There is no significant difference in most cases. Anyhow, you can carry out both analysis as a robustness check.

Regarding the second question, it depends on what you are going to study. In any case, if you are using econometric software to compute marginal effects, make sure you know what are the values for the rest of the variables.

The preference to use logit vs. probit vs. tobit has to do with functional form and disciplinary preferences. When in doubt, plot your data to see which one has a better fit. See http://www.columbia.edu/~so33/SusDev/Lecture_9.pdf

In general I use both methods to show that my results are not driven by the choice of the link function.

Thanks friends for your response

There is no significant difference in most cases. Only in the case of a very small fraction of 0 or 1 there are some differences. Generally it is better for robustness checks of the results to compare one of the modells with the semi-nonparametric or the semiparametric maximum likelihood estimators. These approaches differ from the probit and logit estimators because they can handle a broader class of error distributions.

Gabler/Laisney/Lechner adopted the semi-nonparametric approach of Gallant and Nychka (1987) for univariate and bivariate binary-choice models.

Gabler, S., F. Laisney, and M. Lechner. 1993. Seminonparametric estimation of binarychoice models with an application to labor-force participation. Journal of Business and Economic Statistics 11: 61–80.

thanks Horst i have gone through the paper you have mentioned it is quite useful

Hi Aasif,

Just to point out that the coefficients from a probit and logit model are not directly comparable(although their marginal effects are). The reason is the variance of the standard normal CDF is 1 ( since we have N(0,1)) while the variance of the logistic distribution is (π^2)/3. Thus to make the coefficients across the models comparable you should divide the coefficients from the logit model by√((π^2)/3)≈1.8. However, Amemiya (1981) and others noticed that dividing by 1.6 tends to give a better approximation for most of the range of the curve so you should use that instead.

Hence you may find that the coefficients appear to be different although in fact the two models are telling you the same thing. For predictions/marginal effects however this is not an issue.

Respected Stephen

yes i know in order to compare these models we have to make some adjustments as your good self rightly mentioned div.1.8 or 1.6

Sir in my model when i run the probit in stata i found all the variables except one were significant and when i run the marginal effects i found in one of the variable it is 1.29e-07 which almost means zero sir. Sir may your good self help me what shall i do in this case is their any problem in my model or i have to to interpet it only on the basis of its sign.....

Regards

Aasif

Yes this variable seems to have a very small (if any) effect. You can interpret the sign as the point estimate is your best guess but at the same time the insignificance means that you can't be very confident that this is the true direction of the effect.

If you look at the summary statistics for that variable it may take very large values (e.g. GDP in $) in which case it may be sensible to express it differently (e.g. millions of $ by dividing GDP by a million -> scaling down the X scales up the Beta so B*X is unchanged). I believe software can have problems dealing with very large or very small number so this may be worth doing, however I suspect it won't change your results very much if at all.

More generally, insignificant variables don't necessarily indicate a problem. Although if you had a very strong belief that this variable should have an effect after controlling for the other variables, you may wonder whether there is endogeneity / measurement error, or whether you have included it in the correct form.

Also, as you no doubt know, in a non-linear model like probit/logit, the partial effects of your variable depend on the values of the X variables at which they are evaluated. If you are looking at the marginal effect at the mean this effect may be 0, but nonetheless for other individuals this may actually be having a significant effect. (If you use stata the command - margins - would let you explore this).

Respected stephen

I am very much thankful for you good self for replying an helping me.

