1.15 Moments

Four questions describe the shape of any distribution, and each is answered by a moment:

1st.
Where is it? – the mean.
2nd.
How spread out is it? – the variance.
3rd.
Is it lopsided, and which way? – the skewness.
4th.
How heavy are the tails? – the kurtosis.

That is the whole idea. A moment is what you get by raising each deviation to a power and averaging; the power decides which question is answered. Higher powers give more weight to observations far from the centre, which is why the third and fourth moments say something about the tails while the first two do not.

1.15.1 Raw moments (about the origin)

Definition 1.44. The \(r\)th raw moment, taken about zero, is \[\mu '_r=\frac {\sum X^r}{n} \qquad \text {or, for a frequency distribution,}\qquad \mu '_r=\frac {\sum fX^r}{\sum f}.\]

The first raw moment is the mean: \[\mu '_1=\frac {\sum X}{n}=\overline {X}.\] (The zeroth is \(\mu '_0=1\), since \(X^0=1\) for every observation. It is worth stating once and then forgetting.)

1.15.2 Central moments (about the mean)

Raw moments answer the first question but not the others, because they are measured from zero, which is an arbitrary place. To describe shape we measure from the centre.

Definition 1.45. The \(r\)th central moment is \[\mu _r=\frac {\sum (X-\overline {X})^r}{n} \qquad \text {or}\qquad \mu _r=\frac {\sum f(X-\overline {X})^r}{\sum f}.\]

Take them in turn:

  • \(\mu _1=\frac {\sum (X-\overline {X})}{n}=0\) always – the deviations from the mean cancel by definition. This is not a useless fact: it is the reason the second moment has to square them.
  • \(\mu _2=\sigma ^2\), the variance.
  • \(\mu _3\) measures lopsidedness. Cubing preserves sign, so observations far to the right contribute large positive amounts and those far to the left large negative ones. If the two sides balance, \(\mu _3=0\).
  • \(\mu _4\) measures tail weight. The fourth power is large only for observations far from the mean, so \(\mu _4\) is dominated by the extremes.

1.15.3 Converting between them

Central moments are usually computed from raw ones, because \(\sum X^r\) is easier to accumulate than \(\sum (X-\overline {X})^r\): \[\mu _2=\mu '_2-\left (\mu '_1\right )^2,\] \[\mu _3=\mu '_3-3\mu '_2\mu '_1+2\left (\mu '_1\right )^3,\] \[\mu _4=\mu '_4-4\mu '_3\mu '_1+6\mu '_2\left (\mu '_1\right )^2-3\left (\mu '_1\right )^4.\] The first of these is the familiar \(\sigma ^2=\overline {X^2}-\overline {X}^2\). Note that the subtracted term is the raw first moment \(\mu '_1=\overline {X}\); the central first moment \(\mu _1\) is zero and would subtract nothing.

Where those three come from. They are not separate results to be memorised. Expand \((X-\overline {X})^r\) by the binomial theorem, remembering that \(\overline {X}=\mu '_1\) is a constant, and average term by term: \[\mu _r=\sum _{j=0}^{r}(-1)^j\binom {r}{j}\,\mu '_{r-j}\left (\mu '_1\right )^{j}.\] The alternating sign comes from the minus in \((X-\overline {X})\), and \(\binom {r}{j}\) is the ordinary binomial coefficient.

Taking \(r=0,1,2\) in turn, and using \(\mu '_0=1\): \begin {align*} \mu _0 &=\binom {0}{0}\mu '_0 &&=1\\[2pt] \mu _1 &=\binom {1}{0}\mu '_1-\binom {1}{1}\mu '_0\,\mu '_1 &&=\mu '_1-\mu '_1=0\\[2pt] \mu _2 &=\binom {2}{0}\mu '_2-\binom {2}{1}\mu '_1\,\mu '_1+\binom {2}{2}\mu '_0\left (\mu '_1\right )^2 &&=\mu '_2-2\left (\mu '_1\right )^2+\left (\mu '_1\right )^2\\ & &&=\mu '_2-\left (\mu '_1\right )^2 \end {align*}

and \(r=3\) and \(r=4\) give the third and fourth formulas above in the same way.

The first two are worth reading rather than passing over. \(\mu _0=1\) always, because averaging \((X-\overline {X})^0=1\) over \(n\) observations gives \(1\); it carries no information. And \(\mu _1=0\) always — the two terms cancel exactly. Neither depends on the data at all, which is why the useful moments start at \(\mu _2\).

1.15.4 Making them comparable

\(\mu _3\) has the cube of the original units and \(\mu _4\) the fourth power, so neither can be compared between data sets measured in different units – or even judged large or small on its own. Dividing by the appropriate power of \(\sigma \) removes the units entirely: \[\text {Skewness}=\frac {\mu _3}{\sigma ^3},\qquad \text {Kurtosis}=\frac {\mu _4}{\sigma ^4}.\] These are pure numbers, and can be read directly:

  • Skewness \(>0\): the right tail is longer – the distribution leans left with a tail running right. Skewness \(<0\): the reverse. Skewness \(=0\): symmetric.
  • Kurtosis \(=3\) for a normal distribution. Above 3 means heavier tails than normal (more extreme values than one would expect); below 3 means lighter. Because 3 is the reference, the excess kurtosis, \(\frac {\mu _4}{\sigma ^4}-3\), is often quoted instead, so that the normal sits at zero.

