2

I am encountering some problems with setting up a multinomial logistics regression (stepwise).

My data/model is the following:

  • Dependent variable: strategy1/strategy2/no preference
  • "Indenpendent" variable: risk perception value (prob*impact in percentage)
  • Moderator: risk profile (likert scale)
  • +some other control variables (demogrpahics etc.)

Since the risk perception value is given, there is no variance in the values. I basically have two scenarios, each with different risk peception values (high_prob*high_impact and low_prob*high_impact) and these values are the same for all respondents. This is because the choice/preference for a certain strategy is supposed to be dependent on the (given) risk perception). However, when I run the regression for ONE of the scenarios with only this variable it gives me the following output:

No significance level for the varaible but only for the intercept. For the significance level of the variable it only shows a "." (dot).

So, SPSS does not give me a significance level for the variable (probably because of no variance), only for the intercept. Do I just have to accept it or is there any way to tackle this problem? Is this output even useful? Would it be better to generate a random perceived value in the range of 0.9 to 0.99 for e.g. high probability to get some variance? And then cluster them in two categories, thereby obtaining two perceived values for each scenario?

Thank you very much in advance.

user18399
  • 41

2 Answers2

2

First of all, stepwise is problematic. Search variable selection or model selection for lots of discussions here and elsewhere.

Second, if you have a variable that doesn't vary then it cannot have an effect.

What to do next? Well, sorry, but I'd say "start over". Either you need different data or a different model. You cannot investigate this model with these data.

To see this most blatantly, suppose you are trying to see if 8th grade boys are taller than 8th grade girls, but you only collect data on 8th grade boys.

Peter Flom
  • 119,535
  • 36
  • 175
  • 383
  • Thanks for your reply, even though it sounds frustrating. However, collecting new data wouldnt help me since I would still have the same scenarios determining the risk perception. I just dont know how to quantify them and incorporate them into the regression. For example scenario 1 has a risk perception value of 0,09 (prob*impact), whereas scenario 2 has a risk perception value of 0,01. The problem is that this value is the same for all respondents since they had to select a certain strategy according to the scenario they read. I want to find out, what strategy is chosen given the scenario. – user18399 Jan 06 '13 at 21:31
0

Ok, I got valuable input from by supervisor:

Since I have have two scenarios (perceptions) per respondent, I can use binary coding for the scenarios to distinguish between them. By doing that, I obtain varying independent variables (perceptions) and can run one regression for both scenarios at once.

user18399
  • 41