What logit and probit coefficients mean for the probability of a crime.
Published
September 20, 2026
Updated
November 15, 2026
When the outcome is a binary variable of 0s and 1s, the predicted probabilities you get from running OLS can be negative or above 100%. So you need to use a logit or probit model, which uses a function that keeps predicted probabilities between 0 and 100%. I show you how to run them in Python, what their coefficients mean and why you need a threshold to decide which announcements to investigate. I simulate data on market crimes, where 1 indicates a crime, and 0 indicates no crime.
Simulating the data
The market crime is illegal insider trading before corporate announcements. The announcements are unscheduled, such as takeovers and mergers, so only insiders know about them in advance. For each announcement, I simulate one variable, information_value, which measures how much the price moves on the announcement date, in standard deviations from its average.
Crimes happen at random, but the probability of a crime increases quickly with information_value. I choose the two coefficients that set this probability somewhat arbitrarily. Since I know these true coefficients, I can compare them with the logit estimates.
NoteHow the announcements are simulated
The true score of announcement \(i\) uses the two coefficients I set,
\[\eta_i = -1.0 + 2.5\,x_i,\]
where \(x_i\) is information_value, a standard normal draw. The logit model turns each score into a probability of a crime, and then I draw whether a crime happens.
I draw 30,000 announcements and fit the models to the first 20,000. A second post tests the logit model on the last 10,000.
Of the 30,000 announcements, 11,112 are crimes (37.0%), with 7,409 in the first 20,000 and 3,703 in the last 10,000.
Why not OLS
I first run OLS on the 0/1 outcome, the linear probability model. It uses the score, \(\eta = \beta_0 + \beta_1 x\), directly as the predicted probability,
\[p = \beta_0 + \beta_1 x, \tag{1}\]
where \(\beta_0\) is the intercept, and \(\beta_1\) is the coefficient on information_value (\(x\)).
NoteCode: the linear probability model
import statsmodels.api as smdesign = sm.add_constant(variables)fitting_design = design.loc[fitting_sample]fitting_outcome = is_crime.loc[fitting_sample]linear = sm.OLS(fitting_outcome, fitting_design).fit()linear_prediction = linear.predict(fitting_design)print(linear.params.round(4).to_string())print(f"\npredicted values range from {linear_prediction.min():.3f} to "f"{linear_prediction.max():.3f}")print(f"below 0: {(linear_prediction <0).sum():,} of {len(linear_prediction):,} "f"({(linear_prediction <0).mean():.1%})")print(f"above 1: {(linear_prediction >1).sum():,} ({(linear_prediction >1).mean():.1%})")
const 0.3697
information_value 0.3074
predicted values range from -0.770 to 1.548
below 0: 2,274 of 20,000 (11.4%)
above 1: 409 (2.0%)
In OLS, each coefficient is a fixed change in the predicted probability. The coefficient on information_value, 0.3074, says that one more standard deviation of information_value adds 0.3074, or 30.7 percentage points, to the predicted probability of a crime at each announcement. The predicted probabilities range from −77.0% to 154.8%, and 2,274 of the 20,000 are below zero and 409 above 100%, which is not possible for a probability.
The logit and probit models
The logit and probit models start with the same score as OLS and convert it into a probability between 0 and 100%.
The logit model treats the score as the log odds of a crime. The odds of a crime are the number of crimes for every announcement without a crime. If 20 of 100 announcements are crimes, the odds are 20/80 = 0.25, one crime for every four announcements without one. In terms of the probability \(p\), the odds are \(p/(1 - p)\), here 0.2/0.8 = 0.25. The table below shows the odds and the log odds (the log of the odds) for six probabilities.
