Sorting and Grading
Extracted source text · arXiv 2208.10894
\documentclass[12pt]{article}
\usepackage{amsfonts, amssymb, amsmath, amsthm}
\usepackage[onehalfspacing]{setspace}
\usepackage[round, longnamesfirst]{natbib}
\usepackage{tikz}
\usepackage{epigraph}
\usepackage{tikzpeople}
\usepackage{multirow}
\usepackage{comment}
\usepackage[english]{babel}
\usepackage{amsmath,amsfonts,amsthm,bm}
% \usepackage{ctable}
% \usetikzlibrary{automata,positioning,arrows}
\usepackage{array}
\newcolumntype{?}{!{\vrule width 2pt}}
\newlength{\Oldarrayrulewidth}
\newcommand{\Cline}[2]{%
\noalign{\global\setlength{\Oldarrayrulewidth}{\arrayrulewidth}}%
\noalign{\global\setlength{\arrayrulewidth}{#1}}\cline{#2}%
\noalign{\global\setlength{\arrayrulewidth}{\Oldarrayrulewidth}}}
\usepackage[justification=centering,textfont={sc},labelfont={sc}]{caption,subcaption}
\newcommand{\interior}[1]{%
{\kern0pt#1}^{\mathrm{o}}%
}
\newcommand{\costpar}{r}
\newcommand{\dempar}{d}
\newcommand{\cbar}{\costpar^{\dagger}(\dempar)}
\newcommand{\cbara}{\costpar^{\dagger\dagger}(\dempar)}
\newcommand{\cbarb}{\costpar^{\dagger\dagger\dagger}(\dempar)}
\newcommand{\cs}{\costpar^{*}(\dempar)} % Stackelberg cutoff
\newcommand{\css}{\costpar^{**}(\dempar)} %
\newcommand{\csss}{\costpar^{***}(\dempar)} % IP to P cutoff
%\newcommand{\overline{x}^{\gK}}{x^{\textrm{max}}}
\newcommand{\xmin}{x^{\textrm{min}}}
\newcommand{\gP}{\mathcal{P}}
\newcommand{\gK}{\mathcal{K}}
\newcommand{\gIP}{\gP^I}
\newcommand{\gQIP}{\gP^{I+}}
\newcommand{\gIK}{\gK^I}
\newcommand{\gXIP}{\mathcal{X}^{\gIP}}
\newcommand{\gXQIP}{\mathcal{X}^{\gQIP}}
\newcommand{\gXP}{\mathcal{X}^{\gP}}
\newcommand{\gXIK}{\mathcal{X}^{\gIK}}
\newcommand{\gXK}{\mathcal{X}^{\gK}}
\newcommand{\ds}{\dempar^{*}}
\newcommand{\dss}{\dempar^{**}}
\newcommand{\xUB}{\overline{\overline{x}}}
\newcommand{\xLB}{\underline{\underline{x}}}
\DeclareMathOperator{\rem}{rem}
\usepackage[capposition=bottom]{floatrow}
\newtheorem{theorem}{Theorem}
\newtheorem{claim}{Claim}
\newtheorem{observation}{Observation}
\newtheorem{corollary}{Corollary}
\newtheorem{remark}{Remark}
\newtheorem{definition}{Definition}
\newtheorem{lemma}{Lemma}
\newtheorem{proposition}{Proposition}
\newtheorem{condition}{Condition}
\usepackage{apptools}
\AtAppendix{
\counterwithin{lemma}{section}
\counterwithin{remark}{section}
\counterwithin{corollary}{section}
\counterwithin{definition}{section}
\counterwithin{proposition}{section}
}
\setlength{\textwidth}{6.6in} \setlength{\evensidemargin}{0in}
\setlength{\oddsidemargin}{0in} \setlength{\textheight}{8.4in}
\setlength{\topmargin}{-.0in} \setlength{\headheight}{0in}
\usepackage{pgfplots}
\usepgfplotslibrary{fillbetween}
\usetikzlibrary{intersections}
\pgfplotsset{compat=1.13}
\makeatletter
\renewenvironment{proof}[1][\proofname]{\par
\pushQED{$\blacksquare$}%
\normalfont \topsep6\p@\@plus6\p@\relax
\trivlist
\item[\hskip\labelsep
\bfseries
#1\@addpunct{:}]\ignorespaces
}{%
\hfill \popQED\endtrivlist\@endpefalse
}
\makeatother
\newcommand{\ubeta}{\underline{\beta}}
\renewcommand{\interior}[1]{\text{int}({#1})}
\usepackage[hidelinks, colorlinks = true,
linkcolor = black,urlcolor = blue,
citecolor = black,
anchorcolor = blue]{hyperref}
\usepackage{cleveref}
\crefname{subsection}{subsection}{subsections}
