Project

General

Profile

Support #1394

Follow up on having a “unified” weighting variable for analyzing

Added by Abigail Dumalus 4 months ago. Updated about 2 months ago.

Status:
Feedback
Priority:
Normal
Category:
Weights
Start date:
08/11/2020
% Done:

80%


Files

UserForum_Issue1394_public.do (1.03 KB) UserForum_Issue1394_public.do Alita Nandi, 10/14/2020 03:38 PM
#1

Updated by Abigail Dumalus 4 months ago

Hello Alita,

I wanted to follow up on my query regarding having a “unified” weighting variable for analysing interviews from 2002 to 2018. I remember you mentioned that people interviewed from 2002 to 2008 (BHPS) are not the same ones interviewed from 2009 to 2018 (UKHLS). You said that we are essentially looking at two different UK populations since migration has had an impact on its composition. Has there been a resolution on this?

#2

Updated by Alita Nandi 4 months ago

  • Status changed from New to Feedback
  • Assignee changed from Alita Nandi to Abigail Dumalus
  • % Done changed from 0 to 50

Hi Abigail,

As this is a longitudinal survey, the same sample members are interviewed every year. Part of the BHPS sample, after the BHPS ended in 2018, continued to be interviewed from onwards 2010 as part of Understanding Society. You can check the PIDP values and you will find many respondents interviewed during 2002-2008 were also interviewed from 2010 onwards. But what I meant was that if you used the longitudinal weights bw_indin01_lw then the weighted estimates will be representative of the 2001 UK population while if you used the longitudinal weights w_indinub_lw then the weighted estimates will be representative of the 2010 UK population.

Hope this helps.

Best wishes,
Alita

#3

Updated by Alita Nandi 4 months ago

  • % Done changed from 50 to 80
  • Private changed from Yes to No

Hi Abigail,

Last time we spoke you said you wanted to estimate models using fixed effects using data from all 18 waves of BHPS and 9 waves of UKHLS and wanted to know which set of weights to use. This is the response based on advice from our survey statistician.

As NI was not included in the BHPS until 2001, if you want to look at the entire period 1991-2018, you should exclude NI from all analysis. Otherwise you should analyse BHPS 1-11 and rest of the waves separately, where your estimates for the first analysis will be representative of GB, for the latter... UK.

Including NI
BHPS Waves 11-18, use br_indin01_lw
UKHLS Waves 1-2, use b_indinui_lw
UKHLS Waves 2-6, use f_indinub_lw
UKHLS Waves 6-9, use i_indinui_lw

Excluding NI
BHPS Waves 1-11, use bi_indin91_lw
BHPS Waves 11-18, use br_indin01_lw
UKHLS Waves 1-2, use b_indinui_lw
UKHLS Waves 2-6, use f_indinub_lw
UKHLS Waves 6-9, use i_indinui_lw

Note Stata does not allow svyset with xtreg but it allows weights and clustering.
xtreg depvar indepvars [pw=weight], fe vce(cluster psu)

Stratification can be ommitted - the CIs will be slightly conservative but otherwise correct. To include correct clustering you should have the following structure: wave information nested within pidp, then within hidp (unless longitudinal analysis is used; if so hidp should be omitted), and then within psu. You can omit any levels except for psu. Youcan add extra levels higher up too if you like. And the weight is of course should be appropriate to your analysis.

#4

Updated by Abigail Dumalus 4 months ago

Hello Alita,

I’ll take note of these helpful details. Would it be easier if I can attach the respective weights before merging the waves together in one longitudinal data file? Or can I program this in a syntax on Stata after merging? Waves 2 and 6 have multiple (i.e. more than one) weights, since they overlap?

UKHLS Waves 1-2, use b_indinui_lw
UKHLS Waves 2-6, use f_indinub_lw
UKHLS Waves 6-9, use i_indinui_lw

I still need to drop wave 1 of UKHLS if I’m analysing waves 11-18 of the BHPS along with waves 2-9 of UKHLS longitudinally speaking, correct? Thanks so much. I am sorry if I keep misunderstanding any of these.

#5

Updated by Alita Nandi 4 months ago

My suggestion would be to add the weights to the indresp files in before merging them. But remember to merge the relevant weight to all the waves it needs to be applied to. For example, attach br_indin01_lw to bl_indresp bm_indrep.... br_indresp.

#6

Updated by Abigail Dumalus 3 months ago

Hello Alita,

My understanding of your weighting assignment is that Waves 2 and 6 have multiple (i.e. more than one) weights, since they overlap (i.e. 1-2, 2-6, 6-9, seen below). Is this a wrong reading from my end?

