OT suggested areas of research that could bring differst stat cultures together

[Art Kendall]

There are very different cultural silos in data analysis and statistics.

Years ago there were different kind of correlations depending on the measurement quality of the variables.
Someone figured out that many kinds of correlation coefficients were shortcuts to simplify hand calculation.  Using the same Pearson correlation software produced identical correlation coefficients. No more need for different hand formulae or different software for many coefficients.

Also years ago, there were two kinds of researchers "good" researchers who did ANOVA and "poor" researchers who used correlations/regression. It was found that ANOVA was a special case of correlation/regression. Now partially based on Doug Carroll's work on n-battery canonical correlations, the General Liner(ized) Model is routinely taught. It subsumed a wide array of analyses that were thought of as different things.

After receiving an email from Bruce Weaver, it occurred to me that meta analysis has strong logical and epistemological  parallels with complex samples and multi-level modeling.  One thing I had intended to do when I retired in 2001 was to compare and contrast complex sampling and  multilevel modelling.
I shall not have the opportunity to do it but perhaps some list members could try running their data (or text book data) on all 3: complex samples, multilevel modelling, and meta-analysis software.
After all these days that hard part is gathering the data and getting it entered into the computer.  Running all 3 approaches on simulated sets of data could also be very beneficial.

I hope that someone can use these ideas and perhaps even get funded to explore this area.  Much statistical research has tended to become narrower and narrower in focus.  Stepping back and taking a wider view could create new insights.

[Bruce Weaver]

Here's another example (I think) of what Art is talking about.  Some years ago, I wrote a syntax file to compare results from the Cochran-Mantel-Haenszel method to results from logistic regression.  You can download the file here:

http://www.angelfire.com/wv/bwhomedir/spss/pooling_odds_ratios.txt

I thought I had also included in the comments pooled odds ratios computed via standard meta-analytic methods, but do not see them there now.  As I recall, the pooled odds ratio computed via standard meta-analysis was nearly identical to the ORs from CMH and logistic regression.  

Cheers,
Bruce

Art Kendall wrote

There are very different cultural silos in data analysis and statistics.

Years ago there were different kind of correlations depending on the measurement quality of the variables.
Someone figured out that many kinds of correlation coefficients were shortcuts to simplify hand calculation.  Using the same Pearson correlation software produced identical correlation coefficients. No more need for different hand formulae or different software for many coefficients.

Also years ago, there were two kinds of researchers "good" researchers who did ANOVA and "poor" researchers who used correlations/regression. It was found that ANOVA was a special case of correlation/regression. Now partially based on Doug Carroll's work on n-battery canonical correlations, the General Liner(ized) Model is routinely taught. It subsumed a wide array of analyses that were thought of as different things.

After receiving an email from Bruce Weaver, it occurred to me that meta analysis has strong logical and epistemological  parallels with complex samples and multi-level modeling.  One thing I had intended to do when I retired in 2001 was to compare and contrast complex sampling and  multilevel modelling.
I shall not have the opportunity to do it but perhaps some list members could try running their data (or text book data) on all 3: complex samples, multilevel modelling, and meta-analysis software.
After all these days that hard part is gathering the data and getting it entered into the computer.  Running all 3 approaches on simulated sets of data could also be very beneficial.

I hope that someone can use these ideas and perhaps even get funded to explore this area.  Much statistical research has tended to become narrower and narrower in focus.  Stepping back and taking a wider view could create new insights.