THE CLASSIC ASSUMPTION TEST (AUTOCORRELATION, HETEROSCEDASTICITIY, MULTICOLINERITY AND NORMALITY) FOR PANEL DATA (WITH SPSS, EVIEWS AND STATA)

Before I speak further, first you must promise to read this article in the DATE and in order that you will not get lost! I am pretty sure you are now in a very difficult position, difficult to find tutorials both on blog and on youtube about classical assumption test tutorial for panel data you have. This is for one basic reason: "CLASSIC ASSUMPTION TEST" LOGIC THAT YOU NOT FULLY UNDERSTAND. This may be due to several reasons: your lecturer never teach you or your lecturer teach you well but instead you open Instagram in the class’ corner, Haha.

Okay, before we talk about the classical assumption test, first i have to make sure that the data you have is the data panel. As we know, panel data is a combination of cross section data and time series data. As an example of my data as follows:

The data above is an example of panel data, with cross section units in the form of bank’s stock codes (AGRO, BABP, BACA, etc.) and with time series units in the form of years (2010-2014). The CAR, NPL, LDR, TDR and EAR are independent variables while the dependent variable is ROA. For those of you who come from the economics department, it is certainly not foreign to these variables, yes they are the variables of the banking performance measurement. But, to make it easier, I replaced the independent variables with X1, X2, X3, X4 and X5 while I will replace the dependent variable with Y.

If you want to test the classic assumptions and test panel data regression with both SPSS and Eviews, you "MUST" follow the data format as above. My advice is that you type the data in Microsoft Excel, so you can easily copy it to SPSS or Eviews. Okay, after you make sure that the data you have is a data panel, then the next step is to test the classical assumption. Yes, before we conduct a regression test, free from the violation of the classic assumption test is a "mandatory". Just focus here first!

1. Autocorrelation Test
What is an autocorrelation test? What is the purpose of autocorrelation test? Autocorrelation test is intended to see whether observations in year t are affected by the previous year (t-1). For example by using my data, then the analogy is the 2012 data gaining influence from 2011? This is what the autocorrelation test tries to answer. If there is any influence, then it is said that there is an autocorrelation problem. Then, we go back to the panel data. Try to observe my sample data carefully! The 2014 data from the AGRO bank is directly adjacent to 2010 data from the BABP bank. Can we compare data from two different companies? Absolutely NOT. It looks like you think I'm making it up. All you need to know is "THERE IS NOTHING A CLASSIC ASSUMPTION TEST FOR PANEL DATA". To test the classical assumption, then the position of your data must be clear, whether the time series model or cross section model!!! Then what about my data example? Try to look closer to my data! More similiarly to what, time series model or cross section model? Yes, "Panel data is the same as cross section data". The answer is simple, time series data will not repeat periods! View my data! 2010, 2011, 2012, 2013, 2014, then ... back to 2010 ... Then what about the autocorrelation test? I ask you once again, what is the purpose of the autocorrelation test? "To find out whether the data in the previous period affected data in the current period", then can we compare AGRO data in 2014 with BABP data in 2010? Or in other words, can we say that the data released by the BABP in 2010 was influenced by the data released by AGRO in 2014 ??? It's making no sense, right?. Therefore, Autocorrelation Test is only intended for time series data! That it makes sense if I compare the 2010 AGRO’s data to 2011 AGRO’s data and so on. Right? So, the answer to all of your struggle regarding the autocorrelation test on the panel data is ... NO AUTOCORELATION TEST ON THE PANEL DATA !!! Even if there are theses or journals that perform autocorrelation tests, the test results "have no meaning at all" or just to add to their paper page thickness ... Haha

2. Heteroscedasticity test
Back to the same question as the autocorrelation test ... What is a heteroscedasticity test? What's the purpose? This test is done to see whether the variable variance in the regression model is same or not. If it is same, it is called homoskedasticity, if not, it called heteroscedasticity. Still confused??? Just ask Google then haha…, honestly I was also still confused with it’s definition and purpose. My goal is to give you an understanding of the relationship of classic assumption tests with panel data. In essence, a heteroscedasticity test is “mandatory” to be used, whether you use time series data, cross section and panel data though! Have you tried and failed? Keep calm, there are many heteroscedasticity test methods, if one fails you can find another methods until it succeeds. Some of the methods include: Graph analysis, Glejser method, Park method, White method, Rank Spearman method and Goldfeld-Quandt method. And you need to remember, when you want to do this test in the Eviews, your data form may not be "Pooled" but with a regular model. How? Just googling it, there are a lot of tutorial on Youtube.

3. Multicollinearity test

Immediately, a multicollinearity test is conducted to test whether there is a near perfect linear correlation between more than two independent variables. So essentially, the comparison is the variable, not the data! So, the multicollinearity test is the same as the heteroscedasticity test and it is also "MANDATORY" to do! This test is done by looking at the VIF (variance inflation factor) value and the Tolerance (TOL) value of each variable. All you need to know, is that the value of VIF = 1 / TOL. So if you get TOL = 2, then how much its VIF value? Yes, 0.5 aka ½. The standard of your variable so it is said to be free of multicollinearity problems if the VIF value is less than 10 and / or its TOL value is greater than 0.10 (0.10 is not derived from alpha values, such as 0.1, 0.05, or 0.01) . How? Please googling, there are hundreds of its tutorials.

4. Test for Normality

Well here is one of the classic assumption tests that you must understand its logic same as the autocorrelation test. Normality test and autocorrelation test are “partner in crime” against PANEL DATA. Before, you have to know exactly what is the purpose of the normality test? Normality tests are carried out in order to find out whether your data is normally distributed or not. If there are extreme values ​​in your data for example like -876 or +876 while your average data ranges from +/- 20, then your data is not normally distributed. Then, what about the company's financial ratios? We can’t blame a company for issuing data that is far different from the average value of other companies' data right?. We return to the purpose of normality test, the goal is to find out "DOES THE DATA DO NOT PLAN TOO MUCH FROM THE AVERAGE VALUE". So, if the data you used comes from secondary data, with a panel model and the samples of more than one company, then you are "recommended" not to do the normality test. The reason is the same as the autocorrelation test, the results will have no meaning at all !!! Therefore, normality test is usually use because you use primary data, go directly to the field, use questionnaires and so on. Then the question is, do all secondary data don’t have to test for its normality ??? Don't make quick conclusions first. You are still required to do normality test if you observe only one company, even though the data is secondary.

Okay, as conclusion, for those of you who have panel data and want to test classical assumptions, then what is required for you is a heteroscedasticity test and multicollinearity test. For autocorrelation and normality tests it should not be done, because the results do not give any meaning at all. And again, there is no classical assumption test specifically for panel data, because the classical assumption test is only for data whose position is clear whether the time series or cross section, while the panel data itself is more similar to cross section data. As for the regression test, please use panel data regression provided by the Eviews application. How? Please googling it haha..

Thank You

Akhmad Azhari

Economic & Business Faculty

Hasanuddin University

Indonesia

Bahasa Version written on February 26^th , 2016

Translated on English on July 25^th, 2018