ACFRAN

To test for the presence of serial correlation in a regular time series using serial correlation coefficients and the Von Neumann ratio. Significance is determined using both normal approximations and random-ization.

Calling statement

acfran c1 ;

nran k1 ;

nlag k1 (10).

Input

c1 A column of numeric data.

Missing values : Allowed. Observations with missing values are simply ignored.

Subcommands

nlag - the maximum number of lags which should be considered. Must be less than the number of

observations.

Output

For each lag from 1 to nlag, we present

·  Observed autocorrelation at that lag

·  Standardised autocorrelation and approximate p-value using a normal approximation

·  Randomisation p-values for the autocorrelation coefficient

In addition, we present

·  Observed Von Neumann ratio for the data

·  Standardised VN ratio and approximate p-value using a normal approximation

·  Randomisation p-values for the VN ratio

Speed of macro : FAST

Notes

If the number of observations is less than 10, then the default lag automatically changes from 10 to one less than the number of available observations.

Reference : Manly, F.J. (1997) Randomization, bootstrap and Monte Carlo methods in biology,

Chapman and Hall, London.

Null hypothesis : We test the null hypothesis that the series is random.

Alternative hypothesis :

·  For the kth serial correlation coefficient : There is kth autocorrelation (short-term dependence) within the series.

·  For the Von Neumann ratio : The series is a simple random walk.

Test-statistics : We use the kth sample serial correlations (for k = 1,…,nlag) as test-statistics. If the kth sample serial correlation value is significantly different from zero then this provides evidence of auto-correlation at the kth lag.

In addition, we use the Von Neumann ratio as an overall test for the presence autocorrelation. This ratio tests the null hypothesis that the series is random against the alternative hypothesis that it is a simple random walk. The Von Neumann ratio is of the form

, where xi is the ith data value.

Randomization procedure : We randomize the order of the points in the observed series, since under the null hypothesis this will random.

Standard procedure

The function
% acf c1

computes serial correlation coefficients, and produces p-values using normal approximations: If rk is the kth sample serial correlation, then under the null hypothesis the standardized version

zk = [rk + 1/(n -1)] / sqrt{1/n}

has an approximate standard normal distribution for sufficiently large n. We also include p-values based upon these normal approximations within our macro output.

Within the macro output, we provide a p-value for the Von Neumann ratio based upon a normal approximation. Under the null hypothesis, for sufficiently large n, the Von Neumann ratio has a normal distribution with mean 2 and variance 4(n-2)/(n2-1).

WORKED EXAMPLE FOR ACFRAN

Name of dataset

PROLOCULI

Description

The data are mean diameters of megalospheric proloculi of the Cretaceous bolivinid foraminifer Afrobolivina afra from 92 levels in a borehole in Gbekebo, Ondo State, Nigeria. The rank of the depth is recorded, and provides a measure of the age of the sample (1 = oldest, corresponding to late Cretaceous, 92 = youngest, corresponding to early Palaeocene). Diameters are recorded, but interest really lies in the 91 differences between diameters from adjacent depths.

Our source

Manly, F.J. (1997) Randomization, bootstrap and Monte Carlo methods in biology,

Chapman and Hall, London.

Original source

Reyment, R.A. (1982), Phenotypic evolution in a Cretaceous foraminifer, Evolution, 36, pp. 1182-1199.

Data

Number of observations = 92

Number of variables = 3

For each observation, sample number (which corresponds to rank depth) (left), diameter (middle) and difference from previous diameter (left) are given.