Example 1.46. Calculate \(\mu _1, \mu _2, \mu _3, \mu _4\) for the data given below
Direct Method

\(X\) \(X-\overline {X}\) \((X-\overline {X})^2\) \((X-\overline {X})^3\) \((X-\overline {X}^4\)
10 \(-20\) 400 \(-8 000\) 160, 000
20 \(-10\) 100 \(-1 000\) 10 000
30 0 0 0 0
40 10 100 1 000 10 000
50 20 400 8 000 160 000
150 0 1 000 0 340 000

Solution. \[\overline {X}=\frac {1}{n}\sum X=\frac {1}{5}(150)\hspace {0.5cm}\implies \hspace {0.5cm}\overline {X}=30\]
\[\mu _1=\frac {\sum (X-\overline {X})}{n}=\frac {0}{5}=0\]
\[\mu _2=\frac {\sum (X-\overline {X})^2}{n}=\frac {1000}{5}=200\]
\[\mu _3=\frac {\sum (X-\overline {X})^3}{n}=\frac {0}{5}=0\]
\[\mu _4=\frac {\sum (X-\overline {X})^4}{n}=\frac {340,000}{5}=68, 000\]

Indirect Method

\(X\) \(D_X=X-a\) \(D_X^2\) \(D^3_X\) \(D_X^4\)
10 \(-20\) 400 \(-8 000\) 160, 000
20 \(-10\) 100 \(-1 000\) 10, 000
30 0 00 0 0
40 10 100 1 000 10, 000
50 20 400 8 000 160, 000
0 1 000 0 340,000

\[\mu '_1=\frac {\sum D_X}{n},\hspace {0.8cm} \mu '_2=\frac {\sum D^2_X}{n},\hspace {0.8cm}\mu '_3=\frac {\sum D^3_X}{n},\hspace {0.8cm}\mu '_4=\frac {\sum D^4_X}{n}\]
\[\mu _1=0\hspace {0.5cm}\text {always}\]
\[\mu _2=\mu '_2-\left (\mu '_1\right )^2=\frac {1000}{5}-0^2=200\]
\[\mu _3=\mu '_3-3\mu '_2\mu '_1+2\mu ^{\prime 3}_1=0-3(200)(0)+2(0)^3=0\]
\[\mu _4=\mu '_4-4\mu '_3\mu '_1+6\mu '_2\mu ^{\prime 2}_1-3\mu ^{\prime 4}_1=68,000-4(0)(0)+6(200)(0)^2-3(0)^4=68,000\]

Note 1.47. The two methods gave identical numbers at every step, which makes the indirect method look pointless. That is only because \(a=30\) was chosen here, and \(30\) happens to be the mean, so every \(\mu '_r\) was already a central moment and each correction term collapsed to zero. Choose \(a\) anywhere else — as in the next example, where \(a=7\) but the mean is \(6.9\) — and the correction terms do real work.

1.15.5 Moment Ratios

\[\beta _1=\frac {\mu ^2_3}{\mu ^3_2}\hspace {1cm}\text {and}\hspace {1cm} \beta _2=\frac {\mu _4}{\mu ^2_2}\]

1.15.6 Skewness

Skewness is defined as lack of symmetry of the values about some central value i.e. mean, mode, medians

NMMoeoradmnea,lMeddisitarinbution N ∽ (0,1)

Figure 12: A symmetric distribution: mean, median and mode coincide.

pMMooesidatenive Skewness

Figure 13: Positive skew. The long tail runs to the right and the mean is pulled above the mode.

nMMeoegdaaentive Skewness

Figure 14: Negative skew. The long tail runs to the left and the mean is pulled below the mode.

Coefficient of skewness measures the degree of lack of symmetry \[SK=\frac {\text {Mean}\hspace {0.2cm}-\hspace {0.2cm}\text {Mode}}{\sigma }\]
\[SK=\frac {3(\text {Mean}\hspace {0.2cm}-\hspace {0.2cm} \text {Median})}{\sigma }\]
\[SK=\frac {Q_1+Q_3-2\text {Median}}{Q_3-Q_1}\]
\[SK=\frac {\sqrt {\beta _1}(\beta _2+3)}{2(5\beta _2-6\beta _1-a)}\]

Interpretation

1.
If \(SK=0\), then the distribution symmetric.
2.
If \(SK>0\), then the distribution is positively skewed.
3.
If \(SK<0\), then the distribution is negatively skewed.

KURTOSIS is a measure of the degree of peakedness.

βββ222 > = < 333 N(0,1)

Figure 15: Kurtosis. \(\beta _2=3\) is the normal curve; above 3 the tails are heavier, below 3 lighter.

Coefficient of Kurtosis is calculated as: \[CK=\frac {\text {Semi Interquartile Range}}{P_{90}-P_{10}}\] The distribution is normal if \(CK=0.263\), otherwise the distribution is not normal.

