The SSM Procedure (Experimental)

Example 27.6 Model with Multiple ARIMA Components

This example shows how you can fit the REGCOMPONENT models in Bell (2011) by using the SSM procedure. The following DATA step generates the data used in the last example of this article (Example 6: Modeling a time series with a sampling error component). The variable y in this data set contains monthly values of the VIP series (value of construction put in place), a U.S. Census Bureau publication that measures the value of construction installed or erected at construction sites during a given month. The values of y are known to be contaminated with heterogeneous sampling errors; the variable hwt in the data set is a proxy for this sampling error in the log scale. The variable hwt is treated as a weight variable for the noise component in the model.

data test;
  input y hwt;
  date = intnx('month', '01jan1997'd, _n_-1 );
  format date date.;
  logy = log(y);
  label logy = 'Log value of construction put in place';
  datalines;
115.2    0.042
110.4    0.042
111.5    0.067
127.9    0.122
150.0    0.129
149.5    0.135
139.5    0.152
144.6    0.168
176.0    0.173

   ... more lines ...   

The article proposes the following model for the log VIP series:

\[  \log (y) = \mu _{t} + hwt * \eta _{t}  \]

where $\mu _{t}$ follows an ARIMA(0,1,1)${\times }$(0,1,1)$_{\mi {12}}$ model and $\eta _{t}$ is a zero-mean, AR(2) error process. The following statements specify this model for logy:

 proc ssm data=test;
    id date interval=month;
    trend airlineTrend(arma(d=1 sd=1 q=1 sq=1 s=12));
    trend ar2Noise(arma(p=2)) cross=(hwt);
    model logy = airlineTrend ar2Noise;
    output outfor=for;
 run;

Output 27.6.1: Estimates of the ARIMA Components Model

The SSM Procedure

Model Parameter Estimates
Component Type Parameter Estimate Standard Error
airlineTrend ARMA Trend Error Variance 0.000768 0.00717
airlineTrend ARMA Trend MA_1 0.342128 0.38575
airlineTrend ARMA Trend SMA_1 -0.999000 9.30218
ar2Noise ARMA Trend Error Variance 0.410905 0.09947
ar2Noise ARMA Trend AR_1 0.484630 0.06758
ar2Noise ARMA Trend AR_2 0.439580 0.23570


The ARIMA(0,1,1)${\times }$(0,1,1)$_{\mi {12}}$ trend $\mu _{t}$ is named airlineTrend and the zero-mean, AR(2) error process $\eta _{t}$ is named ar2Noise. See the TREND statement for more information about the ARIMA notation. The estimates of model parameters are shown in Output 27.6.1. These estimates are different from the estimates given in the article; however, the estimated trend and noise series are qualitatively similar. The article uses an estimation procedure that consists of a sequence of alternating steps in which one subset of parameters is held fixed and the remaining parameters are estimated by MLE in each step. This process continues until all the estimates stabilize. The SSM procedure estimates all the parameters simultaneously by MLE.

The following statements produce the plot of the estimate of the airlineTrend component (shown in Output 27.6.2). This plot is very similar to the trend plot shown in the article (the article plots are in the antilog scale).

 proc sgplot data=for;
     title "Smoothed Estimate of the ARIMA(0,1,1)(0,1,1)12 Trend";
     series x= date y=smoothed_airlineTrend;
     scatter x= date y=logy;
 run;

Output 27.6.2: Estimate of the airlineTrend Component

Estimate of the airlineTrend Component


The following statements produce the plot of the estimate of the ar2Noise component (shown in Output 27.6.3). This plot is also very similar to the noise plot shown in the article (once again, the article plots are in the antilog scale).

 proc sgplot data=for;
     title "Smoothed Estimate of the AR(2) Noise";
     series x= date y=smoothed_ar2Noise;
     refline 0;
 run;

Output 27.6.3: Estimate of the ar2Noise Component

Estimate of the ar2Noise Component