Sample Diam Diff Sample Diam Diff

1 156 * 47 201 -2

2 146 -10 48 261 60

3 136 -10 49 262 1

4 152 16 50 271 9

5 147 -5 51 202 -69

6 190 43 52 235 33

7 169 -21 53 214 -21

8 170 1 54 212 -2

9 179 9 55 210 -2

10 176 -3 56 241 31

11 184 8 57 211 -30

12 162 -22 58 247 36

13 155 -7 59 238 -9

14 154 -1 60 235 -3

15 151 -3 61 227 -8

16 150 -1 62 236 9

17 187 37 63 230 -6

18 220 33 64 241 11

19 205 -15 65 232 -9

20 194 -11 66 230 -2

21 221 27 67 238 8

22 185 -36 68 234 -4

23 171 -14 69 230 -4

24 177 6 70 254 24

25 194 17 71 256 2

26 176 -18 72 210 -46

27 170 -6 73 230 20

28 178 8 74 231 1

29 177 -1 75 225 -6

30 168 -9 76 227 2

31 176 8 77 226 -1

32 209 33 78 237 11

33 172 -37 79 250 13

34 195 23 80 226 -24

35 169 -26 81 229 3

36 173 4 82 240 11

37 156 -17 83 205 -35

38 161 5 84 221 16

39 161 0 85 208 -13

40 147 -14 86 207 -1

41 158 11 87 215 8

42 162 4 88 233 18

43 233 71 89 210 -23

44 184 -49 90 213 3

45 205 21 91 198 -15

46 203 -2 92 213 15


Plot

Worksheet

C1 Rank depth

C2 Diameter

C3 Distance from previous diameter

Aims of analysis

To investigate whether or not stage-to-stage differences in diameters suffer from autocorrelation.

Standard procedure

Welcome to Minitab, press F1 for help.

MTB > Retrieve "N:\resampling\Examples\Proloculi.MTW".

Retrieving worksheet from file: N:\resampling\Examples\Proloculi.MTW

# Worksheet was saved on 07/08/01 11:51:17

Results for: Proloculi.MTW

MTB > ACF 8 c3 c4.

Randomization procedure

MTB > % N:\resampling\library\acfran c3 ;

SUBC> nran 999 ;

SUBC> nlag 8 ;

SUBC> autocor m1 ;

SUBC> vonneu c5.

Executing from file: N:\resampling\library\acfran.MAC

Randomization tests for autocorrelation in a univariate time series

SERIAL CORRELATION COEFFICIENTS

for lags of 1 to k

* KEY *

> row = lag, j

> corr = observed autocorrelation at the j^th lag

> z-value = standardized autocorrelation

> normal p = p-value using normal approximation

> 1s - ran p = one-sided randomization p-value, H1: negative correlation

> 1s + ran p = one-sided randomization p-value, H1: positive correlation

> 2s ran p = two-sided randomization p-value

Data Display

Row corr z-value normal p 1s - ran p 1s + ran p 2s ran p

1 -0.421503 -3.91489 0.000090 0.001 1.000 0.002

2 0.100958 1.06907 0.285038 0.871 0.130 0.260

3 -0.102290 -0.86979 0.384417 0.182 0.819 0.364

4 0.060894 0.68689 0.492152 0.760 0.241 0.482

5 -0.135631 -1.18784 0.234895 0.123 0.878 0.246

6 0.004072 0.14484 0.884836 0.564 0.437 0.874

7 0.106034 1.11750 0.263782 0.883 0.118 0.236

8 -0.155653 -1.37885 0.167942 0.068 0.933 0.136

* NOTE * The interpretion of serial correlation p-values:

Significance levels should be reduced appropriately to

account for the effects of multiple testing. The simplest

procedure is to divide the p-value by the number of lags

being considered (this is conservative). For example, if

a 95% significance level is required and lags up to 10

are of interest, then individual p-values smaller than

0.05/10 = 0.005 are taken to be significant.

VON-NEUMANN RATIO

Data Display (WRITE)

Observed von-neumann ratio 2.835

Standardized von-neumann ratio 4.029

P-value using normal approximation 0.0001

One-sided randomization p-values 1.0000

0.0010

Two-sided randomization p-value 0.0020

* NOTE * The use of the Von-Neumann ratio :

This ratio provides a test of randomness within a time

series. It tests the null hypothesis that the observed

series is random against the alternative hypothesis that

it is a simple random walk (in which the value

at a particular point in time is partly determined by

the value at the previous time point).

Modified worksheet

C5 Column containing 999 Von Neumann ratios, one for each randomized dataset

M1 A 999*8 matrix. The kth column contains 999 serial correlation coefficients at lag k,

one for each randomzied dataset.

Discussion

Randomization and standard methods present a very similar picture. There is strong evidence of autocorrelation at lag 1 (p-value = 0.000 by standard methods, 0.002 by randomization), but no evidence of autocorrelation at any other lag. The Von-Neumann ratio also provides clear evidence against this being a random series (p-value = 0.000 by standard methods, 0.002 by randomization).

6