Monday, 12 February 2018

Time Series Analysis (Lecture 2): Choosing Optimal Lags in EViews

General Overview on Lag Selection

Since this blog is tailored for beginners in econometrics, I will not be engaging an advanced discussion on the topic but an introductory approach by which a beginner can understand the essence of using lags in a model and the pitfalls that may occur if lags are excessively used. Interested readers who require more advanced information on lag selection can consult appropriate econometric textbooks. Having said that, in economics the dependence of a variable Y (outcome variable or regressand) on another variable(s) X (the predictor variable or regressor) is rarely instantaneous. Very often, Y responds to X with a lapse of time. Such a lapse of time is called a lag. Therefore, in time series analysis, some level of care must be exercised when including lags in a model.

So how many lags should be used in a model? There is no hard-and-fast-rule on the choice of lag length. It is basically an empirical issue. As noted in Damodar Gujarati Basic Econometrics, there is no a priori guide as to what the maximum length of the lag should be. The researcher must bear in mind that, as one estimates successive lags, there are fewer degrees of freedom left, making statistical inference somewhat unstable. Economists are usually not that lucky to have a long series of data so that they can go on estimating numerous lags. More importantly, in economic time series data, successive values (lags) tend to be highly correlated increasing the likelihood of multicollinearity in the model.

Also, from Jeffery Wooldridge’s Introductory Econometrics: A Modern Approach with annual data, the number of lags is typically small, 1 or 2 lags in order not to lose degrees of freedom. With quarterly data, 1 to 8 lags is appropriate, and for monthly data, 6, 12 or 24 lags can be used given sufficient data points. Again, in the words of Damodar Gujarati Basic Econometrics “the sequential search for the lag length opens the researcher to the charge of data mining”. He further stated that the nominal and true level of significance to test statistical hypotheses becomes an important issue in such sequential searches”. For instance, if the lag length, k, is incorrectly specified, the researcher will have to contend with the problem of misspecification errors. In addition, because of the lags involved, distributed and or autoregressive models raise the topic of causality in economic variables.

Hence, before you estimate a time series equation, it is necessary to decide on the maximum lag length. Like I mentioned earlier, this is purely an empirical question. Suppose there are 40 observations in all, by including too many lagged values, your model consumes degrees of freedom, not to mention introducing the likelihood of multicollinearity occurring. As noted in my previous tutorial on multicollinearity, it leads to imprecise estimation; that is, the standard errors tend to be inflated in relation to the estimated coefficients. As a result, based on the routinely computed t ratios, we may tend to declare (erroneously), that a lagged coefficient(s) is statistically insignificant. In the same vein, including too few lags will lead to specification errors. The easiest way out of this quagmire, is to decide using a criterion like the Akaike or Schwarz and choose that model that gives the lowest values of these criteria. Most econometric packages easily compute these optimal lag length but note some trial and error is inevitable.

Choosing Optimal Lags in EViews

For instance, if there are limited observations in a vector autoregressive (VAR) estimation, it is often advised to use the Akaike Selection Criterion (AIC) in selecting the lag length that "prefers" the more parsimonious models. However, the information criterion with the smallest criterion value evidences the most ideal lag length to employ. Most researchers prefer using the Akaike information criterion (AIC) but my valuable advice is always to select that criterion with the smallest value, because that ensures the model will be stable. Let us begin by showing how you can select the optimal lag order for your model and variables using the EViews analytical package. 

Please note that in EViews, the procedure is simply to run an initial VAR on the variables at level with the default settings and obtain the results. I will go through the steps in detail.

For this tutorial, I will extract data from Gujarati and Porter Table 21.1 dataset. It is a quarterly data on United States from 1970 to 1991, which is 88 observations. The variables are gdp (gross domestic product), pdi (personal disposable income) and pce (personal consumption expenditure).

Step 1: Load Data into EViews
To import the Excel file into EViews, go to: File >> Import >> Import from file >> Next >> Finish. If it is correctly done, you obtain:
EViews Workfile from
EViews Workfile
Source: CrunchEconometrix
From the EViews interface, the three variables gdp, pce and pdi are individually shown. Double-clicking on each variable shows them in separate sheets, like is:

EViews Creating Group Data from
EViews Creating Group Data
Source: CrunchEconometrix

Step 2: Create Group Data
But because I need to obtain the optimal lag for the model, it becomes necessary to open this data as a GROUP by putting all three variables in a worksheet. To do that: Press down the Cntrl key >> click on gdp, pce and pdi >> Right click on any part of the screen >> Open >> as Group:

EViews - Open as Group Data from
EViews - Open as Group Data
Source: CrunchEconometrix

When you click "as Group", you should have this:

EViews Group Data from cruncheconometrix
EViews Group Data
Source: CrunchEconometrix

Step 3: Run Unrestricted VAR model
Now that our variables are grouped, next is to run an unrestricted VAR model with the level of the variables and taking different lags before deciding which model is the best. Remember, I am using quarterly data which allows me to use up to 8 lags. But if yours is a yearly data you can use 2 lags at the most in order not to lose too many degrees of freedom or if monthly data, up to 24 lags. The unrestricted VAR is chosen only on the assumption that the three variables are not co-integrated.

