If we perform linear regression, and the result is that there is no linear relationship between the independent variable and the dependent variable, we should suspect whether the relationship is not linear? Threshold Panel Regression is a development of panel regression with the assumption that the relationship between the dependent and independent variables is not linear. For example, a non-linear relationship is the relationship between inflation and economic growth. If we look for a linear relationship between the two variables, we will not get the value (either positive or negative). However, the relationship is positive/negative/unrelated to a certain value (threshold) and negative after a certain value (threshold).

Threshold panel regression is a novelty in determining the relationship between the independent variable and the dependent variable. Panel threshold regression analysis is a development of linear panel regression analysis, which essentially divides the estimation unit into two or more regimes. To estimate the model, there are two ways that can be done, the first through ordinary least squares (OLS) if the threshold value is known and the second with conditional least squares if the threshold value is not known. The principle of conditional least squares is to find the threshold value and the slope parameter value together. Hansen (1997) recommends that the selected model is a model with a minimum residual sum of squares (RSS) value. Enders (2004) suggests that around 15% of the minimum value and 15% of the maximum value for the threshold variable are not included as candidates for the threshold value to ensure the adequacy of degrees of freedom in each regime.

Testing the stationarity assumption on panel data was proposed by Baltagi (2005). Unit root test on panel data can use Levin, Lin and Chu (LLC) Test, Im, Pesaran and Shin Test (IPS), Breitung's Test, p-value Combination Test, and Residual Based LM Test.

The hypothesis used in testing panel data is the same as for time series data. The null hypothesis used is the data has a unit root.