0
$\begingroup$

My dataset consists of 4,000 companies from 24 countries. Each company has certain features, the IV's. Consider the variable list below: green is the DV, red the IV's and blue are dummies (one for each country a company might be from) for which I want to control the OLS output. There are 24 countries entered.

enter image description here

However, once I let SPSS perform an OLS on these variables, it throws out most of the IV's.

enter image description here

The resulting model is almost useless in explaining any of the variance in the DV. My questions are:

  1. Why is SPSS throwing out these variables?
  2. How can I prevent this from happening?

enter image description here

[edit]

Added regression model in response to @Penguin_Night's and @NickCox's answers:

enter image description here

$\endgroup$
  • 5
    $\begingroup$ Presumably the predictors omitted are collinear with some of those included, but that's SPSS doing good on your behalf, not being difficult. Tastes and standards vary, but in my field firing 25 predictors at 475 observations would not be regarded as statistics worth much discussion. I'd advise against fitting anything like so complicated a model to a modest dataset unless you have a really good theory-based story to back it up. Incidentally, if you have 4000 companies, what happened to the other 3525? $\endgroup$ – Nick Cox May 7 '13 at 12:20
  • $\begingroup$ This part of the analysis is based on 475 financial firms...the other 3525 firms are from different industries. $\endgroup$ – Pr0no May 7 '13 at 13:12
  • $\begingroup$ The problem I'm having is as follows: based on theory, companies from countries with english legal origin (UK, US, Australia, Canada, Ireland, etc.) should have lower CSR scores than companies from countries with french, german, or scandinavian legal origin. However, in my regression (shown here: stats.stackexchange.com/questions/58304/…) the opposite is true. I want to know the inner workings and that's why I thought of inserting country dummy variables for each company. $\endgroup$ – Pr0no May 7 '13 at 13:19
  • $\begingroup$ The reasoning is that of the 475 companies, 170 are US-based. Other companies from english legal origin are UK (23 companies), Singapore (13), Hong Kong (24), Austraial (24) and Canada (41). It actually shows in the correlation table that Australian and UK companies have a positive relation to CSR score, but US companies negative. Could it be that because of the relatively high number of US companies, the coefficient for aggregated english-law companies becomes negative? $\endgroup$ – Pr0no May 7 '13 at 13:24
  • 2
    $\begingroup$ My dislike of "dummy" as a term is personal, but not directed against you, but please set it to one side. My main point was that if your hypothesis is that there are two fundamentally different kinds of legal system, then you only need one indicator. If you want to look at matters in more detail, that's commendable as social science, but trickier statistically when it means estimating lots of parameters. Again, if you want substantive advice on this kind of data, I can't provide it. $\endgroup$ – Nick Cox May 7 '13 at 13:33
8
$\begingroup$

It is very difficult to diagnose without knowing which one is categorical, binary, and continuous.

If a categorical variable has k levels, you only need (k-1) dummies to capture the information. If you fit all k of them, SPSS will throw one out. The reason is that if all k are present, the whole set are perfectly collinear, standard errors will then get too big; the model, fail.

Another reason for some others being excluded is probably their information was already perfectly captured by other variables. For instance, you "LAW" categorical variable may be already perfectly explained by the country dummies because it's very unlikely that a country will have two equally superior official languages or law systems.

I think overall, the analysis presented here is too hasty. I'd suggest checking some tabulations between predictors to better understand if there is any exclusiveness going on. Fit the model step by step manually to observe any change.


The regression in the OP shows that US-based companies have a negative relation to CSR score, and UK-based companies have a positive relation. In my dataset, 170 out of 475 companies are US-based. All other english-law based companies have only about 100 companies between them. How can it be that if most of english law-based countries have a negative relation to CSR score, in the regression I linked to, the english-law variable is actually positive (.740)?

There can be many reasons.

  1. Regression takes into account of group size AND the actual data each case brings to the table. A smaller group like UK can still drag the average if their firms has a very large and positive score, if the US firms' scores are mediocre.

  2. The reference groups in the two models are different. In the linked output, the reference group is all non-English countries. In the above output, the reference group are all other 18 countries that were not included in the model. Since UK and US are probably not the only two countries practicing the English system, changes in their mean estimate may be possible.

  3. Different control variables can also change the estimates, an exception is when the control variables are totally independent from the main predictor, which in this case is unlikely.

$\endgroup$
  • $\begingroup$ Not my wording! OP explained in a comment that he is just focusing on financial firms. $\endgroup$ – Nick Cox May 7 '13 at 13:26
  • $\begingroup$ @NickCox Revised. I was answering that when the comments are in exchange. Apology. $\endgroup$ – Penguin_Knight May 7 '13 at 13:27
  • $\begingroup$ Fine; no offence taken. $\endgroup$ – Nick Cox May 7 '13 at 13:35
  • $\begingroup$ @NickCox please see my regression model here: stats.stackexchange.com/questions/58304/…. It shows that companies from english-law bases countries outperform companies from countries with french, german or scandinavian legal origins, regarding CSR score. This is contrary to almost every paper out there in the legal field, so I'm trying to look at where this result comes from. I have added in the opening post a small regression with only the english-law based country dummies (and financial returns but nevermind). $\endgroup$ – Pr0no May 7 '13 at 13:43
  • 1
    $\begingroup$ Just highlighting a point made by Penguin_Knight that seems to get lost in the discussion: If, for the most part, all firms from a given country also have the same legal system, you can't put both countries and legal countries as predictor in an OLS regression model. From what you said until now, this seems to be your main problem at the moment. $\endgroup$ – Gala May 7 '13 at 14:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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