Why Weibull Parameter Names Can Mislead
A Weibull model may be described using a shape and a scale, or using a shape and a parameter called a rate. The difficulty is that “rate” does not identify one universal formula: one convention uses the reciprocal scale, while another puts a coefficient directly in the power-law exponent. Those quantities are related, but they are not generally equal. A substitution that forgets the shape-dependent power can change the distribution substantially.
The safest practice is to begin with the survival function, \(1-F(x)\), and write its exponent explicitly. This tutorial makes the three common forms precise, derives how to convert among them, and uses CDF and quantile calculations to check the conversions.
Three Forms of the Weibull Distribution
Let the shape parameter \(k>0\), and let \(X\) be a nonnegative random variable. In the shape-scale convention, the scale \(b>0\) has the same units as \(X\). The distribution is supported on \(x\geq 0\).
The same CDF can be expressed in two rate forms. We use \(\rho\) for the inverse-scale rate and \(\alpha\) for the power-rate coefficient. Stating the formula alongside the parameter name avoids relying on potentially ambiguous terminology.
Some references call \(\rho\) the Weibull rate, while others call \(\alpha\) a rate parameter or rate coefficient. The word alone is therefore insufficient to identify the formula. The defining exponent is decisive: is it \((\rho x)^k\), or is it \(\alpha x^k\)?
Worked Example: Converting a Scale to Both Rate Forms
Suppose \(k=2\) and the scale is \(b=5\) minutes. Convert to the two rate forms and find \(P(X\leq2)\).
For the inverse-scale rate, take the reciprocal of the scale: \(\rho=1/b=1/5=0.2\) per minute. For the power-rate coefficient, raise that reciprocal to the shape: \(\alpha=\rho^k=(0.2)^2=0.04\) per minute squared.
At \(x=2\) minutes, the scale form gives \((x/b)^k=(2/5)^2=0.16\). The inverse-scale form gives \((\rho x)^k=(0.2\cdot2)^2=0.4^2=0.16\), and the power-rate form gives \(\alpha x^k=0.04\cdot2^2=0.04\cdot4=0.16\). Thus all three produce the same probability:
The coefficient \(0.04\) is not the inverse-scale rate \(0.2\). It is the square of that rate because the shape is 2.
Conversion Rules and Their Derivation
The conversion is obtained by equating the exponents. This is more reliable than trying to remember which parameter should be reciprocated.
Proof. For \(x\geq0\), equality of the scale and inverse-scale survival functions requires \(\exp[-(x/b)^k]=\exp[-(\rho x)^k]\). Since the exponential function is one-to-one, their nonnegative exponents must agree: \((x/b)^k=(\rho x)^k\). For any \(x>0\), divide both sides by \(x^k\), obtaining \(b^{-k}=\rho^k\). Both \(b^{-1}\) and \(\rho\) are positive, so taking the positive \(k\)th root gives \(\rho=1/b\).
Equating the scale and power-rate exponents gives \((x/b)^k=\alpha x^k\). For \(x>0\), division by \(x^k\) yields \(\alpha=b^{-k}\). Combining this with \(\rho=1/b\) gives \(\alpha=\rho^k\). Conversely, substituting these relations into either exponent makes the exponents equal for every \(x\geq0\), so the CDFs agree everywhere. Therefore the distributions are identical. \(\square\)
The density conversion follows as well. Differentiating \(1-\exp[-\alpha x^k]\) for \(x>0\) gives \(f(x)=\alpha kx^{k-1}\exp[-\alpha x^k]\). Substituting \(\alpha=b^{-k}\) recovers the shape-scale density, since \(\alpha kx^{k-1}=k x^{k-1}/b^k=(k/b)(x/b)^{k-1}\).
Worked Example: A Nonunit Shape Makes the Power Matter
Let \(k=3\) and \(b=4\) hours. Find both rate parameters, then compute the CDF at \(x=2\) hours.
The inverse-scale rate is \(\rho=1/4=0.25\) per hour. The power-rate coefficient is \(\alpha=b^{-k}=4^{-3}=1/64=0.015625\) per hour cubed. At \(x=2\), the scale exponent is \((2/4)^3=(1/2)^3=0.125\).
In the inverse-scale form, the exponent is \((0.25\cdot2)^3=(0.5)^3=0.125\). In the power-rate form it is \(0.015625\cdot2^3=0.015625\cdot8=0.125\). Hence
The numerical check confirms the conversion: the reciprocal scale is \(0.25\), but the coefficient multiplying \(x^3\) is \(0.25^3=0.015625\), not \(0.25\).
Quantiles Under Each Convention
For a continuous distribution with a strictly increasing CDF on its positive support, the \(p\) quantile \(Q(p)\), for \(0<p<1\), is the unique value satisfying \(F(Q(p))=p\). The Weibull CDF is strictly increasing for \(x>0\), because its density is positive there.
Derivation. Set the CDF equal to \(p\) in the scale convention: \(p=1-\exp[-(Q(p)/b)^k]\). Rearranging gives \(\exp[-(Q(p)/b)^k]=1-p\). Taking logarithms yields \(-(Q(p)/b)^k=\ln(1-p)\), so \((Q(p)/b)^k=-\ln(1-p)\). The right side is positive for \(0<p<1\); taking the positive \(k\)th root and multiplying by \(b\) gives \(Q(p)=b[-\ln(1-p)]^{1/k}\).
