Identifying Distribution Families from Shape and Tails

Identifying a Family Is a Process of Elimination

A distribution family is not identified reliably by a single feature such as a bell-shaped histogram or a long right tail. Several different models can look similar in a finite sample, and a sample histogram is affected by bin width and sampling variation. A useful identification process combines support, symmetry, density shape, quantiles, and tail behavior. These clues narrow the possibilities and suggest checks; they do not, by themselves, prove that a population follows a particular model.

This tutorial focuses on six continuous families: normal, logistic, Laplace, lognormal, Weibull, and Pareto Type I. The goal is to recognize their characteristic signatures and use them coherently. In particular, the distinction between a model on the real line and one restricted to positive values is often more decisive than the apparent shape of a histogram.

Reference Signatures for Six Families

The following parameterizations fix the meaning of the parameters used here. For the lognormal model, $\log X\sim N(\mu,\sigma^2)$; for the Weibull model, $\lambda$ is the scale and $k$ is the shape.

Family	Support and density	Useful identifying clues
Normal	$\mathbb{R}$; $f(x)=\frac{1}{\sigma\sqrt{2\pi}}\exp[-(x-\mu)^2/(2\sigma^2)]$	Symmetric, smooth bell shape; log density is quadratic in $x$.
Logistic	$\mathbb{R}$; $f(x)=\frac{\exp[-(x-\mu)/s]}{s\{1+\exp[-(x-\mu)/s]\}^2}$, $s>0$	Symmetric and smooth, with heavier tails than the normal; logistic quantiles are explicit.
Laplace	$\mathbb{R}$; $f(x)=\frac{1}{2b}\exp[-\|x-\mu\|/b]$, $b>0$	Symmetric with a sharp cusp at its center; each side of the log density is linear.
Lognormal	$(0,\infty)$; $f(x)=\frac{1}{x\sigma\sqrt{2\pi}}\exp[-(\log x-\mu)^2/(2\sigma^2)]$	Positive and typically right-skewed; $\log X$ is normal and symmetric.
Weibull	$(0,\infty)$; $f(x)=\frac{k}{\lambda}(x/\lambda)^{k-1}\exp[-(x/\lambda)^k]$	Positive support; its survival function has a simple power transformation.
Pareto Type I	$[x_m,\infty)$; $f(x)=\alpha x_m^\alpha x^{-(\alpha+1)}$, $x_m,\alpha>0$	Positive threshold and power-law tail; log survival is linear in $\log x$ above $x_m$.

Support is a first filter. A genuinely negative value is incompatible with a lognormal, Weibull, or Pareto Type I model, while a strictly positive measurement does not by itself distinguish among those three. Symmetry is another strong clue: the normal, logistic, and Laplace families are symmetric about $\mu$, whereas the latter three positive-support families are generally right-skewed on the original scale. A data collection rule can truncate or censor observations, however, so apparent support should be interpreted in light of how the data were obtained.

Symmetry, Shape, and Quantiles

For a continuous distribution that is strictly increasing over the range containing its quantiles, define $Q(p)$, for $0<p<1$, as the unique value satisfying $F(Q(p))=p$. If the distribution is symmetric about $\mu$, its quantiles pair around that center.

Theorem: Suppose $F$ is continuous and strictly increasing on the interval containing $Q(p)$ and $Q(1-p)$, and the distribution is symmetric about $\mu$. Then, for every $0<p<1$, $$ Q(1-p)=2\mu-Q(p). $$

To prove the result, symmetry means that $X-\mu$ and $-(X-\mu)$ have the same distribution. Continuity therefore gives $F(2\mu-x)=1-F(x)$. Substitute $x=Q(p)$:

F(2\mu-Q(p))=1-F(Q(p))=1-p.

By strict increase on the relevant interval, there is only one point with CDF value $1-p$, namely $Q(1-p)$. Thus $Q(1-p)=2\mu-Q(p)$. The strict-increase condition matters: a continuous CDF can have flat sections, and then a generalized-inverse quantile need not pair this way. $\square$

In practice, approximate pairing of lower and upper quantiles around a common center supports a symmetric model, but does not distinguish the three symmetric families. The density near the center helps: Laplace has a cusp, whereas normal and logistic densities are smooth. The normal log density has a quadratic form in $x$, the Laplace log density is piecewise linear, and logistic tails have approximately linear log-density decay far from the center. Quantiles can also help distinguish models. For example, the logistic quantile is $Q(p)=\mu+s\log\{p/(1-p)\}$, while the normal quantile is obtained from the standard normal quantile function.

