Dichotomizing when variable is skewed I have a meeting frequency variable and I want to determine whether it has an effect on hypothesized relationship, by 1) controlling for it and 2)examining whether it is a moderator.  Unfortunately, it wasn't a very sensitive measure and distribution is skewed. The breakdown is: 10% daily, 5% twice a week, 52% weekly, 27% fortnightly and 6% monthly. 
Someone suggested that because of the distribution (52% in a single category), I wouldn't get anything meaningful out of using it as a ordinal/likert variable. They said, instead, I should dichotomize it as weekly or more (67%) or less than weekly (33%).  
I'm unclear on whether or not I should do this.  The articles I have read on dichotomizing variables (e.g. MacCallum et al., 2002) say that it is only okay to dichotomize when "the distribution of a count variable is extremely highly skewed".  Meeting frequency is skewed (standardised skewness = ~-2), but it is not extremely highly skewed.  
Thoughts? I'm thinking at the moment, I'm better off treating it like a likert variable, and then in the limitations section of discussion saying that the variable had low variance and was not very sensitive and this limits the validity of the findings. 
Any assistance on this would be great. 
The article I read about this is:
On the practice of dichotomization of quantitative variables.
MacCallum, Robert C.; Zhang, Shaobo; Preacher, Kristopher J.; Rucker, Derek D.
Psychological Methods, Vol 7(1), Mar 2002, 19-40
 A: I think the advice you got is wrong.
First, it's rather strange to think of the skewness of a variable that is ordinal and has 5 categories. Skewness implies interval level data; your scale is not interval level as it stands, but you could make it so by converting it into "times per month met" then
daily = 30,
twice a week = 8, 
weekly = 4,
fortnightly = 2
6% monthly = 1.
But dichotomizing the data makes "monthly" the same as "weekly" and "daily" the same as "twice a week". These aren't the same. Also, dichotomizing differently (e.g. putting "weekly" with "daily" and "2x a week" might give very different results. That can't be a good thing.
A: Check before aggregating. It is not constructive to decide if you should transform a variable without seeing how it is associated/correlated with the outcome: It's not how the responses fell into the categories that matters; it's how each of this categories helps you understand the outcome that matters.
If you were planning to model it as a continuous predictor, at least check if the five levels correlations with your outcome in a linear way. If they do, saving them as is will give you more information.
If you were planning to model it as four binary indicators, then do so and see if there is any trend. If yes, keep them as indicators. But if you see that the estimates of daily and twice a week are close to zero (assuming >1 per day is the reference group), while fortnightly and monthly are close to each other, then dichotomizing may then make sense.
Generally, aggregation throws away information and often causes bias, I'd recommend only using them when there is a practical reason (e.g. clinical definition of having a disease or not based on a biomarker).
