This course will cover the steps used in weighting sample surveys, including methods for adjusting for nonresponse and using data external to the survey for calibration. Among the techniques discussed are adjustments using estimated response propensities, poststratification, raking, and general regression estimation. Alternative techniques for imputing values for missing items will be discussed. For both weighting and imputation, the capabilities of different statistical software packages will be covered, including R®, Stata®, and SAS®.
Nonresponse Adjustments
Loading...
来自 University of Maryland, College Park 的课程
Dealing With Missing Data
77 个评分
Explore Related Videos
At Coursera, you will find the best lectures in the world. Here are some of our personalized recommendations for you
University of Maryland, College Park
77 个评分
课程 5（共 7 门，Specialization Survey Data Collection and Analytics ）
从本节课中
Implementing the Steps
Software is critical to implementing the steps, but the R system is an excellent source of free routines. This module covers several R packages, including sampling, survey, and PracTools that will select samples and compute weights.
与讲师见面
Richard Valliant, Ph.D.Research Professor
Joint Program in Survey Methodology
In this video, let's talk about making non-response adjustments
using response propensity classes.
So the idea here is we've summarized before is estimate
response propensities, and I'll use logistic regression in this case to do it.
You sort the propensities from low to high, and then you divide
the sample into classes with about the same number of cases per class.
All the sample cases, but as the respondents and
non-respondents are classified.
And a critical thing here is equal size classes means
the same total of R + NR is in each class roughly, so
that the same total count of initial sample units.
Now it definitely does not mean the same number of respondents and
non respondents were in each class because we're trying to
create classes where the response propensity varies.
It's fairly different across the classes.
It's just the grand total count of initial sample units,
respondents, plus non-respondents that should be about the same.
So then we adjust the weight for
each respondents in a class and multiplied the base weight or
whatever the incoming weight is by this non-response adjustments.
So we can apply this to probability samples, but
you could use the same sort of algorithm to compute
quasi-randomization weights in a non-probability sample.
So, keep that in mind as we go through the details here.
Now, how will I do this?
I'm going to use the PracTools package, so I require the package, right here,
and I make the data available by saying, data (nhis).
So, this is the same data set that we used in an earlier example,
and now I issue the pclass function call.
Pclass is a function in PracTools, ao it takes a few parameters.
The first one is the formula for the model.
So in this case I'm regressing RESP,
which is a zero one variable for response or not.
On age factor, and I've coerced that to sex,
and I've coerced that to be a categorical factor.
Hispanic, whether you're Hispanic or not coerced to be a categorical variable and
then the variable race, also coerced to be categorical.
The data set is nhis, and I'm running an unweighted regression.
If your input data set has got survey weights or
some sort of input weights you can say weighted here,
and it'll do weighted least or weighted logistic regression fit.
And then the particular link that's being used is legit, so
that says give me a logistic regression.
You can also ask for complimentary log log, or probe it if you like those better,
and then you tell it the function how many clients do I want to create, so
I created five here, you can do more if you like.
Five's certainly, I'd say a minimum number to create.
All right so we save that in the object called out, and
you better look at what you did, so I'm doing a table on out.
P.class is one of the variables or
components of the output p.class.
So, I'm just counting up how many cases were assigned to each propensity class,
and I've got five of them.
So, the table gives you the range of propensities that are in a class,
and the number of cases in each class.
So you can see the range spreads things out pretty well, the algorithm did.
The count is not quite the same, but similar in classes.
There's some, because of the discreteness of the data,
you can't make these exactly equal, and then critically
use this useNA equals always to make sure you got everything assigned to a class.
And sure enough we've got zero missing values, so that's what we want to see.
You can use the summary function on another piece of the output,
which is there's a component called propensities.
And that's within the output object out.
So these go from a minimum of 0.4769 up to a maximum of 0.7765.
So there's some spread the propensities,
which means the classes should have different response rates within them.
Now just to examine things further, I've done a box plot on these propensities,
and I'm doing it by p.class to see what those things look like.
