Analysing Life Satisfaction after Averaging and Collapsing in two Categories

[Parham Abrahan]

Dear all,

for a project I want to examine the relationship between objective city-indicators and life satisfaction (LS). I obtained around 3000 individual LS-scores ranging from 0 to 10 for 15 cities and want to add the respective indicators for each in order to run a regression analysis. To do so, I collapsed the LS in two groups, ranging from 0-5 and 6-10, and calculated the average for each group for all cities before attaching the indicators, giving me 30 data points.

I had two questions:
Is it correct to run the regression for the low and high-LS category separately (otherwise it wouldn't make much sense for me mathematically speaking) and does anyone have worked on something similar and could suggest a good model I could use? I obviously tested some, but reg doesn't give me significant models while logit ologit don't seem to be reasonable choices, albeit the model is significant.

I would appreciate any help.

[Clyde Schechter]

I wouldn't do it this way at all.

First, making a dichotomy out of what is a (quasi-) continuous variable throws away information and distorts things. For example, you are saying that a person with a score of 0 and a person with a score of 5 are really the same thing, but they are both radically different from a person with a score of 6. There are very few things in the world where such a statement is sensible--I'd be very surprised if this is one of them.

Next, by averaging the scores of everybody in each city, you are throwing away information about variation, which is very important for getting good estimates of the uncertainty of your final results. I would do this quite simply as:

regress LS i.city other_indicator_variables

using the original data unmodified.