Why Normal Tail Problems Require Care
A normal probability statement can hide an equation for an unknown mean or standard deviation. The main work is not simply looking up a normal-table value: it is translating the event into a standard normal quantile with the correct direction. In particular, a right-tail probability is not the same thing as the left-tail probability used to define a standard normal quantile.
These problems are aligned with the SOA Exam P (Probability) syllabus. The method applies when the variable is normally distributed with a positive standard deviation. State that model before using a normal quantile; a calculation based on normal quantiles does not establish that an arbitrary variable is normally distributed.
Standardization converts the original variable to a standard normal variable: \(Z=(X-\mu)/\sigma\). Since \(\sigma>0\), dividing an inequality by \(\sigma\) does not reverse its direction. For example, \(P(X\le x)=\Phi((x-\mu)/\sigma)\). For a right tail, \(P(X>x)=1-\Phi((x-\mu)/\sigma)\), so an upper-tail probability \(r\) corresponds to the left-tail quantile \(z_{1-r}\).
A normal distribution is continuous, so \(P(X<x)=P(X\le x)\) and \(P(X>x)=P(X\ge x)\). This means an endpoint does not create a boundary probability. The equality is useful in setting up equations, but it does not remove the need to identify the correct tail.
From a Tail Statement to a Quantile Equation
The following short derivation explains why standardization gives the parameter-recovery equation. If \(X\sim N(\mu,\sigma^2)\), then \(Z=(X-\mu)/\sigma\sim N(0,1)\). For a left-tail statement \(P(X\le x)=p\), applying this transformation gives \(P(Z\le (x-\mu)/\sigma)=p\). By the definition of \(z_p\), the standardized threshold must equal \(z_p\). Rearranging produces \(x=\mu+z_p\sigma\). A right-tail statement \(P(X>x)=r\) first becomes \(P(X\le x)=1-r\), so its quantile equation uses \(z_{1-r}\), not \(z_r\).
Identify \(X\sim N(\mu,\sigma^2)\), specify which parameters are known, and keep \(\sigma>0\).
For a right tail \(P(X>x)=r\), use the equivalent left-tail probability \(P(X\le x)=1-r\).
Set \((x-\mu)/\sigma\) equal to the quantile for the left-tail probability, then solve for the unknown parameter or parameters.
Substitute the recovered parameters into each original standardized threshold. Confirm both the sign and the implied tail probability.
Recovering a Standard Deviation From One Tail
When the mean is known but the standard deviation is not, one probability statement usually supplies an equation for \(\sigma\). The sign of the standardized threshold is informative: if \(x<\mu\), then \((x-\mu)/\sigma\) must be negative for every valid \(\sigma\). A negative answer for \(\sigma\) signals a setup error, not a possible standard deviation.
Worked Example: A Large Probability Above a Low Threshold
A hypothetical processing time \(X\), in minutes, is modeled as \(X\sim N(72,\sigma^2)\), where \(\sigma\) is unknown. Suppose \(P(X>66)=0.9000\). Recover \(\sigma\) and the variance, then check the tail statement. Use \(z_{0.1000}\approx-1.2816\).
Method. Convert the right tail to a left-tail probability before using a quantile. The given statement implies \(P(X\le66)=1-0.9000=0.1000\). Therefore \[ \frac{66-72}{\sigma}=z_{0.1000}\approx-1.2816. \]
The numerator is \(66-72=-6\), so \(-6/\sigma=-1.2816\). Multiplying by \(-1\) gives \(6/\sigma=1.2816\), and hence \[ \sigma=\frac{6}{1.2816}\approx4.6816\text{ minutes}. \] The variance is the square of the standard deviation, not the standard deviation itself: \[ \sigma^2=\left(\frac{6}{1.2816}\right)^2\approx21.9178\text{ minutes}^2. \] The displayed results are rounded to four decimal places using the stated rounded quantile.
