Mean and Standard Deviation
CSIR NET Chemical Sciences data analysis: mean, median, sample vs population standard deviation, RSD, standard error, Gaussian curve, confidence limits, Q-test.
- 18 min
On this page · 15 sections
- Prerequisites
- The Hook: Five Titres of the Same Acid
- Measures of Central Tendency
- Variance and Standard Deviation
- The Standard Deviation of the Mean
- The Gaussian (Normal) Distribution
- Confidence Limits for the True Value
- Rejecting a Suspect Result: the Q-test
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Averages and weighted averages from Percentages, Ratios, Averages and Mixtures. The sample mean is the same arithmetic.
- Centre, spread and the shape of a distribution from Histograms, Distributions, Scatter and Box Plots. That page reads them off graphs; this page computes them from replicate measurements.
- This is the first page of Data Analysis. The next page, Absolute and Relative Errors, uses the standard deviation defined here.
The Hook: Five Titres of the Same Acid
Measures of Central Tendency
For \( n \) replicate measurements \( x_1, x_2, \dots, x_n \):
$$\boxed{\bar{x} = \frac{1}{n}\sum_{i=1}^{n} x_i}$$The arithmetic mean \( \bar{x} \) is the best estimate of the true value when only random errors are present. The population mean \( \mu \) is the mean of an infinite number of measurements; \( \bar{x} \) approaches \( \mu \) as \( n \) grows.
| Measure | Definition | Use |
|---|---|---|
| Mean \( \bar{x} \) | sum ÷ number | default; uses every value |
| Median | middle value of the ordered set; mean of the two middle values when \( n \) is even | resistant to one wild value |
| Mode | most frequent value | grouped or discrete data |
| Range \( w \) | largest − smallest | quick spread; used in the Q-test |
For the five titres, ordered: 24.30, 24.32, 24.34, 24.36, 24.38. Mean = median = 24.34 mL, range = 0.08 mL.
Deviations always sum to zero
The deviation of a result from the mean is \( d_i = x_i - \bar{x} \). By the definition of the mean,
$$\sum_{i=1}^{n} (x_i - \bar{x}) = 0$$This is why the spread cannot be the average deviation (it is always zero). Two ways round it: average the absolute deviations (the mean deviation), or average the squared deviations (the variance). Statistics uses the second, because squares behave well algebraically.
The assumed-mean trick
To average numbers that share most of their digits, subtract a convenient working value, average the small remainders, then add it back. For the titres, subtract 24.30 and multiply by 100: the coded values are 2, 6, 0, 8, 4, with mean 4. Undo the coding: \( 24.30 + 4/100 = 24.34 \) mL. The same trick works for the standard deviation (see below), with one difference: shifting does not change the spread, but scaling does.
Variance and Standard Deviation
Population and sample
$$\boxed{\sigma = \sqrt{\frac{\sum_{i=1}^{N} (x_i - \mu)^2}{N}}} \qquad \boxed{s = \sqrt{\frac{\sum_{i=1}^{n} (x_i - \bar{x})^2}{n - 1}}}$$- \( \sigma \) is the population standard deviation, about the true mean \( \mu \).
- \( s \) is the sample standard deviation, the estimate of \( \sigma \) from \( n \) replicates. Chemists report \( s \) for replicate data.
- \( s^2 \) and \( \sigma^2 \) are the variances.
Why \( n - 1 \)? The deviations are measured from \( \bar{x} \), which was itself computed from the same data. Once \( \bar{x} \) is fixed, only \( n - 1 \) deviations are free; the last is forced by \( \sum d_i = 0 \). So there are \( n - 1 \) degrees of freedom. Dividing by \( n \) would systematically underestimate the spread (strictly, the variance \( \sigma^2 \)), because the data sit closer to their own mean than to \( \mu \). For large \( n \) the difference vanishes.
Computational form
Expanding the square gives a form that needs only two running totals:
$$\boxed{s = \sqrt{\frac{\sum x_i^2 - \frac{\left(\sum x_i\right)^2}{n}}{n - 1}}}$$Use it with coded data. With raw numbers such as 24.32 it subtracts two nearly equal large numbers and loses digits, on a calculator or in your head.
Relative standard deviation
The SD has the units of the data. To compare precision across different quantities or concentration levels, divide by the mean:
$$\boxed{\text{RSD} = \frac{s}{\bar{x}}} \qquad \text{CV} = \frac{s}{\bar{x}} \times 100\% \qquad \text{in ppt: } \frac{s}{\bar{x}} \times 1000$$The coefficient of variation (CV) is the RSD as a percentage. For the titres, RSD = 0.0316/24.34 = 0.0013 = 0.13% = 1.3 ppt (parts per thousand).
