1
$\begingroup$

Suppose you want to predict an outcome using a sample whose N is...

  • 10,000 based on most demographic variables
  • 9,000 based on Survey Question 1
  • 3,000 who answered "Yes" to Question 1 and thus were given Question 1A
  • 700 who answered "Yes" to Question 1A and thus were given Question 1B
  • 4,000 who answered "Yes" to Question 2 and thus were given Question 2A
  • 600 who answered "Yes" to Question 2A and thus were given Question 2B

What methods, if any, have you found that are successful/worthwhile/practical ways of pulling all such information into a single model?

$\endgroup$
  • $\begingroup$ How is the "outcome" related to these subsets? $\endgroup$ – whuber Oct 10 '14 at 0:44
  • $\begingroup$ Every person has a Y value. $\endgroup$ – rolando2 Oct 10 '14 at 0:46
2
+75
$\begingroup$

If I understand your question correctly, there are 2 challenges in your data

  1. 1000 missing value for question 1. There is a whole literature on how to handle missing values in regression. Statistical Analysis with Missing Data by Little and Rubin is a very good reference for this topic. I think this is a minor issue for you.

  2. conditional branching in survey design. I think this is your main problem. I am not sure if there is any statistical model that can handle this. What I usually do in practice is to combine multiple questions. For example, suppose your outcome $Y$ is a numerical scale of depression level and you have the following questions:

a) What is your gender?

b) If female, do you have a new born at home?

Let's say a mother with a new born is often associated with higher chance of depression. What you really need is a binary variable, which equals 1 if the respondent is a mother with new borns and 0 otherwise. However, you don't want to directly ask this question as you will lose many other useful information such as gender(and it is really a strange question and will confuse many!).

So what I would suggest you to do is to go through your survey and think about defining explanatory variables that have direct causal relationship with your outcome. This is often called data pre-processing, sth. I consider far more useful than using complicated statistical models.

Peter

$\endgroup$
  • $\begingroup$ Your part 2 of the answer is good. $\endgroup$ – Subhash C. Davar Oct 15 '14 at 10:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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