Skip to main content

Adjusting with Weights... or Adjusting the Data Collection?

I just got back from JSM where I saw some presentations on responsive/adaptive design. The discussant did a great job summarizing the issues. He raised one of the key questions that always seems to come up for these kinds of designs: If you have those data available for all the cases, why bother changing the data collection when you can just use nonresponse adjustments to account for differences along those dimensions?

This is a big question for these methods. I think there are at least two responses (let me know if you have others).

First, in order for those nonresponse adjustments to be effective, and assuming that we will use weighting cell adjustments (the idea extends easily to propensity modeling), the respondents within any cell need to be equivalent to a random sample of the cell. That is, the respondents and nonrespondents need to have the same mean for the survey variable. A question might be, at what point does that assumption become true? Of course, we don't know. But if we alter our data collection strategy, we will at least strive to empirically verify that over some range of response rates.

Second, this is an empirical question. It would be nice to have studies that looked at this question. Does balancing response along specified dimensions lead to reduced nonresponse bias after adjustment? My hunch is "yes, it does."

Comments

  1. Hi James,
    Not an expert on this, but wouldn't you say that with nonresponse adjustments, you expect the transformation to be linear (or at least parametric), whereas if you adjust the data collection, adjustments can take any form. I agree with you that this may or may not matter in practice, and that that's an empirical question.

    ReplyDelete
  2. Peter, Thanks for the interesting comment. It seems to me that when we make our nonresponse adjustments, after all is said and done, we postulate that our model is correct and that the adjusted measures are unbiased (or as unbiased as can be given the available data). Normally, this assumption can't be tested. If there is an unobserved covariate that is making that assumption invalid, we can't know it.

    When we are collecting data, it is as if we assume that we are breaking (or maybe reducing) the correlation between that covariate and response, conditional on all the observed data.

    My hunch is that trying to break the correlation during data collection is worthwhile. Sadly, in most cases, we may never know. Today we lean pretty heavily on the Pew studies first reported in 2000 as evidence that response rate may not be a good indicator. It would be nice to have similar empirical studies on this question.

    ReplyDelete
  3. I entirely agree, both about reducing the potential for bias during data collection and about the unfortunate lack of evidence.

    To me, the argument for reducing rather than only adjusting using the same available information is one about robustness of the design with respect to nonresponse bias. Relying only on adjustments leaves greater potential for nonresponse bias due to heterogeniety within adjustment cells/groups. Doing something to equalize response rates across cells should reduce this risk. I think this is the same as what you said, stated in a slightly different way.

    One of the challenges with the empirical evidence is that there may not be substantial bias in weighted estimates to begin with, in which case one would erroneously conclude that only adjusting is fine. Yet it is about robustness - reducing the risk of bias, when there is bias...

    ReplyDelete
    Replies
    1. I think you are right that it is more robust to control the data collection in this way.

      I also think it may be the case that we can reduce the variability induced by interviewers by giving more centralized direction. Just conjecture...

      Delete

Post a Comment

Popular posts from this blog

The Cost of a Call Attempt

We recently did an experiment with incentives on a face-to-face survey. As one aspect of the evaluation of the experiment, we looked at the costs associated with each treatment (i.e. different incentive amounts).

The costs are a bit complicated to parse out. The incentive amount is easy, but the interviewer time is hard. Interviewers record their time for at the day level, not at the housing unit level. So it's difficult to determine how much a call attempt costs.

Even if we had accurate data on the time spent making the call attempt, there would still be all the travel time from the interviewer's home to the area segment. If I could accurately calculate that, how would I spread it across the cost of call attempts? This might not matter if all I'm interested in is calculating the marginal cost of adding an attempt to a visit to an area segment. But if I want to evaluate a treatment -- like the incentive experiment -- I need to account for all the interviewer costs, as best…

Goodhart's Law

I enjoy listening to the data skeptic podcast. It's a data science view of statistics, machine learning, etc. They recently discussed Goodhart's Law on the podcast. Goodhart's was an economist. The law that bears his name says that "when a measure becomes a target, then it ceases to be a good measure." People try and find a way to "game" the situation. They maximize the indicator but produce poor quality on other dimensions as a consequence. The classic example is a rat reduction program implemented by a government. They want to motivate the population to destroy rats, so they offer a fee for each rat that is killed. Rather than turn in the rat's body, they just ask for the tail. As a result, some persons decide to breed rats and cut off their tails. The end result... more rats.

I have some mixed feelings about this issue. There are many optimization procedures that require some single measure which can be either maximized or minimized. I think thes…

Training for Paradata

Paradata are messy data. I've been working with paradata for a number of years, and find that there are all kinds of issues. The data aren't always designed with the analyst in mind. They are usually a by-product of a process. The interviewers aren't focused (and rightly so) on generating high-quality paradata. In many situations, they sacrifice the quality of the paradata in order to obtain an interview.

The good thing about paradata is that analysis of paradata is usually done in order to inform specific decisions. How should we design the next survey? What is the problem with this survey? The analysis is effective if the decisions seem correct in retrospect. That is, if the predictions generated by the analysis lead to good decisions.


If students were interested in learning about paradata analysis, then I would suggest that they gain exposure to methods in statistics, machine learning, operations research, and an emerging category "data science." It seems like…