What shifting and scaling do
If every result is transformed as \( y = a x + b \):
| Quantity | Becomes | Example: °C to K (\( a = 1, b = 273.15 \)) |
|---|---|---|
| Mean | \( a\bar{x} + b \) | shifts by 273.15 |
| SD | \( \lvert a \rvert\, s \) | unchanged |
| Variance | \( a^2 s^2 \) | unchanged |
| RSD | changes unless \( b = 0 \) | falls, because the mean grows |
So converting a set of readings from mL to L divides the SD by 1000; converting °C to °F (\( a = 1.8 \)) multiplies it by 1.8 and the variance by 3.24.
Pooling several small sets
When the same method is applied to several samples, each measured a few times, the precision estimates can be pooled (assuming the method has the same \( \sigma \) for all of them):
$$s_{\text{pooled}} = \sqrt{\frac{\sum_{j} (n_j - 1)\, s_j^2}{\sum_{j} (n_j - 1)}}$$Each set contributes its own degrees of freedom. The pooled estimate is more reliable than any one set’s \( s \).
The Standard Deviation of the Mean
A mean of \( n \) results scatters less than single results do. Its standard deviation, the standard error of the mean, is
$$\boxed{s_{\bar{x}} = \frac{s}{\sqrt{n}}}$$For the titres, \( s_{\bar{x}} = 0.0316/\sqrt{5} = 0.0141 \) mL.
The square root is the important part. To halve the uncertainty of a mean you need four times as many replicates; to cut it by ten, a hundred times as many. Precision bought by repetition gets expensive fast, which is why a better method usually beats more replicates.
The Gaussian (Normal) Distribution
Random errors in a well-behaved measurement follow the Gaussian (normal) distribution:
$$y = \frac{1}{\sigma\sqrt{2\pi}} \exp\!\left[-\frac{(x - \mu)^2}{2\sigma^2}\right]$$- Symmetric about \( \mu \), so mean = median = mode.
- Its inflection points are at \( \mu \pm \sigma \).
- The total area under the curve is 1; the area between two values is the probability that a result falls between them.
- A smaller \( \sigma \) gives a taller, narrower curve (better precision).
With \( z = (x - \mu)/\sigma \), the fraction of results expected within a band about the mean is:
| Band | Fraction inside | Fraction outside |
|---|---|---|
| \( \mu \pm 1\sigma \) | 68.3% | 31.7% |
| \( \mu \pm 1.96\sigma \) | 95.0% | 5.0% |
| \( \mu \pm 2\sigma \) | 95.4% | 4.6% |
| \( \mu \pm 3\sigma \) | 99.7% | 0.3% |
graph LR
A["Replicates x1 ... xn"] --> B["Mean x̄"]
A --> C["Sample SD s
(n − 1 in denominator)"]
C --> D["RSD = s / x̄"]
C --> E["Standard error
s / √n"]
E --> F["Confidence interval
x̄ ± t·s/√n"]
B --> FConfidence Limits for the True Value
A confidence interval is a range that, at a stated probability, contains the true mean \( \mu \).
If \( \sigma \) is known (from long experience with the method):
$$\mu = \bar{x} \pm \frac{z\sigma}{\sqrt{n}} \qquad (z = 1.96 \text{ for } 95\%)$$If only \( s \) from a few replicates is available (the usual case), the extra uncertainty in \( s \) is allowed for by Student’s \( t \):
$$\boxed{\mu = \bar{x} \pm \frac{t\, s}{\sqrt{n}}} \qquad t \text{ for } n - 1 \text{ degrees of freedom}$$| Degrees of freedom | 1 | 2 | 3 | 4 | 5 | ∞ |
|---|---|---|---|---|---|---|
| \( t \) at 95% (two-sided) | 12.71 | 4.30 | 3.18 | 2.78 | 2.57 | 1.96 |
\( t \) is large for few degrees of freedom and falls towards \( z = 1.96 \) as \( n \) grows. Two replicates (one degree of freedom) give a very wide interval.
Two quick significance tests
These go one step beyond the syllabus line but appear in data-based questions:
- t-test against a known value. To check a method on a certified reference material of value \( \mu \): \( t_{\text{calc}} = \frac{\lvert \bar{x} - \mu \rvert \sqrt{n}}{s} \). If \( t_{\text{calc}} \) exceeds the tabulated \( t \), the difference is significant and the method probably has a systematic error.
- F-test for two precisions. \( F = s_1^2/s_2^2 \) with the larger variance on top. If \( F \) exceeds the tabulated value for the two sets of degrees of freedom, one method is genuinely more precise.
Both compare a calculated statistic with a table value. In an exam the critical value is given.
