I have used two models: (1) binary logistics regress , when I regress I found some objective based variables are significant and negative against the theory and my hypothesis. The nature of the variables are continuous, dummy and categorical; and , (2) In the second stage, I employed Tobit model to estimate the maximum amount to pay to certain hypothetical market. Likewise, I found same wrong sign of sig variables (variable are, ordinal, categorical, dummy, continuous)
With theory and literature can't heading forward.
Literature says wrong signs brought because of:
(i) Bad Economic Theory; (ii) omitted variable; (iii) high variance; (iv) selection bias, (v) data definition/measurement error; (vi) outliers; (vii) simultaneity/lack of identification; (viii) bad instrument; (ix) specification error; (x) ceteris paribus confusion; (xi) interaction terms; (xii); Regression of the mean; (xiii) common trends, and so on.
So how to do it? Any suggestion to get rid of it?
Hello Haimanot,
To your list of possible explanations for paradoxical regression coefficients (which, by the way, is a great start), I'd add (multi)collinearity. You can certainly check your independent variables / predictors for collinearity as part of the data analysis. If excessive, you could try: (a) combining IVs that are strongly related into composite variables (perhaps via principal component analysis, factor analysis, or some rationally-based approach to definition); or (b) selectively dropping IVs that appear to be redundant with others.
If measurement error is a concern, always opt for measures with higher reliability over those with lower, if all else is equal.
You didn't mention sample size. If too small, it could be simply a matter of inaccurate parameter estimates because of too few cases (e.g., overfitting).
I empathize with your concern; it's no fun when data don't play well together with the (hoped-for) theory/pattern. However, that is often an opportunity to dive a bit deeper and discover why the apparent misfit occurs.
Good luck with your work!
Hi David,
Thank you for your time and details.
I think this is so helpful
FYI, I have gone all procedures that you figured out. Surveyed sample is 398 ; Is this sample small? Multidisciplinary checked and confirmed free. I have deducted some variables after principal component analysis conducted. Also, I have tried other models, for instance, heck-man-two stage sample selection, however, the output is still unpleasant like previous one.
From the very outset, I have designed cragg's double hurdle model: In the first stage I was plan to use Logit/probit; and, secondly, Tobit to estimate the maximum payments done. Yet, sign of the result is against of hypothesis and theory.
Those some of significant variables in both models have contradicting signs.
Scads thanks!
First thing I would ask is why you think the signs are incorrect? Best, D. Booth
Hello David,
Thank you for raising fundamental Q.
To say right or wrong is subjective; that means use of words to describe intuition of something sought care.
To explain it a bit, I presuppose education has a positive effect on acceptance of new idea/technology, but I got the reverse result. So, in this case the sign of education the sign of education is (-ve), which is unjustifiable. It could be a new finding which is being unexpectedly against with empirical study and intention of study
At last I took your message in using appropriate words, because words have power to convey, build, destroy, discomfort, pay attention, etc.
Why couldn't it be the case that this new technology is unhepful. That seemed to be the case for closed circuit television in the early days, esp in mathematics. See the link and the refs from this paper:
Article College student learning from televised versus conventional ...
Best wishes, David Booth
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression