
judge whether a normal model fits a data set, explain why it fails for skewed data, and use and critique the normal approximation to a binomial distribution.
Instruction
A2.DS.B.8 / A2.DS.B.9 ends with a warning: recognize that there are data sets for which such a procedure is not appropriate. A normal model can always be computed from a mean and an SD. Whether it describes the data is a separate question.
A failure: St. Louis snowfall. Calendar-year snowfall at St. Louis Lambert International Airport, 1950-2024, has mean 18.36 in and SD 10.41 in. Model it as N(18.36, 10.41) and check:
- The model gives P(X < 0) ≈ 0.039. It predicts that about 1 year in 26 has negative snowfall, which is impossible.
- The model gives P(X < 5) ≈ 0.100. The data: 3 of 75 years (0.040). The model more than doubles the frequency of very snowless years.
- The model gives P(X ≥ 40) ≈ 0.019. The data: 3 of 75 (0.040). It halves the frequency of very snowy years.
The data are skewed right, and the model is symmetric. It puts too much probability in the lower tail and too little in the upper tail. The within-1-SD percentage (68%) is right on target, which is why checking only the center can mislead you.
How to check normality.
- Graph it. A histogram or dot plot should be roughly symmetric and single-peaked, without heavy tails or clear outliers.
- Compare mean and median. A large gap signals skew (St. Louis snowfall: 18.36 versus 16.3).
- Check the percentages. Compare counts within 1, 2 and 3 SDs and in each tail with 68-95-99.7.
- Normal probability plot. Plot each data value against the z-score a normal distribution would give its rank. Normal data fall close to a straight line; curvature means skew, and bends at the ends mean heavy or light tails. CODAP and most statistics software draw these. In a spreadsheet, sort the data and plot it against NORM.S.INV((rank − 0.5)/n).
The normal approximation to the binomial (PS.9). A binomial count is a sum of n independent 0/1 trials. When n is large, sums of many independent pieces are approximately normal. This is the central limit theorem, previewed here and developed in Lesson 1.4. So Binomial(n, p) is approximately N(np, √(np(1 − p))) when np ≥ 10 and n(1 − p) ≥ 10.
Example 1. Heads in 14 fair tosses: μ = 7, σ = √3.5 ≈ 1.871. Exactly, P(X ≤ 5) = 0.212 (RLP Statistics A). The normal curve with a continuity correction (a bar for 5 covers 4.5 to 5.5, so use 5.5) gives z = (5.5 − 7)/1.871 ≈ −0.80 and an area of 0.211, very close. Without the correction, z = (5 − 7)/1.871 ≈ −1.07 gives 0.143, which is far off. Here np = 7 misses the rule of thumb, yet the approximation works because p = 0.5 makes the binomial symmetric.
Example 2. St. Louis rainy days: in a random sample of 100 days from 2000-2024, p = 0.3117, so X ~ Binomial(100, 0.3117) with μ = 31.17 and σ ≈ 4.63 (np = 31.17 and n(1 − p) = 68.83 both pass). P(X ≥ 40): exact 0.038; normal with continuity correction (39.5) 0.036. The approximation is decent but slightly low, because a binomial with p below one half is still a little skewed right.
For discussion (MP.3). Is “n ≥ 30 makes everything normal” a rule, or a rule of thumb? What should decide whether a normal model is appropriate: a threshold, or a check of the data? Write your answer, or talk it through with a parent or guardian.
📖 Read for this lesson
The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.
Formative check
Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.
A normal model for St. Louis annual snowfall (mean 18.36, SD 10.41) predicts about 3.9% of years below 0 inches. What does this show?
Simulation & tools
Watch a binomial become normal
On the Lab screen, set the binary probability to 0.2.
- With 4 rows, drop 1,000 balls and sketch the histogram. Record its shape.
- Increase to 12 rows, then to the maximum number of rows, dropping 1,000 balls each time. Record the shape, the sim’s mean and SD, and np and √(np(1 − p)) for each.
- What is np at the largest number of rows the sim allows? How many rows would np ≥ 10 need? Does the histogram already look roughly normal before that?
Question: why does the histogram become more symmetric as the number of rows grows, even though p = 0.2 stays fixed? No sim: use BINOM.DIST in a spreadsheet to chart Binomial(4, 0.2), Binomial(12, 0.2) and Binomial(50, 0.2) as column charts.
