The DHS Program User Forum
Discussions regarding The DHS Program data and results
Home » Data » Weighting data » Weighting data after merging survey rounds with different levels of representation
Weighting data after merging survey rounds with different levels of representation [message #11566] Wed, 11 January 2017 01:11 Go to next message
jswindle is currently offline  jswindle
Messages: 5
Registered: September 2016
Member
Hello DHS experts and other forum users,

My problem relates to weighting after merging survey rounds.

I have merged the IR for the Malawi 2000, 2004, and 2010 surveys. When I weight this data, I am unsure how to best create the strata given different levels of representation across the surveys. The 2000 and 2004 surveys were representative across the 10 largest districts and across the three regions. The 2010 survey was representative across all 27 districts, as well as across the three regions.

I am combining these survey data with district level, time-varying data for foreign aid.

Currently I am doing the following, but am unsure about the commands. I am especially looking for guidance about the fifth through seventh lines below, which begin with "egen mw_00_strata..."

Thank you.

* Weight the dataset
generate weight = v005/10000000
recode survey (2000=1) (2004=2) (2010=3)
egen clusters=group(survey v021), label
egen mw_00_strata = group(survey region urban), label
egen mw_04_strata = group(survey region urban), label
egen mw_10_strata = group(survey region district urban), label
gen strata = .
replace strata = mw_00_strata if year==2000
replace strata = mw_04_strata if year==2004
replace strata = mw_10_strata if year==2010
svyset clusters [pweight=weight], strata(strata) singleunit(centered)
Re: Weighting data after merging survey rounds with different levels of representation [message #11575 is a reply to message #11566] Wed, 11 January 2017 18:56 Go to previous messageGo to next message
Bridgette-DHS is currently offline  Bridgette-DHS
Messages: 3017
Registered: February 2013
Senior Member
Following is a response from Senior DHS Stata Specialist, Tom Pullum:

Unfortunately, the stratification variable is often incorrect or labelled incorrectly in surveys conducted before about 2010. The following lines will work if you change the path.

set more off
set maxvar 10000

cd e:\DHS\DHS_data\IR_files

use MWIR41FL.dta, clear
gen survey=1
append using MWIR4DFL.dta
replace survey=2 if survey==.
append using MWIR61FL.dta
replace survey=3 if survey==.

* The strata in MW41 are given by s006
* The strata in MW4D are given by group(sdist v025)
* The strata in MW61 are given by v022

gen mw_00_strata = s006
egen mw_04_strata = group(sdist v025), label
gen mw_10_strata = v022

gen strata_temp=.
replace strata_temp=mw_00_strata if survey==1
replace strata_temp=mw_04_strata if survey==2
replace strata_temp=mw_10_strata if survey==3

egen strata=group(survey strata_temp)

tab strata survey, table clean
Re: Weighting data after merging survey rounds with different levels of representation [message #11576 is a reply to message #11575] Wed, 11 January 2017 21:20 Go to previous messageGo to next message
jswindle is currently offline  jswindle
Messages: 5
Registered: September 2016
Member
Hello Tom and Bridgett,

Thank you for your very helpful and prompt reply.

I ran the code you shared and got some results that surprised me. When I ran the final command of "tab strata survey, table clean" I got an error message saying that I could not use those options. When I instead ran "tab strata survey", I got these interesting results:


