Why Normal Probability Calculations Go Wrong
A normal probability calculation can look convincing and still answer the wrong question. A right-tail probability can be mistaken for a left-tail area, the variance can be used where the standard deviation belongs, or a standardized cutoff can be reported as if it were in the original measurement units. Interval calculations add another risk: subtracting cumulative areas in the wrong order.
A reliable calculation keeps three things connected: the event stated in the problem, the corresponding event on the standard normal scale, and the area that represents that event. This tutorial focuses on checking those connections rather than introducing new ways to look up normal probabilities.
The model should be appropriate for the measurement being studied, and the mean and standard deviation must refer to that same variable and its units. For example, if \(X\) is measured in minutes, then \(\mu\) and \(\sigma\) must also be in minutes. Once these conditions are clear, use \(\sigma\)—not \(\sigma^2\)—to standardize.
A Short Derivation: Keep the Event Attached to the Cutoff
Standardization changes the scale, but it does not change which observations satisfy an event. Because \(\sigma\) is positive, dividing by \(\sigma\) preserves the direction of inequalities. For any original-scale cutoff \(x\),
Thus \(P(X\le x)=\Phi((x-\mu)/\sigma)\). A right-tail event remains a right-tail event after standardization, and an interval must have its lower standardized bound subtracted from its upper standardized bound:
The last equality follows because the cumulative area up to the upper bound contains both the interval and the area up to the lower bound. Subtracting that lower cumulative area leaves only the area between the bounds. For a continuous normal variable, including or excluding either endpoint does not change the probability.
Worked Examples: Matching the Event to the Area
Worked Example: A Right Tail Is Not a Left Tail
Suppose, hypothetically, that a machine’s fill volume \(X\), in millilitres, follows \(X\sim N(70,8^2)\). Find the probability that a fill exceeds 82 millilitres. Assume this normal model and its stated parameters are appropriate for the fills being considered.
The event is \(X>82\), so the requested area is to the right of the cutoff. First standardize 82 millilitres:
The standard normal cumulative area to the left of 1.5 is \(\Phi(1.5)=0.9332\). That is not the requested right-tail area, so take the complement:
The model gives a probability of about 0.0668, or 6.68%, that a fill exceeds 82 millilitres. A common wrong answer is 0.9332: it is a valid normal area, but it describes fills at or below 82 millilitres, the opposite event.
Worked Example: Standardize Both Interval Bounds Correctly
Suppose a hypothetical batch of packages has weights \(X\sim N(500,40^2)\), with weights measured in grams. Under this model, find the probability that a package weighs between 460 and 560 grams. Assume the normal model is appropriate for these package weights.
Standardize each endpoint using the same mean and standard deviation. The lower endpoint gives \(z_L=(460-500)/40=-1\), and the upper endpoint gives \(z_U=(560-500)/40=1.5\). Therefore the original event maps to \(-1<Z<1.5\), in that order.
The probability is about 0.7745, or 77.45%. A frequent error is to subtract the upper cumulative area from the lower one, obtaining \(0.1587-0.9332=-0.7745\). A negative probability signals that the areas have been reversed; the upper cutoff’s cumulative area must be larger.
Another error is to use 40² in the denominator. The variance is \(40^2=1600\) grams squared, but standardization divides by the standard deviation, 40 grams. Using 1600 would not produce a dimensionless z-score and would not correctly express distance from the mean in standard deviation units.
Worked Example: Keep the Original Units in the Conclusion
Suppose a hypothetical commute time \(X\) follows \(X\sim N(30,5^2)\), where times are measured in minutes. Assuming this model is appropriate, find the probability that a commute takes more than 35 minutes.
The cutoff is 35 minutes. Its standardized value is
Since the question asks for a time greater than 35 minutes, the event is the right tail beyond \(z=1\):
The model estimates a 0.1587 probability, or 15.87%, that a commute exceeds 35 minutes. The value \(z=1\) means the cutoff is one standard deviation above the mean; it is not a commute time of 1 minute. The probability describes the event in the original context, more than 35 minutes.
A Practical Error-Checking Sequence
Before accepting a numerical answer, check the calculation in a fixed order. This helps catch errors even when a table lookup or software output is accurate.
Write the requested event using \(X\), such as \(X>c\) or \(a<X<b\). This makes the tail direction and bounds explicit.
Confirm that the stated normal model, mean, and standard deviation apply to the variable in question, and that the cutoff uses the same measurement units.
For each original cutoff \(x\), calculate \(z=(x-\mu)/\sigma\). Use the standard deviation and check that the result is unit-free.
Use a left-tail cumulative area directly for a left-tail event, take its complement for a right-tail event, and subtract lower cumulative area from upper cumulative area for an interval.
State what the probability refers to and restore the original variable’s units when describing cutoffs or measurements. A probability itself has no measurement units.
Common Mistakes and Their Corrections
| Wrong reasoning | Correction |
|---|---|
| “The cutoff has \(z=1.5\), so the probability above it is \(\Phi(1.5)\).” | \(\Phi(1.5)\) is the area to the left. For the area above, use \(1-\Phi(1.5)\). |
| “The standard deviation is 8², so divide by 64 when standardizing.” | The model notation gives variance as \(\sigma^2\), but the z-score denominator is the standard deviation \(\sigma\), here 8. |
| “The z-score is 1, so the cutoff is 1 unit of the original measurement.” | A z-score is measured in standard deviation units. Convert back with \(x=\mu+z\sigma\) if an original-scale cutoff is needed. |
| “For an interval, subtract the smaller cumulative area from the larger one without checking which bound is which.” | Identify the lower and upper cutoffs first. Then calculate \(\Phi(z_{\text{upper}})-\Phi(z_{\text{lower}})\). |
| “The calculator or table gave an area, so it must answer the question.” | Check what area the table or software reports and compare it with the event. A correct lookup can still be the wrong tail. |
The central pitfall is treating a z-score or table entry as though it were already the answer. A z-score locates a cutoff on the standard normal scale, while a cumulative area has a particular direction and boundary. Always translate back to the event before interpreting the result.
Check Your Understanding
For each question, identify the relevant check before calculating or interpreting.
- If \(X\sim N(20,4^2)\), what standard normal area represents \(P(X>24)\): a left tail, a right tail, or an interval?
- A student standardizes \(x=45\) using \((45-\mu)/\sigma^2\). What is wrong with the denominator, and what denominator should be used?
- For \(P(a<X<b)\), with \(a<b\), which cumulative area should be subtracted from which?
- A cutoff has \(z=-0.8\). Does that mean the original measurement is \(-0.8\) units? Explain what the z-score means instead.
- A calculation produces a negative value after subtracting two cumulative areas for an interval. What should be checked first?