Open in a new tab ↗ · PhET Interactive Simulations, University of Colorado Boulder · CC BY 4.0
For heads in 14 fair tosses, the normal approximation without a continuity correction gives P(X ≤ 5) ≈ 0.143, much further from the exact 0.212 than the corrected 0.211.
Record (or write) a 90-second argument about whether a normal model is appropriate for a data set of your choice (NOAA annual temperature, snowfall or rainfall at a Missouri station, or CDC rates). Use at least two checks from the lesson, one of them a graph. Then add a short note on whether the same two checks would support the opposite conclusion for a different data set you have seen in this lesson.
How confident are you that you can decide whether a normal model fits data, and use the normal approximation to the binomial with its conditions?
Practice
Work these on paper or in your notebook, then open Check your answer. Aim for all of Fluency and Application; try at least one Challenge.
Printable version: this unit’s practice workbook (PDF)
Fluency
Build speed and accuracy with the core skill.
- X ~ Binomial(n = 9, p = 0.25). Find P(X = 8), P(X ≤ 8), and the mean and standard deviation of X.
Check your answer
Answer: P(X = 8) ≈ 0.0001; P(X ≤ 8) ≈ 1.0000; mean 2.25; SD ≈ 1.299P(X = k) = C(9, k) pk(1 − p)n−k. - X ~ Binomial(n = 7, p = 0.6). Find P(X = 2), P(X ≤ 2), and the mean and standard deviation of X.
Check your answer
Answer: P(X = 2) ≈ 0.0774; P(X ≤ 2) ≈ 0.0963; mean 4.20; SD ≈ 1.296P(X = k) = C(7, k) pk(1 − p)n−k. - Check the normal-approximation conditions for Binomial(40, 0.1).
Check your answer
Answer: np = 4 < 10, so the normal approximation is not appropriate. - Check the conditions for Binomial(200, 0.3).
Check your answer
Answer: np = 60 and n(1 − p) = 140, both ≥ 10: appropriate. - Binomial(100, 0.5): find the mean and SD of the approximating normal distribution.
Check your answer
Answer: μ = 50, σ = 5. - To approximate P(X ≤ 12) for a binomial count with a normal curve, which boundary should you use?
Check your answer
Answer: 12.5 (continuity correction). - To approximate P(X ≥ 25), which boundary should you use?
Check your answer
Answer: 24.5. - X ~ Binomial(100, 0.3117) (rainy days in a random sample of 100 St. Louis days). Find P(X ≤ 25) exactly and with the normal approximation (continuity correction).
Check your answer
Answer: Exact ≈ 0.1092; normal ≈ 0.1106 (μ = 31.17, σ ≈ 4.632). - X ~ Binomial(50, 0.4). Find P(X ≥ 25) exactly and by the normal approximation.
Check your answer
Answer: Exact ≈ 0.0978; normal ≈ 0.0970 (μ = 20, σ ≈ 3.464).
Application
Use the skill in context. Show your reasoning.
- Springfield calendar-year precipitation, 1950-2024: mean 42.92 in, SD 8.51 in. A normal model gives P(X < 30) ≈ 0.064; the data have 6 of 75 years below 30 inches. The mean (42.92) and median (43.25) are close. Is a normal model reasonable here? Compare with St. Louis snowfall.
Check your answer
Answer: Reasonably: the prediction (4.8 years) is close to the observed 6, and mean ≈ median suggests little skew. St. Louis snowfall, by contrast, is clearly right-skewed, and the normal model misstates both tails (it predicts negative snowfall).Check the center, the tails and a graph. - Springfield annual mean temperature, 1950-2024: 69.3% of years within 1 SD and 94.7% within 2 SD of the mean. The record also has a strong upward trend (see the era comparison in Lesson 3.4). Is a single normal model a good description of these 75 values?
Check your answer
Answer: The percentages look close to 68% and 95%, but the data are not a stable distribution: the trend means early and recent years come from different centers, so a single normal model hides the most important feature. Model the trend first, or describe eras separately. - A student says: ‘n = 30, so the binomial with p = 0.05 is approximately normal.’ Evaluate the claim.
Check your answer
Answer: Wrong: np = 1.5, far below 10. With p this small, the distribution is strongly skewed right, and a normal approximation would give poor (even negative-count) answers.
Challenge
Stretch problems. Expect to think before you write.
- For p = 0.2, what is the smallest n for which both np ≥ 10 and n(1 − p) ≥ 10? Explain why the binomial becomes more normal as n grows.