Example 1.48. Given the data below

C-I \(0-2\) \(2-4\) \(4-6\) \(6-8\) \(8-10\) \(10-12\) \(12-14\)
\(f\) 1 3 5 2 6 2 1
\(X\) 1 3 5 7 9 11 13
Table 11: Frequency distribution for the moments example. \(X\) is the class mid-point.

Using \(a=7\) as an arbitrary value.

1.
Find the first four central moments.
2.
Calculate \(\beta _1\) and \(\beta _2\).
3.
Calculate the coefficient of skewness.
4.
Calculate the coefficient of kurtosis, and comment.

Solution. The classes all have width \(h=2\), so instead of working with \(X\) directly we work with the coded value \[d=\frac {X-a}{h}=\frac {X-7}{2}\] This is not a trick to save arithmetic only. Shifting by \(a\) changes no central moment at all, and dividing by \(h\) scales the \(r\)th one by a known factor, so we can undo both at the end.

C-I \(f\) \(X\) \(d\) \(fd\) \(fd^2\) \(fd^3\) \(fd^4\)
\(0-2\) 1 1 \(-3\) \(-3\) 9 \(-27\) 81
\(2-4\) 3 3 \(-2\) \(-6\) 12 \(-24\) 48
\(4-6\) 5 5 \(-1\) \(-5\) 5 \(-5\) 5
\(6-8\) 2 7 0 0 0 0 0
\(8-10\) 6 9 1 6 6 6 6
\(10-12\) 2 11 2 4 8 16 32
\(12-14\) 1 13 3 3 9 27 81
Total 20 \(-1\) 49 \(-7\) 253
Table 12: Working table for the coded moments. Each column is the one before it multiplied by \(d\).

(1) The first four central moments.
The moments about \(a\), in coded units, are the column totals divided by \(n=20\): \begin {align*} m'_1 &=\frac {\sum fd}{n}=\frac {-1}{20}=-0.05 & m'_2 &=\frac {\sum fd^2}{n}=\frac {49}{20}=2.45\\ m'_3 &=\frac {\sum fd^3}{n}=\frac {-7}{20}=-0.35 & m'_4 &=\frac {\sum fd^4}{n}=\frac {253}{20}=12.65 \end {align*}

The mean follows at once, since \(\overline {X}=a+h\,m'_1\): \[\overline {X}=7+2(-0.05)=6.9\]

Now convert to moments about the mean with the formulas of the previous section: \begin {align*} \mu _2 &=m'_2-(m'_1)^2 = 2.45-(-0.05)^2=2.4475\\ \mu _3 &=m'_3-3m'_1m'_2+2(m'_1)^3 = -0.35-3(-0.05)(2.45)+2(-0.05)^3=0.01725\\ \mu _4 &=m'_4-4m'_1m'_3+6(m'_1)^2m'_2-3(m'_1)^4\\ &=12.65-4(-0.05)(-0.35)+6(-0.05)^2(2.45)-3(-0.05)^4=12.6167 \end {align*}

These are still in coded units. Multiplying the \(r\)th moment by \(h^r=2^r\) puts them back into the original units of \(X\): \begin {align*} \mu _2 &=2.4475\times 2^2 =9.79 & \mu _3 &=0.01725\times 2^3 =0.138\\ \mu _4 &=12.6167\times 2^4 =201.867 & \mu _1 &=0 \quad \text {always} \end {align*}

So the variance is \(9.79\) and the standard deviation \(\sqrt {9.79}=3.129\).

(2) \(\beta _1\) and \(\beta _2\). \begin {align*} \beta _1 &=\frac {\mu _3^{\,2}}{\mu _2^{\,3}}=\frac {(0.138)^2}{(9.79)^3}=\frac {0.019044}{938.31}=0.0000203\\ \beta _2 &=\frac {\mu _4}{\mu _2^{\,2}}=\frac {201.867}{(9.79)^2}=\frac {201.867}{95.844}=2.106 \end {align*}

Both are pure numbers with no units, which is the whole point of dividing by a power of \(\mu _2\). Had we forgotten to convert out of coded units, we would have got the same two answers, because \(\beta _1\) and \(\beta _2\) do not depend on \(h\) at all.

(3) Coefficient of skewness. \[\gamma _1=\sqrt {\beta _1}=\sqrt {0.0000203}=0.0045\] taking the sign of \(\mu _3\), which is positive.

(4) Coefficient of kurtosis. \[\gamma _2=\beta _2-3=2.106-3=-0.894\]

Comment.
\(\gamma _1=0.0045\) is as close to zero as makes no difference, so the distribution is symmetric for practical purposes. There is a slight lean to the right, but a value that small should not be read as evidence of anything.

\(\gamma _2=-0.894\) is negative, so the distribution is platykurtic: flatter in the middle and lighter in the tails than a normal curve. Look back at the frequencies to see why. They run \(1,3,5,2,6,2,1\) — there is no single tall peak, and the values are spread fairly evenly across the range instead of crowding round the centre. That is exactly the shape \(\beta _2<3\) describes.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.