0
$\begingroup$

I have box-office collection data on a number of movies. I also have the production budget, director name, lead actor, actress, language and other meta data related to the movie. I want to know which factors determine the box office collection. If a certain factor (say the director) is correlated with the collection, then I would like to determine its effect.

I was initially thinking of linear regression with dummy variables. However, I realised that I do not want to know whether a certain director, certain actor or certain language play a role, but rather whether direction, actor, etc. have any role in determining the revenue of the movie collectively. I would appreciate any help about the possible directions I should pursue.

Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

It sounds like you're looking to predict an essentially variable (box office collection; might be better modeled as ?) from another essentially continuous variable (production budget; same caveat) and several variables. If this is correct, and if none of your other meta-data that you want to use is , then the general class of analyses you probably want is , in which production budget would be a covariate, and the rest would be categorical factors.

Be careful to mind the assumptions of ANCOVA, as they are often violated. For instance, effects violate the homogeneity assumption of regression slopes for different groups. may also be advisable regarding the assumption of zero . If you suspect in advance that these assumptions will be violated, it's better to begin with a alternative, or at least to interpret those significance estimates first, then test assumptions and try parametric alternatives if is a problem with whatever analysis you'd use otherwise. For some options, see Akritas and Brunner's "Nonparametric Models for ANOVA and ANCOVA: A Review", who also handle ordinal data.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Thanks. In my case there are some additional ordinal variables, e.g, advertisement budget, number of countries of release, etc. Can I still use ANCOVA? $\endgroup$
    – imsc
    Commented Mar 21, 2014 at 19:11
  • $\begingroup$ Why do you say those are ordinal? Both sound like count variables, though budget could be effectively continuous. Note that if you do have ordinal variables, Akritas and Brunner discuss some options for this. $\endgroup$
    – Nick Stauner
    Commented Mar 21, 2014 at 19:13
  • $\begingroup$ My bad. Yes, they are count variables. $\endgroup$
    – imsc
    Commented Mar 21, 2014 at 19:17
1
$\begingroup$

You can still use regression with dummy variables, but instead of looking at significance of individual terms (single dummy variables) you will want to do full/reduced model tests. For example you can look at the model with all the variables you mentioned (as dummy variables), then fit another model without the director variable (all corresponding dummies removed). The full/reduced model test will then compare those 2 models, the null hypothesis is that the simpler model (without director) fits the data just as well as the full model, the alternative is that at least one of the deleted variables (director info) contributes to the fit. If the test is significant then that says that director has an effect (at least one director is different from others, could be many differences). If you don't reject the null hypothesis then that is consistent (but does not necessarily prove, could just be insufficient data) with the idea that director does not matter and the collection numbers are determined by the other terms.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Thanks. few additional question. (i) How do I test for significance, when I removed a variable. (ii) Some movies have multiple directors. My understanding is that the dummy variables need to be non-overlapping. Is it possible to use this approach in this case. $\endgroup$
    – imsc
    Commented Mar 21, 2014 at 19:08
  • $\begingroup$ Further, the number of dependent variables are very large. Can this cause any computational problem? $\endgroup$
    – imsc
    Commented Mar 21, 2014 at 19:16
  • $\begingroup$ @imsc, the common tools for automatically creating dummy variables will create them non-overlapping, but you can create them however you want. Just have dir1 be 1 if that director is involved and 0 otherwise. In R you can do the full reduced test by passing the 2 models to anova, other programs may have something similar, or it is not a hard test to do by hand (need the SSE from each of the models). $\endgroup$
    – Greg Snow
    Commented Mar 22, 2014 at 23:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.