THE DYNAMICS OF POLITICAL THE DYNAMICS OF POLITICAL COMPETITIONCOMPETITION

  

Michael LaverMichael Laver

Trinity College DublinTrinity College Dublin

Party competition is a dynamic systemParty competition is a dynamic system

• Most normal people see political competition as a system in continual motion

• Most see this continual motion as normal, not a sign of a system “out of equilibrium”

• Most see political dynamics as endogenous– The output of cycle c feeds back as input into cycle c + 1

Most models of party competition are staticMost models of party competition are static

• A model identifies key inputs into the system and describes interactions between these

• Specify the model, solve for equilibrium . . .

• . . . describing forecast outputs, conditioned on exogenously determined inputs

• But the whole concept of equilibrium goes against most normal people’s instincts about politics

With traditional static models . . .With traditional static models . . .

• Equilibria change only in response to unforeseen shocks

• Politics thus appears to mutate unpredictably, not evolve endogenously

• Associated with this is an assumption of parties as unitary actors – whose identity is given exogenously

• In my view this “equilibriumist” tradition has run out of theoretical steam

Consider this opinion poll series of party Consider this opinion poll series of party support in Irelandsupport in Ireland

Time (Year/quarter)

972

963

951

942

933

924

921

912

903

894

891

882

873

864

861

Pa

rty s

up

po

rt (

%)

70

60

50

40

30

20

10

0

FF

FG

LAB

PD

WP/DL

Which traditional model can explain …Which traditional model can explain …

• levels and variation in party sizes

• volatility of party sizes over time

• which party loses when another gains

• how parties’ internal rules affect their competitive position

• … while keeping the “shape” of the party system endogenous

It takes a dynamic model to do thisIt takes a dynamic model to do this

• Traditional game theoretic technology is not (yet?) up to the task …

• … of modelling “massively parallel” dynamic interaction between large numbers of individual decision makers

It works right now using agent-based It works right now using agent-based modellingmodelling

• Agents use simple adaptive rules to condition what they do at cycle c +1 on the history of the system up to cycle c

• Simple rules applied recursively in massively parallel interactions can build beautiful behaviours

• Of course we are talking simulations here, but do these give any less intuition than 2- or 3-person game theory?

THE MODEL: 1THE MODEL: 1

• Assume the classical spatial representation of voter preferences and party policy positions

• As with the classical spatial model, assume two “breeds” of agent – voters and party leaders

• Voters’ policy preferences are currently taken as primitives; I will later endogenize these

THE MODEL: 2THE MODEL: 2

• Party leaders’ policy positions evolve continuously in response to voter preferences and the positions of rival leaders, in the following way:

1. Voters (tend to) support the closest party

2. Leaders adapt party policy positions, given the party support profile of all voters

3. The system evolves. Go to 1

• This loop runs forever

THE MODEL: 3THE MODEL: 3

• Currently use exogenously determined party leaders and decision rules; I later endogenize these

– By specifying when agents change breed from voter to party leader (thereby giving birth to new parties)

– By allowing party leaders to replace less successful decision rules with more successful ones

• Voter support for party leaders deterministic or stochastic

• Initial random-normal distribution of voter ideal points; more complex distributions can easily be designed …

• … or read in from real data

ADAPTIVE DECISION RULES FOR ADAPTIVE DECISION RULES FOR PARTY LEADERSPARTY LEADERS: 1: 1

• STICKER. Never change policy– “ideological” party leader

• AGGREGATOR. Set party policy on each policy dimension at the mean preference of all party supporters– “democratic” party leader

• i-HUNTER. If last policy move increased support, make same move; else, make a random move in opposite direction– “insatiable” Pavlovian vote hunter (NB precludes steady state)

ADAPTIVE DECISION RULES FOR ADAPTIVE DECISION RULES FOR PARTY LEADERSPARTY LEADERS: 2: 2

• s-HUNTER. As i-Hunter except, if last policy move left support stable, stand still– “satiable” Pavlovian vote hunter (NB allows steady state)

• PREDATOR. Identify largest party. If this is not you, make policy move towards largest party

THE PROGRAMTHE PROGRAM

• I have programmed various versions of this model using NetLogo 1.3

• For the programs, contact me

• For NetLogo 1.3, download free from:– http://ccl.northwestern.edu/netlogo/

• This is a really powerful and friendly system

RESULTS: 1RESULTS: 1

• The i-HUNTER rule for party leaders is very successful at finding high voter support densities

– Note that this uses very little information about the “geography” of the policy space

– “Win-stick, lose-shift” Pavlovian adaptation (Nowack and Sigmund)

• i-HUNTER beats o-HUNTER

RESULTS: 2RESULTS: 2

• All-AGGREGATOR systems reach steady state

• Conjecture: the result of the locally “granular” quality of the ideal points of a finite set of agents …

• … as opposed to the “infinite” smooth density maps of the traditional spatial model

• But these steady states are easy to perturb and highly path dependent (cf chaos results)

RESULTS: 3RESULTS: 3

• Hunters hunt for support in centrist positions but do not go to the dead centre of the space

• This is realistic, and solves what is a big problem for the traditional spatial model

• The center is a dangerous place in an all-HUNTER system

– cf Schofield (JTP 2003) and valence competition

x dimension

35302520151050-5-10-15-20-25-30-35

y d

ime

nsi

on

35

30

25

20

15

10

5

0

-5

-10

-15

-20

-25

-30

-35

RESULTS: 4RESULTS: 4

• Hunters and lone Predators beat Aggregators

• But 2+ Predators attack each other and don’t necessarily beat Aggregators

• Hunters beat Predators!