\Crefname{subsection}{Subsection}{Subsections}
% \usepackage{todonotes}
\usepackage{enumerate}
\usepackage{appendix}
\newcommand{\TODO}[1]{\textbf{\color{red}TO DO: #1}}
\begin{document}
\begin{titlepage}
\title{Sorting and Grading\footnote{We thank the editor, Nicola Pavoni, and four anonymous referees for comments and suggestions that greatly helped us improve this paper. We
are very grateful to
Arjada Bardhi, Antonio Cabrales,
Matteo Escud\'{e},
B\r{a}rd Harstad, Nenad Kos,
Eirik Kristiansen, Ming Li, Mart\'{i} Mestieri, Eduardo Perez-Richet, Johannes Schneider, Vasiliki Skreta, Takuro Yamashita, and Andy Zapechelnyuk for useful discussions and feedback.}}
\author{Jacopo Bizzotto\thanks{Oslo Business School, OsloMet. Email: \href{mailto:jacopo.bizzotto@oslomet.no}{jacopo.bizzotto@oslomet.no}.} \quad Adrien Vigier\thanks{School of Economics, University of Nottingham. Email: \href{adrien.vigier@nottingham.ac.uk}{adrien.vigier@nottingham.ac.uk.}}}
\date{February 2024}
\maketitle
\begin{abstract}
We
propose a framework to assess how to optimally sort and grade
students of heterogenous ability.
Potential employers face uncertainty regarding an
individual's productive value. Knowing which school an individual went to is useful for two reasons: firstly, average student ability may differ across schools; secondly, different schools may use different grading rules and thus provide varying incentives to exert effort. An optimal school system exhibits coarse stratification with respect to ability, and more lenient grading at the top-tier schools
than at the bottom-tier schools. Our paper contributes to the ongoing policy debate on tracking in secondary schools.
\end{abstract}
\vspace{.2 in}
\noindent \textbf{JEL classification}: D02, D83, I24, J71
\noindent \textbf{Keywords}: Tracking, Grading, School System, Statistical Discrimination.
\thispagestyle{empty}
\end{titlepage}
\newpage
\section{Introduction}
%\epigraph{The average level of educational performance of the population is probably the most decisive factor for the long-term growth of an economy}{Woessmann 2009}
Few debates are more heated than how best to
foster students' achievement.\footnote{While the discussion in this paper focuses on secondary school systems, another application of our analysis concerns public universities.}
Two questions occupy a central place in this debate: (i) should students be sorted based on ability? (ii) how should students be graded?
In this paper, we
study the interplay between sorting and grading, and argue that one reason for sorting students is that, depending on their ability, students ought to be graded differently.
The practice of separating students according to ability, commonly referred to as tracking, typically occurs in high school.\footnote{The age at which students are first tracked varies across countries. See, e.g., \cite{Betts2011Review} or \citet{Woessmann2009Evidence}.}
In England, for example, the public secondary school system includes a limited number of selective grammar schools together with a vast number of comprehensive (i.e. non-selective) schools.\footnote{For an history of tracking in the United Kingdom, see \citet{WhittyPower2015}.}
Countries like Austria, Germany and Switzerland divide students into
academic and vocational high schools.
