Creating Indices with Multiple Tables, Different Observation Sets

[Marcus Chiu]

Dear Statalisters,

I am trying to create indices to express general perceptions towards specific subjects using various opinion polls (merged together as dataset). I have been trying to use factor analysis but, as various questions come from different observation sets, I receive a "no observations" error when I try to factor all the questions together. Is there a way to create such indices to then later regress with other independent variables? Help regarding this matter would be greatly appreciated.


Best,
Marcus Chiu

[Dave Airey]

Unless I'm mistaken, it sounds like you have merged datasets with different measures, and the result is a larger dataset of incomplete cases. I don't think analysis on such a dataset is going to help any substantive questions, and you will have to decide whether questions address the same thing before you merge, before factor analysis.

[Marcus Chiu]

Dear Dave,

I am using various opinion polls that ask similar questions but carried out by, for example, CNN, CBS, Gallup, Pew, with slightly different ranges but each recoded with the same valences (higher the number, the more positive the perception). The questions, while with slight changes, are also repeated for other years. The variables (questions) therefore address the same topic, but they have different observation sets (by poll and by year). Would factor analysis (or any other method to create an index of perceptions towards a topic based on these variables) still not work in this case?



Best,
Marcus

[Dave Airey]

As long as you are happy combining data from two variables in two datasets into one variable that you claim measures the same thing, I'd say you could use factor analysis on several such variables if you have enough complete cases.

[Joe Canner]

Marcus,

Leaving aside for the moment the issue of whether what you are doing is appropriate, and getting back to your original problem, you should probably give a brief summary of how your data is structured and the commands you ran that resulted in the "no observations" error. It's not clear what you mean by "different observation sets", but I suspect that this is the root of your problem. More details will reveal whether there is a solution to this problem or whether what you want to do is impossible.

Regards,
Joe

[Marcus Chiu]

Thank you both for your responses.

To be more specific, I merged several public opinion polls from CNN, CBS, Pew, and others, for the years 2008 to 2012. I recoded all the independent variables appropriately and am not facing any problems with that. I made sure to note all the relevant questions from each poll; questions that asked for general opinions towards China, perceptions of economic threat towards China, and perceptions of military threat towards China.

When I try to perform factor analysis with each variable of one of the three categories (i.e. general opinions towards China for 2008 polls), I receive a "no observations" error from Stata. I do not receive such error when I perform factor analysis with only questions from one poll for one year. I presume this is because the observations (interviewees) from Poll X is different from the observations from Poll Y (and Z and so forth), and hence I cannot factor variables across the polls. I have come across a paper online (http://www.utdallas.edu/~herve/abdi-...S-mfa-2013.pdf) that describes applying multiple factor analysis for multiple datasets with and without the same observations, but have not seen methods to use in Stata.

In short, I am trying to figure out a method to create indices to illustrate, basic perceptions, economic perceptions, and security perceptions of Americans towards China, using multiple polls for various years. As many questions are quite similar (across polls and across years), I would be able to justify combining some variables, but I hope to figure out a better method that allows me to create indices across multiple tables, so that I can then regress the indices with my independent variables. I hope this clarifies my situation. Any feedback on this issue, or on general methodology would be greatly appreciated.


Best,
Marcus