Check your answer
Answer: n = 50 (np = 10, n(1 − p) = 40). A binomial count is a sum of n independent 0/1 trials, and by the central limit theorem sums of many independent terms are approximately normal.
Review
Keep earlier skills sharp.
- (RLP Statistics A) Why might a binomial model fail for rainy days per week at St. Louis?
Check your answer
Answer: Rainy days are not independent (storms last several days), so the counts cluster at 0 and at high values. - (RLP Statistics A) Find the z-score of 39.5 in a distribution with mean 18.36 and SD 10.41.
Check your answer
Answer: z ≈ 2.03.
Seventy-Five Snow Years at St. Louis: The Normal Model That Fails, and How to Show It
NOAA NCEI Global Summary of the Year, element SNOW, station USW00013994 (St. Louis Lambert International Airport), 1950-2024; normal-score and binomial tables computed for this course; the planning memo is an SLS original framing; retrieved 2026-11-10 · Public domain (works of the United States government), with SLS original computed tables
Why this reading. The lesson’s central example is a normal model that fails, fitted to 75 years of St. Louis snowfall, and the discourse task asks you to run at least two checks, one of them a graph, on a data set of your choice. The snowfall years are printed here in full so that the lesson’s every claim about them can be checked, along with the expected z-scores a normal probability plot needs and the exact binomial table for the coin-toss check. It ends with a planning claim to argue about.
Table 1. Calendar-year snowfall at St. Louis Lambert International Airport, 1950–2024 (inches)
NOAA NCEI Global Summary of the Year, element SNOW, station USW00013994. Retrieved 2026-11-10. A calendar year splits each winter in two (January–March of one season and November–December of the next), which is how the product reports it.
| Year | Inches | Year | Inches | Year | Inches | Year | Inches | Year | Inches |
|---|---|---|---|---|---|---|---|---|---|
| 1950 | 11.2 | 1965 | 21.8 | 1980 | 34.5 | 1995 | 14.4 | 2010 | 19.4 |
| 1951 | 27.2 | 1966 | 10.7 | 1981 | 16.9 | 1996 | 18.9 | 2011 | 30.7 |
| 1952 | 10.2 | 1967 | 9.3 | 1982 | 28.7 | 1997 | 24.9 | 2012 | 6.4 |
| 1953 | 6.9 | 1968 | 16.3 | 1983 | 13.9 | 1998 | 14.1 | 2013 | 31.5 |
| 1954 | 1.5 | 1969 | 20.8 | 1984 | 21.0 | 1999 | 14.4 | 2014 | 27.3 |
| 1955 | 5.5 | 1970 | 12.7 | 1985 | 12.1 | 2000 | 25.8 | 2015 | 12.8 |
| 1956 | 23.4 | 1971 | 9.8 | 1986 | 7.4 | 2001 | 2.9 | 2016 | 11.1 |
| 1957 | 8.3 | 1972 | 12.5 | 1987 | 31.5 | 2002 | 23.4 | 2017 | 6.5 |
| 1958 | 30.0 | 1973 | 32.0 | 1988 | 19.8 | 2003 | 23.6 | 2018 | 10.6 |
| 1959 | 5.3 | 1974 | 18.8 | 1989 | 24.1 | 2004 | 5.9 | 2019 | 23.9 |
| 1960 | 40.7 | 1975 | 34.0 | 1990 | 28.8 | 2005 | 14.8 | 2020 | 8.1 |
| 1961 | 26.6 | 1976 | 19.3 | 1991 | 8.0 | 2006 | 11.1 | 2021 | 12.6 |
| 1962 | 18.2 | 1977 | 49.3 | 1992 | 9.9 | 2007 | 16.7 | 2022 | 15.1 |
| 1963 | 12.9 | 1978 | 49.2 | 1993 | 32.4 | 2008 | 26.4 | 2023 | 3.0 |
| 1964 | 28.8 | 1979 | 25.6 | 1994 | 7.8 | 2009 | 12.5 | 2024 | 12.4 |
Sum 1376.8, so x̄ = 18.36 in; s = 10.41 in; median 16.3 in (the 38th value in order). Mean above median by about two inches: the first sign of right skew. Minimum 1.5 (1954), maximum 49.3 (1977).
