is is possible ordered logit or OLS?

[Jang JinHee]

the average value of 8 questions on a 5-point scale was used as dependent variable.
the resulting value are 1, 1.2, 1.3, 2, 2.1, 2.7, ......, 5.
is ordered logit possible at this case?
unable to calculated marginal effect.
should I do OLS? If there is an endogenous variable, should I do IV?
the estimated result is that OLS is significantly better than IV....
is there any basis for using OLS?

[Maarten Buis]

It is fairly normal to use linear regression in this case. It is obviously wrong, but all models are wrong. The definition of a model is that it is a simplification of reality. A simplification is just something that is wrong in a useful way. For example, we could simplify the number $\pi$ with 3.14. This is wrong, but allows us to focus on the three most significant digits. It depends on how we want to use this simplification, whether it is good or bad. For cases like you are describing, a linear model is often good enough.

I don't know why your are not able to calculate marginal effects. You don't say enough for us to diagnose what is going on. I suspect, however, that you can calculate the marginal effect, but could not interpret them. In an ordered logit model there are for a single explanatory variable as many marginal effects as there are outcome categories. I suspect that that is not what you want.

Whether you should use instrumental variables, depends on whether you have an instrumental variable. A good instrument is extremely rare. If you have an endogenous variable, then there is no guarantee that an instrumental variable exists. It is actually very unlikely that a valid instrument is present in your data. Whether you should use instrumental variables depends on whether you have an endogeneity problem and the quality of your instruments. The most common situation is that you have an endogeneity problem but no solution; sh*t happens. In that case you need to be clear about the substantial limitations of your study.

Pedantic note:
Linear regression is the model, and OLS is the algorithm used to estimate its parameters. Since you are asking about he model and not the algorithm (would the answer be different if you estimated the parameters of the linear model with maximum likelihood?), the correct way to refer to it is linear regression and not OLS.