QUESTION:We often have the problem to test, if two or more time series are statistically different (by mean or distribution), in a sense like t-test, U-test or ANOVA. The problem is however, that the data are from a time series, which may be autocorrelated. Thus the measurements of a series are not independent, but the known tests would require this independence. Is there a test available, which, e.g. takes care of the autocorrelation? Or is it possible to "eliminate" the autocorrelation by ARIMA modeling? We build an ARIMA model and try to test with the time series means and the residuals of the ARIMA model. Is this allowed? All hints and suggestions are highly appreciated! ANSWER:The problem you refer to is generally known as pooled cross-sectional time series analysis. It is possible to estimate one set of parameters under the null hypothesis and to compare the resultant error sums of squares with locally estimated error sums of squares. We were asked by a client to incorporate this feature into AUTOBOX. Please visit the web page http://www.autobox.com and download a copy. In specific, AUTOBOX allows you to test the hypothesis that ARIMA models from K groups are equal. Additionally, this test is extended to Transfer Functions. Consider the case where you have n distinct time series (max of 3) and you wish to test the hypothesis that the individual ARIMA models are equal to each other vs. the alternative that at least 1 model differs from the rest. This requires that 1 model be specified for all n and parameter estimation be done locally and compared to a global or generic set of coefficients. A STARTING MODEL MUST EXIST as this will be used. If AUTOMATIC MODELING IS DISABLED and this answer is greater than one (1) the program will : 1. disable all model modification options (sufficiency,necessity etc) 2. expect the time series to be a concatenated series of the n distinct time series and will estimate parameters without using the last set of group i to predict the start of i+1 , where i goes from 1 to 2 (max 3 groups). Hypothesis testing is done by summing the error sum of squares from the n local estimations (done separately) and divide by the total degrees of freedom to obtain a denominator mean square error. The numerator mean square error is the differential error sum of squares (composite estimation less the sum of the locals, divided by the number of groups)( See JOHNSTON : ECONOMETRIC METHODS 1963 Page 137) Following is an example of the approach.
|
|
TESTING THE EQUIVALENCE |
|
OF MODELS BETWEEN |
|
TIME SERIES |
|
POOLED CROSS-SECTIONAL |
|
TIME SERIES |
|
THERE ARE THREE (3) STATES IN A STUDY. WE HAVE BUILT |
|
SEPARATE MODELS FOR EACH OF THE THREE STATES AND ARE |
|
INTERESTED IN COMPARING THESE MODELS AND TESTING THE |
|
HYPOTHESIS THAT THE MODELS HAVE A SET OF COMMON COEFFICIENTS. |
|
FOLLOWING IS THE MODEL FOR NEW JERSEY. |
|
Estimation/Diagnostic Checking for Variable Y = NJ |
|
Number of Residuals (R) =n 256 |
|
Number of Degrees of Freedom =n-m 253 |
|
Residual Mean = R/n 720.571 |
|
Sum of Squares = R .277010E+11 |
|
Variance var= R /(n) .108207E+09 |
|
Adjusted Variance = R /(n-m) .109490E+09 |
|
Standard Deviation = 10463.8 |
|
Standard Error of the Mean = / (n-m) 657.851 |
|
Mean / its Standard Error = /[ / (n-m)] 1.09534 |
|
Mean Absolute Deviation = R /n 8218.96 |
|
AIC Value ( Uses var ) =nln +2m 4741.89 |
|
SBC Value ( Uses var ) =nln +m*lnn 4752.52 |