Sir after calculating my margin effects (probit) i foung the following output. As it is shown in the talbe that all the variables are sig except gender and in case of last variable ie hhincome df/dx is -1.29e-07 and is significant. sir plz help me weather the result is ok or i have any problem. sir as it is generally believed and also found in various empirical studies that when hhincm increases cl decreases hear is the case but regarding marginal effect it shows 1 percent increase in hhincm results in decrease in cl by 1.29e-07 which is almost zero sir plz help what shall i do in this case

obs.1727

LR chi2(6) = 668.31

Prob > chi2 = 0.0000

Log likelihood = -111.87937 Pseudo R2 = 0.7492

cl dF/dx Std. Err. z P>z x-bar [ 95% C.I. ]

age .0000395 .0000402 5.75 0.000 10.8379 -.000039 .000118

gender* -.0000191 .0000342 -0.82 0.413 .555877 -.000086 .000048

region* .000085 .0000918 3.01 0.003 .48176 -.000095 .000265

hhedu -.0000144 .0000155 -5.17 0.000 8.23683 -.000045 .000016

fsize .0000435 .0000458 7.42 0.000 5.02837 -.000046 .000133

hhincm -1.29e-07 1.34e-07 -10.45 0.000 5805.73 -3.9e-07 1.3e-07

obs. P .0718008

pred. P .0000274 (at x-bar)

(*) dF/dx is for discrete change of dummy variable from 0 to 1

z and P>z correspond to the test of the underlying coefficient being 0

Hi Aasif,

It is hard to say from just output but these results seem a bit confusing to me as your p-values are practically 0, yet the confidence intervals all include 0 (which contradicts the p-value). I usually use the margins command in Stata which I don't think is what you have used here so it may be a difference in terms of what output is provided by the package (or perhaps the version of Stata you are using) but it seems strange to me.

The commands I would use after the probit are:

1) for average partial effects: margins ,dydx(*) predict(pr)

2) for marginal effects at the mean: margins ,dydx(*) predict(pr) atmeans

thanks Sir

Dear Aasif,

As others have pointed out already, in the simple binary case, the choice between logit and probit is not an issue. In most practical applications, both will give you the same partial effects. Which brings me to the second part of your question, the one about whether to use partial effects or point estimates. For me, the answer is given, but before reading it, you should ask yourself: What is the aim of my study? Oftentimes, we are more interested in the economic significance of explanatory variables, rather than just the statistical significance. In other words: you should always use the partial effects, or the odds ratios, in applications to real data. The parameter estimates along with their standard errors are perfectly fine for statistical inference, but even if you get the sign right, they are not useful without knowing their magnitude (which the partial effect will give you).

I have attended many research seminars where the presenter performs inference on the parameter estimates and has no idea what they mean in terms of elasticities, marginal effects, or anything useful for an economist. This is bad practice and could spoil an otherwise nice presentation, or paper.

A special case when the parameters are of direct interest is when you run Monte Carlo simulations, but that is a different issue.

Kind regards,

Nikolay

Www.leora-software.com is the new website for the PoloSuite software. The 3rd Edition of Bioassays with Arthropods will be available this fall.

Any one pls help me how could I control variable on stata analysis?

Dear Kefale

May i know your variables and research question?

Ya, sure My research question does credit risk management have effect on profitability of the specified banks I consider in the last 5 year time? and my proposed control variable is natural log of total asset. so my question is what is the command or syntax to run regression under controlled variable.

Although the distribution nature of the functions differ, there is no difference in the results of the two models. The choice will depend on the researcher after checking the robustness of the results.

Depends on your data. You try both and whichever provides the best fit (lowest chi square value) for your data is the correct one. In general, logit is used for non normal data. If there really is no large difference between your chi values, stick with using probit.

The Logit model is better and less complex in interpretation. The log of odds ratio is easy to compute as some softwares such as SPSS gives them to you directly. Since the dependent variable is dummy (binary), the Logit model is user friendly and takes care of it.

Hi everyone, I am new in statistic and working out on goodness of fit my model (Probit model). I run some tests (Post estimation) including wald test, Hosmer Lemeshow and likelihood ratio test and results indicates the goodnesses of fit. I also want to run Lagrange Multiplier Test in STATA 14. Could you please provides me command and interpretation (Null hypothesis) for outcomes?

Thanks

We often use probit and logit models to analyze binary outcomes. A case can be made that the logit model is easier to interpret than the probit model, but Stata’s margins command makes any estimator easy to interpret. Ultimately, estimates from both models produce similar results, and using one or the other is a matter of habit or preference.