Intergenerational persistence
Methodological details
In a nutshell
Intergenerational mobility is usually quantified using its opposite concept: intergenerational persistence. The latter captures the strength of the association in an outcome between generations. We focus on the association between individuals’ income and that of their parents. The measurement of intergenerational income persistence follows a two-step procedure:
- Aggregate the incomes earned by individuals and their parents over their respective life cycles. Income distributions may be transformed into logs or percentile ranks.
- Measure the strength of the association between the two variables with, e.g., a regression coefficient, a correlation coefficient, conditional mean outcomes, or probabilities of ending up in different income quantiles.
The main challenge to the measurement of intergenerational persistence is often to overcome data limitations. Indeed, available data rarely document both the income of individuals and that of their parents throughout their whole life cycle. However, measuring intergenerational persistence based only on a few years of income observation is known to introduce biases (Nybom and Stuhler 2017). Hence, variation in data restrictions between countries can bias cross-country comparisons of intergenerational persistence.
We address this issue with a systematic quantitative review of the literature, compiling estimates from 173 studies along with detailed information on their estimation parameters. This allows us to select the most comparable estimates for each country, and to assess how estimates vary with estimation parameters.
Concepts of intergenerational persistence
Intergenerational persistence measures are broadly categorized into absolute and relative measures. Absolute measures consist of computing average outcomes of individuals whose parents are at a given point on the income distribution (e.g., average income rank if parents are at the 25th income rank, probability of reaching the top income quintile from the bottom parent quintile). Relative measures quantify the expected change in child income for a given variation in parental income. The most commonly used are the intergenerational rank association (IRA) and the intergenerational elasticity (IGE), which are based on different transformations of the income variables:
Intergenerational rank association: Expected change in an individual’s percentile income rank for a 1 percentile rank increase in their parents’ income. This coefficient is obtained by regressing individuals’ percentile income rank on that of their parents, or by computing the correlation between these two variables.
Intergenerational elasticity: Expected percentage increase in an individual’s income for a 1% increase in their parents’ income. This coefficient is obtained by regressing individuals’s log income on that of their parents. Unless specific adjustments are made, this measure can only be estimated on individuals earning strictly positive income.
We use the intergenerational rank association as our headline measure, and also provide intergenerational elasticity estimates in the augmented dataset (see Download section of the website). Absolute mobility measures do not pertain to the scope of this analysis.
Identification and selection of research papers
Step 1: Papers on intergenerational mobility
We rely on the database of academic articles hosted by Semantic Scholar, which allows the automation of queries via an API for free. We use this API to get the reference of all papers which contain both the keyword “intergenerational” and either “mobility” or “persistence” in their title and/or abstract (when available). 5,987 Semantic Scholar references match this criterion. Through the API, we collect their Semantic Scholar identifiers and URL, their title, year of publication, authors, journal, and reference, and, when available, their abstract and open access PDF link.
Step 2: Papers with any mention of IRA and IGE
To narrow down the set of papers, we download the text corresponding to each reference in PDF format. We use the open access link indexed by Semantic Scholar when available. When no open-access link was available we use a standard Google search to find the corresponding PDF, and we rely on institutional access for articles that are not available in open access. Then, we search for any mention of the IRA and IGE estimators in the text in a flexible way (using the regular expressions intergenerational elasticity, intergenerational [a-z]+ elasticity, rank-rank, rank association, and rank correlation). This narrows down the set of relevant academic articles to 1,435.
Step 3: Identification of relevant papers
Finally, we go through each paper to verify whether or not it provides an original national-level IRA or IGE:
| IRA | IGE & IRA | IGE | Duplicate | No IRA or IGE |
|---|---|---|---|---|
| 35 | 64 | 264 | 217 | 855 |
This narrows down the set of relevant academic articles to 363. Indeed, most papers that mention the IRA and IGE estimators actually do not compute one, or not on national-level data. The table below documents the 10 most frequent keyword descriptions of such papers. Also, about 15% of references are actually duplicates, either earlier or versions or preprints of the same paper.
| Keyword | n |
|---|---|
| Education mobility | 157 |
| Discussion/review | 97 |
| Occupation mobility | 71 |
| Other estimator | 61 |
| Local estimates | 38 |
| Theory | 37 |
| Absolute mobility | 34 |
| Chetty et al. (2014) estimates | 33 |
| Wealth mobility | 33 |
| Health mobility | 25 |
Compilation of intergenerational persistence estimates
For the 363 articles identified above, we record each original national-level IGE and IRA as a distinct entry in a database, along with various estimation parameters.
