Bookmark and Share

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

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

Re: st: Regression Across Two Groups

From   Muhammad Anees <[email protected]>
To   [email protected]
Subject   Re: st: Regression Across Two Groups
Date   Wed, 14 Dec 2011 16:12:43 +0500

Dear Maarten, Nick and Others,

Thank you very much for your generous advice. The case of sample
selection sounds the reasonable option so far although all other
material helped me learn issues related to such research questions. I
am confident to do the right thing now.

On Wed, Dec 14, 2011 at 4:01 PM, David Ashcraft
<[email protected]> wrote:
> This seems a sample selection bias case and _heckman_ is a good solution, if the question is about the secondary job e.g. what would be possible level of earning subject to secondary job based on specific traits and one of which can be foreign qualification as a dummy.
> ----- Original Message -----
> From: Nick Cox <[email protected]>
> To: [email protected]
> Cc:
> Sent: Wednesday, December 14, 2011 1:30:35 PM
> Subject: Re: st: Regression Across Two Groups
> From what you say, or from what I understand of it, the Yes or No
> variable is irrelevant to your main question, in one of two senses:
> 1. If people have no earnings, then you could code their earnings as 0
> and include them in the sample.
> 2. If people have no earnings, then they don't belong in your main sample.
> The choice is yours, depending on your research question.
> Either way, you have now got good specific advice from Maarten and
> others which gives you scope for experiment. Either way, this is not
> coming across to me as a logit or probit problem at all.
> Nick
> On Wed, Dec 14, 2011 at 10:16 AM, Muhammad Anees <[email protected]> wrote:
>> There are actually two variables describing earnings, one as a
>> dichomous (recorded in Yes or No) and and other as groups or interval.
>> The objective is to assess if such earnings from secondardy jobs also
>> depend on foreign qualification and if yes, then what is difference
>> between the regressions across the two groups based on foreign
>> qualification or no qualification.
>> On Wed, Dec 14, 2011 at 3:03 PM, Nick Cox <[email protected]> wrote:
>>> Thanks for the example.
>>> I see earnings as a coarsely categorised variable, fit for -intreg- or
>>> -ologit-. In what sense is earnings dichotomous (means, has two
>>> categories)? Why did you say you were interested in logit models?
>>> Nick
>>> On Wed, Dec 14, 2011 at 9:48 AM, Muhammad Anees <[email protected]> wrote:
>>>> Thanks Nick for your suggestion.
>>>> Sure, the data looks like which contains two different sample but
>>>> related. Foreign qualified and no foreign qualification are two
>>>> different dataset earch with the following sample data, only a sample,
>>>> actually the data consist two different samples in two different
>>>> files, which I can deal how to combine in stata for my purpose.
>>>> ear     exper   gender  subject area    language
>>>> 0-5000  20      m       IT      rural   Urdu
>>>> 5000-10000      22      m       ENGINEER        urban   English
>>>> 10001-15000     15      f       ECONOMICS       rural   Urdu
>>>> 5000-10000      10      m       HR      urban   Urdu
>>>> 5000-10000      5       f       STRT MGT        urban   English
>>>> 10001-15000     8       f       MARK    urbna   English
>>>> 0-5000  9       m       SOCIOLOGY       rural   Urdu
>>>> 0-5000  17      m       IT      urban   Urdu
>>>> On Wed, Dec 14, 2011 at 2:39 PM, Nick Cox <[email protected]> wrote:
>>>>> It's categorical/dichotomous, yet the example is [pro]portion of
>>>>> earnings from outside main job. Sounds like a fractional response from
>>>>> the latter. Muhammad: Give us an example of what observations look
>>>>> like before this gets any more obscure, please!
>>>>> On Wed, Dec 14, 2011 at 9:18 AM, Maarten Buis <[email protected]> wrote:
>>>>>> On Wed, Dec 14, 2011 at 6:25 AM, Muhammad Anees wrote:
>>>>>>> Sorry for not clarifying the story about the types of variables, like
>>>>>>> earnings which I have at hand as a categorical/dichotomous variable.
>>>>>>> For example if an individual has a portion of earnings from doing
>>>>>>> consultancies or involved in any R&D organizations beside their normal
>>>>>>> routine jobs. In this case, I was interested in comparing the
>>>>>>> regression models (across foreign qualified and not foreign qualified)
>>>>>>> of earnings on other predictors say experience, research training, job
>>>>>>> nature, industry, region (rural and urban) using logit/probit in case
>>>>>>> of categorical variables and similarly using OLS for continuous
>>>>>>> dependent variable which at least I do not have at this stage.
>>>>>> This is still not clear. The independent/explanatory/right-hand-side/x
>>>>>> variables aren't relevant here, they can be of any type, it is the
>>>>>> type of  the dependent/explained/left-hand-side/y variable that
>>>>>> matters. Earnings is typically collected as either a continuous
>>>>>> variable (how much do you earn?) or as a choice from a set of
>>>>>> intervals (did you earn less than x$, between x$ and y$, etc.?). None
>>>>>> of these are correctly modeled as a logit/probit. In the former case I
>>>>>> would use a -glm- with the -link(log)- option, in the latter case I
>>>>>> would start with assigning each category with a reasonable
>>>>>> representative number and than use -glm- with the -link(log)- option.
>>>>>> There are other solutions for the latter problem, e.g. -intreg-, but
>>>>>> if the underlying distribution is non-normal, which is likely to be
>>>>>> the case with earnings, then it is unclear whether these alternatives
>>>>>> are any better. The comparison is than just a matter of adding the
>>>>>> appropriate dummies and/or interactions.
> *
> *   For searches and help try:
> *
> *
> *
> *
> *   For searches and help try:
> *
> *
> *


Muhammad Anees
Assistant Professor
COMSATS Institute of Information Technology
Attock 43600, Pakistan

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index