|
BIC Value ( Uses var ) =see Wei p153 3727.69 |
|
R Square =1-[ R / (A- A) ] .832228 |
|
THE ESTIMATED MODEL PARAMETERS |
|
MODEL COMPONENT LAG COEFFICIENT STANDARD T-RATIO |
|
# (BOP) ERROR |
|
Lambda Value 1.000000 |
|
Differencing 1 |
|
1 Autoregressive-Factor # 1 1 -.6051146 .611033E-01 -9.903 |
|
2 2 -.3265066 .689884E-01 -4.733 |
|
3 3 -.2133137 .612056E-01 -3.485 |
|
[(1-B**1)]Y(T) = + A(T)[(1+ .605B+.327B**2+ .213B**3)]**-1 |
|
FOLLOWING IS THE MODEL FOR NEW YORK. |
|
Estimation/Diagnostic Checking for Variable Y = NY |
|
Number of Residuals (R) =n 256 |
|
Number of Degrees of Freedom =n-m 253 |
|
Residual Mean = R/n 765.901 |
|
Sum of Squares = R .341596E+11 |
|
Variance var= R /(n) .133436E+09 |
|
Adjusted Variance = R /(n-m) .135018E+09 |
|
Standard Deviation = 11619.7 |
|
Standard Error of the Mean = / (n-m) 730.526 |
|
Mean / its Standard Error = /[ / (n-m)] 1.04842 |
|
Mean Absolute Deviation = R /n 8882.33 |
|
AIC Value ( Uses var ) =nln +2m 4795.54 |
|
SBC Value ( Uses var ) =nln +m*lnn 4806.17 |
|
BIC Value ( Uses var ) =see Wei p153 3862.24 |
|
R Square =1-[ R / (A- A) ] .871268 |
|
THE ESTIMATED MODEL PARAMETERS |
|
MODEL COMPONENT LAG COEFFICIENT STANDARD T-RATIO |
|
# (BOP) ERROR |
|
Lambda Value 1.000000 |
|
Differencing 1 |
|
1 Autoregressive-Factor # 1 1 -.5617603 .621071E-01 -9.045 |
|
2 2 -.3516715 .677407E-01 -5.191 |
|
3 3 -.1608633 .620052E-01 -2.594 |
|
[(1-B**1)]Y(T) = + A(T)[(1+ .562B+.352B**2+ .161B**3)]**-1 |
|
AND OUR THIRD STATE PENNSYLVANIA. |
|
Estimation/Diagnostic Checking for Variable Y = PA |
|
Number of Residuals (R) =n 256 |
|
Number of Degrees of Freedom =n-m 253 |
|
Residual Mean = R/n 1132.07 |
|
Sum of Squares = R .415012E+11 |
|
Variance var= R /(n) .162114E+09 |
|
Adjusted Variance = R /(n-m) .164036E+09 |
|
Standard Deviation = 12807.7 |
|
Standard Error of the Mean = / (n-m) 805.211 |
|
Mean / its Standard Error = /[ / (n-m)] 1.40593 |
|
Mean Absolute Deviation = R /n 9981.22 |
|
AIC Value ( Uses var ) =nln +2m 4845.38 |
|
SBC Value ( Uses var ) =nln +m*lnn 4856.01 |
|
BIC Value ( Uses var ) =see Wei p153 3786.36 |
|
R Square =1-[ R / (A- A) ] .806812 |
|
THE ESTIMATED MODEL PARAMETERS |
|
MODEL COMPONENT LAG COEFFICIENT STANDARD T-RATIO |
|
# (BOP) ERROR |
|
Lambda Value 1.000000 |
|
Differencing 1 |
|
1 Autoregressive-Factor # 1 1 -.7578220 .613060E-01 -12.36 |
|
2 2 -.3965126 .733749E-01 -5.404 |
|
3 3 -.1984544 .613150E-01 -3.237 |
|
[(1-B**1)]Y(T) = + A(T)[(1+ .758B+.397B**2+ .198B**3)]**-1 |
|
TO TEST THE HYPOTHESIS OF A COMMON SET OF COEFFICIENTS |
|
OR PARAMETERS WE NOW SHOW THE RESULT OF ESTIMATION |
|
WHERE THE DATA FROM ALL THREE STATES ARE POOLED AND |
|
USED COLLECTIVELY. |
|
Estimation/Diagnostic Checking UNDER THE NULL HYPOTHESIS |
|
Number of Residuals (R) =n 776 |
|
Number of Degrees of Freedom =n-m 773 |
|
Residual Mean = R/n 863.112 |
|
Sum of Squares = R 0.104376E+12 |
|
Variance var= R /(n) 0.134505E+09 |
|
Adjusted Variance = R /(n-m) 0.135027E+09 |
|
Standard Deviation = 11620.1 |
|
Standard Error of the Mean = / (n-m) 417.946 |
|
Mean / its Standard Error = /[ / (n-m)] 2.06513 |
|
Mean Absolute Deviation = R /n 8964.92 |
|
AIC Value ( Uses var ) =nln +2m 14530.5 |
|
SBC Value ( Uses var ) =nln +m*lnn 14544.4 |
|
BIC Value ( Uses var ) =see Wei p153 10695.1 |
|
R Square =1-[ R / (A- A) ] 0.847911 |
|
THE ESTIMATED MODEL PARAMETERS |
|
MODEL COMPONENT LAG COEFFICIENT STANDARD T-RATIO |
|
# (BOP) ERROR |
|