Full set of variables
- Semantic Scholar identifier
- Corpus the reference comes from
- Title of the paper, as written in the PDF
- Keyword describing the paper when not included (e.g., theory, education mobility, …)
- Full name of the country in English
- Name/acronym of the dataset where parents’ income is observed
- Name/acronym of the dataset where individuals’ income is observed
- Type of the dataset where parents’ income is observed: Admin or Survey
- Type of the dataset where individuals’ income is observed: Admin or Survey
- Sample restrictions, separated by + signs if several
- Earliest (oldest) parents’ birth cohort
- Latest (youngest) parents’ birth cohort
- Earliest year of parents’ income observation
- Latest year of parents’ income observation
- Number of parents’ income observations
- Minimum number of parents’ income observations
- Youngest age of parents’ income observation
- Oldest age of parents’ income observation
- Average age at parents’ income observation
- Parents’ income definition
- If OLS, “observed” - If TSTLS: “predictor1+predictor2+…” - If specific TSTSLS cases “Label of the specification”
- If TSTSLS, first-stage R2
- Level of observation of parents’ income (father/mother/household/family)
- Share of the sample trimmed at the bottom of the parents’ income distribution
- Share of the sample trimmed at the top of the parents’ income distribution
- Individuals’ earliest (oldest) cohort
- Individuals’ latest (youngest) cohort
- Individuals’ earliest year of income observations
- Latest year of individuals’ income observation
- Number of children’s income observations
- Minimum number of children income observations
- Youngest age at individuals’ income observation
- Oldest age at individuals’ income observation
- Average age at individuals’ income observation
- Individuals’ income definition
- Individuals’ income observation level (individual/household/family)
- Estimation on males only, females only, or both
- Share of the sample trimmed at the bottom of the child income distribution
- Share of the sample trimmed at the top of the child income distribution
- Number of observations used in the regression
- Type of estimation (IGE/IRA)
- Value of the coefficient
- Standard error of the coefficient
- p-value of the coefficient
- List of controls (separated by + signs if several) / functional form deviation
- Table of the paper where the coefficient comes from
- Whether the article provides estimates at a finer geography than the country level (T/F)
- Whether article provides estimates per race/immigrant/ethnic/origin group (T/F)
- If the articles contains a table reviewing coefficients in the literature, include table number here
- Does the article provide transition matrices? (T/F)
- Is the paper not yet published in a journal or book?
- Year of publication of the article
- Journal where the paper is published
Current coverage:
- 100% of articles including only an IRA (35 papers)
- 100% of articles including an IRA and an IGE (64 papers)
- 27% of articles including only an IGE, drawn at random (70 papers)
- 4 additional papers that were not on Semantic Scholar
Selection of comparable estimates across countries
The objective of this selection process is to maximize comparability between countries. For some countries, this may not pick up the most up-to-date/best-practice estimate if it is less comparable to the rest of the literature than an ealier one.
The reliability of intergenerational mobility estimates is very much dependent on the use of high-quality data, which are not available in many countries. Hence, researchers are often constrained to estimate intergenerational persistence on data that are sub-optimal in ways that can vary quite a lot from one country to another. For our aim of global comparison, this generates a trade-off between coverage and comparability. We combine a set of minimum requirements and ordered preference rules to narrow down the set of estimates to the most comparable one for each country.
Minimum requirements:
- No co-residence restriction
- Aged at least 25 at income observation
- If parents’ income predicted, at least occupation as a predictor
Preference rules:
- Aged at least 30 at income observation
- Parents income observed > predicted
- All > males > females
- Born within 1960 - 1980
- Both types or Admin data > Survey data
- Parents’ income household > top earner > father > mother
- Child income individual > household > spousal
- Income > earnings > wage
- Highest number of parents’ income observations
- Closest to age 35 at income observation
- Parents’ income predicted with the most variables
- Simplest income definition
- Birth cohort closer to 1970
- Lowest number of sample restrictions
- Parents average age at income observation closest to 35
- Lowest number of control variables
- Largest sample size
Resulting estimates are mapped in the World view section of the website.