Rejecting a Suspect Result: the Q-test
When one result sits apart from the others, Dixon’s Q-test asks whether it can be discarded:
$$\boxed{Q = \frac{\lvert x_{\text{suspect}} - x_{\text{nearest}} \rvert}{w}} \qquad w = \text{range, including the suspect value}$$If \( Q > Q_{\text{crit}} \) (tabulated for \( n \) results and a confidence level), reject the value; otherwise keep it. Test only one value, and only for small sets. If a result fails the test, recompute both the mean and the SD without it.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Sample SD \( s \), divide by \( n - 1 \) | population SD \( \sigma \), divide by \( N \) about \( \mu \) | |
| SD \( s \): scatter of single results | standard error \( s/\sqrt{n} \): scatter of the mean | |
| Variance \( s^2 \) (units squared) | SD \( s \) (same units as data) | |
| RSD = \( s/\bar{x} \) (a fraction) | CV = RSD × 100% | |
| \( z \) (σ known, 1.96 at 95%) | \( t \) (σ estimated, larger, depends on \( n - 1 \)) | |
| Median resists an outlier | mean is pulled by it |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- Why does the sample standard deviation use \( n - 1 \) rather than \( n \)?
- A result is reported as 2.00 ± 0.02 mM (± one SD). What is its RSD in percent?
- Nine replicates give \( s = 0.6 \) mg. What is the standard error of the mean?
Answers
- The deviations are taken from \( \bar{x} \), which was computed from the same data, so only \( n - 1 \) of them are independent (\( \sum d_i = 0 \) fixes the last). Dividing by \( n \) would underestimate \( \sigma \).
- 1%. \( 0.02/2.00 \times 100 \).
- 0.2 mg. \( 0.6/\sqrt{9} = 0.6/3 \).
Level 2: Part B Style
Practice Q1. The sample standard deviation of 10, 12, 14, 16 and 18 is:
(a) 2.83 (b) 3.16 (c) 4.00 (d) 10.0
Answer: (b) 3.16
Mean 14; deviations −4, −2, 0, 2, 4; sum of squares 40.
\( s = \sqrt{40/4} = \sqrt{10} = 3.16 \).
(a) is \( \sqrt{40/5} \), the divide-by-\( n \) value. (d) is the variance.
Practice Q2. A set of results has mean \( \bar{x} \) and standard deviation \( s \). Each result is multiplied by 2 and then 5 is added. The new mean and standard deviation are:
(a) \( 2\bar{x} + 5 \) and \( 2s + 5 \) (b) \( 2\bar{x} + 5 \) and \( 2s \) (c) \( 2\bar{x} \) and \( 4s \) (d) \( 2\bar{x} + 5 \) and \( 4s \)
Answer: (b)
The mean follows the transformation: \( 2\bar{x} + 5 \). Every deviation is doubled and the shift cancels, so the SD becomes \( 2s \).
(d) confuses the SD with the variance, which does become \( 4s^2 \).
Practice Q3. An analyst reports the mean of 5 replicate measurements. To halve the standard error of the mean with the same method, the number of replicates should be:
(a) 10 (b) 15 (c) 20 (d) 25
Answer: (c) 20
\( s_{\bar{x}} = s/\sqrt{n} \). Halving it needs \( \sqrt{n} \) doubled, so \( n \) multiplied by 4: \( 5 \times 4 = 20 \).
Practice Q4. Replicate weighings follow a Gaussian distribution with \( \mu = 50.0 \) mg and \( \sigma = 0.50 \) mg. The percentage of weighings expected outside 49.0–51.0 mg is about:
(a) 0.3% (b) 4.6% (c) 5.0% (d) 31.7%
Answer: (b) 4.6%
49.0 and 51.0 mg are \( \mu \pm 2\sigma \). The area inside is 95.4%, so 4.6% falls outside.
(c) is the value for ±1.96σ; (a) is for ±3σ; (d) is for ±1σ.
Level 3: Part C Style
Practice Q5. Five replicate determinations of chloride in a water sample give a mean of 35.42 mg L⁻¹ with \( s = 0.20 \) mg L⁻¹. Take \( t = 2.776 \) (4 degrees of freedom, 95%) and \( z = 1.96 \).
(i) The 95% confidence interval for the true chloride content. (ii) Treating 0.20 mg L⁻¹ as the known \( \sigma \) of the method, the smallest number of replicates that gives a 95% interval of ±0.10 mg L⁻¹ or better.