By contrast, in the United States and Canada,
tracking takes the form of ability grouping within schools. Yet
other countries choose not to track students at all, including France and Italy. The proponents of tracking
argue that grouping students by ability is socially efficient and
benefits
all categories of students; its critics claim on the other hand that only high-ability students benefit from
tracking and that,
if selection supposedly based on ability ends up capturing students' socio-economic background, tracking might be a factor which contributes to perpetuate inequality across generations.
Empirical evidence on the matter is mixed: \cite{galindo2007heterogeneous} and
\citet{Duflo2011Kenya} report evidence of the positive effects of tracking;
others, including
\cite{Kerr2013Cognitive} and
\cite{matthewes2021better},
document a negative impact overall.
The issue of grading
is no less controversial than that of tracking. While the importance of grading standards for student behavior is generally accepted, disagreement is common regarding what these standards ought to be.
One recurring theme in this debate concerns the merits and flaws of standardized tests.\footnote{Standardized tests are defined as those in which the conditions and content are equal for everyone taking the test, regardless of when, where, or by whom the test was given or graded.} The proponents of such tests emphasize that they are fair and facilitate
cross-student comparisons.
Critics argue that different students ought to be graded differently. For example,
\cite{becker1992learning}
argue that stratifying students into groups in which each has a chance for
success may be preferred to the setting of a single national standard.
We
propose a simple framework to assess how to optimally sort and grade
students.
A population of students is divided across the different schools forming part of a given school system.
Each individual comes out of school with a productive value that can be either high or low. The
distribution of an individual's productive value is determined by a combination of his ability and effort. However,
ability, effort, and productive value are all unobserved
by employers.
Employers only observe the school which the student attended and the grade
which he obtained; they form beliefs about the student's productive value and offer a competitive wage.\footnote{In practice, some individuals seek employment immediately after secondary school while others go to university. What we generically refer to as employers could just as well represent university admissions officers.}
Figure \ref{Figure1ok} summarizes the model.
\begin{figure}
\begin{tikzpicture}[x=3.5cm,y=6.5cm,scale=1, house/.style={
draw, minimum width=12mm, minimum height=10mm,
label={[name=labelaux, inner sep=1mm, distance=0pt]#1},
append after command={%
\pgfextra \draw ([xshift=.5\pgflinewidth]\tikzlastnode.north west)--(labelaux.north)
--([xshift=-.5\pgflinewidth]\tikzlastnode.north east);\endpgfextra
}
},
vertex/.style={draw,circle,minimum size=0.8cm}
,
grade/.style={draw,rectangle, minimum width=8mm, minimum height=12mm,}]
%headings
\draw(0,1) node{STUDENTS};
\draw(1,1) node{SCHOOLS};
\draw(2,1) node{\begin{tabular}{c}
PRODUCTIVE\\
VALUES
\end{tabular}};
\draw(3,1) node{GRADES};
\draw(4,1) node{\begin{tabular}{c}
EMPLOYERS'\\
BELIEFS
\end{tabular}};
%Students
\node[person,shirt=black,scale=1](A) at (0,.9) %{$\theta_1$};
{};
\node[person,shirt=blue,scale=1](B) at (0,.6) %{$\theta_2$};
{};
\node[person,shirt=red,scale=1](C) at (0,.3) %{$\theta_3$};
{};
\node[person,shirt=yellow,scale=1](D) at (0,0) %{$\theta_4$};
{};
%Schools
\node[house](S2) at (1,.1) {};%{$s_2$};
\node[house](S1) at (1,.7) {};%{$s_1$};
%Productive values