In the inverse-scale form, the same rearrangement gives \((\rho Q(p))^k=-\ln(1-p)\), hence \(Q(p)=[-\ln(1-p)]^{1/k}/\rho\). In the power-rate form it gives \(\alpha Q(p)^k=-\ln(1-p)\), hence \(Q(p)=[-\ln(1-p)/\alpha]^{1/k}\). Substituting \(\rho=1/b\) and \(\alpha=b^{-k}\) shows that both rate expressions reduce to the scale expression. \(\square\)
Worked Example: Checking the 90th Percentile in Three Forms
For the \(k=3,\ b=4\)-hour model above, calculate the 90th percentile using each convention.
Since \(p=0.9\), \(-\ln(1-p)=-\ln(0.1)=\ln(10)\approx2.302585\). The scale formula gives \(Q(0.9)=4(2.302585)^{1/3}\approx4(1.32047)=5.2819\) hours.
Using \(\rho=0.25\), the inverse-scale formula gives \(Q(0.9)=(2.302585)^{1/3}/0.25=1.32047/0.25=5.2819\) hours. Using \(\alpha=0.015625\), the power-rate formula gives \(Q(0.9)=(2.302585/0.015625)^{1/3}=(147.3654)^{1/3}\approx5.2819\) hours.
The three calculations agree, including the units of the answer. Substituting the wrong rate quantity would not preserve this agreement.
Common Substitution Errors
The most frequent error is treating \(\rho\) and \(\alpha\) as interchangeable. They are equal only when \(k=1\) (or in the special numerical case where their values happen to coincide); in general, \(\alpha=\rho^k\). A second error is to use the formula from one convention while inserting a parameter from another. Check both the parameter definition and the formula before calculating.
Worked Example: Diagnosing a Reciprocal-Rate Error
Suppose \(k=2\) and \(b=2\) seconds. A calculation uses \(\rho=1/b=0.5\), then incorrectly inserts \(0.5\) as the coefficient in \(1-\exp(-\alpha x^2)\). Compare the correct and incorrect values of \(F(2)\).
The inverse-scale rate is indeed \(\rho=0.5\) per second. But the coefficient required by the power-rate formula is \(\alpha=\rho^k=(0.5)^2=0.25\) per second squared. At \(x=2\), the correct exponent is \(0.25\cdot2^2=0.25\cdot4=1\), so \(F(2)=1-\exp(-1)\approx1-0.3679=0.6321\).
The incorrect substitution \(\alpha=0.5\) gives exponent \(0.5\cdot2^2=2\), and therefore the incorrect CDF value \(1-\exp(-2)\approx1-0.1353=0.8647\). The discrepancy is not rounding error; the wrong coefficient defines a different Weibull distribution.
There is a simple dimensional warning here. For \(k=2\), \(\rho\) has units of inverse seconds, while \(\alpha\) has units of inverse seconds squared. A quantity with units of inverse seconds cannot be substituted for a coefficient multiplying \(x^2\) in a dimensionless exponent.
A Reliable Conversion Check
When given a Weibull parameter pair, do not infer a formula from the parameter labels alone. First identify the expression inside the negative exponential. Then translate the parameters so that this expression is unchanged. The following checks catch most errors before they affect a probability or quantile.
Identify whether its exponent is \((x/b)^k\), \((\rho x)^k\), or \(\alpha x^k\). Confirm the shape \(k\) is the exponent in the model.
Use \(\rho=1/b\) and \(\alpha=b^{-k}=\rho^k\). Do not replace a power-rate coefficient by the reciprocal scale unless \(k=1\).
Evaluate the exponent under both the original and converted forms. Equal exponents give equal survival probabilities and CDF values.
The exponent must be dimensionless, and the three quantile formulas must give the same value after conversion.
The distinction matters in statistical work because a reported parameter value has meaning only together with its convention. Two papers can report different numerical “rate” parameters for the same fitted distribution if one reports \(\rho\) and the other reports \(\alpha\). Conversely, using the same number under different formulas can silently change the model, its predicted probabilities, and its quantiles. Recording the survival function or CDF alongside the parameterization makes the model reproducible.
Common Mistakes
- Taking a reciprocal but forgetting the shape. The scale-to-inverse-scale conversion is \(\rho=1/b\), while the power-rate conversion is \(\alpha=b^{-k}\).
- Using a rate formula with the wrong rate. In \(\exp[-(\rho x)^k]\), use \(\rho\); in \(\exp[-\alpha x^k]\), use \(\alpha\).
- Assuming “rate” has one universal meaning. State the actual survival-function exponent whenever the convention is not specified.
- Checking only parameter labels. Check the exponent, units, and at least one probability or quantile instead.
Check Your Understanding
Use the defining exponent, not the parameter label alone, to answer each question.
- For shape \(k=4\) and scale \(b=3\), find the inverse-scale rate and the power-rate coefficient, including their units if \(X\) is measured in seconds.
- For a model with \(k=2\) and inverse-scale rate \(\rho=0.3\), write the equivalent power-rate coefficient and CDF.
- Derive the \(p\) quantile from the power-rate form \(F(x)=1-\exp(-\alpha x^k)\).
- A calculation inserts \(1/b\) directly as the coefficient of \(x^k\). Explain why this is generally incorrect and identify the special shape for which the numerical substitution is valid.
- For \(k=3,\ b=2\), compare the exponents at \(x=1\) using the scale and power-rate forms. What conversion makes them agree?