I conduct 3 linear correlation tests (serial correlation LM test, runs test and VR test) on a stock market index full sample of 15 years and as well on each year separately for the same period of time (static windows) in order to capture a changing degree of efficiency. Strangely, the full sample results indicate a different state of efficiency than the average of the 15 subsamples results for the same period. In extreme cases the full sample result indicates inefficiency at a significants of 1% whereas the subsample are on average efficient. How can that be? Shouldn’t it be more or less the same?
Any help is very much appreciated...
Hi Peter,
Without knowing your data, I cannot provide a definitive response, but it is certainly possible that the aggregated results differ from the average of subset analyses. It may be a function of the larger sample size in aggregate, or it may be that you did not weight the averages of the subsets by their sample size. You could treat this as a meta-analysis, and look at the pooled effects. That may give you results of the subsets that are closer in line with the aggregated data.
I can't really provide a more comprehensive response without knowing more about your data.
Ariel
Hi Ariel,
thank you very mich for your answer.
My data consists of daily prices (resp. log returns) from the Swedish stock market index OMXS between 01/01/2000 and 31/12/2014. Also other Scandinavian indices are considered and give similar results.
How could I discuss these findings? Which theory can explain these differences? How can I make sense of it? I couldn't find any hints in the literature...
Peter
Hi Peter,
Again, without testing the data myself, I can only posit possible reasons for these differences. One other possibility is that the smaller windows that you created represent more homogeneous behaviors of the data, whereas the aggregated data do not smooth out spikes or trends.
As for the underlying approach, I am not sure that taking windows of observations makes much sense. This is akin to comparing clusters, rather than the underlying observations. You'll have less variability due to the roll-up, but then you are, in effect, measuring the mean of the means.
At the end of the day, you are the one who will need to justify your analytic approach. I would suggest that you stick with the approach that is most consistent with your discipline.
Ariel
I have 400 data-set in which 140 belong to group A, 200 belong to group B and 40 belong to group C. I employed factor analysis on 400 data. Later on, I executed structural equation model on 400 data. However, I want to know the effect of this model at each group level. In other word, I want to do multi-group analysis. Due to small sample of Group C, I am not able to do multi-group mediation through structural equation model. Can I do t-test for each group (considering factors score of each factor ) and later on aggregate the results. Please suggest me appropriate method for this.
- Badges
- Science topic
- Similar topics
- Counseling Psychology
- Confusion