Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 122 views

    Athrho Interpretation

    Uploaded by

    Benjamin Naula
    This document provides a tip for accessing parameters estimated on a transformed scale, such as standard deviations and correlations, from models in Stata. It explains that parameters like standard deviations and correlations are often estimated on a transformed scale to constrain their values to valid ranges. It demonstrates how to recover the original parameter values using the inverse transformations, and how to test hypotheses and impose constraints on the transformed parameters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Athrho Interpretation

    Uploaded by

    Benjamin Naula
    122 views3 pages
    This document provides a tip for accessing parameters estimated on a transformed scale, such as standard deviations and correlations, from models in Stata. It explains that parameters like standard deviations and correlations are often estimated on a transformed scale to constrain their values to valid ranges. It demonstrates how to recover the original parameter values using the inverse transformations, and how to test hypotheses and impose constraints on the transformed parameters.

    Original Description:

    Athro Description

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document provides a tip for accessing parameters estimated on a transformed scale, such as standard deviations and correlations, from models in Stata. It explains that parameters like standard deviations and correlations are often estimated on a transformed scale to constrain their values to valid ranges. It demonstrates how to recover the original parameter values using the inverse transformations, and how to test hypotheses and impose constraints on the transformed parameters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    122 views3 pages

    Athrho Interpretation

    Uploaded by

    Benjamin Naula
    This document provides a tip for accessing parameters estimated on a transformed scale, such as standard deviations and correlations, from models in Stata. It explains that parameters like standard deviations and correlations are often estimated on a transformed scale to constrain their values to valid ranges. It demonstrates how to recover the original parameter values using the inverse transformations, and how to test hypotheses and impose constraints on the transformed parameters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 3

    The Stata Journal (2011)

    11, Number 2, pp. 315–317

    Stata tip 97: Getting at ρ’s and σ’s


    Maarten L. Buis
    Department of Sociology
    Tübingen University
    Tübingen, Germany
    maarten.buis@uni-tuebingen.de

    There are several models in Stata that estimate coefficients other than regular
    “regression-like” coefficients. Often these have an interpretation as a standard devia-
    tion or a correlation of either error terms or random coefficients. Examples are xtmixed
    or heckman (see [XT] xtmixed or [R] heckman). Sometimes, we want access to these
    coefficients to perform a test or to impose a constraint. However, getting access to these
    parameters is not always straightforward. The problem is that Stata programs often
    do not estimate these coefficients directly, but instead estimate a transformed version
    of those parameters. In this Stata tip, I will illustrate how to recover these parameters,
    test hypotheses, and impose constraints.
    The reason for estimating the transformed parameters rather than estimating the
    parameters directly is that not all positive and negative numbers represent valid values
    for correlations or standard deviations, which can make it more difficult to fit the model.
    The transformation is a way to work around this problem because this transformation
    is chosen such that all numbers represent valid values for this parameter. For example,
    a standard deviation can only be larger than or equal to 0; if we model the logarithm of
    the standard deviation instead, then all positive and negative numbers represent valid
    values. Similarly, correlations are often transformed using a Fisher’s z transformation
    (for a discussion of this transformation, see Cox [2008a]). This transformation is rep-
    resented by z = 1/2{ln(1 + ρ) − ln(1 − ρ)}. In Stata, you would compute this as z =
    atanh(ρ). The inverses of these transformations are implemented in Stata as the exp()
    and tanh() functions (see [D] functions).1
    The example below will use the heckman command because heckman will return
    both standard deviations and correlations on a transformed scale. Other commands
    where these strategies can be applied are, for example, heckprob, treatreg, intreg,
    xtmixed, xtlogit,2 xtprobit,2 xtmelogit, and xtmepoisson. Moreover, there are
    some programs that estimate additional parameters on a transformed scale that are not
    standard deviations or correlations but to which this same strategy can also be applied,
    for example, nbreg or streg.
    If we want to recover the values of the parameters, we need to know how Stata is
    calling them. To find that out, we can add the coeflegend option to our estimation
    command. We can then see that neither rho nor sigma has a legend attached to it.
    However, there are parameters called athrho and lnsigma that do have legends attached
    to them, and we can read in the manual (and deduce from their names) that these are the
    1. The Mata equivalents of these functions have the same names and are documented in [M-5] exp( )
    (for exp() and ln()) and [M-5] sin( ) (for atanh() and tanh()).
    2. However, the rho in the output of these commands does not refer to a correlation.

