If we take a sample of observations from a normal distribution, then the t-distribution with degrees of freedom can be defined as the distribution of the location of the sample mean relative to the true mean, divided by the sample standard deviation, after multiplying by the standardizing term . In this way, the t-distribution can be used to construct a confidence interval for the true mean.
The t-distribution is symmetric and bell-shaped, like the normal distribution. However, the t-distribution has heavier tails, meaning that it is more prone to producing values that fall far from its mean. This makes it useful for understanding the statistical behavior of certain types of ratios of random quantities, in which variation in the denominator is amplified and may produce outlying values when the denominator of the ratio falls close to zero. The Student's t-distribution is a special case of the generalised hyperbolic distribution.
History and etymology
Statistician William Sealy Gosset, known as "Student"
In the English-language literature, the distribution takes its name from William Sealy Gosset's 1908 paper in Biometrika under the pseudonym "Student". Gosset worked at the Guinness Brewery in Dublin, Ireland, and was interested in the problems of small samples – for example, the chemical properties of barley where sample sizes might be as few as 3. One version of the origin of the pseudonym is that Gosset's employer preferred staff to use pen names when publishing scientific papers instead of their real name, so he used the name "Student" to hide his identity. Another version is that Guinness did not want their competitors to know that they were using the t-test to determine the quality of raw material.
Gosset's paper refers to the distribution as the "frequency distribution of standard deviations of samples drawn from a normal population". It became well known through the work of Ronald Fisher, who called the distribution "Student's distribution" and represented the test value with the letter t.
How Student's distribution arises from sampling
Let be independently and identically drawn from the distribution , i.e. this is a sample of size from a normally distributed population with expected mean value and variance .
has a standard normal distribution (i.e. normal with expected mean 0 and variance 1), and the random variable
i.e where has been substituted for , has a Student's t-distribution with degrees of freedom. Since has replaced the only unobservable quantity in this expression is so this can be used to derive confidence intervals for The numerator and the denominator in the preceding expression are independent random variables despite being based on the same sample . This can be seen by observing that and recalling that and are both linear combinations of the same set of i.i.d. normally distributed random variables.
where B is the Beta function. In particular for integer valued degrees of freedom we have:
The probability density function is symmetric, and its overall shape resembles the bell shape of a normally distributed variable with mean 0 and variance 1, except that it is a bit lower and wider. As the number of degrees of freedom grows, the t-distribution approaches the normal distribution with mean 0 and variance 1. For this reason is also known as the normality parameter.
The following images show the density of the t-distribution for increasing values of . The normal distribution is shown as a blue line for comparison. Note that the t-distribution (red line) becomes closer to the normal distribution as increases.
Density of the t-distribution (red) for 1, 2, 3, 5, 10, and 30 degrees of freedom compared to the standard normal distribution (blue). Previous plots shown in green.
In Bayesian statistics, a (scaled, shifted) t-distribution arises as the marginal distribution of the unknown mean of a normal distribution, when the dependence on an unknown variance has been marginalized out:
where stands for the data , and represents any other information that may have been used to create the model. The distribution is thus the compounding of the conditional distribution of given the data and with the marginal distribution of given the data.
This can be evaluated by substituting , where , giving
But the z integral is now a standard Gamma integral, which evaluates to a constant, leaving
This is a form of the t-distribution with an explicit scaling and shifting that will be explored in more detail in a further section below. It can be related to the standardized t-distribution by the substitution
The derivation above has been presented for the case of uninformative priors for and ; but it will be apparent that any priors that lead to a normal distribution being compounded with a scaled inverse chi-squared distribution will lead to a t-distribution with scaling and shifting for , although the scaling parameter corresponding to above will then be influenced both by the prior information and the data, rather than just by the data as above.
As the distribution of a test statistic
Student's t-distribution with degrees of freedom can be defined as the distribution of the random variableT with
Suppose X1, ..., Xn are independent realizations of the normally-distributed, random variable X, which has an expected value μ and variance σ2. Let
be the sample mean, and
be an unbiased estimate of the variance from the sample. It can be shown that the random variable
has a chi-squared distribution with degrees of freedom (by Cochran's theorem). It is readily shown that the quantity
is normally distributed with mean 0 and variance 1, since the sample mean is normally distributed with mean μ and variance σ2/n. Moreover, it is possible to show that these two random variables (the normally distributed one Z and the chi-squared-distributed one V) are independent. Consequently[clarification needed] the pivotal quantity
which differs from Z in that the exact standard deviation σ is replaced by the random variable Sn, has a Student's t-distribution as defined above. Notice that the unknown population variance σ2 does not appear in T, since it was in both the numerator and the denominator, so it canceled. Gosset intuitively obtained the probability density function stated above, with equal to n − 1, and Fisher proved it in 1925.
