Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Age-period-cohort models using mixed command with complex survey data

    Hello,

    I am trying to estimate age-period-cohort models in Stata 13.1 (on Windows) using the mixed command. I have read the Stata documentation on mixed effects models (http://www.stata.com/manuals13/me.pdf) , including the section on survey data (beginning on page 323), but am still not sure how to incorporate the survey structure into my APC models.

    The data I am using are repeated cross sections (three waves) from the National Household Transportation Survey. My observations are at the person-level, but the primary sampling unit is households. I have multiple observations for some households. Final person-weights are provided in the data (wtperfin), and are based on standard inverse-probability weights but adjusted for non-response, multiple landlines per household, and oversampling. With all three years combined, I have more than 300,000 observations.

    My dependent variable is log(annual miles driven) (lmidriv12), and the variable of interest in each model is birth year (cohort) or Millennial generation status (millen=1 if birth year is between 1981-1991). I include controls for gender (female), urban/rural status (urbanized) and education. The period variable is the year of the survey, and the age variable is age_cat , which represents 5-year age categories and ranges in value from 1 to 16.

    I would like to estimate the following models with appropriate weights, at appropriate levels, and standard errors that account for multiple observations per household.

    Code:
    mixed lmidriv12 cohort female urbanized edu_2 edu_3 edu_4 edu_5 edu_miss ||period:||age_cat: 
    mixed lmidriv12 millen female urbanized edu_2 edu_3 edu_4 edu_5 edu_miss ||period:||age_cat:
    My specific questions are:

    1) How do I incorporate the weights into the model? I initially tried

    Code:
    mixed lmidriv12 cohort female urbanized edu_2 edu_3 edu_4 edu_5 edu_miss [pw=wtperfin]|| period:|| age_cat:
    but the iterations backed up and would not make progress towards convergence. Even if the model did converge, I'm not sure this is the appropriate use of weights. I don't see, in the documentation, a way to rescale the weights to use in the period and age levels of the hierarchy with a 3-level model. I'm also not sure how I would do that, if I could, or if I need to in the APC model.

    2) How can I account for multiple observations per household? I initially tried

    Code:
    mixed lmidriv12 cohort female urbanized edu_2 edu_3 edu_4 edu_5 edu_miss ||period:||age_cat: , vce(cluster hhid)
    but received the following error:

    highest-level groups are not nested within hhid


    3) If I had a strata variable as well, would there be a way to incorporate this ? In theory, I would do something like this:

    Code:
    svyset hhid [pw=wtperfin], strata(stratavar)
    svy: mixed lmidriv12 cohort female urbanized edu_2 edu_3 edu_4 edu_5 edu_miss ||period:||age_cat:
    Since there is no svy: option for mixed, I'm not sure how I would do this.


    Thank you for reading. I would be grateful for any advice or insight.


  • #2
    I can't give you specific help, but there is a section in the manual on survey data for both mixed and meglm. Survey weights are specified for each level either using the svyset command or the multilevel command.

    Comment

    Working...
    X