So here's my box plot, and
what you see here is I've got five boxes for the five propensity classes.
The higher four or the higher response rates,
response propensities are pretty tight in terms of the spread.
The distance between first and
third quartile is pretty short here on these boxes.
What you do see in the lower category
is a stream of outliers down here.
So, these go down to somewhat below 0.5.
This is the 0.47 number, that minimum down here,
and they go up beyond 0.65.
So, there's a quite a range in this lower category, more so
than in the other propensity classes.
And this is one reason why people gravitate toward a single value adjustment
for within a propensity class rather than using the individual propensities,
because they don't really trust the model as being that perfect.
So you might in this picture use the medium,
which is this black line in each box as your single adjustment.
But there are some alternatives, which I'll show you in a second.
So we could within a propensity class,
do any of several things.
We could do the things listed here.
You could use the unweighted mean response propensity from the model.
You could use a weighted mean response propensity from a model,
with the weights being the base weights or the input weights.
You could use a median, the thing I pointed to in the box plots.
You could use the unweighted response rate.
Within a propensity class,
just the number of respondents divided by the grand total sample size.
You could use a weighted response rate with the base weights.
So you pick one of those, and then you multiply your input weights by the inverse
of whichever one of those you pick out.
Rather than using individual propensities, which as I said,
it seems chancier.
If your model is not that great a fit, then it makes sense to
limit the number of different non response adjustments you're going to use.
So here's some code just to look at alternatives.
What I've down here is I've used the c bind function
to put things together in columns just so it's easy to look at them, and
then I've used the round function to round these things off.
Here's the parameter on round to three places.
So what have we got inside?
This label mean is used by function.
I operate on the propensities.
My indexes for
the by classes are p.class, and the function I'm using is mean.
So one comment just on style, how you do these things in R.
If you've got a function that takes a lot of parameters, I'd recommend actually
putting the parameter name in here followed by equal in value.
So, data indices fun to remind you what those what things stand for.
You can do it without those, I can omit data equals,
I can omit indices equal, I can omit fun equal, and it would work fine.
But, if you forget what order of the parameters have to be in,
then you can really get confused.
So I'd recommend this.
This is a documentation.
Put the actual name of the parameter on the left side of the equal sign followed
by the name in your dataset that you're filling in for that parameter value.
All right, so I do the same thing for the median, and I label that median.
And then in the by function, here I'm calling
the fun equals, and I left out fun equals here.
I'm defining the function, which is going to operate on
this placeholder variable x is actually a built in function called weighted.mean,
so you have to tell weighted.mean what it's going to handle here.
So I'm using the risk value, which is 01 for response.
The weighted value, which is the survey weight.
So this is going to give you a weighted response rate, and
once my input data, the by needs a data frame,
so I'm extracting just the one column,
resp from nhis and just to make sure that by
understands this is a data frame coerce
to be data frame with data.frame.
And then the next parameter I send it is in the data frame.
I bind in the survey weight variable, svyweight, which is on that NHIS dataset.
These are the indices here, out$p.class, and so you see I violated my own
advice by not putting in the cz equal in front of this thing.
So that's weighted response rate, and here's what comes out of that.
I've got in this column mean, in this column median,
in this column weighted response rate, and I've got those separately for
each of the five response propensity classes.
And you scan across here there's some difference, but
they're not tremendous, so the biggest
difference is in that first one that had a string of outliers in there.
So you might in this case pick
maybe the median response rate, and it certainly be
a intermediate choice between weighted response rate in the mean,
at least in that first response propensity class.
But, that, it's partly a matter of taste.
There is some theory, or some empirical evidence that says
you might just want to use the underweighted response rate.
But, it's a choice that you make as an analyst, so
that covers what I wanted to say about non-response adjustments in
the non-response software that's available to do it.
The other algorithm I mentioned earlier was regression trees, so
I haven't given you a numerical or software example of that.
But that's certainly still good option, and you should keep that in mind.