The distribution of the test statistic T depends on , but not μ or σ; the lack of dependence on μ and σ is what makes the t-distribution important in both theory and practice.
There are various approaches to constructing random samples from the Student's t-distribution. The matter depends on whether the samples are required on a stand-alone basis, or are to be constructed by application of a quantile function to uniform samples; e.g., in the multi-dimensional applications basis of copula-dependency. In the case of stand-alone sampling, an extension of the Box–Muller method and its polar form is easily deployed. It has the merit that it applies equally well to all real positive degrees of freedom, ν, while many other candidate methods fail if ν is close to zero.
Integral of Student's probability density function and p-value
The function A(t | ν) is the integral of Student's probability density function, f(t) between −t and t, for t ≥ 0. It thus gives the probability that a value of t less than that calculated from observed data would occur by chance. Therefore, the function A(t | ν) can be used when testing whether the difference between the means of two sets of data is statistically significant, by calculating the corresponding value of t and the probability of its occurrence if the two sets of data were drawn from the same population. This is used in a variety of situations, particularly in t-tests. For the statistic t, with ν degrees of freedom, A(t | ν) is the probability that t would be less than the observed value if the two means were the same (provided that the smaller mean is subtracted from the larger, so that t ≥ 0). It can be easily calculated from the cumulative distribution functionFν(t) of the t-distribution:
This means that has a classic Student's t distribution with degrees of freedom.
The resulting non-standardized Student's t-distribution has a density defined by:
Here, does not correspond to a standard deviation: it is not the standard deviation of the scaled t distribution, which may not even exist; nor is it the standard deviation of the underlying normal distribution, which is unknown. simply sets the overall scaling of the distribution. In the Bayesian derivation of the marginal distribution of an unknown normal mean above, as used here corresponds to the quantity , where
Equivalently, the distribution can be written in terms of , the square of this scale parameter:
Other properties of this version of the distribution are:
This distribution results from compounding a Gaussian distribution (normal distribution) with mean and unknown variance, with an inverse gamma distribution placed over the variance with parameters and . In other words, the random variableX is assumed to have a Gaussian distribution with an unknown variance distributed as inverse gamma, and then the variance is marginalized out (integrated out). The reason for the usefulness of this characterization is that the inverse gamma distribution is the conjugate prior distribution of the variance of a Gaussian distribution. As a result, the non-standardized Student's t-distribution arises naturally in many Bayesian inference problems. See below.
Equivalently, this distribution results from compounding a Gaussian distribution with a scaled-inverse-chi-squared distribution with parameters and . The scaled-inverse-chi-squared distribution is exactly the same distribution as the inverse gamma distribution, but with a different parameterization, i.e. .
In terms of inverse scaling parameter λ
An alternative parameterization in terms of an inverse scaling parameter (analogous to the way precision is the reciprocal of variance), defined by the relation . The density is then given by:
Other properties of this version of the distribution are:
If has a Student's t-distribution with degree of freedom then X2 has an F-distribution:
The noncentral t-distribution generalizes the t-distribution to include a location parameter. Unlike the nonstandardized t-distributions, the noncentral distributions are not symmetric (the median is not the same as the mode).
Here a, b, and k are parameters. This distribution arises from the construction of a system of discrete distributions similar to that of the Pearson distributions for continuous distributions.
One can generate Student-t samples by taking the ratio of variables from the normal distribution and the square-root of χ2-distribution. If we use instead of the normal distribution, e.g., the Irwin–Hall distribution, we obtain over-all a symmetric 4-parameter distribution, which includes the normal, the uniform, the triangular, the Student-t and the Cauchy distribution. This is also more flexible than some other symmetric generalizations of the normal distribution.
Student's t-distribution arises in a variety of statistical estimation problems where the goal is to estimate an unknown parameter, such as a mean value, in a setting where the data are observed with additive errors. If (as in nearly all practical statistical work) the population standard deviation of these errors is unknown and has to be estimated from the data, the t-distribution is often used to account for the extra uncertainty that results from this estimation. In most such problems, if the standard deviation of the errors were known, a normal distribution would be used instead of the t-distribution.
