| Comparing variables across different years [message #4142] |
Tue, 07 April 2015 16:18  |
UAB_user
Messages: 21
Registered: September 2014
Location: Alabama
|
Member |
|
|
Hello,
For my project I am looking at a husband's migration across the 2006 and 2011 Nepal DHS. Is it possible to compare variables across years and test if they are significantly different? I know the PSUs and study design differ between years, so I am assuming that its not correct to do something as simple as conduct a Chi-Sq in a 2x2 tables of Year and another variable (example: Year x Sex).
How do you test if variables differ across years in a univariate analysis?
Thank you very much
Derek
|
|
|
|
| Re: Comparing variables across different years [message #4144 is a reply to message #4142] |
Tue, 07 April 2015 18:38  |
Reduced-For(u)m
Messages: 292
Registered: March 2013
|
Senior Member |
|
|
One way to think about it: you could pool the datasets (append them together) and regress on a constant and a dummy variable for "surveyed in 2011". The coefficient on the survey year dummy is the difference between the two means and you could do a t-test of the coefficient equals 0 (you could also run the regression with no constant and just a dummy for each survey year and those would give you the means of the two survey years, and then you can test the equality of the coefficients with an F-test). That would let you use account for survey design in both years (provided you fixed your PSUs to be survey-specific and de-normalized your weights appropriately, though the re-normalizing may be less important for two surveys from the same country if they are similarly sized samples and similar designs).
You could do this separately for men/women, or you could make a dummy for each group (genderXyear) and put them all in one regression and then test the coefficients of interest with post estimation commands (you would use "test" in Stata).
|
|
|
|
|
|
| Re: Comparing variables across different years [message #4146 is a reply to message #4145] |
Tue, 07 April 2015 19:27 |
Reduced-For(u)m
Messages: 292
Registered: March 2013
|
Senior Member |
|
|
You can add covariates to the model in exactly the same manner. Just instead of running a least-squares regression you run a logit:
local X "list variables here"
svy: logit Y `X' i.year, noconstant
Something like that. Or, if you wanted to compare changes in the X's over time, you could interact the variable with the survey year dummies, and run tests of statistical significance on the coefficient from the interaction term.
For your PSUs, just generate a new PSU variable that is PSU*10000+SurveyYear, and in your svyset command, set your cluster to that new PSU variable.
|
|
|
|
|
|
|
|
|
|
The DHS Program User Forum
Members
Search
Help
Register
Login
Home
