Hello again, Khondokar,

I hope it's OK me contacting you again.  I really want to help you if at all possible. 

I notice that you have been advised to test your data for normality and then determine whether you should use nonparametric or parametric analyses. In relation to that, you might or might not be aware, but the kind of analysis that I have been recommending, namely, multiple regression, is very much a parametric statistic. In fact, I don't know of any nonparametric statistic that could be used for your kind of question. 

There are a number of assumptions that need to be met for multiple regression - and there is some debate about whether normality of data is one of them.  I have attached a document that I put together some time ago in an attempt to summarise those assumptions.  Some assumptions, such as no outliers, could be very important.    If there appear to be problems regarding the assumptions, you could try to address them in some way in your data (for example by Winsorising outliers) so that you can go ahead with multiple regression.

I mention this because, given my understanding of what you want to analyse, I think that your most appropriate method is multiple regression.  I could still be wrong about that, however, because I'm not 100% sure about what you're trying to do.

Feel free to get back with any questions. I'm only too happy to try to keep helping you.

There are some quite good YouTube sites on which different types of multiple regression are described - usually in SPSS.

Robt.

You can employ none parametric ANOVA (via minitab)

Your dependent variable is categorical so, you can use

- Fisher test with categorical independent variables (table 4X4).

- ANOVA test for continuous for normally distributed data

- Non-parametric K-independent test for data which are NOT normally distributed (you can also use this test if you are not sure about your data distribution).

In the case of Dependent variable, I have several questions for each capacity level and I used 4-points scale. Now, I can compute the response and can make a new variable. And then it is not categorical in nature. In that case what should be the analysis?

Since you said you want measure causation, it is important first to make sure that random allocation was used.  Was it? You mention that you have several other variables which suggests that you may be trying matching or something like that, or looking to see if the Treatment effect is moderated by one or more of these. Is this true? Answers to these will allow people to suggest tests (I assume the responders above made some assumptions).

Hello Khondokar,

I think that Daniel (above) has identified a major problem facing those of us who would like to help you, namely that you seem not to have provided enough information and therefore people need to make assumptions.  From what you wrote above, for example, I am not clear whether you have more than one independent variable and I am also not clear about how you obtained your dependent variable(s).  Could you be more specific please?  For example, how many independent variables do you have, and how many levels are there in each of them?  Also, did you add up some of your original dependent variables to obtain a smaller number of dependent variables?

It is only when those kinds of things are really clear that others can help you with appropriate advice.

Robt.

Thank you Robert,

Yes, I have 9 independent variables which are measured through nominal, ordinal and interval scale. I have one dependent variable but dependent variable has four dimensions. Therefore, It can be said that I have four dependent variables. several questions were asked under each dependent variable and four points scale were used to measure it. Now, I can sum-up my dependent variable by calculating the total score for each respondent.

Hope this might clarify your all questions. 

Hello Khondokar,

Thank you for the extra clarification. There a still several things that I am unclear about, but it looks as if you want to see whether a number of independent variables are related to four dependent variables, or perhaps a single dependent variable.

First, I think you need to decide whether you have four or one dependent variable(s).  That will be determined by what your initial four dependent variables are and whether or not they can be combined.  You can't just add them up to a total score if they are conceptually or psychometrically different from each other or you would be adding up things like apples, oranges, Volkswagens, and amount of money in people's wallets.  Get it?  It is quite possible / easy to add things up, but the total must be meaningful.  So that's your first problem:  deciding whether or not you can add up the four "sub" dependent variables.

Second, and I am not sure about this, but with a number of independent variables that you want to analyse simultaneously as having some prospective influence on your dependent variable(s), I wonder whether you should be considering multiple regression in which you use some of your independent variables (the nominal ones) as dummy variables and the other variables as continuous.  Multiple regression isn't difficult for people who know what they are doing, but it certainly does require some statistical sophistication and care.

If you used multiple regression, you should be able to enter all of your independent variables into an analysis at the one time on one side of an equation, and your dependent variable(s) on the other side of the equation.  If you have only one dependent variable, you could do the analysis in one hit.  If you had four dependent variables, you would need to run four multiple regressions, one for each dependent variable (with all of the independent variables included each time).

There are different forms of multiple regression:  simultaneous, hierarchical, and stepwise.  Which one you might use would depend on the nature of your original research question(s), and you probably need a good statistician to help you with that - someone who takes responsibility for understanding the detailed nature of your investigation and how your research question(s) should be addressed.