– Unexpected. Simple Pavlovian adaptation very effective against superficially more “rational” predatory behaviour

EXTENSIONS: 1EXTENSIONS: 1

• “Switching horizon” below which voters do not see – or react to – small differences in party distances – Damps “excessive” support switching

• Stochastic voter support. Voters switch party affiliation with a probability that is a function of:

– “distance gain” from switching, d, (ideal point to closest party distance minus ideal point to current party distance);

– “switching horizon”, h, reflecting the agent’s ability to “see” a given distance gain

– “switching sensitivity”, s, reflecting the agent’s willingness to switch given a particular “seen” distance gain

EXTENSIONS: 2EXTENSIONS: 2• THE BIRTH OF NEW PARTIES

– If some voter finds no party leader “close enough” in some parameterised sense, s/he changes breed to become a new party leader

– Note there is no concept of “entry” in a fully endogenous dynamic model

– All agents are already in the system, they just change breed

– The concept of party system entry is an old-fashioned “equilibriumist” notion!

– And in the real world new parties are never formed by politicians who “enter” the system from a parallel universe – or even from another country!

THE MODEL THE MODEL && THE THE ““REALREAL” ” WORLDWORLD

• “Statistical signatures” (Scott Moss 2001). Model outputs should have the same statistical distributions as those of the “target” system in the real world

• In this case, e.g. . . .

– cross-sectional variation in party sizes and N of parties– time series variation in party sizes and N of parties– size of largest/smallest party– Etc.

• So let’s model a real party system time series– Though what is “really” real?

Ireland 1986-97: raw opinion poll dataIreland 1986-97: raw opinion poll data

• These poll data subject to both random error and systematic bias

Time (Year/quarter)

972

963

951

942

933

924

921

912

903

894

891

882

873

864

861

Pa

rty s

up

po

rt (

%)

70

60

50

40

30

20

10

0

FF

FG

LAB

PD

WP/DL

Ireland 1986-97: smoothed poll dataIreland 1986-97: smoothed poll data

• Simulate removal of random error with moving average smoother

Time [Year/quarter]

962

953

944

941

932

923

914

911

902

893

884

881

872

863

Pa

rty s

up

po

rt (

5-p

oin

t m

ovin

g a

ve

rag

e %

)

60

50

40

30

20

10

0

FF

FG

Labour

PD

WP/DL

Ireland 1986-97: poll data & real votingIreland 1986-97: poll data & real voting

• Broken lines = poll data; solid lines = real voting (FF green, FG blue)• Clear evidence of pro-FF and anti-FG bias

Time [Year/quarter]

972924892871

Part

y s

up

po

rt (

%)

60

50

40

30

20

10

0

Ireland 1986-97: simulated time seriesIreland 1986-97: simulated time series

• Assumes FF and FG are i-Hunters;

• and PD, Lab, DL are Stickers

Time (simulation cycles)

193

190

187

184

181

178

175

172

169

166

163

160

157

154

151

Pa

rty s

up

po

rt (

%)

50

40

30

20

10

0

PD

FF

FG

LAB

WP/DL

Comparing statistical signaturesComparing statistical signatures

 

Opinion polls unsmoothed

Opinion polls smoothed

Real elections

Simulation

Mean SD Mean SD Mean SD Mean SD

Party size

FF 51 3.3 51 1.8 46 2.0 42 2.3

FG 25 2.9 25 2.1 30 2.6 28 3.3

LAB 12 4.8 12 4.2 13 6.3 17 2.5

PD 8 4.1 8 3.3 7 3.6 9 1.3

WP/DL

3 1.4 3 1.2 4 1.1 3 0.1

Summaries

SD SIZE 19 1.2 20 1.0 18 1.0 16 0.5

MAX SIZE 51 3.3 52 1.8 46 2.0 42 2.3

MIN SIZE 3 1.3 3 1.2 4 1.1 3 0.1

MAX - MIN 48 3.5 48 2.1 42 1.5 39 2.3

Generic 5-Hunter systemGeneric 5-Hunter system

• Statistical signature completely different

Time (simulation cycles)

199

196

193

190

187

184

181

178

175

172

169

166

163

160

157

154

151

Part

y s

upport

(%

)50

40

30

20

10

0

PARTY1

PARTY2

PARTY3

PARTY4

PARTY5

ONWARD AND UPWARDONWARD AND UPWARD

• Explore the birth and death of parties

• Enable endogenous choice of decision rules

• “Grow” new decision rules using genetic algorithm– Sexual reproduction of the fittest rules plus random mutation

• Enable endogenous evolution of voter preferences– Or reading in of real preference data

• Refine ideas on statistical signatures