UKHLS Waves 1-2, use b_indinui_lw
UKHLS Waves 2-6, use f_indinub_lw
UKHLS Waves 6-9, use i_indinui_lw

Isn’t i_indinui_lw the “combined UKHLS+BHPS+IEMB longitudinal adult main interview weight“ covering waves 7-9 of the UKHLS only? Isn’t f_indinub_lw the “longitudinal adult main interview weight” covering waves 3-9 of the UKHLS? I am basing these from the variable search results on the Understanding Society website.

I still need to drop wave 1 of UKHLS if I’m analysing waves 11-18 of the BHPS along with waves 2-9 of UKHLS longitudinally speaking. Wouldn’t it be easier to use i_indinus_lw for UKHLS Waves 2-9 as this is the “longitudinal adult main interview weight” covered in the indresp files for waves 2-9 of the UKHLS? Please let me know if I am on the right track.

As you have noted previously: if I used the longitudinal weights bw_indin01_lw then the weighted estimates will be representative of the 2001 UK population while if I used the longitudinal weights w_indinub_lw then the weighted estimates will be representative of the 2010 UK population. This tells me that it won’t make sense to compare these distinct UK populations over time from 2002 to 2018. Does this mean that it is impossible to follow those from the 2001 UK population during the period between 2010 and 2018 since there is an entirely different set of individuals who comprise the 2010 UK population?

Thanks for you time and consideration!

#7

Updated by Alita Nandi 2 months ago

  • Assignee changed from Abigail Dumalus to Alita Nandi
#8

Updated by Alita Nandi 2 months ago

  • Assignee changed from Alita Nandi to Olena Kaminska

Sorry there was a typo: UKHLS Waves 1-2, use b_indinus_lw

And you can use i_indinub_lw for Waves 2-9.

And you can use i_indin01_lw for BHPS Waves 11-18+UKHLS Waves 2-9 for the BHPS sample only.

#9

Updated by Olena Kaminska 2 months ago

Abigail,

Just to add, if you are concerned about the differences in the populations, then you can use full samples but restrict to those who lived in this country in 1991 (or were born to people who lived here in 1991) or 2001.

And just to clarify Alita's last comment you can't use BHPS with BHPS weight alongside UKHLS with UKHLS weight - this will be wrong. You either should use all the data available with 'ub' or 'ui' weight or restrict your analysis to BHPS data and use BHPS weights.

Hope this helps,
Olena

#10

Updated by Abigail Dumalus about 2 months ago

Hello Alita,

When I am restricting the analysis to Waves 11-18 of the BHPS, using indin01_lw as weighting variable results in zero weighted frequencies for life satisfaction. What am I misunderstanding here from what was recommended above? I have no problem with waves 2-9 of the UKHLS. Could it be that the respective xx_indin01_lw for waves 11-18 have been overwritten as missing when the indresp database files were merged? I was checking the data editor and found missing values for waves before wave 2 of the UKHLS. Kindly let me know what must I do to rectify this situation. Thanks so much.

Kind Regards,
Abigail

#11

Updated by Alita Nandi about 2 months ago

  • Assignee changed from Olena Kaminska to Abigail Dumalus

Hi Abigail,

It is not your mistake, it is mine. Very sorry. The names of the weight variables for BHPS have not yet been harmonised with the UKHLS naming conventions (it will be done at this release or the next). Once the names are harmonised the variable name I said will work. Basically for BHPS W11-18 the weight variable for longitudinal analysis is br_lrwtuk1 (this will be renamed to br_indin01_lw later).

Again sorry about that.
Best wishes,
Alita

#12

Updated by Abigail Dumalus about 2 months ago

Hello Alita,

If br_lrwtuk1 is the weight variable for BHPS W11-18 for longitudinal analysis, then I have to generate a new weight variable that “unifies” this with indin01_lw in the meantime, don’t I? As of now, using indin01_lw alone results in attrition of these 8 BHPS waves and the data starts from 2010 to 2018, covering UKHLS W2-9.

If I move back as far as 1991, does this mean that the respective xx_indin91_lw for BHPS waves before 2001 are also not yet harmonised with the indin91_lw variable? Can you please confirm that lrwght is the weight variable for longitudinal analysis for BHPS W2-18? If yes, does it make sense to “unify” this with indin91_lw variable for now?

#13

Updated by Alita Nandi about 2 months ago

  • Assignee changed from Abigail Dumalus to Olena Kaminska
#14

Updated by Olena Kaminska about 2 months ago

Abigail,

The data for indin01_lw starts with 2001, and the data for indin91_lw starts in 1991. While the variable names are not harmonized between UKHLS and BHPS yet (they will be in the upcoming release) the values are harmonized - so the weights have identical concept - at the moment they are just under different names.

