The DHS Program User Forum: Dataset use in Stata

Home » Data » Dataset use in Stata » Dropping observations

Show: Today's Messages :: Show Polls :: Message Navigator

Dropping observations [message #10923]

Thu, 06 October 2016 19:00

shayankhan
Messages: 16
Registered: September 2016
Location: Karachi, Pakistan

Member

I want to restrict my analysis to children under the age of 5 years using the Pakistan DHS of 2012-13. So, I am using the command of drop to just keep the observations of children that are under 5 years of age and am doing my analysis on them. Is this the valid approach?

Report message to a moderator

Re: Dropping observations [message #10924 is a reply to message #10923]

Fri, 07 October 2016 07:38

Bridgette-DHS
Messages: 3190
Registered: February 2013

Senior Member

Following is a response from Senior DHS Stata Specialist, Tom Pullum:

Quote:

Yes, that's valid. But if you are using the KR file, which contains most of the child-level variables, that file is already limited to children born in the past five years (both alive and dead), so you don't need to do anything.

Report message to a moderator

Re: Dropping observations [message #10925 is a reply to message #10924]

Fri, 07 October 2016 07:45

shayankhan
Messages: 16
Registered: September 2016
Location: Karachi, Pakistan

Member

Yeah, I guess I took the long route on that. But, now using the KR file. I will now just merge the variables that I need from HR and IR files to KR file and use that for my analysis. I'd appreciate if you could also let me know the consequences of not using svy commands in my analysis. I mean I know it is important to use it for representativeness but I would like to know that is there any situation where not using svy command is justified?

Thanks.

Report message to a moderator

Re: Dropping observations [message #10930 is a reply to message #10925]

Sun, 09 October 2016 19:52

Trevor-DHS
Messages: 802
Registered: January 2013

Senior Member

You can not use svy if all you want is a point estimate. If you want confidence intervals or significance estimates then you need to use the svy commands. The point estimate will be the same whether you use the svy commands or not, but confidence intervals and tests of significance will be wrong if you do not use the svy commands

Report message to a moderator

Re: Dropping observations [message #10931 is a reply to message #10930]

Mon, 10 October 2016 04:08

shayankhan
Messages: 16
Registered: September 2016
Location: Karachi, Pakistan

Member

Hi Trevor,

What do you mean by point estimate? Are you talking about descriptive statistics? I did my logit model both by weights and unweighted and the significance differs as well as the coefficients.

Report message to a moderator

Re: Dropping observations [message #10933 is a reply to message #10931]

Mon, 10 October 2016 07:29

Bridgette-DHS
Messages: 3190
Registered: February 2013

Senior Member

Another response from Tom Pullum:

You can omit the svy correction for data quality checks, for confirming the construction of recodes, and for a preliminary or tentative analysis. However, your final results should always include the svy adjustments. The weight adjustment is most important. The cluster and stratum adjustments will only affect the standard errors and confidence intervals. They will not affect cross tabs or estimates of proportions, means, or rates. But if you want to construct confidence intervals or do tests, you really must use the cluster and stratum adjustments.

Before you try to merge IR and HR data onto the KR file, you should check which variables you need. Almost all of the relevant characteristics of the mother (from the IR file) are already included on the child's record. There may be some HR variables worth merging, but first check whether they are already in the KR file.

Report message to a moderator

Re: Dropping observations [message #10934 is a reply to message #10933]

Mon, 10 October 2016 07:36

shayankhan
Messages: 16
Registered: September 2016
Location: Karachi, Pakistan

Member

Thanks Bridgette. Yes, I got that much that I have to use weights with the survey command before any analysis. There is a recent paper by Woolrdidge by the title "What are we weighting for" in which the authors also mention exceptions for not using weights in regressions. The best way they say is to mention both the results, weighted as well as unweighted so that the reader can judge the difference in estimates.

What I don't understand though is that what is the difference between number of observations and population size. For example, when i run the svy command, i get number of observations and also the population size and both values are different. Am I to anlayze descriptives using observation size or population size?

Report message to a moderator

Re: Dropping observations [message #10938 is a reply to message #10934]

Mon, 10 October 2016 08:19

Bridgette-DHS
Messages: 3190
Registered: February 2013

Senior Member

Another response from Tom Pullum:

Quote:

For every parameter there is a point estimate and an interval estimate. I would not describe the point estimate as a descriptive statistic but you can if you want. I have nothing to add to what I said before.

Report message to a moderator

Re: Dropping observations [message #10939 is a reply to message #10938]

Mon, 10 October 2016 09:04

shayankhan
Messages: 16
Registered: September 2016
Location: Karachi, Pakistan

Member

Thanks a lot Tom.

Report message to a moderator

Previous Topic:	Performing a manual backward stepwise logistic regression in Stata
Next Topic:	svy command and descriptives

Goto Forum:

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Sat Nov 9 13:11:09 Coordinated Universal Time 2024