Check. Using \(\sigma=6/1.2816\), the standardized threshold is \((66-72)/\sigma=-6/(6/1.2816)=-1.2816\). Thus \(P(X\le66)\approx0.1000\), and \(P(X>66)\approx1-0.1000=0.9000\), as required.
Common mistake avoided. Using \(z_{0.9000}=1.2816\) directly for the right-tail probability would give \(-6/\sigma=1.2816\), implying a negative \(\sigma\). The correction is to complement the right-tail probability first. Also, \(4.6816\) is \(\sigma\); the variance is about \(21.9178\), not \(4.6816\).
Recovering Both the Mean and the Standard Deviation
Two distinct probability statements can determine both \(\mu\) and \(\sigma\), provided they give independent information. Each statement supplies a quantile equation of the form \(x=\mu+z_p\sigma\). Solving the pair of equations is often clearer than trying to infer the parameters separately by intuition. When table probabilities are rounded, the resulting parameters are approximate; verify them against the same rounded quantiles used in the calculation.
Worked Example: Two Tail Statements Determine Two Parameters
A hypothetical measurement \(X\) is modeled as \(N(\mu,\sigma^2)\), with both parameters unknown. Suppose \(P(X<80)=0.1587\) and \(P(X>104)=0.0228\). Use \(z_{0.1587}\approx-1\) and \(z_{0.9772}\approx2\) to recover \(\mu\) and \(\sigma\).
Method. The first statement is already a left-tail probability, so \[ \frac{80-\mu}{\sigma}\approx-1, \qquad\text{or}\qquad 80\approx\mu-\sigma. \] For the second statement, complement the upper tail: \(P(X\le104)=1-0.0228=0.9772\). Its quantile equation is \[ \frac{104-\mu}{\sigma}\approx2, \qquad\text{or}\qquad 104\approx\mu+2\sigma. \]
Subtract the first equation from the second: \(104-80\approx(\mu+2\sigma)-(\mu-\sigma)\), so \(24\approx3\sigma\), giving \(\sigma\approx8\). Substitute into \(80\approx\mu-\sigma\): \(80\approx\mu-8\), so \(\mu\approx88\). The recovered model is \(X\sim N(88,8^2)=N(88,64)\).
Check. At 80, the standardized value is \((80-88)/8=-1\), giving \(P(X<80)\approx0.1587\). At 104, it is \((104-88)/8=2\), giving \(P(X>104)\approx1-0.9772=0.0228\). Both statements agree with the recovered parameters.
Common mistake avoided. Treating \(0.0228\) as a left-tail probability would assign a negative quantile to 104, even though 104 is above the recovered mean. The upper-tail statement must instead use the complementary left-tail probability \(0.9772\).
Using Opposite Tails and Checking Inequality Direction
When one threshold lies below the mean and another lies above it, their standardized values should have opposite signs. More generally, the ordering of thresholds must agree with the ordering of their standardized values when \(\sigma>0\). This gives a useful check before or after solving. A sign mismatch often reveals that a tail was translated incorrectly.
Worked Example: Recovering Parameters From Unequal Tail Areas
A hypothetical claim amount \(X\), in thousands of dollars, is modeled as \(N(\mu,\sigma^2)\). Suppose \(P(X<40)=0.2000\) and \(P(X>70)=0.1000\), with unknown \(\mu\) and \(\sigma\). Use \(z_{0.2000}\approx-0.8416\) and \(z_{0.9000}\approx1.2816\). Recover both parameters and check the original tail probabilities.
