Hi! I'm a novice at this, so please explain as if I were a kid. I receive results in a Logit regression where Log likelyhood is zero. What does this mean? Is the probability zero, or what?
Which software package are you using? Do you get any error messages such as "variable X was dropped because it perfectly predicts Y"? Do you get any coefficients on any of the covariates in your model? Did you check via cross-tabulation the relationship between the outcome and the various covariates? If so, were there any covariates that had either all outcomes = 1 or all outcomes = 0?
Hi, and thanks for the answer. I use STATA. No, I don't receive and answer like that. I get ciefficients, yes. When I do a cross-tabulation, one observation of the independent variable differs from the others. The problem arises when I add lags in other variables.
Hi Lars. First off, please note that it is spelled "Stata" not "STATA" (it is not an acronym).
Can you post your log file so that I can see the output from your regression and your tab? If could be that you are running into issues of autocorrelation with your lags that logit doesn't like. In general, if you have longitudinal data you should be using xtlogit
xtlogit is a longitudinal model for binary outcome variables. I am attaching a link the Stata manual entry for xtlogit. I can answer more specific questions after you've read it and tried out the example. I think you'll find this model appropriate for your data
No, you do not need to use lags in longitudinal data, unless you believe that past values are impacting future values. What percent of missing values do you have? Are values missing for all variables?
Well, that is my hypothesis: That lags (in this case bank credit) affects propability of financial crisis (binary dependent variable). Forget about the interpolation question, that was for an other question here on research gate, sry. this is sort of what I want to do: http://www.nber.org/papers/w15512.pdf
You can use lagged data in longitudinal models, but you can't use the AR correlation, unless your observation periods are evenly-spaced out. If your data are evenly spaced out, you can run your model with no problem. To lag the variable you are interested in, simply use the "L." prefix, such as L1.X (which indicates a lag 1 of X).
I'm sorry, but now I don't follow you at all. Why would I need multiple periods to get lags? I have one period, where all previous years are lags. If I take the year 1978 for instance, then 1977 is lag one, 1976 is lag two, and so forth. Or is your definition of a period different? I produced the data myself, and it will be published shortly in a journal. I can't leak it to anyone just yet.
I'm suspecting that we are talking about different things. Anyways, this is what I want to do: I want to examine how lags of bank credit are related to the occurence of financial cisis. In other words, I'm doing time series regression with a binary dependent variable, financial crisis.
Your previous descriptions of your data have not been clear, and therefore my responses have not been clear. Your last two posts have been the most clear so far. You DO have time series data, not one time-point. If you had only one time point, you would not be able to get lagged values. Now that you clarify that you indeed have data stretching back annually, it makes more sense.
As for the model: you can run the data through xtlogit where you have Y regressed on X and/or lagged X. To lag a variable in Stata you simply add the prefix "L." as in L1.X.
Your data will be autocorrelated so you need to use a modeling approach that handles that. Autocorrelation occurs when you measure something repeatedly. Current values will be correlated to past values, to a diminishing degree, as you go further out. If you don't account for autocorrelation, your standard errors will be incorrect. xtlogit is specifically designed for binary outcomes.
What did your test for autocorrelation tell you? Do you have autocorrelation, and if so, at how many lags?
In general, xtlogit will handle the problem, but depending on what you are specifically interested in addressing, you could go further by specifying the corr(ar) option with the PA_option.
Run the model in the usual manner and see what the results show. We can discuss further refinements as necessary.
you have to choose xtset or tsset, but not both. If have more than one group you are analyzing you'll need to specify the variable for that group. I suggest that you go with
I don't understand what you mean by full list of commands? Do you mean to write out the model? If so, you need to give me the names of your Y and X variables, the name of your group ID variable, and the name of your year variable. Also you'll have to tell me which of the X variables you want lagged,
This is clearly over my head. I don't have sufficient knowledge in neither Stata or econometrics theory to perform this. No matter what I do I come up with new errors. Is it ok just to make a logit regression with lags of Xs (first differences), or will the journal I will send this to likely reject it if I do?
I suggest that you hire a statistician. It is unlikely that a peer-reviewer will not notice that you have longitudinal data and not using the correct model