tab strata survey

group(surv
ey
strata_tem survey
p) 1 2 3 Total

1 226 0 0 226
2 416 0 0 416
3 607 0 0 607
4 320 0 0 320
5 17 0 0 17
6 253 0 0 253
7 190 0 0 190
8 470 0 0 470
9 27 0 0 27
10 447 0 0 447
11 767 0 0 767
12 174 0 0 174
13 556 0 0 556
14 172 0 0 172
15 438 0 0 438
16 433 0 0 433
17 611 0 0 611
18 187 0 0 187
19 459 0 0 459
20 195 0 0 195
21 340 0 0 340
22 800 0 0 800
23 105 0 0 105
24 121 0 0 121
25 450 0 0 450
26 331 0 0 331
27 186 0 0 186
28 193 0 0 193
29 28 0 0 28
30 188 0 0 188
31 435 0 0 435
32 185 0 0 185
33 239 0 0 239
34 67 0 0 67
35 22 0 0 22
36 591 0 0 591
37 193 0 0 193
38 787 0 0 787
39 95 0 0 95
40 614 0 0 614
41 285 0 0 285
42 0 420 0 420
43 0 283 0 283
44 0 47 0 47
45 0 850 0 850
46 0 40 0 40
47 0 732 0 732
48 0 81 0 81
49 0 693 0 693
50 0 263 0 263
51 0 690 0 690
52 0 78 0 78
53 0 625 0 625
54 0 31 0 31
55 0 789 0 789
56 0 101 0 101
57 0 705 0 705
58 0 307 0 307
59 0 403 0 403
60 0 42 0 42
61 0 735 0 735
62 0 230 0 230
63 0 3,553 0 3,553
64 0 0 92 92
65 0 0 754 754
66 0 0 825 825
67 0 0 318 318
68 0 0 33 33
69 0 0 789 789
70 0 0 35 35
71 0 0 786 786
72 0 0 60 60
73 0 0 718 718
74 0 0 45 45
75 0 0 821 821
76 0 0 32 32
77 0 0 781 781
78 0 0 138 138
79 0 0 650 650
80 0 0 76 76
81 0 0 832 832
82 0 0 480 480
83 0 0 646 646
84 0 0 53 53
85 0 0 723 723
86 0 0 55 55
87 0 0 746 746
88 0 0 44 44
89 0 0 786 786
90 0 0 41 41
91 0 0 823 823
92 0 0 127 127
93 0 0 668 668
94 0 0 197 197
95 0 0 755 755
96 0 0 29 29
97 0 0 706 706
98 0 0 42 42
99 0 0 778 778
100 0 0 70 70
101 0 0 747 747
102 0 0 81 81
103 0 0 737 737
104 0 0 63 63
105 0 0 831 831
106 0 0 37 37
107 0 0 782 782
108 0 0 35 35
109 0 0 767 767
110 0 0 90 90
111 0 0 761 761
112 0 0 66 66
113 0 0 723 723
114 0 0 85 85
115 0 0 778 778
116 0 0 137 137
117 0 0 746 746

Total 13,220 11,698 23,020 47,938


The part of these results that I found surprising is that the number of strata per survey vary in strange way. There are 41 categories for 2000, 22 categories for 2004, and 54 categories for 2010. The result for 2010 makes sense; there were 27 districts and when stratified by urban/rural you get 54. The result for 2004, I believe comes from 11 districts categories stratified by urban/rural; those 11 district categories are the ten largest districts that were sampled in a representative manner and then there is one big catch-all for the other 17 districts, hence the huge total of 3,553 respondents in the catch-all rural category (at least that is my guess). The 2000 results are perplexing. From what I can gather in the final report for the 2000 Malawi DHS, the sampling was done in the same manner as the 2004 survey, so I'm not sure why there are 41 categories here. Thoughts?

Once I have calculate the strata correctly, would the rest of this code (pasted below) work to appropriately survey set the data?

generate weight = v005/10000000
egen clusters=group(survey v021), label
svyset clusters [pweight=weight], strata(strata) singleunit(centered)

Or would you simply do:

generate weight = v005/10000000
svyset [pweight=weight], psu(v021) strata(strata)

In case it is relevant for deciding how to svyset the data, my ultimate goal is to do a three-level mixed effects model with the higher orders being the districtyear and district variables.

A final issue I am facing if I do this sort of mixed effects model is whether the 2000 and 2004 data from the 17 districts that are not sampled sufficiently to be representative could be appropriately incorporated into such a model. I realize that is outside the purvue of the DHS surveys, but I'm guessing you have faced these types of issue before in your own research. Any thoughts?

thank you kindly,
Jeff
Re: Weighting data after merging survey rounds with different levels of representation [message #11593 is a reply to message #11576] Fri, 13 January 2017 11:45 Go to previous message
Bridgette-DHS is currently offline  Bridgette-DHS
Messages: 3017
Registered: February 2013
Senior Member
Following is a response from Senior DHS Stata Specialist, Tom Pullum:

Sorry about that mistake--"table clean" is an option with "list" and not with "tab". Don't know why I added that--it definitely would not run.

You definitely need something like "egen clusters=group(survey v021)" to get unique identifiers for the clusters across the three surveys and the strata across the three surveys.

The change you observe in the number of strata from one survey to the next is not implausible (although usually the definitions are the same from one survey to the next.) In general, the number of strata is in the range of 20 to 60. I suggest you look at the report. I HOPE it will confirm what I passed on to you.

DHS estimates at the stratum level are always representative, in terms of being unbiased. It is true that some older documentation mentioned a lack of representativeness, but that actually refers just to higher standard errors when there are fewer cases. That can be an issue, but bias is NOT an issue. The sampling is designed so that small strata tend to be over-sampled (large strata correspondingly tend to be under-sampled) in order to get more stable estimates.

The generic term we use for the first national sub-division is "region". The generic term for the second level is "district". In general the strata are the combinations of region and urban/rural.
Previous Topic: Post-stratification for DHS data
Next Topic: Disaggregating to lower Administrative Division
Goto Forum:
  


Current Time: Thu Mar 28 14:08:29 Coordinated Universal Time 2024