Overdispersed and Markovian Children

This paper challenges the common assumption that children's genders follow a simple independent coin-toss model by demonstrating through data that birth genders exhibit slight imbalances, family-specific variations, sequential dependencies, and overdispersion, while also illustrating how sample sizes impact statistical detection power.

The Gist

The Big Picture: Are We Just Flipping Coins?

Imagine you are walking down the street or looking at your own family. You see boys and girls. It feels like nature is flipping a fair coin for every baby: Heads = Boy, Tails = Girl. If this were true, every family would be a perfect roll of the dice. If you had 8 kids, you'd expect a mix, and families with only boys or only girls would be incredibly rare.

But Nils Lid Hjort, a statistician from Oslo, took a giant magnifying glass to some very old data from 19th-century Germany (Saxony). He looked at nearly 38,500 families that had at least 8 children. What he found was that the "coin" nature flips isn't actually perfectly fair, and it doesn't behave the same way for every family.

Here are the four main discoveries from the paper, explained simply:

---

1. The Coin is Slightly Weighted (The "0.485" Secret)

The Analogy: Imagine a casino where the house always wins just a tiny bit.
The Reality: We often think the chance of having a girl is exactly 50/50. Hjort's data shows it's actually about 48.5%.
Why it matters: To prove this tiny difference (48.5% vs. 50%) isn't just a fluke, you need a massive amount of data. It's like trying to tell if a coin is slightly bent. If you flip it 10 times, you might get 6 heads and 4 tails and think, "No big deal." But if you flip it 15,000 times and get that same ratio every time, you know the coin is bent. The paper shows that with modern data, we can finally be 100% sure the "girl coin" is slightly heavier than the "boy coin."

2. The "Family Personality" (Overdispersion)

The Analogy: Imagine a classroom of students taking a test.

• The Simple View (Binomial): Every student has the exact same chance of getting a question right (say, 50%).
• The Real View (Overdispersion): Some students are naturally "lucky" or "unlucky." In some families, the "girl probability" is naturally higher (maybe 55%), and in others, it's lower (maybe 40%).

The Discovery: The data showed way more "all-boy" and "all-girl" families than a simple coin flip would predict.

• Simple Math: If you flip a fair coin 8 times, getting 8 heads is very rare.
• Real Life: It happens more often than math predicts.
  Why? Because some families just have a "genetic tendency" toward one gender. It's not that the coin changes during the family; it's that every family has its own unique coin that is slightly different from its neighbor's. Hjort calls this Overdispersion—the data is "spread out" more than the simple model allows.

3. The "Streak" Effect (Markovian Children)

The Analogy: Imagine a basketball player on a hot streak. If they make a shot, they are slightly more likely to make the next one.
The Reality: Hjort wondered: Does having a boy make it slightly more likely to have another boy?
The Discovery: Yes, but only a tiny bit. If a family just had a girl, the chance of the next child being a girl goes up slightly (from 48.5% to maybe 49%).
The Catch: The data didn't tell him the order of the children (e.g., Girl-Boy-Girl), only the total count. So, he had to use a computer to simulate millions of possible birth orders to see if a "streak" model fit the data better than the "random family coin" model. It turned out the "streak" effect exists, but it's very subtle.

4. The Power of Big Numbers (Sample Size)

The Analogy: Finding a needle in a haystack.
The Reality: The paper spends a lot of time talking about Sample Size.

• If you look at a small group of families (say, 500), the "all-boy" families might just look like random luck. You can't prove anything.
• But if you look at 38,000 families (like in this study), those "all-boy" families stop looking like luck and start looking like a pattern.

Hjort explains that with small data, you might miss the truth. With huge data, you can detect tiny, almost invisible differences. It's like hearing a whisper in a quiet room vs. a whisper in a stadium. In a stadium (huge data), you need a very loud whisper to be heard, but once you hear it, you know it's real.

The "Royal Flush" Families

The paper notes something funny: There are more families with 8 girls or 8 boys than simple math predicts.

• Simple Math: Predicts about 117 all-girl families.
• Real Data: Found 161.
• The Fix: When you account for the fact that some families are "girl-heavy" and some are "boy-heavy" (the Overdispersion), the math finally matches reality. The "Royal Flush" (all one gender) happens more often because some families are just naturally stacked that way.

Conclusion: Nature is Messy (and Interesting)

The paper concludes that while the world of babies looks like a simple coin toss at first glance, it's actually a complex mix of:

1. A slightly biased coin (slightly more boys).
2. Different coins for different families (some families lean toward girls, others toward boys).
3. Tiny streaks (having a boy makes the next one slightly more likely to be a boy).

The Takeaway: We are all the result of a "hierarchical cascade" of coin tosses, but the coins aren't perfect, they aren't all the same, and they sometimes remember what they just flipped. And to see these tiny secrets, you need to look at a lot of data.

Technical Summary

1. Problem Statement

The paper investigates the statistical distribution of gender in human families. While the "Natural Start Assumption" suggests that gender is determined by independent, balanced coin tosses ($p = P(\text{girl}) = 0.50$), empirical data suggests deviations from this simple Binomial model. The author aims to:

1. Demonstrate that the true probability of a girl is slightly less than 0.50.
2. Quantify overdispersion: the phenomenon where the variance in the number of girls per family exceeds what is predicted by a standard Binomial distribution with a fixed probability $p$.
3. Investigate whether the gender of a child depends on the gender of the preceding child (Markovian dependence).
4. Analyze how sample size influences statistical power and the detection of these subtle effects.

The primary dataset used is the Geißler (1889) data from Saxony, comprising $n = 38,495$ families, each with at least $m=8$ children.