The log odds are 0 at a probability of 50%, negative below it, and positive above it so that they can take any value, like the score. The logit model sets the log odds equal to the score:
Solving (2) for \(p\) gives the logistic function, which converts the score into a probability,
\[p = \frac{e^{\eta}}{1 + e^{\eta}}. \tag{3}\]
In (3), \(e^{\eta}\) is positive and smaller than \(1 + e^{\eta}\), so \(p\) is between 0 and 1 for every score. The animation below shows what the logistic function does to the fitted values.
The probit model uses the standard normal distribution instead, so the predicted probability of a crime is the area under the normal curve to the left of the score,
\[p = \Phi(\eta). \tag{4}\]
The total area under the normal curve is 1, so \(p\) is also between 0 and 1 for every score.
I fit both models with the statsmodels library, which chooses the coefficients that make the observed outcomes most likely (maximum likelihood estimation).
coefficient I set logit probit logit / probit
intercept -1.0 -1.0001 -0.5697 1.7553
information_value 2.5 2.4498 1.3897 1.7629
predicted probabilities differ by 0.00843 on average
The logit estimates, −1.000 and 2.450, are close to the −1.0 and 2.5 I set.
The logit coefficients are 1.76 times the probit ones, but the predicted probabilities differ by only 0.84 percentage points on average. So the two models give almost the same probabilities, and a logit coefficient cannot be compared directly with a probit one. The animation below shows why.
The chart shows each model’s predicted value as information_value changes.
NoteCode: the chart
import matplotlib.pyplot as pltRED, TEAL, ORANGE, GREY ="#C0392B", "#17868A", "#D2822B", "#888888"BG, INK, GRID, SPINE, AXIS_TEXT ="#FCEFE3", "#1f1f1f", "#EADCCC", "#D5C6B4", "#4a4a4a"def style_chart(figure, axis): figure.patch.set_facecolor(BG) axis.set_facecolor(BG) axis.grid(True, axis="y", color=GRID, lw=1.0) axis.set_axisbelow(True)for side in ("top", "right"): axis.spines[side].set_visible(False)for side in ("left", "bottom"): axis.spines[side].set_color(SPINE) axis.tick_params(axis="both", length=0, colors=INK, pad=6) axis.xaxis.label.set_color(AXIS_TEXT) axis.yaxis.label.set_color(AXIS_TEXT) axis.title.set_color(INK) legend = axis.get_legend()for text, line inzip(legend.get_texts(), legend.get_lines()): text.set_color(line.get_color())low, high = np.percentile(information_value[fitting_sample], [0.5, 99.5])grid = np.linspace(low, high, 300)curve_input = pd.DataFrame({"const": 1.0, "information_value": grid})figure, axis = plt.subplots(figsize=(9, 5.2))linear_curve = linear.predict(curve_input)axis.plot(grid, linear_curve, color=RED, lw=2.4, label="linear regression")axis.plot(grid, logit.predict(curve_input), color=TEAL, lw=2.8, label="logistic")axis.plot(grid, probit.predict(curve_input), color=ORANGE, lw=2.2, ls=(0, (4, 3)), label="probit")axis.fill_between(grid, linear_curve, 0, where=linear_curve <0, color=RED, alpha=0.15)axis.fill_between(grid, linear_curve, 1, where=linear_curve >1, color=RED, alpha=0.15)axis.axhline(0, color=GREY, lw=1.2)axis.axhline(1, color=GREY, lw=1.2)axis.set_xlabel("Information value of the announcement (standard deviations)")axis.set_ylabel("Predicted value")axis.set_title("The linear model predicts below zero and above one")axis.legend(frameon=False)style_chart(figure, axis)plt.tight_layout()plt.savefig("probability_curves.png", dpi=140, bbox_inches="tight", facecolor=BG)plt.show()
What the coefficients mean
Since information_value is measured in standard deviations, each coefficient indicates the effect of a one-standard-deviation increase in information_value. In OLS, one more standard deviation raises the probability of a crime by 0.3074, or 30.7 percentage points. In the logit model, one more standard deviation raises the log odds of a crime by 2.4498, as (2) shows, so the odds are multiplied by \(e^{2.4498} = 11.59\), the odds ratio. In the probit model, one more standard deviation moves the score 1.3897 to the right on the standard normal curve in (4), increasing the area to its left (the probability). Only the OLS coefficient is a change in the probability, so the logit and probit coefficients cannot be compared with it directly.
