On December 19 & Earlier, Pura Rayco-Salon <[email protected]>
wrote:
>
> We did a reseach using a controlled before-and-after
> quasi-experimental design, but with a different set of
> subjects at endline. How do I analyse this data?
>
> Two provinces were chosen, one as control area and the
> other as experimental area.
> From each province, 3 cities were chosen.
> From each city, 3 villages were chosen.
> From each village, a random sample of women were
> chosen -
>
> A representative health care worker from
> each intervention village was trained on the new
> health care delivery system and they in turn trained
> the other health care workers in their village
> (between 15 to 30 village health care workers per
> village). No such training was given to the control
> areas.
Pura:
I am not expert in xtmixed, so more knowledgeable members of the
list might be able to give you better or more correct formulations.
Do not forget to add individual level covariates like age, pregnancy
history, and weight to your models! I like BIC for comparing
models, but other may disagree. Similarly, you may want to put "ml"
as an option after the xtmixed commands to better compare
likelihoods. (however these comparisons are very dependent on
normality).
c_id unique to the six study cities
v_id unique to the 18 villages
city_type 1= Upper Class 2= Middle Class 3= Lower Class
village_type 1= Coastal 2= Inland 3= Mountain
For interactions it is better to have short names.
gen g= group==2 //0-1 group indicator
gen s= survey==2
gen gs= g*s
label gs "Group-Survey Impact of Intervention"
gen v= village_type
gen c = city_type
* Gives 18 unique group city village combinations: effectively a
village ID
/************************
*********************************************************
* 2. gs varies with city type but not village type *
*********************************************************
xi: reg y g s gs i.v i.gs*i.c
test gs _IgsXc_1_2 _IgsXc_1_3
estat ic
/
************************************************************************
************************************************************************
***************************************************************
Fixed Effects Models Only. WRONG MODELS
The problem with these is that they use the between-person variation
to estimate the precision of the group X survey term, which is the
measure of impact of the intervention. They also ignore the fact
that measurements in one city and village may be more alike than
measures in different cities and village. So, you cannot use the
standard error to predict effects in OTHER villages.
You can use the fixed effects models to do tests of fit, residuals
plots, for your models.
************************************************************************
************************************************************************
************************************************************/
*************************************************
* 1. gs does not vary with city and village type*
*************************************************
xi: reg y g s gs i.c*i.v
estat ic
*********************************************************
* 2. gs varies with city type but not village type *
*********************************************************
xi: reg y g s gs i.v i.gs*i.c
test gs _IgsXc_1_2 _IgsXc_1_3
estat ic
************************************************************************
*************************
* 2a. Fits a separate parameter to each village . Compare to Model 3*
************************************************************************
*************************
xi: reg y g s gs i.v_id
/* RIGHT MODELS: Random Effects Models for xtmixed: Preferred, but have
small degrees for estimating the random variances and therefore large
standard error. I'm not sure that all of these models can be fit! */
***********************************************************
*3. One level nesting--ignore that villages are nested *
* within city and ignore city-type village-type structure. *
***********************************************************
xi: xtmixed y g s gs || v_id:
estat ic
********************************************************
* 4. Two-level nesting: ignore city-type village-type *
* structure. This is the most conservative approach. *
********************************************************
xtmixed y g s gs || v_id: || cIty_id:
estat ic
***********************************************
* 5 Use nesting AND structure, if possible *
***********************************************
xi: xtmixed y g s gs i.c i.v || v_id: || c_id:
estat ic
************************************************************
* 6 Use nesting: interaction of city-type & intervention *
************************************************************
xi: xtmixed y g s gs i.v i.c i.gs*i.c || v_id: || c_id:
test gs _IgsXc_1_2 _IgsXc_1_3
estat ic
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