Worked Example: A Symmetric Density With a Central Cusp

Suppose a proposed density is $f(x)=\frac{1}{6}\exp(-|x-2|/3)$, for $-\infty<x<\infty$. Identify its family and parameters, and describe the checks that support the identification.

The density matches the Laplace form $f(x)=\frac{1}{2b}\exp[-|x-\mu|/b]$. Comparing the absolute-value term gives $\mu=2$ and $b=3$; the leading constant also agrees because $1/(2b)=1/6$.

The density is symmetric about 2, and it has its sharp maximum there. For $x>2$, its log density is $-\log 6-(x-2)/3$, which is linear in $x$. For $x<2$, it is $-\log 6+(x-2)/3$, also linear but with the opposite slope. These features distinguish the Laplace shape from a smooth normal or logistic peak. The proposed function is normalized: integrating on either side of 2 gives $1/2$, since $\int_0^\infty (1/6)e^{-u/3}\,du=1/2$.

Positive-Support Families and Transformations

When observations are positive and strongly right-skewed, inspect transformed scales and survival behavior rather than relying only on the original-scale histogram. The lognormal model is symmetric after taking logarithms: if $Y=\log X$, then $Y\sim N(\mu,\sigma^2)$. Its quantile function follows directly from the transformation:

Q_X(p)=\exp\{\mu+\sigma\Phi^{-1}(p)\},

where $\Phi$ is the standard normal CDF. Thus the logarithms of opposite-tail quantiles add to $2\mu$. The original-scale median is $e^\mu$, but the original distribution is not symmetric about this value.

For the Weibull family, the survival function is $P(X>x)=\exp[-(x/\lambda)^k]$ for $x>0$. This gives an exact identification transformation:

Result: If $X$ has a Weibull distribution with scale $\lambda>0$ and shape $k>0$, then $$ \log\{-\log P(X>x)\}=k\log x-k\log\lambda,\qquad x>0. $$

Indeed, take the negative logarithm of the survival function to get $-\log P(X>x)=(x/\lambda)^k$. Taking logarithms again gives $k(\log x-\log\lambda)$, which is the stated linear relationship. Its slope in $\log x$ is $k$, and its intercept is $-k\log\lambda$. This is a population identity; a plot based on estimated survival probabilities is only a diagnostic and can be noisy, particularly in the far tail.

For Pareto Type I, $P(X>x)=(x_m/x)^\alpha$ for $x\geq x_m$. Consequently,

\log P(X>x)=\alpha\log x_m-\alpha\log x,\qquad x\geq x_m.

The negative slope on a log-survival versus log-$x$ scale is $-\alpha$, and the support begins at $x_m$. This contrasts with the Weibull transformation, which linearizes $\log\{-\log P(X>x)\}$ against $\log x$. A lognormal tail does not have either of these exact linear relationships over its entire support.

Worked Example: Reading a Weibull Shape and Scale

A positive lifetime is modeled with density $f(x)=\frac{2x}{9}\exp[-(x/3)^2]$, $x>0$. Determine whether it is Weibull and identify its parameters.

The Weibull density is $\frac{k}{\lambda}(x/\lambda)^{k-1}\exp[-(x/\lambda)^k]$. The exponent in the proposed density is $-(x/3)^2$, so $k=2$ and $\lambda=3$. Substitution into the leading factor gives $\frac{2}{3}(x/3)=2x/9$, exactly as proposed.

The survival function is therefore $\exp[-(x/3)^2]$, and the transformed survival expression is $\log\{-\log P(X>x)\}=2\log x-2\log 3$. Its slope is 2 and its intercept is $-2\log 3$, agreeing with the identified shape and scale. The density's mode is also a useful shape check: differentiating its log gives $1/x-2x/9=0$, hence $x=3/\sqrt{2}$, a positive interior mode.

Worked Example: Distinguishing Lognormal and Pareto Clues

Consider two candidate models for a positive measurement. Model A has $\log X\sim N(0,0.5^2)$. Model B is Pareto Type I with $x_m=1$ and $\alpha=2$. Compare their supports, medians, and central quantiles.

Both models have positive support, so positivity alone cannot choose between them. For Model A, $Q(p)=\exp\{0.5\Phi^{-1}(p)\}$, and its median is $Q(0.5)=1$. Using $\Phi^{-1}(0.25)\approx-0.6745$ and $\Phi^{-1}(0.75)\approx0.6745$, the quartiles are approximately $0.714$ and $1.400$.