    
    c 2011 StataCorp LP st0230
    316 Stata tip 97

    Fisher’s z transformed correlation and the natural logarithm of the standard deviation,
    respectively. So we can recover these parameters as follows:

    . http://www.stata-press.com/data/r11/womenwk
    . heckman wage educ, select(married children educ) nolog coeflegend
    Heckman selection model Number of obs = 2000
    (regression model with sample selection) Censored obs = 657
    Uncensored obs = 1343
    Wald chi2(1) = 403.39
    Log likelihood = -5250.348 Prob > chi2 = 0.0000

    wage Coef. Legend

    wage
    education 1.099506 _b[wage:education]
    _cons 7.042147 _b[wage:_cons]

    select
    married .5420304 _b[select:married]
    children .4409418 _b[select:children]
    education .0722993 _b[select:education]
    _cons -1.473038 _b[select:_cons]

    /athrho .8081049 _b[athrho:_cons]


    /lnsigma 1.807547 _b[lnsigma:_cons]

    rho .6685435
    sigma 6.095479
    lambda 4.075093

    LR test of indep. eqns. (rho = 0): chi2(1) = 47.02 Prob > chi2 = 0.0000

    . // standard deviation of the residual of the wage equation


    . di exp([lnsigma]_b[_cons])
    6.0954785
    . //correlation between residuals of the wage and selection equation
    . di tanh([athrho]_b[_cons])
    .6685435

    If we have specific hypotheses, then one way of testing these is to rephrase these hy-
    potheses in terms of the transformed metric. Here I compute the transformed standard
    deviations and correlations on the fly by using the trick to enter the computations as
    ‘= exp’ (Cox 2008b). This way, exp is immediately evaluated, and test only sees the
    number that is the result of that computation. Notice that the transformations are not
    defined for standard deviations of 0 or correlations of −1 or 1. This is another way in
    which we can see that one needs to be careful when testing hypotheses on the boundary
    of the parameter space (for example, see Gutierrez, Carter, and Drukker [2001]).
    M. L. Buis 317

    . test ([lnsigma]_b[_cons] = `= ln(6)´ )


    > ([athrho]_b[_cons] = `= atanh(.7)´)
    ( 1) [lnsigma]_cons = 1.791759
    ( 2) [athrho]_cons = .8673005
    chi2( 2) = 2.29
    Prob > chi2 = 0.3177

    Using similar tricks, we can also impose constraints on these transformed auxiliary
    parameters.
    . constraint 1 [lnsigma]_cons = `= ln(6)´
    . constraint 2 [athrho]_cons = `= atanh(.7)´
    . heckman wage educ, select(married children educ) constraint(1 2) nolog
    Heckman selection model Number of obs = 2000
    (regression model with sample selection) Censored obs = 657
    Uncensored obs = 1343
    Wald chi2(1) = 465.26
    Log likelihood = -5251.522 Prob > chi2 = 0.0000
    ( 1) [lnsigma]_cons = 1.791759
    ( 2) [athrho]_cons = .8673005

    wage Coef. Std. Err. z P>|z| [95% Conf. Interval]

    wage
    education 1.104284 .0511957 21.57 0.000 1.003943 1.204626
    _cons 6.911303 .7037765 9.82 0.000 5.531927 8.29068

    select
    married .5405253 .0639579 8.45 0.000 .4151702 .6658805
    children .4405534 .0258773 17.02 0.000 .3898347 .491272
    education .0725339 .010469 6.93 0.000 .0520151 .0930527
    _cons -1.467094 .1436783 -10.21 0.000 -1.748699 -1.18549

    /athrho .8673005 . . . . .
    /lnsigma 1.791759 . . . . .

    rho .7 . -1 1
    sigma 6 . . .
    lambda 4.2 . . .

    LR test of indep. eqns. (rho = 0): chi2(1) = 44.67 Prob > chi2 = 0.0000

    References
    Cox, N. J. 2008a. Speaking Stata: Correlation with confidence, or Fisher’s z revisited.
    Stata Journal 8: 413–439.
    ———. 2008b. Stata tip 59: Plotting on any transformed scale. Stata Journal 8:
    142–145.
    Gutierrez, R. G., S. Carter, and D. M. Drukker. 2001. sg160: On boundary-value
    likelihood-ratio tests. Stata Technical Bulletin 60: 15–18. Reprinted in Stata Technical
    Bulletin Reprints, vol. 10, pp. 269–273. College Station, TX: Stata Press.

    You might also like