T-Transformation & Scales? Transforming 3 different variables into one variable

#1
Hello,

If anyone could help me with this SPSS question I will be forever grateful. :)

I have three measures in my survey regarding drinking behavior and after conducting a chronbach alpha test they are essentially the same question asking the same thing (alpha= 0.582).

I would like to transform these three separate drinking behavior variables into one drinking behavior variable, however, they are all scaled differently (one is drinks in the past two weeks, one is drinks within two hours, and one is peak drinks in one occasion).

How can I make sure they are all scaled the same? I understand how to compute the three variables into one (using the addition function), but I am unsure how to weight the variables the same.

Help? Thank you!
 

CB

Super Moderator
#2
Hi there,

I have three measures in my survey regarding drinking behavior and after conducting a chronbach alpha test they are essentially the same question asking the same thing (alpha= 0.582).
Hmm. That's quite a low Cronbach's alpha (though your number of items is low, which is probably the reason). But maybe more importantly, Cronbach's alpha doesn't tell you whether a set of items measures the same thing. It is an estimate of reliability in the classical test theory sense, but it does not directly tell you very much about the internal structure of the test. See On the use, the misuse, and the very limited usefulness of Cronbach’s alpha

I would like to transform these three separate drinking behavior variables into one drinking behavior variable, however, they are all scaled differently (one is drinks in the past two weeks, one is drinks within two hours, and one is peak drinks in one occasion).
Do you mean number of drinks in the past two weeks(/two hours/peak occasion)? If so, don't the second two items just tell you information that's captured in the first item?

If you wanted to make sure each item has the same influence on the total score, you could standardise each item before summing. But if the first item captures all the info in the whole set, then just using that in your main analyses might be better, and would mean you have a variable whose scale has intuitive meaning.
 
#3
Hi CowboyBear,

Thank you so much for your response, I really, really, really appreciate it. Yes, it is quite low indeed. My thesis advisor suggested I run the chronbach's alpha test and should I find poor internal consistency (which I did), to then transform each of the 3 drinking variables into one drinking variable. FYI I have a very small sample size (N=29) and 3 DVs, which I'm hoping to transform into 1 and then conduct a hierarchical MLR.

The reason I had three 3 DV on drinking is because there is no consensus on the definition of binge drinking, so I wanted to include the different definitions of each. For that reason, I asked:

1) In the past two weeks, on how many occasions have you had more than 4 drinks in a row (if you are female) or 5 drinks in a row (if you are male)?
0 times (recoded as 1)
1-3 times (recoded as 2)
4-6 times (recoded as 3)
7-9 times (recoded as 4)
10+ times (recoded as 5)

2) During a typical drinking session, how many drinks do you typically consume within 2 hours?
0 drinks (recoded as 0)
1-3 drinks (recoded as 2)
4-6 drinks (recoded as 5)
7-9 drinks (recoded as 8)
10+ drinks (recoded as 11)

3) In the last 30 days, what is the highest number of drinks you have consumed in one drinking session? (this was filled in by respondents and ranges)


Now, since the questions capture the same information, or essentially because there is this latent variable of being a risky drinker as deemed by my advisor, I'm hoping to create one variable.

Do you have any suggestions on how to do this in SPSS? Would standardization as you mentioned work in this case?

Thank you again.

Yasmine
 

CB

Super Moderator
#4
Hi again,

My thesis advisor suggested I run the chronbach's alpha test and should I find poor internal consistency (which I did), to then transform each of the 3 drinking variables into one drinking variable.
I don't really follow why a low Cronbach's alpha implies you should bundle the items together, but since your advisor recommends it I won't focus on that.

The reason I had three 3 DV on drinking is because there is no consensus on the definition of binge drinking, so I wanted to include the different definitions of each. For that reason, I asked:

1) In the past two weeks, on how many occasions have you had more than 4 drinks in a row (if you are female) or 5 drinks in a row (if you are male)? ....

2) During a typical drinking session, how many drinks do you typically consume within 2 hours?....

3) In the last 30 days, what is the highest number of drinks you have consumed in one drinking session? (this was filled in by respondents and ranges)
Ah, ok. So your data is in the form of behaviourally anchored rating scales, and not in the form of actual counts of drinks.

It looks like your first and second items both have 5 possible response options. Is this also the case for the third item (after recoding into ranges)? If so, all your items have the same number of response options - i.e. similar scaling. The conventional thing to do would be just to use the same coding scheme for response options for all items (i.e. options of 1-5; you would need to change item 2's coding). You could then just sum the responses for all 3 items.
 
#5
Hi again,

Thank you for your response!

It looks like your first and second items both have 5 possible response options. Is this also the case for the third item (after recoding into ranges)? If so, all your items have the same number of response options - i.e. similar scaling. The conventional thing to do would be just to use the same coding scheme for response options for all items (i.e. options of 1-5; you would need to change item 2's coding). You could then just sum the responses for all 3 items.
The third question was a fill in the blank, so it ranges a lot, min 0 max 16. Would it be incorrect to create z scores for each of the three, create a composite, create a z score of the composite and then a t score and use that as the new variable for analysis?

Thanks!
Yasmine
 

CB

Super Moderator
#6
Creating a z score for each item and then summing sounds reasonable. I'm not sure why you'd create a z score and then a t score of the composite too though, surely one or the other would do the job?