Confidence intervals and hypothesis tests are two statistical procedures in which the quantiles of the sampling distribution of a particular statistic (e.g. the standard score) are required. In any situation where this statistic is a linear function of the data, divided by the usual estimate of the standard deviation, the resulting quantity can be rescaled and centered to follow Student's t-distribution. Statistical analyses involving means, weighted means, and regression coefficients all lead to statistics having this form.
Quite often, textbook problems will treat the population standard deviation as if it were known and thereby avoid the need to use the Student's t-distribution. These problems are generally of two kinds: (1) those in which the sample size is so large that one may treat a data-based estimate of the variance as if it were certain, and (2) those that illustrate mathematical reasoning, in which the problem of estimating the standard deviation is temporarily ignored because that is not the point that the author or instructor is then explaining.
A number of statistics can be shown to have t-distributions for samples of moderate size under null hypotheses that are of interest, so that the t-distribution forms the basis for significance tests. For example, the distribution of Spearman's rank correlation coefficientρ, in the null case (zero correlation) is well approximated by the t distribution for sample sizes above about 20.
Suppose the number A is so chosen that
when T has a t-distribution with n − 1 degrees of freedom. By symmetry, this is the same as saying that A satisfies
so A is the "95th percentile" of this probability distribution, or . Then
and this is equivalent to
Therefore, the interval whose endpoints are
is a 90% confidence interval for μ. Therefore, if we find the mean of a set of observations that we can reasonably expect to have a normal distribution, we can use the t-distribution to examine whether the confidence limits on that mean include some theoretically predicted value – such as the value predicted on a null hypothesis.
It is this result that is used in the Student's t-tests: since the difference between the means of samples from two normal distributions is itself distributed normally, the t-distribution can be used to examine whether that difference can reasonably be supposed to be zero.
If the data are normally distributed, the one-sided (1 − α)-upper confidence limit (UCL) of the mean, can be calculated using the following equation:
The resulting UCL will be the greatest average value that will occur for a given confidence interval and population size. In other words, being the mean of the set of observations, the probability that the mean of the distribution is inferior to UCL1−α is equal to the confidence level 1 − α.
The t-distribution can be used to construct a prediction interval for an unobserved sample from a normal distribution with unknown mean and variance.
The t-distribution is often used as an alternative to the normal distribution as a model for data, which often has heavier tails than the normal distribution allows for; see e.g. Lange et al. The classical approach was to identify outliers (e.g., using Grubbs's test) and exclude or downweight them in some way. However, it is not always easy to identify outliers (especially in high dimensions), and the t-distribution is a natural choice of model for such data and provides a parametric approach to robust statistics.
A Bayesian account can be found in Gelman et al. The degrees of freedom parameter controls the kurtosis of the distribution and is correlated with the scale parameter. The likelihood can have multiple local maxima and, as such, it is often necessary to fix the degrees of freedom at a fairly low value and estimate the other parameters taking this as given. Some authors report that values between 3 and 9 are often good choices. Venables and Ripley suggest that a value of 5 is often a good choice.
For practical regression and prediction needs, Student's t-processes were introduced, that are generalisations of the Student t-distributions for functions. A Student's t-process is constructed from the Student t-distributions like a Gaussian process is constructed from the Gaussian distributions. For a Gaussian process, all sets of values have a multidimensional Gaussian distribution. Analogiusly, is a Student t-process on an interval if the correspondent values of the process () have a joint multivariate Student t-distribution. These processes are used for regression, prediction, Bayesian optimization and related problems. For multivariate regression and multi-output prediction, the multivariate Student t-processes are introduced and used.
Table of selected values
The following table lists values for t-distributions with ν degrees of freedom for a range of one-sided or two-sided critical regions. The first column is ν, the percentages along the top are confidence levels, and the numbers in the body of the table are the factors described in the section on confidence intervals.
Note that the last row with infinite ν gives critical points for a normal distribution since a t-distribution with infinitely many degrees of freedom is a normal distribution. (See Related distributions above).
Calculating the confidence interval
Let's say we have a sample with size 11, sample mean 10, and sample variance 2. For 90% confidence with 10 degrees of freedom, the one-sided t-value from the table is 1.372. Then with confidence interval calculated from
we determine that with 90% confidence we have a true mean lying below
In other words, 90% of the times that an upper threshold is calculated by this method from particular samples, this upper threshold exceeds the true mean.
And with 90% confidence we have a true mean lying above
In other words, 90% of the times that a lower threshold is calculated by this method from particular samples, this lower threshold lies below the true mean.
So that at 80% confidence (calculated from 100% − 2 × (1 − 90%) = 80%), we have a true mean lying within the interval
Saying that 80% of the times that upper and lower thresholds are calculated by this method from a given sample, the true mean is both below the upper threshold and above the lower threshold is not the same as saying that there is an 80% probability that the true mean lies between a particular pair of upper and lower thresholds that have been calculated by this method; see confidence interval and prosecutor's fallacy.
^Hurst, Simon. "The Characteristic Function of the Student t Distribution". Financial Mathematics Research Report No. FMRR006-95, Statistics Research Report No. SRR044-95. Archived from the original on February 18, 2010.
^Helmert FR (1875). "Über die Berechnung des wahrscheinlichen Fehlers aus einer endlichen Anzahl wahrer Beobachtungsfehler". Z. Math. U. Physik. 20: 300–3.
^Helmert FR (1876). "Über die Wahrscheinlichkeit der Potenzsummen der Beobachtungsfehler und uber einige damit in Zusammenhang stehende Fragen". Z. Math. Phys. 21: 192–218.
^Helmert FR (1876). "Die Genauigkeit der Formel von Peters zur Berechnung des wahrscheinlichen Beobachtungsfehlers directer Beobachtungen gleicher Genauigkeit" [The accuracy of Peters' formula for calculating the probable observation error of direct observations of the same accuracy] (PDF). Astron. Nachr. (in German). 88 (8–9): 113–132. Bibcode:1876AN.....88..113H. doi:10.1002/asna.18760880802.
^Lüroth J (1876). "Vergleichung von zwei Werten des wahrscheinlichen Fehlers". Astron. Nachr.87 (14): 209–20. Bibcode:1876AN.....87..209L. doi:10.1002/asna.18760871402.
^Pfanzagl J, Sheynin O (1996). "Studies in the history of probability and statistics. XLIV. A forerunner of the t-distribution". Biometrika. 83 (4): 891–898. doi:10.1093/biomet/83.4.891. MR 1766040.
^Pearson, K. (1895-01-01). "Contributions to the Mathematical Theory of Evolution. II. Skew Variation in Homogeneous Material". Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences. 186: 343–414 (374). doi:10.1098/rsta.1895.0010. ISSN 1364-503X.
^Park SY, Bera AK (2009). "Maximum entropy autoregressive conditional heteroskedasticity model". J. Econom.150 (2): 219–230. doi:10.1016/j.jeconom.2008.12.014.
^Casella G, Berger RL (1990). Statistical Inference. Duxbury Resource Center. p. 56. ISBN 9780534119584.
^ abBailey RW (1994). "Polar Generation of Random Variates with the t-Distribution". Math. Comput.62 (206): 779–781. doi:10.2307/2153537. JSTOR 2153537.
^ abJackman, S. (2009). Bayesian Analysis for the Social Sciences. Wiley. p. 507. doi:10.1002/9780470686621. ISBN 9780470011546.
^ abBishop, C.M. (2006). Pattern Recognition and Machine Learning. New York, NY: Springer. ISBN 9780387310732.
^Ord JK (1972). Families of Frequency Distributions. London: Griffin. ISBN 9780852641378. See Table 5.1.CS1 maint: postscript (link)
^Ord JK (1972). "Chapter 5". Families of frequency distributions. London: Griffin. ISBN 9780852641378.
^Lange KL, Little RJ, Taylor JM (1989). "Robust Statistical Modeling Using the t Distribution" (PDF). J. Am. Stat. Assoc.84 (408): 881–896. doi:10.1080/01621459.1989.10478852. JSTOR 2290063.
^Gelman AB, Carlin JB, Stern HS, et al. (2014). "Computationally efficient Markov chain simulation". Bayesian Data Analysis. Boca Raton, Florida: CRC Press. p. 293. ISBN 9781439898208.
^Shah, Amar; Wilson, Andrew Gordon; Ghahramani, Zoubin (2014). "Student-t processes as alternatives to Gaussian processes" (PDF). JMLR. 33 (Proceedings of the 17th International Conference on Artificial Intelligence and Statistics (AISTATS) 2014, Reykjavik, Iceland): 877–885.
^Chen, Zexun; Wang, Bo; Gorban, Alexander N. (2019). "Multivariate Gaussian and Student-t process regression for multi-output prediction". Neural Computing and Applications. arXiv:1703.04455. doi:10.1007/s00521-019-04687-8.