ILCS Raking
Motivate Need and Illustrate Basic Approach
Dr. Ali Mushtaq
July 3, 2009(for academic purposes only)
RCRC
What is Raking?• A way to Adjust Survey totals “t” to
Independent Controls “T”• Takes existing Survey Weights,
usually wij = 1/pij, where pij is probability of selection
• Ratios them up to each total T in turn, until results are as close as wanted
What is the Value?• Can increase stability of survey
resultsReduce Sample Variance
• Get results that are close to desired outcomes
Reduce bias arising from minor operational errors
What Results to Expect?
• If Controls are Reasonable, Raking Process will converge
(“Hit” all controls)
• And improve survey results related to Control Totals
More Information Quality
• Only Weights are Changed by Raking, not Survey Data
• Data Quality is thus unchanged
• But Information Quality is usually Improved
What Does Raking Cost?
• Usually Done quickly on a PC• Independent Controls Need to be
consistent with each other• Sample must be reasonably large
for Raking to be Safely Applied• Some Costs incurred to explain
Method
Raking Made Simple
• “Fudge” Factor Intuition
• Develop a ratio of target total divided by sample total
• Repeat this process with each of the controls in turn
NSS Example from ILCS
While the NSS RA survey is raked across 4 dimensions (age, gender, marz and urban/rural), the example we’ll use here will just use two dimensions.
Table 1. Raking Example – Source Survey Data
Table 2: Desired Marginals
First Ratio Adjustment
Second Ratio Adjustment
After Second Iteration
ISLS Benefits Achieved
• Reduction in Bias
• Reduction (hopefully) in Variance
• Survey Results are Consistent with Census Projections
Again Many Thanks
Data Quality and Record Linkage Techniques
Springer 2007