Strata omitted because they contain no subpopulation members.

[Yawo Kokuvi]

Hello, Folks,

I have posted here before regarding this subject. To recap, I am using the National Health Interview Survey (NHIS) to fit an SEM model. I restricted my analysis to a sub-population (called afim1, with about 1700 unweighted cases - see data extract below). Results from my analysis continually state that a lot of strata (over 200) were "omitted because they contain no subpopulation members". I used svydes if afim1==1 to check the extent of singleton units in the data, and part of the results are posted below:

Survey: Describing stage 1 sampling units

      pweight: sampwt1
          VCE: linearized
  Single unit: scaled
     Strata 1: strata
         SU 1: psu
        FPC 1: <zero>

                                      #Obs per Unit
                              ----------------------------
Stratum    #Units     #Obs      min       mean      max   
--------  --------  --------  --------  --------  --------
    5002         1*        1         1       1.0         1
    5007         1*        5         5       5.0         5
    5008         2         3         1       1.5         2
    5009         1*        2         2       2.0         2
    5011         2         8         4       4.0         4
    5012         2         5         1       2.5         4
    5013         2         9         1       4.5         8
    5014         1*        1         1       1.0         1
    5016         1*        1         1       1.0         1
    5017         1*        1         1       1.0         1
    5020         2         5         1       2.5         4
    5021         2         2         1       1.0         1
    5022         2         4         2       2.0         2
    5024         2         3         1       1.5         2
    5027         2        24         7      12.0        17
    5028         1*        1         1       1.0         1
    5030         1*        1         1       1.0         1
    5031         2         2         1       1.0         1
    5033         1*        1         1       1.0         1
    5034         1*        3         3       3.0         3

    

    6261         1*        1         1       1.0         1
    6262         1*        2         2       2.0         2
    6264         2        15         5       7.5        10
    6265         2        12         4       6.0         8
    6266         2         2         1       1.0         1
    6267         1*        2         2       2.0         2
    6268         1*        1         1       1.0         1
    6271         1*        6         6       6.0         6
    6272         2        11         4       5.5         7
    6273         1*        1         1       1.0         1
    6274         1*        1         1       1.0         1
    6275         1*        3         3       3.0         3
    6276         2         4         1       2.0         3
    6278         2         3         1       1.5         2
    6279         1*        2         2       2.0         2
    6281         1*        2         2       2.0         2
    6283         1*        2         2       2.0         2
    6284         1*        1         1       1.0         1
    6285         1*        2         2       2.0         2
    6286         1*        1         1       1.0         1
    6287         1*        1         1       1.0         1
    6289         1*        2         2       2.0         2
    6290         1*        1         1       1.0         1
    6292         1*        1         1       1.0         1
    6293         1*        1         1       1.0         1
--------  --------  --------  --------  --------  --------
     441       644     1,720         1       2.7        25

                     312,967 = #Obs with missing values in the 
                    --------   survey characteristics
                     314,687

I realize that there are a number of PSU's with few observations (the ones with 1*). I know others have suggested a number of options including modifying the svy singleunit option, but none of this have worked. I need to focus my analysis on this subpopulation, hence I thinking of a way to combine some of the singleton strata with adjoining ones? I've produced an extract of the dataset with relevant variables - strata, psu etc below. So my question is wow wii I got about doing this merging adjoining single psu with others?

Sorry for the long post - thanks - cY

scratch of the dataset:

