Identify movers in SIPP

Question

Hi everyone, I’m working with SIPP to look at people who move during the year. According to the Census definitions, if TMOVER is between 2 and 7 the person is a mover, and the month they start their new residence is given by ERH_BMONTH.

Using these variables, I get a sensible hump-shaped seasonal pattern in most years. Fewer moves in winter, more in summer. It looks good for 2018 through 2024.

But when I use the same approach on the 2014 SIPP, Waves 2–4, the distribution looks strange: huge spikes in January (~20%), really heavy numbers in March/April, and very low values in autumn. It makes no sense.

Since TMOVER and ERH_BMONTH have the same definitions in those waves, it’s not a coding or variable issue. Could this be a sample design problem with those panels? Has anyone else seen this before or figured out what’s going on? Maybe I am missing something straightforward.

Answer 1

I dont know the answer but this is what AI said.

I would not assume this is purely a coding mistake, but I also would not jump straight to “sample design problem” without checking a few SIPP-specific issues.

The main thing I would check is whether you are using ERH_BMONTH as the move month in a way that creates a January/start-of-spell artifact. ERH_BMONTH is the beginning month of the residence spell, while TMOVER is the monthly mover flag. For identifying the actual month of a move, I would first try using the month of the person-month record where TMOVER is 2–7, rather than using ERH_BMONTH as the timing variable. Then compare that to ERH_BMONTH as a diagnostic.

I would also split the results by wave rather than pooling Waves 2–4. In the 2014 panel, Wave 2 covers 2014, Wave 3 covers 2015, and Wave 4 covers 2016. If the January/March/April pattern is concentrated in one wave, that points toward a wave-specific processing, recall, or weighting issue rather than a general mover definition issue.

A few diagnostics I would run:

1. Tabulate movers by month separately for Wave 2, Wave 3, and Wave 4.

2. Compare three definitions:

   * TMOVER = 2–7 by the monthly record month;
   * change in ERESIDENCEID from month t to t+1;
   * ERH_BMONTH for the residence spell.

3. Check whether the January spike is mostly returning sample members, new respondents, imputed cases, or Type 2/part-year residents.

4. Check whether the spike survives using unweighted counts, person weights, and normalized within-wave weights.

5. Check whether ERH_BMONTH = 1 is picking up people whose spell began before the reference year or at the first observed month, rather than people who truly moved in January.

6. Drop January as a sensitivity test and see whether the remaining seasonal pattern becomes plausible.

My guess is that the issue is related to the redesigned 2014 SIPP’s annual reference period and residence-spell construction, not a change in the formal definition of TMOVER. The 2014 panel moved to a one-year reference period, and that redesign created known concerns about recall, transitions, and panel attrition. So identical variable names and labels do not necessarily mean the month-of-move distribution is comparable to later SIPP files without additional checks.

The key distinction is this:

TMOVER asks whether the respondent moved in a given month.

ERH_BMONTH identifies the beginning month of the residence spell.

Those are usually consistent, but they are not conceptually identical. If you are trying to produce a seasonal distribution of moves, I would treat TMOVER by person-month as the primary measure and ERH_BMONTH as a spell-level consistency check.

So I would not say “the 2014 data are unusable.” I would say: first rule out a spell-start artifact, then rule out pooling/weighting across waves, then check whether the remaining pattern is a real 2014-panel redesign artifact.