1
$\begingroup$

I am working on a regression and I have a factor variable "Marital Status"

Marital status has 5 levels: Single, Married, Divored, Widowed, Other (don't ask me what constitutes someone being an 'other')

When I put this into R it will turn this one variable into 5-1=4 dummy variables and will give me this output

output

with divorced as the baseline.

My question is, can I only use the significant dummy variables? I would drop the variables for 'other' and 'widowed' and keep 'married' and 'single'

Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ Normally you would not drop the insignificant category-labels unless you had a good a priori reason to combine them with baseline (outside the information above). If you'd used a different baseline, you'd have different things getting combined together. $\endgroup$
    – Glen_b
    Commented Apr 15, 2015 at 3:39
  • $\begingroup$ See here, here, here, here, & here. $\endgroup$
    – Scortchi ♦
    Commented Apr 15, 2015 at 9:31
  • $\begingroup$ @Glen_b When should I combine category-labels? When they have the same coeff and confident intervals? When there are not enough observations? $\endgroup$
    – Metariat
    Commented Mar 10, 2016 at 13:40

2 Answers 2

Reset to default
5
$\begingroup$

You should reorder the variable, using "divorced" as reference category seems strange --- and, no, you should keep the original levels! Remember that the "significances" you have listed refers to comparisons with the reference level, so you are comparing, say, "widowed" to "divorced", is that what you want?

It seems more natural to compare "widowed" to "married", for example. You should investigate the comparisons that interest you, not only those your program choose to show you ... That being said, as long as you keep the original levels, it does not matter (algebraically, formally, ...) which level you use as reference level, the estimated model is the same, but standard output from standard routines present the results differently, and that can matter, practically!

Improve this answer
$\endgroup$
2
  • 1
    $\begingroup$ Now that you point it out, it does seem strange. I should keep all 5 levels of the variable though? $\endgroup$
    – Joel Sinofsky
    Commented Apr 14, 2015 at 17:42
  • 2
    $\begingroup$ Yes. Tha variables should be defined before starting the analysis, and then not be changed. $\endgroup$
    – kjetil b halvorsen ♦
    Commented Apr 14, 2015 at 17:43
1
$\begingroup$

Another option that you have is to create binary variables for the "significant" dummy variables or any of the other dummy variables that you are interested in. For example, observations that are "married", would receive a 1, and all others would receive a 0 for a variable that you could call "married_flag".

I would like to point out that just because something is "statistically significant" does not infer that it is practically useful. It is concerning that you don't know what this "Other" category refers to. I have been in this situation before, and I would highly recommend asking someone who might know the answer to this question. You may want to exclude records that selected "Other" if "Other" refers to subjects that simply did not answer the question. It is always helpful to be aware of as many aspects of the data as possible.

Improve this answer
$\endgroup$

Not the answer you're looking for? Browse other questions tagged or ask your own question.