Questions about Logit regression applicable in my research design.

[Bing Song]

Hello, everyone.

I am not new to use Statalist, but this is the first time I am not sure I find an answer from Statalist. So, I would like to ask all resourceful experts my question. And I am going to try to make it understandable.

I have not had my dataset, I am still trying to identify the ideal methodology for my research question. For now, I can just tell that logit is what I sense the most suitable method.

In my research question, the dependent variable (LHS) is ' investment behavior', i.e. buying, selling, shortselling and saving. The explanatory side (RHS) will have at least 10 variables (majority of them will be dummy and ordinal). The dataset will be definitely a panel. My problem exists for LHS now.

Since people can trade simultaneously (e.g. buying and selling at the same time; shortselling, buying and saving at the same time), so the ordered logit and multinomial logit should be excluded from my options, right?

Now, I think that will be make sense if I construct the dependent variables like the below:
Only Buy, Only Sell, Only Shortsell, Only Save, Buy+Sell, Buy+Shortsell, Buy+Save ... and Buy+Sell+Shortsell+Save. In total, I will have 15 binary outcome as my dependent variable.

So I try to find the command in stata to run these dependent variable simultaneously in logit regression , but I didn't find the option. Or should I run all these 15 dependent variables separately?Also, when I search online, almost all results are showing multinomial logit regression.

I am confused and do not know if this variable construction is applicable for any logit model or not. If it does not work this way, any suggestions to construct my dependent variable?

Many thanks in advance,
Bing

[Clyde Schechter]

Well, if you are looking for a Stata command that will analyze 15 mutually exclusive and exhaustive dichotomous outcomes simultaneously like this, take a look at -mlogit-.

Whether this is an appropriate approach to your problem, or whether something else is needed (and I don't have any concrete ideas about what that something else might be) is a question that requires expertise in the field of investment behavior. As I know nothing at all in that field, I will not pretend to advise you on that.

[Bing Song]

Well, if you are looking for a Stata command that will analyze 15 mutually exclusive and exhaustive dichotomous outcomes simultaneously like this, take a look at -mlogit-.

Whether this is an appropriate approach to your problem, or whether something else is needed (and I don't have any concrete ideas about what that something else might be) is a question that requires expertise in the field of investment behavior. As I know nothing at all in that field, I will not pretend to advise you on that.

Thanks a lot for the response. But, I thought the -mlogit- command is to conduct the multinomial logit regression, since the assumption underlying the multinomial logit is that the categories in dependent variable cannot violate independence of irrelevant alternatives (IIA). In my case, if I run the command -mlogit-, can I say that I can also use 1, 2, 3, 4, ..., 15 to categorize the options?
However, this kind of categorization is not violating the IIA assumption?

Looking forwards to response.

[Clyde Schechter]

The IIA criterion is not about how you numerically label the categories of your variable. It is a substantive criterion regarding the actual behaviors under different circumstances. You are correct that some experts will not accept the use of -mlogit- results if the IIA criterion is not met in the data. But that doesn't preclude you from trying -mlogit-. You would then have to test whether the criterion is actually met in your data. But not everybody agrees completely with this concept anyway. See, for example, https://statisticalhorizons.com/iia.

So yes, if you use -mlogit- you would assign numbers 1 through 15 in some order to the 15 possible behaviors. Which behavior you assign to which number is not important from a statistical perspective, although for purposes of wrapping your mind about the results, it might make sense to give particular assignments to numbers in a way that makes it easier to understand.

[Bing Song]

The IIA criterion is not about how you numerically label the categories of your variable. It is a substantive criterion regarding the actual behaviors under different circumstances. You are correct that some experts will not accept the use of -mlogit- results if the IIA criterion is not met in the data. But that doesn't preclude you from trying -mlogit-. You would then have to test whether the criterion is actually met in your data. But not everybody agrees completely with this concept anyway. See, for example, https://statisticalhorizons.com/iia.

So yes, if you use -mlogit- you would assign numbers 1 through 15 in some order to the 15 possible behaviors. Which behavior you assign to which number is not important from a statistical perspective, although for purposes of wrapping your mind about the results, it might make sense to give particular assignments to numbers in a way that makes it easier to understand.

Thanks for the thorough answer. Now, I am clear about the dependent variable if I use logit regression.

But RHS variables become another problem, because they are going to be either categorical variables or binary variables, which lead my problem to model choice again. Based on this feature of independent variables, I should use is ANOVA or MANOVA. However, either requires the response (dependent variable) being continuous (interval or ratio) level of measurement.

So, there is no model for nominal dependent variable(s) and also categorical or/and binary independent variables, right?

I feel this is a stupid question, since I design something that cannot be done statistically...

Thanks for your patience!

Best Regards,
Bing

[Clyde Schechter]

-mlogit- is appropriate when the outcome variable is categorical, and works with dichotomous, categorical and continuous predictor variables. So it is OK here.

[Bing Song]

-mlogit- is appropriate when the outcome variable is categorical, and works with dichotomous, categorical and continuous predictor variables. So it is OK here.

Thanks a lot! Sincerely appreciate your help and clear answer!

[Marcos Almeida]

Just beware that, having several levels for the Yvar, say, 15 categories (as stated in #1), will provide a quite complex output.

[Bing Song]

Just beware that, having several levels for the Yvar, say, 15 categories (as stated in #1), will provide a quite complex output.

Yes, I understand that the complexity will be a challenge. But I need to have them to see all possibilities before I get my data. After I get my data, I may can reduce the categories. And then, I guess I will need a lot of help for result interpretion or error correction etc. from Statalist!

Thank you for your warning!!!