2
$\begingroup$

Currently iam working on my master thesis which is about risk adjusted returns in the Asia Pacific REIT market. The goal of the paper is to determine/find variables that conceive explanatory power over the IV.

To determine this I also performed a FE regression based on countries. The result are more or less in line with FE for the whole sample (with firmID as dummy) and OLS regression. The suprising thing actually is that for the whole sample period (so still with country dummies) some of the variables seem to be insignificant, while on the contrary, if the sample is split up in 2 sub-periods, they show to have a significant effect in both the periods.

I find this rather strange that a variable is significant for both the sub-periods, but fail to do so for the whole sample period. What could be the reason the reason for this? Is this a statistical problem or more should I be looking more in the finance direction? For those with an finance background, the variables which show this pattern are market capitalization, price-to-book ratio and dividend yield.

I've uploaded the FE regression result for the Australian market. Hopefully this works clarifying.

Help is highly appreciated!

PS. Lagrange and Hausman test favored FE model.

Whole sample regression; http://www.freeimagehosting.net/newuploads/pyjot.png

Sub period 1; http://www.freeimagehosting.net/newuploads/d3ull.png

Sub perdiod 2; http://www.freeimagehosting.net/newuploads/betwt.png

$\endgroup$
1

2 Answers 2

2
$\begingroup$

When you compute separate models for different subsets of your data you get different models. For example, factor mc is significant in both sub-periods but not the whole period. The crucial thing, imho, is that the coefficients the different models predict for this factor differ. So for mc in the first sub-period, the coefficient -0.00014 adequately accounts for your data. In the second sub-period, the coefficient is positive, 0.00015, and also adequate for that time period. Since this factor has a negative influence in the first but a positive influence in the second sub-period, if you look at the whole period the coefficient is close to 0 and cannot account well for the data.

If it makes sense to you that this factor should influence the outcome differently in the two sub-periods then treating them differently would be useful. Then you could stay with two separate models or use a regression method that allows you to specify breakpoints. If assuming different effects in the sub-periods is theoretically implausible you should rather stay with the model for the whole period. So I would not call this a statistical problem, it has more to do with the motivation behind such an analysis and its interpretation.

$\endgroup$
2
  • $\begingroup$ Thanks for answer, this is in line what @maarten said. I quess I indeed have to motivate it $\endgroup$
    – Henk
    Commented Jul 12, 2013 at 12:46
  • $\begingroup$ That's true, I didn't see his answer until after I posted mine. If you find the answers satisfactory, could you accept one of them? $\endgroup$
    – robert
    Commented Jul 12, 2013 at 12:56
1
$\begingroup$

Lets take your first explanatory variable mc, which is insignificant in your enter period, but significant in bot sub-periods. In the first sub-period it is negative and your second sub-period it is positive. So you did not find a significant effect in the whole period because the different effects in the different periods canceled out.

$\endgroup$
1
  • $\begingroup$ Ah oke, that sounds plausible! Actually never thought about it this way. So it is correct to say that when one find an opiste effect in the sub periods, the effect cancels out..? $\endgroup$
    – Henk
    Commented Jul 12, 2013 at 12:41

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.