Testing for normal distribution in STATA

Use a histogram and a normal distribution curve

hist X, normal

check whether the curve follows a normal distribution

Using the Shapiro-Wilk Test

swilk x

When your data is not normally distributed use non-parametric test. Non parametric tests disregards
the normality assumption.

Non-parametric test

When testing equality of between two means

Mann-Witney U Test/Wilcoxon rank-sum test

ranksum x, by()

Time series analysis

Time series data are data collected on the same observational unit at multiple time periods.

Times series data is used for the following functions

1. Forecasting models eg forecast the inflation levels, interest rate levels etc using the following
models
a. Autoregressive (AR) models
b. Autoregressive distributed lag (ADL) models
2. To estimate dynamic causal effects. The concern if to know the effect of a policy shift over time.

Time series models raises new technical issues

1. Time lags
2. Correlation over time (serial correlation or autocorrelation)
3.

Step 1

Issue the command

tsset year telling Stata that year is your time scale variable

Converting a string to numeric or a numeric to string in Stata(encode; decode)

String to numeric (encode)

Means creating a numeric variable new_X

Numeric to string (decode)

Create a new string variable for new_X

decode X, gen(new_X)

Plotting a two way graph

twoway line X Y [time variable]

If your time variable is year

twoway line X Y Year

you can also do it for 3 variables

twoway line X Y Z Year

Correlation coefficient

If the data is normally distributed use the Pearson correlation coefficient

correlate var1 var2 var3

Pairwise correlation and use stars to show the correlations that are significant at less than 1% level

pwcorr, star(0.01)

For nonnormally distributed data you use Spearman Correlation.

Stationarity in time series

The OLS requires that your time series be stationary. It is therefore, important to determine stationarity
in time series data. You need to deal with stationary time series processes. A stationary process is
characterized by time-invariant mean, variance and autocovariance.

Some time series data tend to have a stochastic trend commonly known as a unit root. A stochastic
trend or deterministic trend also known as a unit root. Regressing non-stationary series will result is
spurious regression

Commands for testing

Testing for Unit root using the Augmented -Fuller test, you are testing whether the process is either a
random walk or a random walk with a drift. There are variant forms of the ADF depending on whether
you allow for a non-zero constant/or a deterministic trend.

Augmented Dickey-Fuller test for the presence of a unit root in X with a trend term

dfuller X, trend

Augmented Dickey-Fuller test for the presence of a unit root in X with a drift term
dfuller X, drift

Augmented Dickey-Fuller test for the presence of a unit root in X with a drift term and two lags

dfuller X, drift lags(2) regress

You can also use the Phillip Perron test by issuing the command

pperron X, trend

Differenced models

Generating differences for variable X. You can generate the differenced series by issuing the command
d2. The prefix “d2” means second differencing, if you want third differencing you issue “d3” etc

gen dX = d2.X

Autoregressive models

Autoregressive model using lags for forecasting. AR is another form of a time series model. It uses a
variable’s own history to predict its future. In an auto regressive model no other variable is needed.
Regressors in autoregressive consists of lagged dependent variables only. Written as AR(p), means the
pth order autoregressive model. Autoregression is based on the premise that time series are serially
correlated i.e. current level of most series are correlated with previous or earlier levels.

First order Autoregressive AR(1)

reg Y L.Y

Second order Autoregressive AR(2)

reg Y L.Y L2.Y

Third order Autoregressive AR(3)

reg Y L.Y L2.Y L3.Y

Distributed Lagged models

The models uses the lag value of X to account for the lag effect. You might want to see if previous values
have any impact on current values.

y=β 0 + β 1 X t + β 2 X t−1 + β 3 X t−2 +ε t

How to generate the lagged values of X. You can issue the commands

gen xlag1= x[_n-1] to generate the first lag values of X

gen xlag2= x[_n-2] to generate the second lagged values of X

Alternatively you can use the lag operator L to perform the task after declaring tsset command

L.X for first lag of X

L2.X for second lag of X

reg Y X L.X L2.X

reg Y L(1/2). X L(1/2).X instruct STATA to use lags 1-2 of the variable X as the
regresors. L1 up to L2 refers to lag 1, lag 2.

F test for joint significance of X and its lags

test X L.X L2.X

Autoregressive Distributed Lag Models (ARDL)

This is the major achievement in dynamic single equation models.

You need to first install a package for the ARDL model by issuing the command. The package ardl will
automatically determine the optimum lag length.

ssc install ardl

to regress issue the command

ardl y x z y is the dependent variable, x and z are explanatory variables

Autoregressive Distributed Lag Modelling approach to Cointegration analysis

ARDL can be used in long-run relationship analysis when the variables are integrated I(.). The model can
produce consistent estimates of the long-run coefficients. The ARDL helps to study the long-run
relationships among time series variables.

In the presence of cointegration the traditional ARDL model is in applicable, hence the need to correct
for long-run relationship. You use an error-correction style ARDL model in the presence of cointegration

ardl y x z, ec

the postestimation command when ec is used;

estat ectest

You can also issue the command

ardl y x z, ec

the postestimation command when ec is used is

estat ectest

Cointegration and error correction model

It is possible for time series to be integrated (move together) in a non-stationary way. Such series will
follow a common stochastic trend. These time series are said to be cointegrated. If variables X and Y are
cointegrated, we cannot rely on OLS estimates. We can apply the differenced equation to remove the
history of cointegration or co-movement.

 If we difference the series we will eliminate the history of how Y will be pulled into a long-run
relationship with X
 If we estimate the variables in levels, our test statistics are unreliable because the variables are
non-stationary.

The appropriate model for a cointegrated case is the error-correction model (ECM) of Hendry and
Sargan

Stata has a suite of commands for fitting, forecasting, interpreting and perfroming inference on vector
error-correction models (VECMs) with cointegrating variables. After fiitng a VECM, the irf commands can
be used to obtain impulse-response functions (IRFs) and forecast-error variance decompositions
(FEVDs).

To select the optimum lag length use the command varsoc

Varsoc x y

To do the vector error correction model use the command vec

vec xy

Testing for Cointegration

You can use the Engle and Granger’s extension of the ADF test

If X and Y are integrated that is I(1) or I(2) etc, that is integrated of order 1 or integrated of order 2, then
a two variable error correction model is necessary. With two variables there can only be one
cointegrating relationship linking the long run paths. With m variables there can be m-1 cointegrating
relationships

Run the OLS model

regress y x

Get the residuals

predict e, resid

Run a unit root test on the residuals

dfuller e, lags(10)

Vector Autoregressive models (VAR)

These models have been developed to characterize the joint time-series of a set (vector) of variables
without making the restrictive assumptions that would allow the identification of structural dynamic
models. VAR is a simple generalization of predicting a variable based on its own lagged variable. It also
involves predicting a vector of variables based on lags of all variables.

y=α 0 + α 1 y t−1 + β 0 X 1+ β 2 X t −1 + ε ty
 x
x=θ 0+θ 1 x t −1 +δ 0 y 1+ δ 2 y t−1 +ε t

The is called a vector autoregressive model because y is auto regressive on y t −1 , in the same way x is
autoregressed on its lagged values. The variable y t −1 and x t−1 are the lagged values of y and x
respectively

On VAR models we can do Granger causality test. The null hypothesis is that the lagged values of x does
not help to predict y given the presents of lagged values of y and vice versa.

It is also possible to do the impulse response function and the variance decompositions. This will help to
identify the structural shocks affecting our model.