Healthcare Premium Analysis

MCA Semester I · 26MCAC103 Business Statistics using Python

What decides the price of a health insurance policy?

A statistical analysis of 1,676 premium quotes. Age sets most of the price, cover sets some of it, and gender and BMI set almost none.

MIT ADT University · College of Management and Computer Applications

Team: Aditi Naik · Aaisha Khan · Shrinidhi Hegde · Yogesh Kurdule · Siddharth Mane

Guide: Prof. Pradnya Mulay

r = 0.80

Age drives the price

Flat until 35, then it climbs fast. Age carries 88% of the importance in the best model.

Rs 416

Bigger cover costs less per lakh

Rs 1 crore cover costs Rs 416 per lakh. Rs 5 lakh costs Rs 3,614.

p = 0.95

Gender does not matter

Men and women pay the same. BMI has no link either once age is removed.

R² 0.995

Random Forest predicts best

Average error of only Rs 296. Log-linear is the one that is easiest to explain.

The question

What goes into a yearly premium?

The premium is the yearly price of a policy. We wanted to know which facts about a person change it, and whether a computer model can predict it.

  • Does age change the price?
  • Do gender, height and weight change it?
  • Does a bigger cover (sum insured) change it?
  • Can we predict the price with a computer model?

Data and method

1,676 quotes, six steps

1,676price quotes (rows)
94different ages
9cover amounts
744different people

What we did, step by step

  1. Cleaned the dataFilled 13 missing values
  2. Described itAverages and charts
  3. Tested ideasHypothesis tests
  4. Built modelsTo predict the price
  5. Compared pricesUsing index numbers
  6. Gave adviceRecommendations

Key words

Six terms in plain words

Average (mean)
Total divided by count.
Median
The middle value.
Correlation (r)
How strongly two things move together, from 0 to 1.
p-value
Below 0.05 means the result is real, not luck.
R-squared (R²)
How much a model explains. 1 is perfect.
Index
A comparison where the starting group equals 100.

Finding 1

Price is low until 35, then rises fast

The average premium is Rs 30,300 and the middle premium (median) is Rs 18,225. The average is higher because a few very costly quotes pull it up.

Average premium by age group

Rs per year. Bar length is proportional to the value.

0–17
Rs 13,773
18–25
Rs 13,488
26–35
Rs 13,506
36–45
Rs 15,645
46–55
Rs 26,550
56–65
Rs 38,887
66–75
Rs 58,604
76–85
Rs 83,557

What moves with price?

Correlation with premium (0 to 1)

Age
r = 0.80
Cover
r = 0.27
BMI
r = 0.03*

*After removing the effect of age. BMI (Body Mass Index) = weight ÷ height², with height in metres.

Age is a very strong link, cover is weak to medium, and BMI shows no link.

Hypothesis tests

Which differences are real?

A hypothesis test starts by assuming nothing is different, then checks whether the data proves otherwise. The p-value is the chance the result is just luck. Below 0.05 is a real difference. Above 0.05 is no proof of a difference.

QuestionAnswerp-value
Do men and women pay differently?No0.95
Does age group change the price?Yes, a lot< 0.001
Does cover change the price?Yes, but less< 0.001
Do older people pay more?Yes (r = 0.80)< 0.001
People 65+ pay about Rs 69,777 on average. Under 65, the average is Rs 18,803.

Prediction

Four models, one winner for each job

A model is a formula that predicts price from age, cover and the rest. We built four and compared them to find the one that predicts best and is still easy to explain.

Models 1 and 2 · Regression

Straight line

Finds how much price changes when an input changes. The problem is that price grows in a curve, not a line.

+1 year of age ≈ Rs 1,028 more
+1 lakh of cover ≈ Rs 215 more
R² = 0.83
Model 3 · Regression

Log-linear (curved)

Fixes the curve by using age groups and the log of price. Easy to explain in plain words.

R² = 0.97
Doubling the cover raises price about 23%
Model 4 · Machine learning

Random Forest

Builds 300 small decision trees. Each asks yes/no questions such as age above 50 or cover above 10 lakh. All trees vote and the average is the price. Needs no formula and handles curves.

R² = 0.995
Average error only Rs 296
How we use them

Quote with one, explain with the other

Random Forest is the most accurate. Log-linear is second and the best one we can still explain to people. The straight line is the weakest.

Quote: Random Forest
Explain: log-linear

Which model won?

Test score (R²). Closer to 1 is better.

Straight line
0.809
Log-linear
0.940
Random Forest
0.995

What really decides the price

Importance in the Random Forest

Age
88.32%
Cover
11.46%
BMI
0.15%
Gender
0.07%

Age 88%, cover 11.5%, BMI and gender almost 0. This matches the hypothesis tests.

Index numbers

Comparing every age with one starting group

An index sets the starting group to 100 and shows every other group as a number against it. The data has no dates, so we compare across age instead of across time.

Index = (group price ÷ starting group price) × 100

Example: ages 46–55 pay Rs 14,693 and ages 18–25 pay Rs 7,015. So 14,693 ÷ 7,015 × 100 = 209, which means about 2 times the price.

The price is the same up to age 35. Ages 46–55 pay about 2 times, and ages 76–85 pay about 7.7 times.

Price index by age

Ages 18–25 = 100

0–17
100
18–25
100
26–35
100
36–45
115
46–55
209
56–65
322
66–75
547
76–85
773

Finding 2

Bigger cover is cheaper per lakh

When the cover doubles from Rs 5 lakh to Rs 10 lakh, the price rises only 35%.

Rs 5 lakh
Rs 18,071
Rs 10 lakh
Rs 24,359
Rs 1 crore
Rs 41,616

Average yearly premium in Rs.

Recommendations

What we recommend

  1. Buy cover before age 35Quotes are lowest.
  2. Choose a bigger coverIt is cheaper per lakh.
  3. Price only by age and coverGender and BMI do not change the price.
  4. Quote with Random Forest, explain with log-linearOne for accuracy, one for clarity.
  5. Collect health, smoking and date dataThe current data cannot show these effects.

Limits of our work

What to keep in mind

  • The data looks like a fixed price list, so accuracy is very high.
  • The same people repeat, so p-values are a guide.
  • There are no health or smoking details.
  • There are no dates, so no forecast over years.

Conclusion

Price depends mostly on age, then cover

It is flat to 35 and reaches about 7.7 times by ages 76–85.