\node[vertex](pv1) at(2,.8){1};
\node[vertex](pv2) at(2,.6){0};
\node[vertex](pv3) at(2,.2){1};
\node[vertex](pv4) at(2,.0){0};
%Grades
\node[grade](g1) at(3,.8){pass};
\node[grade](g2) at(3,.6){\hspace{.8mm}fail\hspace{.8mm}};
\node[grade](g3) at(3,.2){pass};
\node[grade](g4) at(3,.0){\hspace{.8mm}fail\hspace{.8mm}};
%Beliefs
\draw[|-|,color=black] (4,0) -- (4,.1);
\draw[-|,color=black] (4,0.1) -- (4,0.3);
\draw[-|,color=black] (4,0.3) -- (4,0.5);
\draw[-|,color=black] (4,0.5) -- (4,0.6);
\draw[-|,color=black] (4,0.6) -- (4,0.85);
\node at(3.95,0){0};
\node at(3.95,.85){1};
\node(B4) at(4,0.1){};
\node(B3) at(4,.3){};
\node(B2) at(4,.5){};
\node(B1) at(4,.6){};
%\node at(4.1,0.1){$q^-_{s_2}$};
%\node at(4.1,.3){$q^-_{s_1}$};
%\node at(4.1,.5){$q^+_{s_2}$};
%\node at(4.1,.6){$q^+_{s_1}$};
%arrows
\path[->, auto] (A) edge node {} (S1)
(B) edge node {} (S1)
(D) edge node {} (S2);
\path[->] (C) edge node {} (S2);
\path[->, dashed] (S1) edge node {} (pv1)
edge node {} (pv2)
(S2) edge node {} (pv3)
edge node {} (pv4);
\path[->, dashed] (pv1) edge node {} (g1)
edge node {} (g2)
(pv2) edge node {} (g1)
edge node {} (g2)
(pv3) edge node {} (g3)
edge node {} (g4)
(pv4) edge node {} (g3)
edge node {} (g4);
\path[->, dashed] (pv1) edge node { } (g1)
edge node {} (g2)
(pv2) edge node { } (g1)
edge node {} (g2)
(pv3) edge node {} (g3)
edge node {} (g4)
(pv4) edge node {} (g3)
edge node {} (g4);
\path[->, auto] (g1) edge node { } (B1)
(g2) edge node { } (B3)
(g3) edge node {} (B2)
(g4) edge node {} (B4);
\end{tikzpicture}\caption{The Setting}\label{Figure1ok}
\end{figure}
The assignment of students to schools
matters for two reasons:
(a) the school that a student went to conveys information
regarding his productive value; (b) different schools may use different grading rules.
We study the problem of a planner designing the school system so as to
maximize social welfare. Since due to moral hazard each student's equilibrium
effort is below the socially optimal level, the problem of the planner in essence amounts to maximizing total effort.
The basic question we address is the following: should students be sorted according to ability and, if so, how?
The trade-off is as follows:
\begin{itemize}
\item
On the one hand,
sorting students enables tailored incentives. This is valuable, as the optimal grading rule for any given school depends on the average ability of the students assigned to that school. The basic idea
is simple.
For a student who is regarded as ``bad", the value of obtaining a good grade is very high
as long as this grade
is reliable
(i.e., as long as the grading rule generates few false positives).
By contrast, in the case of students who are regarded as ``good", the logic is reversed.
In consequence, a one-size-fits-all approach to grading might end up depressing
all students' incentives to exert effort in school.
\item
On the other hand, two schools with different average student ability and the same grading rule should be merged. The logic is as follows.
Mixing
students increases the uncertainty faced by employers, which, in turn, augments the
weight of school grades in employers' inferences. Yet,
when grades matter more, top marks
become worth the effort. Thus, mixing students indirectly favors their effort.
\end{itemize}
At an optimum,
the aforementioned trade-off is resolved as follows:
(i) students whose ability is above a cutoff are separated from those whose ability is below;
(ii) on each side of the separating cutoff,
all students
are pooled.
Bottom-tier students are evaluated through a tough grading rule (i.e., one generating
few false positives), while top-tier students are evaluated through a lenient grading rule
(i.e., one generating few false negatives).