Note: if the variables are cointegrated, you should run the vector error correction model

To run the unrestricted VAR model, go to: Quick >> Estimate VAR >> Dialog box opens:

EViews VAR Specification from
EViews VAR Specification
Source: CrunchEconometrix

Type in all the variables names in the Endogenous variables box (note under VAR, there is no exogenous variable, all variables are endogenous). Since between 1 to 8 lags can be used because I am using a quarterly data, I begin with 4 lags before deciding which model is the best.

Click OK….here is the output (to save space only relevant part shown):

EViews Regression Output from
EViews Regression Output
Source: CrunchEconometrix
The EViews output reports among others, the AIC and Schwarz criterion. You will also observe that the output returned 2 sets of results, those identified by red bracket are for the respective endogenous variables with each column representing the result for gdp, pce and pdi in that order. But the results we are most interested in are those identified by the blue bracket. These are the estimates for the VAR system. However, at this moment, we are only interested in the criterion. Hence, between the AIC and Schwartz, the former’s criterion of 26.85144 is lower than that of Schwartz at 27.98004. Therefore, we conclude based on this output that the lag selection must be based on the AIC.

Step 4: Choose Optimal Lag length for the Model
However, we cannot be running the unrestricted VAR model using different lag lengths before deciding on the best model to adopt, there is a simplified way of obtaining the optimal lag structure at once given a variety of information criteria. To do that, click on View >> Lag Structure >> Lag Length Criteria >> the Lag Specification dialog box opens:

EViews LagSpecification Dialogue Box from
EViews Lag Specification Dialogue Box
Note: I put in 8 lags because I am at liberty to use up to 8 lags due to the nature of my data (quarterly). So, if yours is a yearly data, you may put in 2.

Click OK to obtain the various information criterion from lag 0 to 8 shown below:

EViews Lag Structure for the Model from
EViews Model Lag Structure
Source: CrunchEconometrix
From the output, the selected lag order is indicated by an asterisk sign (*) which is distributed between lags 1 and 2, but mostly on lag order 2. The rule-of-thumb is to select the criterion with the lowest value which again is the AIC at 26.90693 this is because the lower the value, the better the model. We can conclude that the optimal lag length for the model is 2 and the best criterion to adopt for the model is AIC.

The same procedure can be adopted in obtaining the respective lags for each variable. For instance to obtain for gdp:
1.   Double click on gdp >> Quick >> Run the unrestricted VAR >> OK >> Obtain the output
2.   Click View >> Lag Structure >> Lag Length Criteria >> Lag Specification dialog box opens >> OK

…and you obtain this:

EViews Lag Structure for gdp from
EViews - Lag Structure for gdp
Source: CrunchEconometrix

From the output, the best criterion that fits the gdp model is the AIC with the lowest figure of 9.937278 meaning that the optimal lag length for gdp is 2.

Doing the same procedure for pce, here is the result:
EViews Lag Structure for pce from
EViews - Lag Structure for pce
Source: CrunchEconometrix
From the output, the optimal lag length for pce model is 4 given the AIC value at 8.698617 which the lowest among the criterion, hence it is the best criterion for the pce model. For pdi, the optimal lag length is 1 given the AIC value at 9.602079 shown below:
EViews Lag Structure for pdi from
EViews - Lag Structure for pdi
Source: CrunchEconometrix
Caveat: There are also cases where the used lag length is that which is most selected by the criterion named after the econometricians who developed them, like HQ, SIC, AIC and LR, etc. Some researchers prefer Schwartz criterion when the variables are more than 4 and use the AIC when the variables are less than 4. As, mentioned in the introductory part of this tutorial, the decision on the choice of lag is purely an empirical issue. Generally, we choose the lag length for which the values of most of these lag length criteria are minimised, indicated by asterisks in the EViews output.

[Watch video tutorial on optimal lag selection using EViews]

Having gone through this tutorial, it will be easy to understand and know how to determine the optimal lags for a model regardless of the analytical package used. Remember that the “Lag length criteria” indicates a definite way of selecting the optimal lags after estimating the initial VAR model. Also VAR and ARDL models are susceptible to arbitrary use of lags as this may erode the degrees of freedom, weaken the significance of the coefficients, may induce auto-correlation and weaken the strength of diagnostic tests.

Try the outlined steps on your models and if there are further and comments, do post them below…..

No comments:

Post a Comment