# Need help doing multiple linear regression with several independent variables that are of differing levels of measurement

I am attempting to predict levels of body dissatisfaction from a number of independent variables.

The dependant variable is Body dissatisfaction = overall score on a body dissatisfaction scale.

The independent variables (predictors):

1. Self esteem levels = overall score on Rosenberg Self Esteem Scale

2. Drive for muscularity levels = overall score on Drive for Muscularity scale

3. Age

4. Current involvement in sport and exercise = yes/no answer

5. Previous involvement in sport and exercise = yes/no answer

6. Masculine gender role identification = overall score on Bem Sex Role Inventory scale

7. Fitness magazine exposure = amount of fitness magazines read + time spent reading these magazines.

8. Perfectionism levels = overall score on frost multidimensional perfectionism scale.

9. Levels of appearance conversations regarding muscularity = overall score on Male Body talk scale.

So basically I need to know:

• Can I do multiple regression when I have different IV's with different levels of measurement? If not, what other type of test could I use?

• Do I need to transform the nominal variables (current and previous sport and exercise involvement) in dummy variables even though there is only two possible answers (yes = 1, no = 2)?

• Can the overall score of scale items be considered interval data? If not, how can I alter it e.g. Rasch modelling

I can provide more information as needed, thanks for any help in advance!

## 2 Answers

You can certainly do multivariate regression with different IVs with different measurements. You just have to interpret your results with that specific measurement in mind (an X increase in that unit of measurement leads to an X decrease/increase in the outcome...a one-level increase on a Likert scale, a one pound increase in weight, etc.).

I would recommend recoding your variables with yes/no outcomes into 0 and 1 answers (this can be done very quickly in Excel or Stata; most likely other software, too). Remember that your results will represent the effect of your 1s, relative to your 0s. That is, if current involvement in sports is the variable and 1 is a yes, and you get a significant beta of, say, 2.3, that means that current involvement in sports has a significant likelihood of increasing your dependent variable by 2.3 units relative to those who do not exercise.

Interval data: If it's, say, an infinite range of numbers where moving a point any direction is relatively the same as moving a point at any other point in the scale (so, a change from 49 to 50 is likely the same as a change from 72 to 73, for example), you can call that interval. Think temperature. Technically, there is no limit either direction, and a one-degree change always represents the same measurement. If the differences between points are not precisely known--for example, you have a self-esteem scale that runs from 1 to 5, representing survey answers of "hate my body, dislike my body, neutral, like my body, love my body", that's ordinal data. You can't say that the move from hate to dislike is equivalent to the move from dislike to neutral. There's no real mathematical measurement. Does that make sense? There's also ratio data, which has the properties of ordinal (it has a logical increasing/decreasing order) and interval (equal units), but is also bounded by zero. An example of this would be age--you can't be any negative number of years old, but a one-year difference always represents the same amount of time, whether you are old or young.

You can check this out for more examples: http://psychology.ucdavis.edu/faculty_sites/sommerb/sommerdemo/scaling/levels.htm

However, I am NOT familiar with Rasch modeling, so I'm not entirely sure how to put this part of my answer in context for what you need. I yield the floor at this point.

Just want to add one point to @ShannonC 's comment on interval data: although scales are not technically interval data, in some cases it's appropriate to treat them as such. Especially if you have taken the mean or created a factor from several items, such that your values for each individual are continuous and normally distributed within the bounds of the scale. For example, although Jim's responses to the questions about masculinity may be 2, 3, 2, 3 his average score is 2.5. Sarah's average score is 2.3, and Bob's is 2.7. You can then use a histogram to see if the answers are normally distributed. If they are, it's probably safe to treat the scales as a continuous variable.