1
$\begingroup$

I'm working on some of my analysis section for my Master's Thesis in MPA, and while I have consulted my professors, I'd just like to double check with some external eyes (As obviously they are on my committee, so they can't over-input into my writing, and not all of them use SPSS) While I realize interpreting results is more involved than a brief glance, I'd thought I'd put it out there.

I'm running a Binary Logistic Regression in SPSS on the 2012 National Financial Capability Survey (NFCS) from FINRA. The model is running on a sample of 16768 cases.

My research question regards home ownership and student loan debt, which for this dataset is specifically whether the presence of student loan debt amongst respondents with some college education negatively impacts the likelihood of owning a home.

This output is from the model after all variables have been added. For background:

The DV is a binary (Yes/No) question "Do you or your spouse/partner currently own a home?" with 1 being yes and 2 being no.

The IV in question (Highlighted in Red below) is a similar binary question "Do you have student loans" with again 1 being yes and 2 being no.

Variables are set to reference(Last), therefore "No" would be the reference case for the IV in question.

NFCS Model Printout

Based upon this, I'm leaning towards the conclusion that holding other factors constant, having student loan debt makes a respondent 1.650 times more likely to not own a home (I'm aware of the problems of confusing phrasing), thus concluding that while significant, it is perhaps not as significant a factor as income (Variable A8) (Which had odd-ratios that were much higher)

Based on this, can my evidence support this conclusion, in your opinions? Am I missing anything? I've not had much experience working with Logistic regression, so I'm keen to get this right (And not embarrass myself in front of my committee).

Thanks for your time, and let me know if you need more information.

Note: Just to clarify, I realize its a big ask, but keep in mind I'm not asking for completely definitive judgements. You don't have to critique it to the degree of my thesis committee. I would just like some extra (yet still critical) eyes on what I might be doing wrong, missing, or need to drill down on, and if this seems like a professionally (Which is subjective) defensible conclusion to reach. If you saw this data and conclusion in a paper, what would jump into your mind?

$\endgroup$

2 Answers 2

1
$\begingroup$

Your conclusion seems correct to me. I think you cannot compare different coefficients in a model directly if they are measure different things. You can make a trick here, change your income unit from such as 10k dollar to one dollar then the coefficient will become very very very small. 1 dollar's difference in income will not have too much effect on the probability to buy a house even the p value will not change.

$\endgroup$
3
  • $\begingroup$ Thank you. So, what you're saying is that I need to be cautious when I list other coefficients, such as income, as the 1 unit change in income will not be the same as the on/off state for the student loan variable, which will be something I'll need to mention when I write my analysis? $\endgroup$
    – Eric Click
    Aug 27, 2015 at 17:28
  • $\begingroup$ Just for clarification, the income is being measured in an ordinal scale with roughly evenly spaced categories (Banded incomes) hence, I treated it categorically in the regression model (There were only 11 categories. My faculty have suggested treating it continuously, but I would have to research some methods for doing so (Midpoint looks promising). $\endgroup$
    – Eric Click
    Aug 27, 2015 at 18:29
  • $\begingroup$ I am not very familiar with SPSS, I think you can do this way in SPSS if you treat income as a "factor" i.e you treat it as a categorical variable, then all other variables will compare with a reference level (probably least income group). If you put income as a covariate in SPSS then it will be a continuous variable and you only get one $\beta$ which means one level increase how much the odds of buying a home increase. $\endgroup$
    – Deep North
    Aug 28, 2015 at 2:10
1
$\begingroup$

Bear in mind that the logistic equation is linear in the log odds, so the actual changes in the probabilities will depend on all the variables. The actual change in p, dp/dx, is p(1-p)*beta. One more caveat. If there are any interaction terms in the equation involving this variable, you would need to take those into account, too.

$\endgroup$
1
  • $\begingroup$ There aren't any interaction terms, as my knowledge of statistics isn't very advanced - thus one of the pitfalls I will be listing in my thesis. In regards to the first point, do you mean the difference between (B) and Exp(B) or that the probabilities are relative to each other? $\endgroup$
    – Eric Click
    Aug 27, 2015 at 17:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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