Method. The lower-tail statement gives \[ \frac{40-\mu}{\sigma}\approx-0.8416, \qquad\text{so}\qquad 40\approx\mu-0.8416\sigma. \] For the upper tail, \(P(X\le70)=1-0.1000=0.9000\), so \[ \frac{70-\mu}{\sigma}\approx1.2816, \qquad\text{or}\qquad 70\approx\mu+1.2816\sigma. \]
Subtract the first threshold equation from the second: \[ 30\approx(1.2816+0.8416)\sigma=2.1232\sigma. \] Therefore \(\sigma\approx30/2.1232\approx14.1296\) thousand dollars. Substitute into the first equation: \[ \mu\approx40+0.8416(14.1296157)\approx40+11.8915=51.8915 \] thousand dollars. The parameters are reported to four decimal places, using the stated rounded quantiles.
Check. With these parameter values, \((40-\mu)/\sigma\approx-0.8416\), so \(P(X<40)\approx0.2000\). Also, \((70-\mu)/\sigma\approx1.2816\), so \(P(X>70)\approx1-\Phi(1.2816)\approx0.1000\). The lower threshold has a negative standardized value and the higher threshold a positive one, as expected.
Common mistake avoided. The \(0.1000\) in the second statement is an upper-tail area, not the left-tail area at 70. Using \(z_{0.1000}\approx-1.2816\) there would contradict the fact that 70 is greater than 40 and, for this solution, greater than the mean.
Common Traps and a Reliable Check
Normal tail problems often require only a few equations, but small translation errors can change the sign of a recovered parameter or produce an answer that appears plausible while failing the original condition. Before finalizing a solution, check all of the following:
- Identify which tail is stated. A statement \(P(X>x)=r\) corresponds to \(z_{1-r}\); a statement \(P(X\le x)=p\) corresponds to \(z_p\).
- Respect the sign. If \(x<\mu\), then \((x-\mu)/\sigma<0\); if \(x>\mu\), it is positive. These signs hold because \(\sigma>0\).
- Keep standard deviation and variance distinct. The quantile equation solves for \(\sigma\). Report variance only after squaring: \(\operatorname{Var}(X)=\sigma^2\).
- Check every original statement. Re-standardize each threshold and verify whether the resulting probability is a left tail or a right tail.
- Keep the precision consistent. If a normal table supplies rounded quantiles, parameter values calculated from them are approximate. Do not claim more precision than the quantiles support.
These checks are especially useful when the problem asks for a probability involving a threshold on the opposite side of the mean from the event stated in the data. The safest approach is always to write the event, convert it to a left-tail probability if needed, and only then select the quantile.
Check Your Understanding
For each problem, state the quantile equation before solving. Answers use standard normal table values rounded to the precision shown.
- A hypothetical reading \(X\sim N(50,10^2)\) is recorded. Find the value \(c\) such that \(P(X>c)=0.1000\). (Answer: \(c\approx62.82\))
- A hypothetical duration \(X\sim N(\mu,5^2)\) satisfies \(P(X<37)=0.8413\). Recover \(\mu\). (Answer: \(\mu\approx32\))
- A hypothetical measurement \(X\sim N(100,\sigma^2)\) satisfies \(P(X>92)=0.9000\). Recover \(\sigma\), using \(z_{0.1000}\approx-1.2816\). (Answer: \(\sigma=8/1.2816\approx6.2422\))
- A hypothetical score \(X\sim N(\mu,\sigma^2)\) satisfies \(P(X<80)=0.1587\) and \(P(X>104)=0.0228\). Recover both parameters, using \(z_{0.1587}\approx-1\) and \(z_{0.9772}\approx2\). (Answer: \(\mu\approx88,\ \sigma\approx8\))
- A hypothetical loss amount \(X\), in thousands of dollars, is normal with unknown parameters. Suppose \(P(X<47)=0.1587\) and \(P(X>71)=0.0228\). Recover \(\mu,\sigma\), then find \(P(55<X<67)\). Use \(z_{0.1587}\approx-1\), \(z_{0.9772}\approx2\), \(\Phi(1.5)\approx0.9332\), and \(\Phi(0)=0.5000\). (Answer: \(\mu\approx55,\ \sigma\approx8,\ P(55<X<67)\approx0.4332\))