Nature of outcome variable and predictors in regression analyses Let’s say I have a questionnaire that measures whether someone is depressed. It ranges from 1 to 50.
I ask people to report stressful life events in a 30 days period.
I hypothesize that stressful life events can predict the value of this score.
I do a regression analyses and find a coefficient value of 0.1 for my predictor that represents the number of stressful life events. It is highly significant and the average number of reported stressful life events is 100.
Let’s say I have an individual who reported 1000 stressful life events, then the resulting value of the regression model would exceed the limits of my questionnaire.
Can I hence do a regression analyses where my outcome variable is fixed between two values (e.g. 1 to 50), and predictors do not have clear minima and maxima?
 A: Item Response Theory (IRT; wiki) is a very well-established approach for making sense of data from tests and questionnaires. I can't exactly go into the details here, but the main idea in this case is that you try to model the probability of a participant agreeing or disagreeing with individual items on your questionnaire. In your case, you're interested in whether the number of stressful life events predicts the probability of agreeing with depression-related items.
IRT can be quite complicated. For example, do you assume that all of the items on your questionnaire are equally related to stressful events, or this relationship differ between items? Or are people equally likely to agree with each item, or are some items more likely to be endorsed than others?
More concretely, you would model the probability of participant $p$ agreeing with item $i$ as a function of their level of depression, $\theta_p$ using a logistic regression model, and model the relationship between the number of stressful events $x_p$ and $\theta_p$. This can be done as a multilevel model, for instance using the lme4 package, but I'm not going to go step-by-step through how to do this.

Another solution, which is much easier but somewhat less reliable, is to just transform your questionnaire scores using, for example, a scaled logit transform, to put the scores on a scale from $-\infty$ to $+\infty$
$$
\text{Score}_{\text{Transformed}} = 
log(\frac{\frac{\text{Score}}{50}}
              {1 - \frac{\text{Score}}{50}})
$$
and doing linear regression on the transformed scores.
