SCCS test of hypotheses
From InterSciWiki
See the computer instructions for how to make tables, the same as on the course home page.
They key is to use ALL THE BEST VARIABLES (or a sample thereof) for each topic to make sure you test different variations for measuring you key relationships. E.g., If you are testing whether hunting societies are more likely to have menstrual taboos, don't use just one variable, use all for variables on taboos. It might not be number of taboos but just presence/absence for example.
You choose from the <SCCS index of variables> variables for a pair of topics that have sufficient coded data. You can use one or more constructed factors in the database or that you have constructed (explain how). Propose a general problem and viewpoint as to how these variables might be related in a non obvious way (i.e., avoiding two measures of the same thing, or connected by definition). Propose which variables you will use, and present hypotheses as to how they might be related.
Construct tables to show how they are actually related. Interpret the results.
Draw conclusions.
Outline 1st Draft Cross-Tab Paper
Contents |
[edit] Examples
- Some sample cross-tabs
- Christina's 2006 Five Step Process to your Final Paper (ppt)
- Warfare and Political Leadership (pdf)
- Sample commentary on an excellent paper
- Sample commentary on a not so good paper
- see formats: Condition 1 (N=C1) ==> Condition 2 (N=C2) (Exceptions: N=C1 not C2) (correl, signif, Fisher Exact) Reminder: C1 is less frequent than C2
[edit] Maps
[edit] Cross-Tabs and Graphs
- Statistical Analysis of Cross-Tabs
- Cross-Tabs in Spss
- A simple way to make Graphs from Spss Cross-tabs
Interpreting the cross-tab-- some caveats here:
- compare percents that are significantly greater than total percents over the whole table.
- If you want to compare, say, two same-column row percentages (percentaged on rows) then also use the Fisher Exact 2x2 test to compare raw frequencies of cases for row a freq / row a total / row b freq / row b total. Percent differences dont mean much if the samples are small.
- if you have more that 3 x 3 or 8-9 cells total Cramer's wont be valid because the expected values will be too close to zero. Then rely on tau-B or on a graphic result, and use the extrapolaion line in the graph.
- If you do a cross tab that includes a FACTOR SCORE,
- use only the Tau-b statistic, not Cramer's V (there are too many cells for it to be valid)
- include only the statistic results DO NOT include the cross-tab itself in the paper
- it would be better only to CORRELATE and show results graphically than do the cross-tab
- 5 If few cells and Cramer's V is much greater than tau-b in absolute value it is evidence of non-ordinal ("non-linear") relationship. Look to your percentages or graphs
- 6 Never include the Case Processing Summary in a paper!
[edit] Additional Stats
- Fisher Exact 2x2 test
- Another Fisher Exact 2x2 test
- Exact rxc Contingency test Exact analysis of large contingency tables used to solve the problem of sparsely populated tables when your cross-tab cells are not all larger than 4.
- Ordinal contingency tests
- Dave Garson's Stats page
- Selective statistics
[edit] Sample table
set your browser view -- text size -- so that the columns of numbers are vertical here
As discussed in class, 10/30/2007: Crosstabulation (Pasted directly from Spss, adding tabs and a space or two - put SPACE in first column of every line!). To copy and paste from an Spss output file,... and then control-A then copy, paste.
v775 (Low) Compliance of individuals w/ community norms (Columns)
*v774 (No) External Warfare (Rows)
1 High 2 Moderate 3 Highly variable Total Row %s
Row 1 Col% of Row 1 total
1 Frequent, yearly 20 15 9 =75% of 12 44
2 Common 5 6 1 12
3 Occasional 2 3 1 6
4 Rare or never 15 4 1 20
4 Row% of Col 3 Total =75% of 20 =100%
Total 42 28 12 82
Col Total % =100%
Cramer's V = .224 sig. = .223
Kendall's Tau-b = -.194 sig. = .043 (linear? contingent?)
Interpretation of Percentages is "Contingent Col 3 ==> Row 1 (75%);
Row 4 ==> Col 1 (75%)" !!!
CONTINGENCY RESULTS: Low Compliance (N=12) ==> Frequent External War (N=44) Exceptions: 3 p=.04 for Tau (ordinal) -.194
CONVERSELY Rare External War (N=20) ==> High Norm Compliance (N=42) Exceptions: 5
On 10/30/2007 we discussed whether a correlation or contingent pattern (as above) would REPLICATE. I thought it interesting to see if the cross-tabs would show the same pattern as above, dividing the sample into a Pre-State/State dichotomy as a layer variable, first recoding a political levels variable this way. And indeed, the pattern DID replicate, p=.01 (stronger than the original sig.=.04 even with a smaller sample), but only for the PRESTATE sample. The reason the STATE level societies did NOT show significance (and did show a positive not negative correlation) was that they almost all had FREQUENT EXTERNAL WAR, and only in 2 cases was there RARE EXTERNAL WAR so basically: this variable did not vary much, and could not form a significant correlation.
Nov 1/2007 we say how to get a similar result without recoding, just using variable 83 as the layer variable (1-4 levels of sovereignty).
We also discussed how to use the Fisher Exact tests to test significance of results with selected contrasts within cross-tab results.