The model against the data, cell by cell
| Event | Normal model N(18.36, 10.41) | The 75 years | Which years |
|---|---|---|---|
| Below 0 in | Φ((0 − 18.36)/10.41) = Φ(-1.76) = 0.039 | 0 of 75 | impossible |
| Below 5 in | Φ(-1.28) = 0.100 | 3 of 75 = 0.040 | 1954, 2001, 2023 |
| At or above 40 in | 1 − Φ(2.08) = 0.019 | 3 of 75 = 0.040 | 1960, 1977, 1978 |
| Within ±1 SD | 0.683 | 51 of 75 = 0.680 | the center passes |
These are the lesson’s figures. The center passes and both tails fail: the model predicts negative snowfall, too many very low years, and too few very high ones, which is what a symmetric curve does to a skewed pile.
The graph: a normal probability plot by hand
Sort the 75 values. Plot the kth smallest against the z-score a normal distribution assigns to rank k of 75, which is Φ−1((k − 0.5)/75), the spreadsheet’s NORM.S.INV. The lower half of the ranks is tabulated below; rank 76 − k takes the same z with the opposite sign.
| Rank | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| z | -2.47 | -2.05 | -1.83 | -1.68 | -1.55 | -1.45 | -1.36 | -1.28 | -1.21 | -1.14 | -1.08 | -1.02 | -0.97 | -0.92 | -0.87 | -0.82 | -0.77 | -0.73 | -0.69 |
| Rank | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| z | -0.64 | -0.60 | -0.56 | -0.52 | -0.49 | -0.45 | -0.41 | -0.38 | -0.34 | -0.31 | -0.27 | -0.24 | -0.20 | -0.17 | -0.13 | -0.10 | -0.07 | -0.03 | 0.00 |
Six points fix the shape: rank 1 (1.5 in) against -2.47, rank 10 (7.4) against -1.14, rank 38 (16.3) against 0, rank 66 (30.7) against +1.14, rank 72 (34.5) against +1.68, rank 75 (49.3) against +2.47. Normal data sit on a line. Here the low end sits above the line (a normal curve would reach below zero) and the high end climbs away from it: the signature of right skew. Sketch it and say that in words.
Table 2. Exact Binomial(14, 0.5), for the coin-toss check
| k | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| P(X = k) | 0.0001 | 0.0009 | 0.0056 | 0.0222 | 0.0611 | 0.1222 | 0.1833 | 0.2095 | 0.1833 | 0.1222 | 0.0611 | 0.0222 | 0.0056 | 0.0009 | 0.0001 |
P(X ≤ 5) = 0.0001 + 0.0009 + 0.0056 + 0.0222 + 0.0611 + 0.1222 = 0.2120, the lesson’s exact value. With the continuity correction, z = (5.5 − 7)/1.871 = −0.80 and Φ(−0.80) = 1 − 0.7881 = 0.2119 from Table 1 of the Lesson 1.2 reading. Without it, z = −1.07 gives 0.1423. You can now see why: the uncorrected version throws away half of the 0.1222 bar at k = 5.
The claim to argue about (SLS original framing, real data)
A planning memo, written in the style such memos are written in: “St. Louis averages 18 inches of snow a year with a standard deviation of about 10. A winter above 40 inches is therefore a 2-in-100 event, and the salt and plowing budget, sized for 40 inches, will be exhausted roughly twice a century.”
Take it apart with the tools of this lesson. The model’s figure is right as arithmetic: z = (40 − 18.36)/10.41 = 2.08, and 1 − Φ(2.08) = 0.019, about 2 in 100. The record: 3 of 75 years exceeded 40 inches (1960, 1977, 1978), which is 4.0%, or about 1 year in 25. The normal model understates the snowy tail by a factor of two, because it was never a model of this shape. For the dry side, the memo’s logic would say that a year under 5 inches is a 1-in-10 event; the record has 3 in 75 (1954, 2001, 2023), 1 in 25, so the model overstates the dry tail by the same factor. And its mathematics permits a year with less than zero inches, which no budget can use. Sums of many small influences are normal; a total that is pinned at zero on one side and has a few huge storms on the other is not. Two more cautions: a record of 75 years cannot distinguish a 1-in-50 from a 1-in-25 event with any confidence (3 events is a small count), and 1977 and 1978 are consecutive, so the three exceedances are not three independent draws.
Sources: NOAA NCEI, Global Summary of the Year data search (ncei.noaa.gov); NOAA NCEI Access Data Service (the query used here) (ncei.noaa.gov)