The logit model is ECS4863’s qualitative response topic in its examinable form: a binary outcome, an estimated equation in log-odds, and three tasks that follow from it. Question 3 of UNISA’s January/February 2025 ECS4863 Advanced Econometrics paper, worth 16 marks, asks for all three: interpret the coefficients, derive the probability, and compute it for one person.
The question is described in our own words below; the answers are TheronGroup’s.
The question
A property development agency wants to know what drives the probability that a person in Gauteng buys a house. From a sample of 120 000 people, a logit model is estimated with H = 1 if the person buys a house and H = 0 otherwise. The explanatory variables are the natural log of income (lnY), the natural log of savings (lnS), and a dummy DP equal to 1 if the person has dependants in the household. The estimated log-odds are:
L̂i = −2,77 + 0,347 lnY + 0,122 lnS + 0,1 DP
se (0,8275) (0,0858) (0,004) (0,004)
3.1 (6 marks). Interpret the coefficients on lnY, lnS and DP.
3.2 (7 marks). Write down the expression for the probability that a person buys a house.
3.3 (3 marks). Find that probability for a person with income of R100 000, savings of R70 000 and no family members.
3.1 Interpreting the coefficients
The left-hand side is the log of the odds of buying a house, L = ln[P ÷ (1 − P)], so every coefficient is a change in log-odds, not in probability. The standard errors are read in order: intercept, lnY, lnS, DP. With 120 000 observations the ratio of coefficient to standard error is a z statistic, compared with 1,96 at the 5% level.
lnY: z = 0,347 ÷ 0,0858 = 4,04
lnS: z = 0,122 ÷ 0,004 = 30,50
DP: z = 0,1 ÷ 0,004 = 25,00
| Variable | Coefficient | z | Interpretation |
|---|---|---|---|
| lnY | 0,347 | 4,04 | Positive, as expected, and significant at 5%. Income is in logs and the dependent variable is log-odds, so 0,347 is the elasticity of the odds with respect to income: a 1% rise in income raises the odds of buying a house by about 0,35%, savings and dependants held constant. |
| lnS | 0,122 | 30,50 | Positive, as expected, and significant at 5%. A 1% rise in savings raises the odds of buying by about 0,12%, income and dependants held constant. A deposit comes out of savings, so the sign is the one theory predicts. |
| DP | 0,1 | 25,00 | Positive and significant at 5%. A person with dependants has odds of buying that are e^0,1 = 1,105 times those of a person without, about 10,5% higher, at the same income and savings. A household with children needs space and places more value on owning it. |
Where the marks are lost. A logit coefficient is not a change in probability. Writing “a 1% rise in income raises the probability of buying by 0,347” loses the interpretation marks, because the effect on the probability depends on where on the S-curve the person sits. Say “odds” or “log-odds”, and for the dummy give the odds ratio e^0,1 = 1,105 rather than reading 0,1 straight off as 10%.
3.2 The probability expression
The logit sets the log of the odds equal to the linear index, and the probability is found by solving for P.
ln[Pi ÷ (1 − Pi)] = Li = −2,77 + 0,347 lnYi + 0,122 lnSi + 0,1 DPi
Step 1, exponentiate both sides:
Pi ÷ (1 − Pi) = e^Li
Step 2, multiply out:
Pi = e^Li − Pi × e^Li
Pi + Pi × e^Li = e^Li
Pi × (1 + e^Li) = e^Li
Step 3, divide:
Pi = e^Li ÷ (1 + e^Li)
= 1 ÷ (1 + e^−Li)
So:
P̂i = 1 ÷ [1 + e^−(−2,77 + 0,347 lnYi + 0,122 lnSi + 0,1 DPi)]
This is the logistic distribution function. As L runs from minus infinity to plus infinity, P runs from 0 to 1 along an S-shaped curve, so a fitted probability can never fall outside the unit interval, which is the defect of the linear probability model that the logit removes. The same function gives the marginal effect of any regressor on the probability, ∂P ÷ ∂X = β × P × (1 − P), which is largest at P = 0,5 and shrinks towards either tail.
Where the marks are lost. Seven marks are for the derivation, not the final line. Show the odds, the exponentiation and the algebra that isolates P; a script that writes down P = 1 ÷ (1 + e^−L) with no steps earns a fraction of the marks. Substitute the estimated equation into the final expression, since the question asks for this model’s probability, not the general form.
3.3 The probability for one person
The model has no family-size variable, so “no family members” is read as no dependants: DP = 0. The regressors are natural logs.
Step 1, take natural logs:
lnY = ln 100 000 = 11,5129
lnS = ln 70 000 = 11,1563
Step 2, compute the logit:
L̂ = −2,77 + 0,347 × 11,5129 + 0,122 × 11,1563 + 0,1 × 0
= −2,77 + 3,9950 + 1,3611 + 0
= 2,5860
Step 3, compute the probability:
P̂ = 1 ÷ (1 + e^−2,5860)
= 1 ÷ (1 + 0,07532)
= 0,9300
Odds = e^2,5860 = 13,28, that is, 13,28 to 1
The probability that this person buys a house is 0,93, or 93%. With dependants (DP = 1) the logit rises to 2,6860 and the probability to 0,9362, a small change, because at a probability this high the S-curve is already flat.
Where the marks are lost. The logs are natural logs. Using base-10 logs gives log Y = 5 and log S = 4,8451, a logit of −0,4439 and a probability of 0,39, which is wrong by more than half. The other trap is the wording: the question mentions family members, but the model has only a dependants dummy, so state that DP = 0 is the reading taken before substituting.
Studying ECS4863
Logit and probit questions turn on the same three steps each time: log-odds interpretation, the probability transformation and a substitution. The ECS4863 course holds an eight-part course and four topic parts on time series, panel data and qualitative response, 19 examination sittings worked from 2016 to 2026, the assignments worked, and a master revision reference and formula sheet, at R3 000 a year.