To convey the paper's main insights in the simplest possible way and ensure the tractability of the model,
the framework we propose makes several assumptions of note.
Firstly, we ignore the fact that some schools may be intrinsically better than others, perhaps because they have more resources, better teachers, or a combination of both.
Secondly, we ignore
``direct" peer effects.
For example, if students learn from their peers, one's educational attainment
should be positively affected by an increase in the mean student ability of the school to which he was assigned. Alternatively, students may have homophilous preferences, and learn best when grouped with students of similar ability as their own.\footnote{It is well known
that, in the presence of
direct
peer effects, the
optimal allocation of students depends in a
complex way on the curvature properties of the educational production
function. See, e.g., \cite{arnott1987peer}. }
Thirdly, we assume that an individual's productive value is binary.\footnote{An example with three productive values is presented in Section \ref{sec:morethantwo}. }
Enriching the model along any of the aforementioned dimensions
would evidently improve the realism of our framework.
Last but not least,
one critical premise of our model, is that the planner can assign students to schools on the basis of their ability. Austria, England, Germany and Switzerland
provide examples of this kind. However, in countries like
France and Italy, students are typically free to choose their secondary school.\footnote{In France and Italy, this choice is somewhat constrained by a student's place of residence.}
More generally,
social planners may either lack information regarding students' ability,
or be unable to use that information (e.g., for political reasons). We therefore examine a variant of the model in which students can choose the school they go to. In this variant of the model, a school system needs to be incentive compatible for all students in order to be implementable.
We show first that without transfers (e.g., differentiated school fees or scholarships) the planner can do no better than randomly allocating students to schools.
We then show that, with transfers, any constrained-optimal school system has the same qualitative features as when there are no incentive compatibility constraints: students whose ability is above a cutoff are separated from those whose ability is below, and, on each side of the separating cutoff,
all students
are pooled.
However, the separating cutoff is larger.
In other words, the incentive compatibility constraints reduce the size of the top tier while increasing that of the bottom tier.
The rest of the paper is organized as follows. We next discuss the related literature. The model is presented in
Section
\ref{sectionmodel}. Section \ref{section:eqeffort}
contains the preliminary analysis.
Our main result is presented in Section \ref{sec:mainresult}. Section
\ref{sec:IC} studies incentive compatibility.
Section
\ref{sec:morethantwo} examines
an example with
three productive values, while Section
\ref{sec:conc} concludes.
\paragraph{Related Literature.}
We bridge two strands of literature: the literature on tracking, and that on grading.
The (mostly empirical) literature on tracking
goes at least as far back as
\cite{cook1924study} and \cite{whipple1936grouping}.
The arguments in favor of tracking are typically twofold (see, e.g., \cite{slavin1990achievement} or \cite{Betts2011Review}). Firstly, more homogeneous groups of students allow teachers to tailor their pedagogical strategies.
Secondly, with tracking, a
students' performance is only compared to that of similar-ability peers, preventing a possible lowering of self-esteem that could result from comparisons with the work of higher ability students, while allowing highly gifted students to be appropriately challenged and to view their own abilities more realistically.
We
contribute to the aforementioned literature by
arguing that another natural reason for which tracking could be efficient is that tracking allows incentives to be tailored to students' characteristics, via targeted grading policies.\footnote{A different argument in support of tracking is as follows. Suppose firms are heterogenous and the efficient matching between firms and students is assortative.
Then tracking could help achieve the efficient matching.
We view all of these arguments as complementary.}
The literature on grading can be divided in two strands.
A first strand studies the link between grading standards and effort.
\citet{becker1992learning} and
\citet{Betts98} argue
that a rise in
grading standards has ambiguous effects on overall achievement. While more able students increase effort
in order to reach the new standard, less able students, who must increase effort by a greater amount to
increase their achievement, may give up instead.
