I am writing my master thesis about how the difference between majority's (Belgians) opinion about socially relevant topics and majority's perception of minority's (Turks and Moroccan's) opinion about socially relevant topics predicts intergroup hostility. For example, the data I am using has three questions about Belgians opinion about the role of women in society (e.g. "To what extent do you agree that men are more apt to lead than women"). I want to somehow compose those three questions to get an independent variable that would measure Belgians opinion about the role of women in society. I am confused about which approach would be more suitable for this; making latent variables or composite variables?
You say you have three questions about feelings toward women aimed at uncovering hostility, by description this is a latent variable analysis.
It is good way to uncover opinions, especially ones where people might have a reason to be elusive about how they actually feel.
There are very specific methods for doing this well and extracting meaningful inference. One of the more robust is using a Likert Analysis (see video here).
In a Likert analysis you have several questions aimed at discovering the level of a SINGLE latent variable. The creation of the scales & questions is non-trivial.
Although you say you have questions already, you may need to revisit and tune those questions to create meaningful variables with as little bias as possible.
The most important thing to understand is that all of the questions are reflective of a single variable and they must all be contributive in some way. But from your own description, this sounds like the way to go.
A good course (at your school or online) in the creation and evaluation of surveys might be a solid place to start!
Ultimately it is up to you. Statisticians and psychometricians would likely suggest you at least consider a latent variable approach because such an approach (among other things) would allow you to (1) investigate the properties of the items (e.g., at what areas of the latent trait does each item provide the most statistical information) and (2) statistically adjust regression estimates for attenuation. Investigation of item properties (i.e., 1) can be done using item response theory (IRT), and correction for attenuation (i.e., 2) can be done using structural equation modeling (SEM). Given your interest is to treat the latent variable as an independent variable, SEM may be the latent variable model most useful to your application. Though, as @bethanyP points out, the use of latent variable models is "non-trivial".
Because this is for a master's thesis, I would think your advisor would be ok with the composite variable approach (especially because, without prior experience, learning how to use and interpret latent variable methods such as SEM could take weeks). Just keep in mind that the use of composite variables may attenuate (i.e., artificially shrink) regression parameters involving latent variables.