where
${displaystyle infty$.
and the cumulative distribution function cannot be integrated into a single expression.
Normal distribution with parameters μ and σ is denoted as
${displaystyle N(mu ,sigma )}$. If the rv X is normally distributed with expectation μ and standard deviation σ, one denotes:
${displaystyle !,Xsim N(mu ,sigma )}$
Probability mass function[edit]
To verify that f(x) is a valid pmf we must verify that (1) it is nonnegative everywhere, and (2) that the total integral is equal to 1. The first is obvious, so we move on to verify the second.
Now let
${displaystyle w={xmu over sigma {sqrt {2}}}}$. We see that
${displaystyle dw={dx over sigma {sqrt {2}}}}$.
Now we use the Gaussian integral that
${displaystyle int _{infty }^{infty }e^{w^{2}},dw={sqrt {pi }}}$
Mean[edit]
We derive the mean as follows
We now see that the right integral is the complete integral over a normal pmf. This is therefore 1.
Variance[edit]
We let
${displaystyle w={xmu over sigma {sqrt {2}}}}$
We now use integration by parts with u=w and v=(1/2)e^(w^2)
We see that the bracketed term is zero by L’Hôpital’s rule.
Now we use the Gaussian integral again
External links[edit]
Exponential Distribution[edit]
Probability density function 

Cumulative distribution function 

Parameters  λ > 0 rate, or inverse scale 

Support  x ∈ [0, ∞) 
λ e^{−λx}  
CDF  1 − e^{−λx} 
Mean  λ^{−1} 
Median  λ^{−1} ln 2 
Mode  0 
Variance  λ^{−2} 
Skewness  2 
Ex. kurtosis  6 
Entropy  1 − ln(λ) 
MGF  
CF 
Exponential distribution refers to a statistical distribution used to model the time between independent events that happen at a constant average rate λ. Some examples of this distribution are:
 The distance between one car passing by after the previous one.
 The rate at which radioactive particles decay.
For the stochastic variable X, probability distribution function of it is:
${displaystyle f_{x}(x)={begin{cases}lambda e^{lambda x},&{mbox{if }}xgeq 0\0,&{mbox{if }}x<0end{cases}}}$
and the cumulative distribution function is:
${displaystyle F_{x}(x)={begin{cases}0,&{mbox{if }}x<0\{1e^{lambda x}},&{mbox{if }}xgeq 0end{cases}}}$
Exponential distribution is denoted as
${displaystyle Xin {mbox{Exp(m)}}}$, where m is the average number of events within a given time period. So if m=3 per minute, i.e. there are three events per minute, then λ=1/3, i.e. one event is expected on average to take place every 20 seconds.
Mean[edit]
We derive the mean as follows.
We will use integration by parts with u=−x and v=e^{−λx}. We see that du=1 and dv=−λe^{−λx}.
Variance[edit]
We use the following formula for the variance.
We’ll use integration by parts with
${displaystyle u=x^{2}}$and
${displaystyle v=e^{2x}}$. From this we have
${displaystyle du=2x}$and
${displaystyle v=2e^{2x}}$.
We see that the integral is just
${displaystyle operatorname {E} [X]}$which we solved for above.
External links[edit]
Gamma Distribution[edit]
Probability density function 

Cumulative distribution function 

Parameters 


CDF  
Mean  
Median  No simple closed form 
Mode  
Skewness  
Ex. kurtosis  
Entropy 
The Gamma distribution is very important for technical reasons, since it is the parent of the exponential distribution and can explain many other distributions.
The probability distribution function is:
${displaystyle f_{x}(x)={begin{cases}{frac {1}{a^{p}Gamma (p)}}x^{p1}e^{x/a},&{mbox{if }}xgeq 0\0,&{mbox{if }}x<0end{cases}}quad a,p>0}$
${displaystyle Gamma (p)=int _{0}^{infty }t^{p1}e^{t},dt,}$
is the Gamma function. The cumulative distribution function cannot be found unless p=1, in which case the Gamma distribution becomes the exponential distribution. The Gamma distribution of the stochastic variable X is denoted as
${displaystyle Xin Gamma (p,a)}$.
Alternatively, the gamma distribution can be parameterized in terms of a shape parameter
${displaystyle alpha =k}$and an inverse scale parameter
${displaystyle beta =1/theta }$, called a rate parameter: