coefficient of logit

[ALKEBSEE RADWAN]

hi every one
please i have an issue
when i run logistic or logit command the coefficient comes out over 1 (such as 2.3 or 3.5)
please is that ok or there is a mistake

[William Lisowski]

That is OK. The logisitic model fits the log of the odds ratio - which can then be transformed into the probability - and the log of the odds ratio takes values from minus infintity to plus infinity, so large coefficients do not imply a probability bigger than 1 or less than 0.

[Phil Bromiley]

As William notes, these are quite reasonable parameters for either estimator.

You may find it easiest to pick either logit (if you like probabilities) or logistic (if you like odds ratios) rather than worry with both. You will find it helpful to use margins and possibly after the logit or logistic to help understand the output.

[Richard Williams]

Coefficients are always dependent on variable scaling. If you estimate a model using income in dollars, and then reestimate the same model using income in thousands of dollars, the coefficient for income will be a thousand times larger ($1,000 has 1,000 times as much impact as does $1). You can easily tell if a coefficient is statistically significant or has a positive or negative effect, but other means are usually needed to get a better feel for the substantive significance of findings.

[Clyde Schechter]

Let me strike a note of partial dissent from what others have written above.

It is certainly true that for continuous predictor variables, the coefficient can be of any magnitude because scaling of the predictor is a major determinant. But if you have a dichotomous predictor, you should be quite suspicious if you get logistic regression coefficients that are outside the range between -1.4 and 1.4. The reason I say this is that while such coefficients are theoretically possible, in the real world it is very unusual for associations that strong to exist. A logistic regression coefficient of 1.4 corresponds to an odds ratio of exp(1.4) which is just a bit larger than 4. There simply aren't that many real world situations where an odds ratio that large (or smaller than 1/4) actually happens. There are a few that come to mind: the association of smoking with risk of lung cancer, or the association of asbestos with risk of mesothelioma. But those are quite exceptional, and most things in the real world are so multiply-determined and subject to so many different influences that associations as strong as an odds ratio of more than 4 seldom happen.

[William Lisowski]

I will agree with Clyde. In my response at post #2, based on the brevity of the explanation in post #1 - no explanation of why "over 1" should be a concern - and on previous posts, I assumed the concern was that the coefficient was in some way functionally limited by the choice of a logistic model, as opposed to empirically limited by the unlikelihood of such a strong outcome. Clyde provides a good rule of thumb, and perhaps the concern in post #1 is warranted, but in the absence of any substantive information about the problem, the data, and the model, it's impossible to address whether empirical concern would be warranted or not.

[ALKEBSEE RADWAN]

@William Lisowski, @ Clyde Schechter, @ Richard Williams, @ Phil Bromiley
i m grateful for all of u
thanks for answering me
i got it completely