Troubleshooting
Problem
I am performing a Logistic Regression analysis, either Binary or Multinomial in SPSS. I am sure that one of my independent variables is significant, but the odds ratio reported by SPSS as exp(B) is very close to 1.000. Likewise, the parameter value is very close to 0.000. What is wrong?
Resolving The Problem
Perhaps the scale of the variable is inappropriate.
The odds ratio can be interpreted as the multiplicative adjustment to the odds of the outcome, given a *unit* change in the independent variable. If the unit of measurement is very small compared to the size of a meaningful change, the odds ratio will be very close to one.
To be concrete, consider an analysis which uses annual income, measured in dollars, as an independent variable, with typical incomes in the tens of thousands of dollars, and a range of tens of thousands of dollars as well. The outcome we are trying to predict might be the purchase of an automobile. We will not be surprised to find that if we offer an individual an additional dollar, the odds that she will buy a new car does not change noticeably. A more meaningful unit of change would be a thousand dollars. Thus, dividing the original variable by 1000, so that income is now measured in thousands of dollars, will likely provide a value for exp(B) which has a meaningful interpretation. We might find that our hypothetical exp(B) is now 1.01, which we would interpret to mean that each additional thousand dollars in income results in a 1% increase in the odds of an automobile purchase. This is compounded: for each thousand dollars, we again multiply by 1.01, so that a five thousand dollar increase would result in an increase of (1.01)^5 = 1.0510100501, in excess of 5.1%.
Of course, we could instead compute exp(B*1000) ourselves. This is the same as (exp(B)^1000), that is, exp(B) raised to the one-thousandth power. This computation is mathematically correct, but care must be taken to obtain the most accurate estimate available of the coefficient B, or of exp(B). Any rounding errors will be enormously magnified. For most users, rescaling the independent variable after choosing a scale for which a unit change is meaningful will be less trouble. If the units are irrelevant, or in the absence of any obvious scale, simply standardize the variable, so that a unit change is one standard deviation. (This can conveniently be done by choosing Analyze->Descriptives, place a check in the box "Save standardized values as variables".)
Similar considerations apply when exp(B) is enormous. If a unit change is many times larger than would ever be observed, exp(B) will be astronomical if B is positive, or close to zero if B is negative. It may even be missing because of overflow.
Related Information
Historical Number
25957
Was this topic helpful?
Document Information
Modified date:
16 April 2020
UID
swg21479543