* Example generated by -dataex-. To install: ssc install dataex
clear
input int year byte quarter int(strata psu) byte region double nhispid byte hivtest float sampwt1
2000 1 5292 1 2 20000001890102 1  727.1
2000 1 5105 1 2 20000001940101 1  269.9
2000 1 5183 2 4 20000002250101 0   1592
2000 1 5069 2 3 20000004150102 1  832.2
2000 1 5168 1 2 20000007320101 0  523.4
2000 1 5088 2 4 20000007380102 0  534.3
2000 1 5124 2 3 20000018940101 1 1775.6
2000 1 5272 1 2 20000019750102 1  922.6
2000 1 5173 1 3 20000020010102 0  760.4
2000 1 5212 2 3 20000022420102 1  701.2
2000 1 5021 1 3 20000023270101 0  211.2
2000 1 5207 1 1 20000035740101 0  539.7
2000 1 5089 1 3 20000037490101 1  295.9
2000 1 5159 1 2 20000039530101 1  549.2
2000 1 5266 1 3 20000039840101 0  387.7
2000 1 5179 1 1 20000067460101 0    553
2000 1 5027 1 1 20000067500101 0  272.5
2000 1 5061 1 1 20000079580101 .  431.9
2000 1 5105 1 2 20000081530101 1  443.1
2000 1 5118 2 3 20000086190101 1  709.6
2000 1 5166 2 4 20000087520101 1  330.8
2000 1 5049 2 1 20000096220101 0    292
2000 1 5286 1 3 20000100210101 .  887.8
2000 1 5219 1 3 20000102900102 1  892.9
2000 1 5147 1 4 20000113690102 0  419.5
2000 1 5084 2 1 20000119200101 1 1163.5
2000 1 5187 1 1 20000129100101 1  620.2
2000 1 5256 1 2 20000129850102 0  847.3
2000 1 5255 2 2 20000129900101 1  478.5
2000 1 5255 2 2 20000129900202 1  478.5
2000 1 5283 2 3 20000130740101 0  907.9
2000 1 5047 1 1 20000133990101 .    246
2000 1 5314 1 3 20000147660101 1  633.6
2000 1 5186 1 1 20000149220101 1    538
2000 2 5260 2 4 20000161220101 1  239.8
2000 2 5117 1 1 20000162840102 1    708
2000 2 5126 1 1 20000168750202 1  331.2
2000 2 5040 1 2 20000169250202 0  193.8
2000 2 5011 1 3 20000180850101 . 1594.8
2000 2 5163 1 4 20000182660101 1    211
2000 2 5107 1 4 20000197770101 . 1006.4
2000 2 5116 1 1 20000200800102 1 1124.9
2000 2 5219 1 3 20000204440101 1  582.6
2000 2 5011 1 3 20000218700102 0  661.9
2000 2 5108 1 3 20000219900101 1    375
2000 2 5214 2 1 20000220760102 1  667.7
2000 2 5253 2 3 20000220770101 0  535.5
2000 2 5242 1 3 20000235040101 0  357.9
2000 2 5177 1 1 20000241600101 1    419
2000 2 5219 2 3 20000243240101 1  299.1
2000 2 5247 2 1 20000254700101 .  325.7
2000 2 5119 1 3 20000256260102 1  430.9
2000 2 5155 2 4 20000257490103 1  490.1
2000 2 5017 2 3 20000262050102 0  617.9
2000 2 5034 1 2 20000275620102 0  736.6
2000 2 5243 1 2 20000276300203 0  353.4
2000 2 5315 1 3 20000278050101 1  352.1
2000 2 5207 1 1 20000280380101 0  554.6
2000 2 5207 1 1 20000296170101 1  286.1
2000 2 5226 2 2 20000297600101 1  464.6
2000 2 5073 1 3 20000300320101 0  556.9
2000 2 5169 1 1 20000312550101 0  350.3
2000 2 5034 1 2 20000313190101 1  843.3
2000 3 5073 2 3 20000348130101 1  255.4
2000 3 5008 2 4 20000351460101 0  310.1
2000 3 5277 2 4 20000352650102 0  336.6
2000 3 5027 2 1 20000352960104 0  637.8
2000 3 5103 2 3 20000368190102 .  381.3
2000 3 5293 2 2 20000370910101 1  622.4
2000 3 5260 2 4 20000372120101 1  325.5
2000 3 5059 1 4 20000372610101 1  722.4
2000 3 5214 2 1 20000377200101 1  730.5
2000 3 5073 1 3 20000392570101 0  555.4
2000 3 5066 2 2 20000395560101 1    616
2000 3 5020 2 4 20000407840101 1  206.4
2000 3 5160 2 1 20000408940101 0  299.6
2000 3 5207 1 1 20000410710101 1  560.8
2000 3 5187 1 1 20000410800103 1  601.9
2000 3 5322 1 3 20000411280102 1  600.9
2000 3 5195 1 4 20000415230101 0  413.4
2000 3 5102 1 3 20000415390101 1 1781.5
2000 3 5042 2 3 20000415960101 1    377
2000 3 5071 2 4 20000428620101 1  588.1
2000 3 5073 2 3 20000429040101 1  392.2
2000 3 5082 1 4 20000430830101 1  349.8
2000 3 5289 2 1 20000431010104 0 1931.7
2000 3 5239 2 3 20000432690106 1   2670
2000 3 5027 2 1 20000433890101 .  288.9
2000 3 5085 2 3 20000435480101 1  504.6
2000 3 5059 1 4 20000446080101 .  294.6
2000 3 5293 2 2 20000448520101 1  423.3
2000 3 5207 1 1 20000449450102 1  433.2
2000 3 5216 1 1 20000452410101 1  337.5
2000 3 5068 1 1 20000452450102 0  706.7
2000 3 5160 1 1 20000452460101 1  358.5
2000 3 5066 2 2 20000453240102 0  649.9
2000 3 5177 2 1 20000453640101 0    848
2000 3 5270 2 3 20000454010101 0  481.1
2000 3 5040 1 2 20000454490101 1  738.7
2000 3 5177 2 1 20000473700101 0  682.1
end
label values quarter QUARTER
label def QUARTER 1 "Quarter 1", modify
label def QUARTER 2 "Quarter 2", modify
label def QUARTER 3 "Quarter 3", modify
label values region REGION
label def REGION 1 "Northeast", modify
label def REGION 2 "North Central/Midwest", modify
label def REGION 3 "South", modify
label def REGION 4 "West", modify
label values hivtest yesno_10
label def yesno_10 0 "No", modify
label def yesno_10 1 "Yes", modify

[Steve Samuels]

The strata with neighboring numbers are not actually geographic neighbors, so merging them has little justification. The Variance Estimation Guidance, NHIS 2016 , page 3, has this to say about the stratum and PSU variables (PSTRAT and PPSU):

As discussed above, in order to mask true geographical locations, the PSTRAT and PPSU levels are pseudo-levels or simplified versions of the true NHIS sample design variables. Analysts are cautioned that these simplified design structures are not designed to support geographical analyses below the Census Region level.

To address the single-unit problem, Stata provides the singleunit() option in svyset. I would choose singleunit(scaled), but you might try the others to see how how much of a difference the choice makes. (Opinions differ: see this recent post.) . Of course you will also need the subpop() option to restrict analyses to the "afm1" subpopulation.

[Yawo Kokuvi]

thanks very much - I will try that. best - cY