Behavioural Microsimulation and Policy Analysis Microsimulation as a tool for policy formulation and evaluation Reykjavik, November 11 2016 Behavioural Microsimulation and Policy Analysis Ugo Colombino Dept. of Economics and Statistics, Torino, Italy

Outline Microsimulation was born behavioural Microsimulation made it as non-behavioural Behavioural strikes back What kind of behavioural model? Two examples

Microsimulation was born behavioural Guy Orcutt, A new type of socio-economic system, Review of Economics and Statistics,1957 “The most distinctive feature of this new type of model is the key role played by actual decision making units of the real world such as the individual, the household and the firm… …This new type of model consists of various sorts of interacting units which receive inputs and generate outputs.

Microsimulation “made it” as non-behavioural The times were not ready for what Orcutt had in mind, despite early very interesting experiments e.g. Orcutt et al., Microanalysis of socioeconomic systems: A simulation study, New York, 1961. During the ’60s and ’70s, big advances in computational power and data collecting management made it possible in the following decades to build large «arithmetic» microsimulation models

Behavioural strikes back The re-marriage of microsimulation with behavioural modelling at the policy level starts with the new millenium… emergence of new policy issues where questions about behaviour are crucial… …especially issues related to incentive effects of taxes and transfers …development of appropriate microeconometric tools

Focus on labour supply I focus on labour supply choices: Most common applications of microsimulation are tax-benefit reforms. The main interest is in the effects on incentives to work more or less, with the implications on tax revenue, efficiency, income distribution etc. However, most of what I say carries over to other choices: retirement, consumption, savings, child-care, educational choices…

What kind of behavioural model? “Sufficient statistics” Structural

The “sufficient statistics” approach If the reform consists of “marginal” changes in wages or incomes, the static microsimulation itself might be sufficient to make some inferences on incentives behaviour welfare

The “sufficient statistics” approach (incentives) Arithmetic microsimulation produces estimates of local changes in the budget sets due to taxes, transfers, deductions etc. Local changes in the budget sets might indicate the direction of change in incentives (and suggest changes in behavior)

The “sufficient statistics” approach (behaviour) Suppose the reform consists in a small changes in net wage rate W and net unearned income Y. For each household, the static microsimulation computes percentage changes %W and %Y. If we have some local estimate of wage elasticity ηW and of income elasticity ηY of labour supply, the behavioural response can be approximated by ηW(%W) + ηY(%Y) = % change in labour supply

The “sufficient statistics” approach (welfare) Moreover, it turns out that the change in the budget hdW +dY can be interpreted as the change in money-metric utility Note: h is the current value of hours of work

What if the reform is going from A to B? Local information might be misleading There is no simple way to represent the reform as a marginal change in wage rate and income A disposable income B hours of work

What if the reform is going from A to B? To tell that the «green» guy doesn’t work under rule A but works und rule B we need an explicit representation of preferences and budget constraint disposable income A Choice under rule B B Choice under rule A hours of work

The structural approach More precisely (e.g. Marshack, Hurwicz, Lucas): we need a model that contains explicit representation of: Preferences (invariant w.r.t. policies) Constraints (that change in a known way w.r.t. policies)

The Conditional Logit Model The CLM provides a very convenient approach for specifying, estimating and simulating a structural model It can be applied: With truly discrete alternatives, e.g. modelling the choice between economics and statistics As a discrete approximation to a – in principle – continuous choice

Approximating a continuous opportunity set… h

…with a discrete one c h

…with a discrete one c h

…with a discrete one c (h2, c2) (h1, c1) (h7, c7) h

The choice probability Random utility: ε is extreme-value distributed θ = preference parameters Behavioural rule: Choice probability (McFadden): where: c = disposable income = f(gross incomes at hi) f( ) = tax-benefit rule that transforms gross incomes into disposable income

Convenient features of the CLM However complicated the utility v( ) and the tax-benefit rule f( ) might be, the choice probability has always the same form This feature makes both estimation and simulation very convenient h could be a vector, e.g. the labour supply choices of different household members In general: the same approach carries over to simultaneous choices, e.g. labour supply and educational choices, fertility and child care mode etc.

Extensions of CLM: The RURO Model There might be many opportunities of each type:

Extensions of CLM: The Random Utility Random Opportunities (RURO) Model Choice probability in the RURO model:

A road-map for CLM Specify the set of discrete alternatives Specify v( ) Estimate θ Validate the model Simulate a reform f1 Evaluate a reform f1

Example 1: simulating a specific reform Shifting Taxes from Labour to Property A Simulation under Market Equilibrium Flavia Coda Moscarola - CeRP-Collegio Carlo Alberto Ugo Colombino - University of Turin Francesco Figari - University of Insubria Marilena Locatelli - University of Turin

Motivation More efficiency: reduce taxes on labor & increase taxes on property More equity: increase income support for low-income households

Method EU-SILC 2008 Italy EUROMOD RURO model Fiscal neutrality Simulation under labour market equilibrium We compare arithmetic and behavioral simulation

Pre- and Post-reform budget (earnings)

Effects on labour supply

Effects on income distribution

Example 2: identifying an optimal reform Five crossroads on the way to basic income: An Italian tour Ugo Colombino - University of Turin Italian Economic Journal, 2015