Hope this helps,
Olena

#15

Updated by Abigail Dumalus about 2 months ago

Hello Olena,

Many thanks for your response. I tried generating a single variable to combine these two weighting variables. When I did, I used that to run the xtreg command with weighting but I keep getting this error: “weight must be constant within pidp”. I used “xtset pidp” before this regression, by the way. Please let me know what I should be doing, instead, so that the aforementioned error is avoided. What am I missing here?

#16

Updated by Alita Nandi about 2 months ago

  • Assignee changed from Olena Kaminska to Abigail Dumalus

Hello Abigail,

If you are using xtreg you can only use one weight for the entire set of waves - as it is the same set of people you are following over that period. So, if you are using xtreg for BHPS wave 1 to UKHLS W9 - then use i_indin91_lw and carry it over to all previous waves.But if you are using xtreg for analysing data from BHPS waves 1 -10, then use j_lrwtuk1 from BHPS W10 and apply that to all waves prior. Basically, you need to decide which sample you are going to analyse. You can analyse separate sets of samples - but as we discussed earlier you will need to analyse them separately.

#17

Updated by Abigail Dumalus about 2 months ago

Hello Alita,

I am using xtreg for W11-18 of BHPS along with W2-9 of UKHLS, so I am looking at the “harmonisation” of lrwtuk1 for the former and indin01_lw for the latter set of waves. I combined this under one variable but I keep getting the error above. Do I need to drop all weights variables first then merge these two weights with the longitudinal data files from BHPS W11-18 to UKHLS W2-9? I do understand that we are following the same people from 2001 to 2018. It is just frustrating that this xtreg error keeps emerging. Is there a better way around this challenge? Thanks so much for your consideration.

#18

Updated by Alita Nandi about 2 months ago

Option 1: Append individual level files (indresp) from W11 of BHPS until W18, and W2 UKHLS until W9 into a long format file. Then (many to one) merge i_indin01_lw from i_indresp onto this long file.

Option 2: Append individual level files (indresp) from W11 of BHPS until W18 into a long format file. Then (many to one ) merge br_lrwtuk1 from br_indresp onto this long file. Repeat the same with UKHLS W2-W9 UKHLS indresp files & i_indin01_lw from i_indresp. Analyse the two parts separately.

Preference is for Option 1.

#19

Updated by Abigail Dumalus about 2 months ago

Hello Alita,

For option 1, does this mean I can disregard the weight variable lrwtuk1 for BHPS W11-18 and simply focus on indin01_lw? As you mentioned before, lrwtuk1 corresponds with the aforementioned BHPS waves. Wouldn’t I be getting zero weights for these BHPS waves when using indin01_lw alone when merging from the latest indresp data file? My understanding is that I am analysing the same people from 2001 onwards. I mentioned in a previous update that I am getting zero weights for 2001 to 2008 when I look at indin01_lw. Is my understanding mistaken?

#20

Updated by Alita Nandi about 2 months ago

Hi Abigail - I am providing a Stata syntax file to show you what I mean.

#21

Updated by Abigail Dumalus about 2 months ago

Hello Alita,

Thanks so much for the syntax file that you sent. I may be misunderstanding what you said here:

“ The names of the weight variables for BHPS have not yet been harmonised with the UKHLS naming conventions (it will be done at this release or the next). Once the names are harmonised the variable name I said will work. Basically for BHPS W11-18 the weight variable for longitudinal analysis is br_lrwtuk1 (this will be renamed to br_indin01_lw later).”

My understanding is that using indin01_lw indicates that we are following the 2001 UK population until wave 9 of the UKHLS. How should I consider lrwtuk1 in relation to indin01_lw then? Is lrwtuk1 representative of another UK population? I apologise that I am still confused why using indin01_lw results in zero weights for the period covering W11-18 of BHPS. I am deeply sorry for misunderstanding this one bit.

#22

Updated by Alita Nandi about 2 months ago

No same population. This is how longitudinal weights in this survey are computed: Iniital wave cross sectional weight is multiplied by non-response weight between two consecutive waves. So, br_lrwtuk1= bk cross-sectional weight X non-response weight to correct for non-response between Wave 11 & 12 X non-response weight to correct for non-response between Wave 12 & 13.....non-response weight to correct for non-response between Wave 17 & 18

So, if you were to use the 2001 BHPS sample and
- follow them until Wave 18, use br_lrwtuk1
- follow them until UKHLS W9 then use i_indin01_lw
- follow them until BHPS Wave 12, use bl_lrwtuk1
- follow them until UKHLS W2, use b_indin01_lw
.....

Also available in: Atom PDF