(a) 35.42 ± 0.25 mg L⁻¹; 16 (b) 35.42 ± 0.18 mg L⁻¹; 16 (c) 35.42 ± 0.25 mg L⁻¹; 15 (d) 35.42 ± 0.56 mg L⁻¹; 4
Answer: (a)
(i) \( \pm \frac{t s}{\sqrt{n}} = \pm \frac{2.776 \times 0.20}{\sqrt{5}} = \pm 0.248 \approx \pm 0.25 \) mg L⁻¹. Option (b)’s ±0.18 uses 1.96 instead of \( t \); (d)’s ±0.56 forgets \( \sqrt{n} \).
(ii) With \( \sigma \) known: \( \frac{1.96 \times 0.20}{\sqrt{n}} \le 0.10 \Rightarrow \sqrt{n} \ge 3.92 \Rightarrow n \ge 15.37 \). Round up: 16 replicates. 15 gives \( 1.96 \times 0.20/\sqrt{15} = 0.101 \), just outside the target.
Practice Q6. Five standardisations of a NaOH solution give 0.1012, 0.1014, 0.1016, 0.1018 and 0.1034 M. Take \( Q_{\text{crit}} = 0.71 \) for \( n = 5 \) at 95% confidence. Consider the statements:
- The Q value for 0.1034 M is 0.73 and the result should be rejected.
- After rejection, the mean is 0.1015 M.
- After rejection, the RSD is about 0.25%.
- The median of the original five results is unchanged by the rejection.
Which statements are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
- \( Q = (0.1034 - 0.1018)/(0.1034 - 0.1012) = 0.0016/0.0022 = 0.727 > 0.71 \): reject. Correct.
- Mean of 0.1012, 0.1014, 0.1016, 0.1018 = 0.1015 M. Correct.
- Deviations ±0.0001 and ±0.0003; sum of squares \( 2.0 \times 10^{-7} \); \( s = \sqrt{2.0 \times 10^{-7}/3} = 2.58 \times 10^{-4} \) M; RSD = \( 2.58 \times 10^{-4}/0.1015 = 0.25\% \). Correct.
- The median of five was 0.1016 M; the median of the remaining four is \( (0.1014 + 0.1016)/2 = 0.1015 \) M. It changed (slightly). Incorrect.
Practice Q7. A thermostat bath is logged ten times; the readings have a sample SD of 0.40 °C. Consider:
- Expressed in kelvin, the SD of the readings is 0.40 K.
- Expressed in °F (\( F = 1.8\,C + 32 \)), the SD is 0.72 °F.
- Expressed in °F, the variance is 0.288 °F².
- The RSD of the readings is the same whether they are expressed in °C or in K.
Which statements are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 1, 2 and 4 only (d) 2 and 3 only
Answer: (a) 1 and 2 only
- Adding 273.15 shifts every value; the SD is unchanged. Correct.
- Scaling by 1.8 multiplies the SD by 1.8: \( 0.40 \times 1.8 = 0.72 \) °F. The +32 has no effect. Correct.
- The variance scales by \( 1.8^2 = 3.24 \): \( 0.16 \times 3.24 = 0.518 \) °F². 0.288 is \( 0.16 \times 1.8 \), the wrong power. Incorrect.
- The SD is the same in °C and K but the mean is not (for example 25 °C vs 298.15 K), so the RSD differs. Incorrect.
Quick Revision Table
| Quantity | Formula or fact |
|---|---|
| Mean | \( \bar{x} = \sum x_i / n \) |
| Sum of deviations | always 0 |
| Sample SD | \( s = \sqrt{\sum (x_i - \bar{x})^2/(n - 1)} \) |
| Computational form | \( s^2 = [\sum x_i^2 - (\sum x_i)^2/n]/(n - 1) \) |
| Population SD | \( \sigma = \sqrt{\sum (x_i - \mu)^2/N} \) |
| RSD, CV | \( s/\bar{x} \); × 100 for %, × 1000 for ppt |
| \( y = ax + b \) | mean \( a\bar{x} + b \); SD \( \lvert a \rvert s \); variance \( a^2 s^2 \) |
| Standard error | \( s/\sqrt{n} \) |
| Gaussian bands | ±1σ 68.3%, ±2σ 95.4%, ±3σ 99.7%; ±1.96σ = 95% |
| Confidence interval | \( \bar{x} \pm t s/\sqrt{n} \), \( t \) for \( n - 1 \) d.f. |
| Q-test | \( Q = \) gap/range; reject if \( Q > Q_{\text{crit}} \) |
Key Takeaways
What’s Next
- Absolute and Relative Errors: accuracy vs precision, systematic vs random error, significant figures, and how errors propagate into a computed result.
- Linear Regression, Covariance and Correlation: fitting straight lines to calibration and kinetic data.
- Experimental Design, Sampling and Data: the Part A view of replicates, sampling and hypothesis tests.
- Data Analysis: Quick Revision.