Motivation Look for the best reform of income support in a 5-dimension space: universal vs categorical policy mean-tested vs non mean-tested level of protection as % of poverty line lump-sum transfers vs wage subsidies flat tax vs progressive tax

Main types of reform CBI = Conditional (means tested) Basic Income UBI = Unconditional (non means tested) Basic Income IWB = In-Work Benefits CBI+IWB UBI+IWB

Method EU-SILC 2008 Italy EUROMOD RURO model Fiscal neutrality Simulation under labour market equilibrium Iterate the simulation until an objective function (e.g. a social welfare function) is maximized

Gini-Sen-Poverty rate The two best policies according to different welfare criteria (Pa means the policy P cover a% of the Poverty line) Utilitarian Gini-Sen Gini-Sen-Poverty rate IWB75 UBI33+IWB75 UBI100 IWB50 UBI50 UBI75

Effects of the two best policies according to Gini-Sen Policy Female employment Male employment Fisc. Neutr. TMTR Winners UBI33+IWB75 - 0.41% - 0.25% 51% 69% UBI50 - 0.80% 66%

CBI vs UBI: the «poverty trap» Policy Female employment Male employment Average transfer per household UBI75 - 1.77% - 0.41% 9800 CBI75 - 1.67% - 0.25% 3300

Recent exercises that identify optimal tax rules Blundell & Shephard, Employment, Hours of Work and the Optimal Taxation of Low-Income Families, Review of Economic Studies, 2012 Aaberge & Colombino, Designing Optimal Taxes with a Microeconometric Model of Household Labour Supply, Scandinavian Journal of Economics, 2013

Surveys on microsimulation and policy analysis O’Donoghue (ed.), Handbook of Microsimulation Modelling, Emerald 2014 Figari & Paulus & Sutherland, Microsimulation and Policy Analysis, ISER wp No. 2014-23 Bourguignon & Spadaro, Microsimulation as a tool for evaluating redistribution policies, Journal of Economic Inequality 2006 Creedy & Duncan, Behavoural microsimulation with labour supply responses, Journal of Economic Surveys 2002

Technical stuff

Sufficient statistics: welfare

1) Specify the set of alternatives Two ways: a) fixed alternatives b) alternatives sampled from a density Alternative (b) is more cumbersome but more efficient Few (3 – 6) alternatives usually work well for estimation For simulation, a larger set might be desirable See Aaberge & Colombino & Wennemo, Evaluating Alternative Representations of the Choice Sets in Models of Labour Supply, Journal of Economic Surveys, 2009

2) Specify v( ) Two most common specifications: quadratic or log-quadratic (linear-in-parameters) Box-Cox (easy to interpret; convenient to test or impose regularity conditions)

3) Estimate θ A microsimulation algorithm (e.g. EUROMOD) computes, for each household, the net income available at each alternative, according to the current tax-benefit rule f0:

3) Estimate θ (cont’d) Then, if household is observed at alternative K, the probability that this household makes this choice given parameters θ is: These probabilities are the components of the likelihood function.

3) Estimate θ (cont’d) Maximum Likelihood Gauss, MatLab, Gretl, R etc. provide (also) pre-canned procedures Also STATA but only with linear-in-parameters v( ); otherwise you must write your own program (maybe use MATA plus the in-built optimization procedure)

4) Validate the model For complex behavioural models,the main validation procedure is out-of-sample prediction… …e.g. predicting the effects of an implemented policy with a model estimated under the pre-policy regime.

4) Simulate a reform f1 As a basis, for each household, compute:

4) Simulate a reform f1 (cont’d) Suppose we are interested in simulating hours of work… Method 1 – Simulate the chosen alternative For each household, draw Then compute The chosen alternative is Then look at the corresponding value of h.

4) Simulate a reform f1 (cont’d) Method 2 – Simulate expected chosen values:

4) Simulate a reform f1 (cont’d) Fiscal neutrality In standard cases, we want to simulate the effects of a specific reform proposal, including its effect on the public budget Sometimes we are interested in a class of reforms and we want to simulate the effects under the constraint of fiscal neutrality… it requires iterating the simulation over the parameters of the reform… there might be more than one solution

4) Simulate a reform f1 (cont’d) Equilibrium Effects Linking household microsimulation to CGE (e.g. Savall in O’Donoghue (ed.) Handbook of Microsimulation Modelling, Elsevier, 2015 Simulating RURO models under labour market equilibrium (e.g. Colombino, A new equilibrium simulation procedure with discrete choice models, International Journal of Microsimulation, 2013

5) Evaluate a reform f1 Measuring welfare Two methods to compute the utility level attained by the household: Method 1 Pick the simulated maximum utility level (corresponding to the simulated) optimal choice, i.e.

5) Evaluate a reform f1 (cont’d) Measuring welfare Method 2 Compute the expected maximum utility (McFadden):

5) Evaluate a reform f1 (cont’d) Measuring welfare Aggregating the household welfare levels into a social welfare index…many problems and approaches… Interpersonal comparability Efficiency and equity effects etc. See for example Aaberge & Colombino, Labour Supply Models in O’Donoghue (ed.) Handbook of Microsimulation Modelling, Elsevier, 2015