For Model B, solving $p=1-(x_m/Q(p))^\alpha$ gives $Q(p)=x_m(1-p)^{-1/\alpha}$. Its median is $Q(0.5)=1(0.5)^{-1/2}\approx1.414$. Its first and third quartiles are $Q(0.25)=1(0.75)^{-1/2}\approx1.155$ and $Q(0.75)=1(0.25)^{-1/2}=2$.

These candidate models therefore predict substantially different central quantiles despite sharing positive support. The lognormal predicts log-scale symmetry around zero; the Pareto predicts a threshold at 1 and a power-law survival relationship above it. A finite sample need not match either set exactly, so these calculations are checks against candidate models rather than a proof of which model generated the data.

A Practical Identification Sequence

Check the support.
Ask whether values may be negative, whether zero is possible, and whether there is a positive lower threshold. Use the measurement process to assess whether observed limits are structural or caused by censoring or truncation.

Assess symmetry and skewness.
Compare opposite-tail quantiles and inspect whether symmetry is plausible on the original scale or only after taking logarithms.

Inspect density shape.
For real-line symmetric candidates, look for a Laplace cusp versus smooth normal or logistic centers. Treat histogram appearance as suggestive, not definitive.

Check family-specific relationships.
Compare quantiles or transformed survival behavior with the relevant model identity. A coherent set of clues is more informative than one isolated feature.

Common Mistakes and Why Identification Matters

Calling every bell-shaped sample normal. Logistic and Laplace distributions are also symmetric and unimodal. Examine the center, tails, and quantile pattern rather than the overall silhouette alone.
Using symmetry without a quantile condition. A continuous CDF can have flat sections. The quantile pairing theorem above requires strict increase over the relevant range when quantiles are defined as unique inverse-CDF values.
Getting the Pareto log-tail slope backward. Since $\log P(X>x)=\alpha\log x_m-\alpha\log x$, the coefficient of $\log x$ is $-\alpha$, not $+\alpha$.
Treating a diagnostic plot as a proof. Estimated tail probabilities are uncertain, and a short sample may not reveal asymptotic behavior. Model identification should combine subject-matter constraints with several distributional checks.

Family identification matters because a model's support and tail behavior influence what summaries and probability calculations are credible. A normal model permits negative values, while a positive-support model does not; a Pareto model assigns a power-law tail, unlike the normal. These differences can be consequential in applications involving lifetimes, costs, or extreme measurements. The next step after proposing a family is to check its adequacy and calculate quantities under it, keeping model uncertainty visible.

Check Your Understanding

Use the family signatures and reasoning above to answer each question.

A distribution is symmetric on the real line and has a sharp cusp at its center. Which of the six families is the leading candidate, and what feature rules out a smooth normal peak?
For a Weibull model, what are the slope and intercept when $\log\{-\log P(X>x)\}$ is expressed as a linear function of $\log x$?
A candidate distribution has a positive threshold and a straight log-survival relationship against $\log x$ above that threshold. Which family is suggested, and what is the sign of the slope?
Why does symmetry alone not always guarantee $Q(1-p)=2\mu-Q(p)$ under a generalized-inverse quantile convention?
Explain why positive support alone cannot distinguish a lognormal, Weibull, and Pareto Type I model.

Family	Support and density	Useful identifying clues
Normal	\(\mathbb{R}\); \(f(x)=\frac{1}{\sigma\sqrt{2\pi}}\exp[-(x-\mu)^2/(2\sigma^2)]\)	Symmetric, smooth bell shape; log density is quadratic in \(x\).
Logistic	\(\mathbb{R}\); \(f(x)=\frac{\exp[-(x-\mu)/s]}{s\{1+\exp[-(x-\mu)/s]\}^2}\), \(s>0\)	Symmetric and smooth, with heavier tails than the normal; logistic quantiles are explicit.
Laplace	\(\mathbb{R}\); \(f(x)=\frac{1}{2b}\exp[-\|x-\mu\|/b]\), \(b>0\)	Symmetric with a sharp cusp at its center; each side of the log density is linear.
Lognormal	\((0,\infty)\); \(f(x)=\frac{1}{x\sigma\sqrt{2\pi}}\exp[-(\log x-\mu)^2/(2\sigma^2)]\)	Positive and typically right-skewed; \(\log X\) is normal and symmetric.
Weibull	\((0,\infty)\); \(f(x)=\frac{k}{\lambda}(x/\lambda)^{k-1}\exp[-(x/\lambda)^k]\)	Positive support; its survival function has a simple power transformation.
Pareto Type I	\([x_m,\infty)\); \(f(x)=\alpha x_m^\alpha x^{-(\alpha+1)}\), \(x_m,\alpha>0\)	Positive threshold and power-law tail; log survival is linear in \(\log x\) above \(x_m\).