I hope that helps.  And, if anyone would like to contribute to this string by offering different, even contradictory, advice, it would not trouble me.  I would welcome it.  Of course, it would be reassuring if others agreed with me.

Part of my problem is that I still don't have a clear idea about what you did.  When you refer to a treatment group, presumably your treatment was a dichotomous (nominal) variable.  I am also assuming that your other independent variables are ones on which you were not able to exert control across your treatment and control groups. Is that correct?  If you could indicate what your treatment was and what (at least some of) your independent variables were, and what your four initial dependent variables are, it should be possible to help you more.

Robt.

Hello again, Khondokar,

I have been doing more thinking about your research and what I wrote above.  First, Daniel's points up above have made me think that we need to know more about your research design.  Were you essentially conducting an experiment or quasi-experiment?  By your saying that you had treatment and control groups, one of those seems to be the case, but I would like to be sure.  Also, did you take measurements on your dependent variables for both groups before administering the treatment to the treatment group as well as after the treatment had been administered to the one group?  Or did you simply take measurements on your dependent variables after the treatment was administered?

If you took measurements on the dependent variables before and after, are you working with the differences between scores at Time 1 and Time 2, or just the scores taken at Time 2?

Daniel, above, raises an important point concerning research design, namely whether or not you used random allocation when determining the composition of your treatment and control groups.  At a guess, I suspect that you did not, and that has several consequences.  First, your design would seem to be a quasi-experiment, not a "true" experiment, or RCT. That's a bit unfortunate, but the stats, if appropriately chosen, should be able to overcome the lack of randomisation.  Second, I think that what you are referring to as independent variables might, apart from what your treatment is, be regarded as confounding variables - i.e., variables that might be having an effect on the dependent variable(s) in addition to any effect that the treatment would have. Third, if this is the case (i.e., eight of your independent variables are really to be regarded as confounders), it will have implications for the kind of regression analysis that you do.

Feel free to get back to me about the above, but I think you need to have access to a really good statistician if you want to do your analyses properly.  There are things like Akaike's information criterion (AIC) that should probably be employed, and there are some initial checks on collinearity among your variables that should be conducted.

Finally, as Paulo pointed out above, the extent of effort you put into this might depend on the level of research you are conducting.  If it is undergraduate, you could do something quite simple (with lots of problems in it!), but if it is PhD or higher I think you really should aim to do the analyses properly.  Even for honours or masters, I think you should try to do the appropriate stats, even if you get a lot of help from outside. I just hope it's available if you need it.

Robt.

First, perform the normality test, and decide which statistics; non-parametric or parametric can be applied

Hello again, Khondokar,

I hope it's OK me contacting you again.  I really want to help you if at all possible. 

I notice that you have been advised to test your data for normality and then determine whether you should use nonparametric or parametric analyses. In relation to that, you might or might not be aware, but the kind of analysis that I have been recommending, namely, multiple regression, is very much a parametric statistic. In fact, I don't know of any nonparametric statistic that could be used for your kind of question. 

There are a number of assumptions that need to be met for multiple regression - and there is some debate about whether normality of data is one of them.  I have attached a document that I put together some time ago in an attempt to summarise those assumptions.  Some assumptions, such as no outliers, could be very important.    If there appear to be problems regarding the assumptions, you could try to address them in some way in your data (for example by Winsorising outliers) so that you can go ahead with multiple regression.

I mention this because, given my understanding of what you want to analyse, I think that your most appropriate method is multiple regression.  I could still be wrong about that, however, because I'm not 100% sure about what you're trying to do.

Feel free to get back with any questions. I'm only too happy to try to keep helping you.

There are some quite good YouTube sites on which different types of multiple regression are described - usually in SPSS.

Robt.

Dear Robert,

Thank you very much for your kind information and help.

I think I have to run the regression analysis that you mentioned earlier and other things I will keep in my mind before conducting the final analysis.

Thank you again for your kind cooperation and time. 

Dear Khondokar Humayun Kabir 

For the T-tests, Regression, and other necessary tests, please check these resources

Best Regards

www.statstutor.ac.uk/resources/uploaded/coventrypairedttests.pdf

www.cal.org/twi/EvalToolkit/appendix/toolkit13_sec9.pdf

https://students.shu.ac.uk/lits/it/documents/pdf/analysing_data_using_spss.pdf

www.stat-help.com/ANOVA%202006-01-11.pdf

https://www.shu.edu/offices/technology/upload/Data-analysis-process.pdf

Dear Khondokar,

It's my pleasure to have helped you.  Please feel free to get back with any questions.

Robt.