Announcement

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

  • Opposite signs on the coefficient estimates of the observed variables in the bidirectional relationship in nonrecursive sem

    Hi,

    I am running nonrecursive models using the sem command. I have encountered a problem with the results, and I am hoping somebody can help me with it:

    The observed variables in the bidirectional relationship in my sem model are:

    V1: % of population that are obese; and
    V2: % of population who took walking trips.

    Both V1 and V2 equations also include various other explanatory variables.

    This is what confuses me:
    The model gives a negative sign for the coefficient estimate of obesity in the walking model (meaning that higher obesity percentages are correlated with lower percentages of walking trips (which I kind of expected because declined health due to obesity may prevent people from making walking trips). However, the model returns a positive sign for the coefficient estimate of walking trips in the obesity model meaning that higher percentages of walking trips are correlated with higher percentages of obesity . This is not logical to me ; I expected to see that more walking trips are correlated with lower obesity rates (and not higher as the model results suggest).

    My question is:
    Is it possible to get opposite signs on the coefficient estimates of the observed variables in the bidirectional relationship in nonrecursive sem models? How could in a nonrecursive model, variable V1 affect V2 in a negative way, while V2 affects V1 in a positive way? If not, what could be wrong in my model? Every example I have seen (including Example 7 - Page 187 of the Stata SEM Reference Manual Release 13 (https://www.stata.com/manuals13/sem.pdf)) has results with similar signs for the coefficient estimates of the observed variables in the bidirectional relationship.

    Any help is appreciated!

  • #2
    It is definitely possible to get opposite signs, and that in and of itself is not a reason to be worried. However, you are worried that the sign makes no substantive sense, and that is a cause for concern. I can think of two explanations:
    1. I assume you are aware of the ecological fallacy (https://en.wikipedia.org/wiki/Ecological_fallacy), i.e. that you cannot conclude that people who are obese walk less because a high percentage of obese people in a region is associated with a low percentage of people walking in a region. This could be the first explanation: aggregate percentages can behave very differently from individual level effects.
    2. Such bidirectional estimates require some pretty strong assumptions in order for your model to be identified. If those assumptions are wrong, then anything can happen.
    ---------------------------------
    Maarten L. Buis
    University of Konstanz
    Department of history and sociology
    box 40
    78457 Konstanz
    Germany
    http://www.maartenbuis.nl
    ---------------------------------

    Comment


    • #3
      Thank you Maarten! Your answers to my questions are definitely helpful. I will have to over the assumptions one more time and check if my data/analysis meet them.

      Comment

      Working...
      X