To compare them with OLS, I turn them into changes in the probability. In the logit model, this change is the slope of the logit curve in the chart above, the marginal effect. Differentiating (3) with respect to \(x\) gives
\[\frac{dp}{dx} = \beta_1\, p\,(1 - p). \tag{5}\]
The slope is largest at a probability of 50%, where it is \(\beta_1/4 = 0.61\), and close to 0 near 0% and 100%, where the logistic curve is flat. In the probit model, differentiating (4) gives
\[\frac{dp}{dx} = \beta_1\, \phi(\eta), \tag{6}\]
where \(\phi(\eta)\) is the height of the standard normal curve at the score.
Because the slope differs across announcements, I average it over the 20,000 announcements. This average marginal effect is a change in the probability, like the OLS coefficient, so the two can be compared. The statsmodels library computes it with get_margeff.
NoteCode: the average marginal effects
average_marginal_effects = pd.Series({"OLS": linear.params.loc["information_value"],"logit": logit.get_margeff().margeff[0],"probit": probit.get_margeff().margeff[0]})print("slope of the predicted probability, averaged over the announcements:")print(average_marginal_effects.round(4).to_string())
slope of the predicted probability, averaged over the announcements:
OLS 0.3074
logit 0.3096
probit 0.3081
The average marginal effect is 0.3096 in the logit model and 0.3081 in the probit model, close to the OLS coefficient of 0.3074. So in all three models, the probability of a crime rises by about 31 percentage points per standard deviation of information_value, on average. The models differ in the predicted probabilities, since only logit and probit keep each one between 0 and 100%.
The OLS coefficient is close to these averages because the OLS line best fits the S-shaped logit curve across the announcements. The animation below shows both slopes at every information value and the average of the logit slope.
Choosing a threshold
The predicted probabilities lie between 0 and 1, while the outcomes are 0s and 1s. So to decide which announcements need further investigation, I need a threshold. I flag every announcement with a predicted probability above it.
NoteCode: three thresholds
crimes = fitting_outcome ==1print(f"crimes among the {len(crimes):,} announcements: {crimes.sum():,}")for threshold in (0.2, 0.5, 0.8): flagged = logit_probability > threshold caught = (flagged & crimes).sum()print(f"threshold {threshold:.0%}: flagged {flagged.sum():,}, crimes caught {caught:,} "f"({caught / crimes.sum():.1%} of all crimes), false alarms {flagged.sum() - caught:,}")
crimes among the 20,000 announcements: 7,409
threshold 20%: flagged 11,345, crimes caught 6,805 (91.8% of all crimes), false alarms 4,540
threshold 50%: flagged 6,799, crimes caught 5,294 (71.5% of all crimes), false alarms 1,505
threshold 80%: flagged 3,300, crimes caught 2,986 (40.3% of all crimes), false alarms 314
With a threshold of 50%, I flag 6,799 of the 20,000 announcements. They include 5,294 of the 7,409 crimes (71.5%) and 1,505 false alarms, flagged announcements without a crime. A threshold of 20% catches 91.8% of the crimes but gives 4,540 false alarms. So a lower threshold catches more crimes but also produces more false alarms.
Conclusion
So when you deal with binary outcomes, you should run a logit or a probit instead of OLS. A logit coefficient has a fixed effect on the log odds, which corresponds to a fixed multiplicative effect on the odds, but it does not have a fixed effect on the probability. To compare it with an OLS coefficient, you use its average marginal effect. Because predicted probabilities lie between 0 and 1, you need a threshold to decide which announcements to investigate.