2. Methodology

Hjort employs a combination of classical statistical inference, simulation-based likelihood estimation, and model comparison.

• Hypothesis Testing for $p$:

  • Tests the null hypothesis $H_0: p = 0.50$ against the alternative $p \approx 0.485$.
  • Calculates required sample sizes ($n$) to achieve specific significance levels ($\alpha$) and statistical power ($1-\beta$) to detect the small deviation from 0.50.
  • Uses the test statistic $V_n = \frac{z - np_0}{\sqrt{np_0q_0}}$ (standard normal approximation).

• Overdispersion Analysis (Beta-Binomial Model):

  • Assumes the probability $p$ varies across families according to a Beta distribution $Beta(a, b)$.
  • Derives the marginal distribution of the number of girls $y$ as a Beta-Binomial distribution.
  • Estimates parameters using the method of moments and minimum Chi-squared ($\chi^2$) fitting.
  • Calculates Pearson residuals ($P(y) = \frac{N(y) - E(y)}{\sqrt{E(y)}}$) and goodness-of-fit statistics ($Z = \sum P(y)^2$) to compare the Binomial vs. Beta-Binomial models.
  • Estimates the extrabinomial standard deviation $\sigma_0$ (the variation of $p$ across families) and constructs confidence curves for it.

• Markovian Dependence Analysis:

  • Proposes a Markov chain model where the gender of the $i$-th child depends on the $(i-1)$-th child.
  • Defines a transition matrix with a tuning parameter $k$ (where $k=1$ implies independence).
  • Challenge: The Geißler data only provides counts of girls/boys, not the order of birth.
  • Solution: Uses Simulated Log-Likelihood. The author simulates $10^5$ gender paths for various $k$ values, aggregates the resulting counts of girls, and fits the implied distribution $f_3(y, \theta)$ to the observed data to estimate the log-likelihood profile.

• Sample Size and Power Analysis:

  • Simulates the distribution of the overdispersion test statistic $W_n$ (ratio of observed variance to model variance) under the null hypothesis for varying $n$.
  • Determines the sample size required to distinguish between a pure Binomial model and an overdispersed model with high power.

3. Key Contributions

• Quantification of Bias: Confirms that $P(\text{girl}) \approx 0.484$ (ratio of boys to girls $\approx 1.062$), a deviation requiring massive sample sizes ($n > 14,000$) to detect with 95% power.
• Evidence of Overdispersion: Demonstrates that the Binomial model is insufficient. The data exhibits "heavy tails" (more all-boy and all-girl families than predicted). The Beta-Binomial model significantly improves the fit, estimating that the underlying $p$ varies across families with a standard deviation $\sigma_0 \approx 0.054$.
• Markovian Dependence: Successfully fits a Markov model to aggregate count data (without order information) using simulation. Finds a weak but statistically significant positive correlation ($\rho \approx 0.044$) between consecutive children of the same gender.
• Family Size Dynamics:
  • Finds that $P(\text{girl})$ remains relatively constant across family sizes ($m=1$ to $12$), contradicting some previous literature.
  • Finds that the level of overdispersion ($\sigma_0$) increases with family size, suggesting that larger families may have stronger dependencies or wider variations in underlying biological probabilities.
• Methodological Insight: Highlights the critical role of sample size in statistical detection. Small deviations from theoretical models (like $p=0.50$ or $\sigma_0=0$) are invisible in small samples but become highly significant with large datasets (e.g., $n=38,495$).

4. Results

• Parameter Estimates:
  • $P(\text{girl}) \approx 0.484$.
  • Overdispersion parameter $\sigma_0 \approx 0.054$ (95% CI: $[0.049, 0.058]$).
  • Markov correlation parameter $k \approx 0.956$ (implying a correlation of $\approx 0.044$ for same-gender siblings).
• Model Fit (Geißler $m=8$ data):
  • Binomial Model: Poor fit. Goodness-of-fit statistic $Z_1 = 159.41$ (far exceeding $\chi^2_7$ critical values).
  • Beta-Binomial Model: Excellent fit. $Z_2 = 13.55$ (close to $\chi^2_6$ expectations).
  • Markov Model: Excellent fit. $Z_3 = 13.17$.
  • Both the Beta-Binomial and Markov models explain the "excess" of all-boy and all-girl families (e.g., observed 264 all-boy families vs. Binomial prediction of 192).
• Sample Size Thresholds:
  • To detect overdispersion ($\sigma_0 \approx 0.05$) with 95% power, approximately 4,000 families of size 8 are required.
  • To detect $p \neq 0.50$ with 95% power, approximately 14,433 children are required.

5. Significance

• Statistical Theory: The paper serves as a pedagogical and rigorous example of overdispersion, illustrating how real-world data often violates the independence assumption of the Binomial distribution. It demonstrates the utility of hierarchical models (Beta-Binomial) and simulation-based inference (Simulated Log-Likelihood) when data is aggregated.
• Biological Insight: It provides robust statistical evidence that human sex determination is not a simple, identical coin toss for every family. There are subtle biological or environmental factors causing families to cluster toward one gender, and these effects are slightly amplified in larger families.
• Historical Context: The author revisits 19th-century data (Geißler, Fisher, Arbuthnot) to show how modern statistical tools can extract deeper insights from historical datasets, validating early intuitions about "imbalanced coins" while refining the magnitude of the bias.
• Practical Implication: The work underscores that in the era of "Big Data," even minute deviations from theoretical norms (like a 1.5% shift in gender ratio or a 4% correlation in siblings) can be detected and must be accounted for in accurate modeling.