In my experiments of trying to measure the sense of social well being I used the list of 38 indicators (offered by some ukrainians sociologists). Each indicator tries to measure one type of social good such as sense of state protection, environmental security, fair government etc. The question in the questionnaire asks whether the respondent has each kind of social good at the enough level. So, the respondent has two choices: enough coded as 1 and not enough coded as 2. (and 99, of course, for those who were unable to answer). Then, frequencies were calculated. Later, the arithmetic mean of percentage of those who chose enough was counted as well as the mean of percentage of those who chose "not enough". The last value was regarded as a negative value. Then both means were summed up. So, at the end, if the final value was below zero I assumed the closer the final value to +100 the higher the level of the sense of social well being. And opposite, if the final value is a negative value heading to -100 then I have to conclude the sense of social well being is below average. Would 0 value be really average level of social well being or maybe +50 should be average?
Each sample times its frequency or weight gives the real mean. It is not necessary to be symmetrical in an interval
In dichotomous answers it is common to set negative (here not enough) to zero and positive answers to 1. This leads to gain more sensible results as the answers directly give the probabilities and proportions. This way 0.5 is hypothetical average (not real average), cos its the center of the probability. Real average is cached based on answers.
I have no Idea of 1, 2 valued answers. If you set the negative to -1 and positive to 1, then 0 is the middle.
Note that for such data the confidence limits are not symmetric near 1 or 0.
A few pointers. The first is that coding missing data as 99 is asking for trouble. Just leave it blank. Even SPSS can manage to read blanks as missing nowadays.
Don't code 1 and 2. Instead, code 1 = enough, 0 = not enough. This allows you to add together the answers to see how many 'enoughs' there were. If you code 1 and 2, then you cannot add up your scale items.
Second point is that you need to validate before you add the answers together. I would suggest that you apply Mokken scaling to the data. This will examine the items to see if there is a subset that conform the properties of a measurement scale, as defined by Guttman. The process is iterative. If not all of the items scale together, then it will try to create a second sub scale with the leftover items, and so on. Some items may not scale at all.
Finally, the sum of the scale items will probably follow something like a Poisson distribution. You can analyse it as a Poisson, but remember to use robust variance estimation, because the scores are likely not to follow an exact Poisson. As Ehsan says, the confidence intervals will not be symmetrical. Don't use normal confidence intervals – this isn't normal data, and you may get nonsense, such as negative values in the lower confidence limit. Even if you don't the intervals are wrong.
All of this is possible in Stata (you will have to download the mokken scaling command, msp, but the rest is built in.
Each sample times its frequency or weight gives the real mean. It is not necessary to be symmetrical in an interval
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics