I am conducting a study to find the impact of digital banking on performance using ATM transactions, Mobile money transactions, POS transactions, Real Time Gross Settlement (RTGS) etc. But the correlation between the variables are too high. I tried excluding one or two variables but the situation persisted.
My question is can I ignore and continue with the analysis? Or is there a way to go about it? Your valuable inputs will be very much appreciated.
The answer depends on the exact problem you are facing:
1 is your model predictive or explanatory? see: https://www.google.com/search?rlz=1C1CHBF_enUS874US874&ei=oWwfXqibEtu2tAbL_ZuoCA&q=multicollinearity+in+predictive+models&oq=multicollinearity+in+predictive+models&gs_l=psy-ab.3..0.406737.433453..434284...1.2..3.230.5125.53j5j1......0....1..gws-wiz.....6..0i71j0i22i30j0i22i10i30j0i362i308i154i357j0i273j0i131j0i67j0i131i67j0i10.y930LeJOZYI&ved=0ahUKEwjo78D6r4bnAhVbG80KHcv-BoUQ4dUDCAs&uact=5
and/or
https://www.google.com/search?rlz=1C1CHBF_enUS874US874&ei=VG4fXozRJZXPtQaAsrPoDA&q=multicollinearity+in+explanatory+models&oq=multicollinearity+in+explanatory+models&gs_l=psy-ab.3..33i299l3.104334.118548..120850...0.2..0.135.2057.21j2......0....1..gws-wiz.......0i71j0i7i30j0i8i7i30.61uCAvDBNWQ&ved=0ahUKEwiMy4rKsYbnAhWVZ80KHQDZDM0Q4dUDCAs&uact=5
At this point you should know the problem that you are facing and have an idea of the solution.
2. If your model is predictive, adaptive lasso variable selection normally does not allow collinear predictors and is optimal for the candidate predictors.(oracle property). See the information here: https://www4.stat.ncsu.edu/~boos/var.select/lasso.adaptive.html
Best wishes, David Booth
Kwadwo Boateng
Multicolleniearity is a sample problem. Some remedies are recomended for this viaolation.
You can change yoor sample (I think it does not work on your research)
You can use principal component analysis or You can use ridge regression.
Best wishes
Cengiz
In accordance with the comments above multicolleniearity is a sample problem, that usually is solved by:
1. Elimination of explanatory variables from the model. 2. Actions on the sample. Introducing new observations or mix cross-sectional and temporal data. These alternatives, however, are often not feasible. 3. Establishment of restrictions on the behavior of population parameters. Using all available extra-sample information, establishing restrictions on the behavior of the model parameters. This would reduce the number of parameters to estimate. 4. Variable transformation. When, above all, time series are handled, multicollinearity can be reduced by transforming the variables with the application of first differences, logarithms ...; However, these modifications, apart from requiring a justification, may have undesirable effects: reduction of degrees of freedom, breach of some basic hypothesis of the model (such as introduction of heteroskedasticity in random disturbance), etc. 5. Multivariate techniques. Frequently, multivariate techniques are used to deal with the problem of multicollinearity, such as factor analysis and that of principal components.
Thank you very much @ Julian Llorent Jurado.
I have opted for Partial Least Square Regression which has Principal Component Analysis factored in the analysis. My problem is can I write down a regression equation from the coefficient output?