One can include a variable even if it is insignificant due to economic intuition. If you believe that this variable is important to your model you should include it.

Allah you can change a proxy variables or interchange the lag or even circumspect the data

Hi Nayon, A good way to decide whether you should include/ exclude a variable to look at the Adjusted R square value of your model. You should run your model with and without that variable. If your adjusted R square is greater without the variable, you'll know that this variable might be irrelevant.

The variable may not be excluded from the model if the probability value is greater than (0.05) if the variable is important with a convincing economic justification.

The p-value is a statistical t-test of means of the means of the variables. It may not be the only test. We have other ways to cross-check - the underlying theory, F-test and the R2. It is possible that multi-collinearity and serial correlation might distort the significance of the p-value through the standard error.

You need to set your model based on the theory And you can’t omit the insignificant variables from model. Moreover, you should employ diagnostic tests to check your results.

Dear Nayon,

I think, 1st you have to screening the data and select best technique (OLS, Joh-Cointegration, VECM,VAR, ARDL/NARDL etc) based upon stationary among data set or in variables. If all variables have I(0) order then you can go for OLS which will lead you proper significance and P value.

Based upon Integral order (Stationarity) I(1) for Johnson Co-Inegration, I(0) and I(1) combination for ARDL to NARDL. You may not need to omit the variable(s). Results will be satisfactory.

If you believe that the variable is important to your model, you should keep it even if it’s not statistically significant. Otherwise, you’ll face an endogeniety problem.

My rule of thumb: If I want to do a forecast, I will only include the significant vars, but if I want to analyze data I can include all vars based on theory.

Obviously, you have no theoretical base for your estimation. But without a carefully specified model, one should not go into a statistical (econometric) estimation. The first problem is, that foreign direct investment - although it is called investment - is not a real investment (it does not increase productive capital) and, I think, that the law of big numbers does not really apply for this variable. There may be numerous small projects, but often their monetary sum may be dominated be one or two purchases with high amounts.

Have you ever thought over, which influences you chosen independent variables should have - and why. For example, one could find reasons, that the exchange rate has positive and negative effects on FDIs. Using both, the real exchange rate and inflation, as explaining variables means that there is a relation (correltion) between two independent variables (i.e. both coefficints are likely biased). Another problem may be the combination of GDP and its growth. These data are principally different in dimension. Therefore, their combination needs a very careful theoretical and mathematical analysis before estimating.

If you have good reasons (arguments), that and why a variable shoud have an (important and clear) effect on FDIs, you should not exclude it, even if it is insignificant. But you should not take the estimated coefficient as true (that is also valid to significant ones, too). If an estimate is far from one would expect from theory, one should not keep it, but take the theoretically "right" value (this should be a guess before estimation), if it is within the significance interval.

Dear Debasmita,

I would want you to consider the following:

1.What does your theory State.

2.Run the analysis with and without the variable. Then consider the adjusted R square. You can then disregard variable where adjusted R square is greater.

3. If the variable is justified economically in u our model, then leave it for your interpretation.

Regards

Dear Kehinde,

May I add:

1. Theory should not only help to choose the relevant variables, but also show which relations between the explaining and dependent variables should be taken for the specification. May be one can derive a range for an acceptable coefficient, too.

2. I would rather recommend one-by-one correlations (or simple OLS) before going into multivariate estimations.

Thank you Prof @Anton Rainer.

1. Theories should explain the relationship between the dependent and independent variable for the model specification.

2. Employ a simple OLS before delvi G into multivariate estimations.

Regards

We exclude the variables in the model if they Are not significant meaning their p value is greater than the rate of accepting errors specified by the researcher. generally we accept errors up to 5% but even if 10% is Not bad depending on the importance of the variable on the target variable. But again. I would like to remind one thing that the estimated model becomes valid after being validated on either just economic plan where we interpret signals carried by the variable on the estimated models if are the expected ones we validate the model otherwise we don’t. On the statistical plan is where we validate model on regarding the individual significance of each variable where we apply t student test or by judging the model as a whole using fisher test as well as the R square which determines the explicative power of independent variables on the dependent one. So my opinion is if the variables are insignificant on both plans is better if they excluded in the model but if on one plan the model is validated there we shouldn’t exclude them even if individually they are insignificant Debasmita Dey

Thanks everyone for all the suggestions. I tried several ways to stick to my model then i added more variables but still there are problems with significance although having a good R2 and adjusted R2. There are high multicollinearity too. Even if i add more observations for example, 40 years data, still the problem stays. So, i decided to change the country and apply the same variables. Lets see what happens now!

You should accept that your model is wrong (because statistcs say that it should be rejected). A wrong specification should (and, in general, can) never be saved by using special methods or by taking other data. You should first think over what a FDI is and which motives may be important. Maybe you could present your specification (with the dimensions and transformations of the variables) that we can tell you, what problems we see.

Take GDP only. Exclude GDP growth. Or take GDP growth and exclude GDP.

agree with Debesh Bhowmik , lag values can "eat up" some of the variances. you can also try to include savings. Debasmita Dey are you have some results of your decision to decided to change the country

One should never exclude a variable from a model. From statistical theory, a " fail to reject" the null hypothesis simply indicates " may be the true" consequently, the insignificant variable should be in the model. Yes, lags can be played with but the contemporaneous term of the variable should be retained.

Chuck, in principle, you are right, but it is too poor a result, if the coefficients of all independent variables are not significant at a sufficient probability level. It means that one cannot even (statistically) say, whether their effect is positive or negative (or no effect). For a carefully specified model, one should even be more ambitious and not only test, if a coefficient is different from zero, but if your expected range are within a hopefully rather narrow significance interval. If e.g. one estimates a propensity of consumption and gets a coefficient of, say, 1.03 with a significance interval of 0,15, I would not hesitate to take 0.9 as the "true" coefficient. It is bad habit among econometricians taking unrounded estitmates (with 5-6 or more digits) without carefully specifying an equation and thinking about what size (range) a coefficient should realistically be before estimation.