Lambda Value 1.000000 |
|
Differencing 1 |
|
1 Autoregressive-Factor # 1 1 -0.6563098 0.353132E-01 -18.59 |
|
2 2 -0.3574422 0.404579E-01 -8.835 |
|
3 3 -0.1929096 0.353165E-01 -5.462 |
|
[(1-B**1)]Y(T) = + A(T)[(1+ 0.656B+ 0.357B**2+ 0.192B**3)]**-1 |
|
MODEL SUMMARY : |
|
MODEL SUMMARY |
|
NJ [(1-B**1)]Y(T)=+ A(T)[(1+ .605B+ .327B**2+ .213B**3)]**-1 |
|
= R /(n-m) = .109490E+09 R = .277010E+11 |
|
NY [(1-B**1)]Y(T)=+ A(T)[(1+ .562B+ .352B**2+ .161B**3)]**-1 |
|
= R /(n-m) = .135018E+09 R = .341596E+11 |
|
PA [(1-B**1)]Y(T)=+ A(T)[(1+ .758B+ .397B**2+ .198B**3)]**-1 |
|
= R /(n-m) = .162114E+09 R = .415012E+11 |
|
GLOBAL [(1-B**1)]Y(T)=+ A(T)[(1+ 0.656B+0.357B**2+ 0.192B**3)]**-1 |
|
= R /(n-m) = .135027E+09 R = .104376E+12 |
|
We have 3 groups and k=3 coefficients are in the model and |
|
we have m observations in group 1 n observations in group 2 |
|
and o observations in group3. In our example m=n=o=260 . |
|
F = [ Q3/k ]/ [ Q2/(m+n+o-3k) ] |
|
1. Pool all observations (m+n+o) where m=260 and n=260 and o=260 |
|
and compute sum of squared residuals from model = .10437E+12 |
|
Q1= .10437E+12 |
|
2. Carry out model estimation locally and total the sums of squared |
|
residuals to obtain Q2 |
|
Q2= .277010E+11 + .341596E+11 + .415012E+11 |
|
Q2= .1034E+12 |
|
3. Compute Q3 = Q1 - Q2 and hence compute F as defined by |
|
Q3= .1044E+12 - .1034E+12 |
|
Q3= .1E+10 |
|
F= [ Q3/k ]/ [ Q2/(m+n+o-2k) ] |
|
F= [.1E+10/3]/[.104E+12/(260+260+260-2*3)] |
|
F= [.033E+10]/[.0134E+11] |
|
F= .3 |
|
with degrees of freedom (k m+n+o-3k) OR (3 774) the tabular |
|
F value is 3.35 at alpha = .01 |
|
4. If F the tabular F then reject the hypothesis of the
equality |
|
of the sets of coefficients otherwise conclude that there is not |
|
enough evidence to state that they are statistically significantly |
|
different from each other. |
|
Since .3 is less than 3.35 we conclude that the groups can not be |
|
statistically proven to be unequal with respect to their parameters |
Consider this ! Motivated by D.Davis@lboro.ac.uk One of the major problems with statistical tests that assume independence is the manner in which degrees of freedom or "information content" is computed. We now present a rather difficult analysis problem where the number of observations is 4 x 10 x 20 x 50 = 4000 but they are not independent random samples. Consider 4 subjects or groups .... for example 4 individuals. (1) Consider 10 conditions or separate experimental conditions. (2) Consider 20 trials or repeat samplings using exactly the same conditions (3) Consider 50 readings taken at fixed intervals over time. (4) (1) (2) (3) |Trial 1 |Trial 2.......Trial 20 (4) |subject|condition |time 1|time 2 ...time 50| | | | | | | A | 1 |......|...... | .. | .. |......|...... | A |10 |......|...... | B | 1 |......|...... | .. | .. |......|...... | B |10 |......|...... | C | 1 |......|...... | .. | .. |......|...... | C |10 |......|...... | D | 1 |......|...... | .. | .. |......|...... | D |10 |......|...... To consider using a 4 way ancova, testing between subjects and conditions, testing within trials; with the 10 conditions nested within the subjects would be downright silly. The obvious problem is that this is time series data and there will be a serious overestimation of the degrees of freedom due to the 50 repeat
measurements over time never mind that the 20 trials or repeat samplings of the individual at the same test conditions.