What is the prediction model to predict the probability?

[Phuc Nguyen]

In a paper, Dasgupta, 2019 used Difference-in-Difference approach to see whether anticollusion laws implemented by different countries (staggered implementation) affect firms financial flexibility.
Dasgupta, 2019, p.2610 used an approach called "prediction model"

by only using pre-leniency observations and predict the probability that the firm will be convicted in the cartel case after the passage of a leniency law.

In particular, what they did is

First, we estimate the propensity of a firm to be convicted in a cartel case. We use a prediction model based on time-varying firm characteristics (asset size, leverage, and ROA), country characteristics (GDP and unemployment), and country fixed effects and three-digit SIC fixed effects.

I do not understand how they calculate the "probability that the firm will be convicted in the cartel case after the passage of a leniency law" like that by using STATA. The one command I can link to is "predict" but it seems not to work in this case.

[daniel klein]

I have not looked into the paper at all, but margins facilitates producing predicted values at specified values of the covariates in the model.

[Fei Wang]

It's very likely that the paper predicted the probability of firms being convicted, using command logit or probit, based on a variety of characteristics in the second quote.

[Phuc Nguyen]

It's very likely that the paper predicted the probability of firms being convicted, using command logit or probit, based on a variety of characteristics in the second quote.

Thanks Fei Wang, I read and understand the command

logit

or

probit

. My concern here is I do not understand how can I assign the dependent variable 0 and 1 in this case. Is there any idea about that? Thanks!

[Fei Wang]

I'm not able to access the paper for now. A general idea is that, we need to observe some firms that have been convicted (1) or have not (0), and run a logit or probit on their characteristics. Then based on the estimates, we predict each firm's probability of being convicted.

[Phuc Nguyen]

Hi Fei Wang

Thank you so much for your following up. I updated the pdf file of this paper in this comment. I read your comment but it seems that we cannot do that because I do not know what is the firm has been convicted.

[Phuc Nguyen]

I'm not able to access the paper for now. A general idea is that, we need to observe some firms that have been convicted (1) or have not (0), and run a logit or probit on their characteristics. Then based on the estimates, we predict each firm's probability of being convicted.

Hi Fei Wang , I emailed the author and they said that they used the Linear Probability model to do so. Is there any suggestion for that?

[Fei Wang]

Hi Fei Wang

Thank you so much for your following up. I updated the pdf file of this paper in this comment. I read your comment but it seems that we cannot do that because I do not know what is the firm has been convicted.

Thanks for the full-text. On pages 2589-2590, the paper says

Using a database on actual cartel convictions, we predict the propensity for a firm to be a member of a convicted cartel based on its industry and country, and other firm characteristics.

On page 2620, the paper says

We use a prediction model, similar to the one in Section 4.1, and we fit it to the data of the U.S. firms that were convicted of being part of a cartel up to year 2003. We then use firm observables in 2004 to predict the likelihood of being convicted in a collusion case.

You can see that they used actual convictions to fit and predict probabilities of conviction. According to #7, they used a linear probability model, which simply -regress- the actual convictions (0 or 1) on various characteristics and make predictions based on the linear regression estimates.

[Phuc Nguyen]

Thanks Fei Wang , I got your point