Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

From |
jpitblado@stata.com (Jeff Pitblado, StataCorp LP) |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: st: margins: (not estimatable) |

Date |
Wed, 13 Jul 2011 14:09:11 -0500 |

Darjusch <altruist81@gmx.de> is using -margins- after a -logit- model that contains the interaction of a single factor variable with the linear, quadratic, and cubic terms of a continuous variable and is getting the '(not estimable)' label in the -margins- output: > I have read a similar thread > (http://www.stata.com/statalist/archive/2011-06/msg00407.html), but the > answer wouldnt solve my problem, I run a logit model where I am > interested in the marginal effect of fail (dummy) on dropout at the > value of mc_1styr_c==0 (the model is based on a regression discontinuity > research design). > > Here is what I do: > > . logit dropout3_en i.fail##(c.mc_1st##c.mc_1st##c.mc_1st) if sex==0, > vce(cluster mc_1st) > > (output omitted) > > . margins ,dydx(fail) at(mc_1st==0) > > Conditional marginal effects Number of obs = > 409 > Model VCE : Robust > > Expression : Pr(dropout3_enrollment), predict() > dy/dx w.r.t. : 1.fail > at : mc_1styr_c~d = 0 > > ------------------------------------------------------------------------------ > | Delta-method > | dy/dx Std. Err. z P>|z| [95% Conf. > Interval] > -------------+---------------------------------------------------------------- > 1.fail | (not estimable) > ------------------------------------------------------------------------------ > Note: dy/dx for factor levels is the discrete change from the base level. > > > > The command: margins ,dydx(fail) at(mc_1st==0) should give me the effect > and the significance level of interest. If I enlarge the sample (lets > say, not condition on sex==0) it works. > > Could someone explain me the core of the problem and how to solve it? The check for 'Estimable functions' performed by -margins- is detailed in the methods and formulas section of -[R] margins-. The basic idea here is that -logit- saves off a hidden matrix, let's call it H, that -margins- uses to check for estimable functions. This H matrix should only contain values -1, 0, and 1. The check is as follows z*b is estimable if z = z*H where b is the coefficient vector, and z a vector of fixed/hypothetical values of the independent variables in the fitted model. -margins- compares z and z*H via relative differences with a numeric tolerance of 1e-5, controlable via the -estimtolerance()- option. We suspect that the cubic polynomial in Darjusch's fitted -logit- model is causing the H matrix to have values outside -1, 0, and 1. This can happen with numerically unstable calculations due to huge scale differences between the independent variables, otherwise caused by the result of the propagation of errors in the finite precision calculations. Darjusch can verify if there are problems with the H matrix by typing the following Stata commands after the -logit- model fit: . matrix H = get(H) . matrix list H If the output has values other than -1, 0, and 1, Darjusch can use the -noestimcheck- option with -margins- to prevent the estimability check. We feel confident in giving Darjusch this advice in this case since there is only one factor variable in the model specification, and the polynomial terms of the continuous variable should not, in theory assuming infinite precision calculations, result in non-estimable functions. --Jeff jpitblado@stata.com * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

**Follow-Ups**:**Re: st: margins: (not estimatable)***From:*D-Ta <altruist81@gmx.de>

- Prev by Date:
**Re: st: if expression involving a variable in a while loop** - Next by Date:
**st: count differents values of a variable** - Previous by thread:
**st: margins: (not estimatable)** - Next by thread:
**Re: st: margins: (not estimatable)** - Index(es):