LIWC-22 Output with Zero Overdispersion

[Eli Sine]

Dear all,

I have analyzed news articles with the text analysis tool LIWC-22. LIWC analyzes the frequency of words in various categories such as emotional expressions and cognitive processes and the values are given in percentage of the total word count (i.e. a value of 7.5 in the category "moral" means that 7.5% of all words in the text are associated with moral discussions). So the values for any variable can be between 0 to 100. My dataset does not have a panel structure.
My research aim is to find out how categories like AI impact categories like moral or negative emotions, etc.

My problem is that for some variables like moral I have to many zero values, so that normal transformation methods like log or sqrt don't work for me. See an overview below:

summarize moral

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
       moral |     22,339    .2066476     .402975          0       7.09


. tabulate moral if moral <=0.1

      Moral |
 Legitimacy |      Freq.     Percent        Cum.
------------+-----------------------------------
          0 |     11,854       93.61       93.61
        .02 |          8        0.06       93.67
        .03 |         28        0.22       93.90
        .04 |         49        0.39       94.28
        .05 |         67        0.53       94.81
        .06 |         90        0.71       95.52
        .07 |        143        1.13       96.65
        .08 |        185        1.46       98.11
        .09 |        239        1.89      100.00
------------+-----------------------------------
      Total |     12,663      100.00

**Questions:**
1. Which transformation methods would be best suited for my dependent variable "moral" to perform an OLS regression?
2. Are there alternative modeling techniques that you would recommend for such a distribution?

Thank you very much for your help!

[George Ford]

poisson estimation?

[Eli Sine]

Hello George,

I was thinking about zero inflated poisson or zero inflated negative binomial models, but they are not suitable for non-integer values. Or am I wrong here?

[George Ford]

Not a problem. Some type of Poisson/robust would do.

[Eli Sine]

Ok, just for understanding, you mean a zero inflated poisson?

[George Ford]

Read this:

https://statisticalhorizons.com/zero-inflated-models/