\cite{costrell1994simple} concludes that social
planners with
egalitarian concerns
should thus lower grading standards, rather than raise them. \cite{himmler2013double} further argue that
schools with disadvantaged students ought to lower standards more than others so as to compensate a lack of incentives to graduate.\footnote{\citet{DubeyGeanakoplos10} and \citet{Zubrickas15} take a different approach and
show that the use of coarse grades is optimal
when students
care primarily about their relative rank in class.}
The second strand of the literature on grading
views grade not primarily as a means to incentivize effort, but as a strategic tool for schools to maximize their
students' payoffs on the job market. In this spirit,
\cite{chan2007signaling}, \cite{ehlers2016honest}, and \cite{tampieri2019students}
develop signaling theories of grade inflation;
\cite{OstrovskySchwarz10} explore information disclosure in matching markets
and argue that schools may suppress some information about students in order
to improve their average job placement;
\cite{BoleslavskyCotton15} study information disclosure in contests and show that, in equilibrium, schools adopt grading policies
that do not perfectly reveal ability.
We contribute to the literature on grading by exploring the interplay between grading design and the allocation of students to schools.
Our paper is intimately connected to
the vast literature studying statistical discrimination in labor markets, understood as employers' reliance
on payoff-irrelevant characteristics
for making inferences about workers' productivity.\footnote{For a detailed review of the literature on the subject, see e.g. \cite{onuchic22}. }
In the Arrowian tradition (\cite{arrow1973}), we study an environment in which productivity is endogenous.
As in \cite{coate1993will},
because in our model employers' prior beliefs affect their interpretation of the productivity signal (in our case, the grade),
employers' conjectures affect
incentives to invest in skill acquisition.
We depart from the literature
following \cite{arrow1973}
in two crucial ways. Firstly, statistical discrimination does not proceed from equilibrium multiplicity in our model: each school induces a unique equilibrium level of effort; said effort, however, varies with the way students are graded. Secondly,
whereas the aforementioned literature takes
the payoff-irrelevant characteristic for making inferences about a worker's productivity
as given
(e.g., race or gender),
in our model said characteristic (namely, the school which a student attended) is either a choice made by the student, or part of the decision maker's design problem.
In this sense, our model is closer to Spence signalling (\cite{spence1973press}). However, the connection with signalling stops here: our model involves both adverse selection and moral hazard, and employers do not care exclusively about students' types, but about a combination of type and effort (productive values).
Our work is also connected to
the literature on mechanism design
with peer effects, comprising
\citet{jehiel1996not,jehiel1999multidimensional}, \cite{jehiel2001efficient},
\cite{board2009monopolistic}, \citet{rayo2013monopolistic},
\citet{rothschild2013redistributive,rothschild2016optimal}, and \cite{yamashita2021large}. Relative to this literature,
the main novelty of our work
is that how much an individual values being part of a certain group is not fixed, but instead depends on a grading rule which policy makers can adjust.\footnote{The problem we analyze can be viewed as an information design problem with constraints. Effectively,
the school system plays the role of an information structure for employers.}
Finally, the problem we analyze in Section \ref{sec:IC} connects our work to
a number of recent papers that share with ours the feature that the planner/designer offers a menu of items comprising a statistical experiment of some kind. This literature includes
\citet{esHo2007optimal}, \cite{kolotilin2017persuasion}, \cite{li2017discriminatory}, \citet{bergemann2018design}, \cite{guo2019interval}, \cite{wei2020reverse}, \cite{yamashita2021type}, and \cite{yang2022selling}.
In all those papers,
the experiment
helps the agent solve a decision problem. Instead,
in our setting
the experiment (i.e., the grading rule) helps potential employers evaluate the productive value of the agent (the student).
%Within this strand of literature, the findings of \citet{popov2013university} resonate with our results. In their model, bottom-tier schools compete for better job assignments by raising the average ability of students who receive ``A" grades, thus setting excessively high grading standards. We show that, in such schools, adopting a tough grading rule can have the additional advantage of incentivizing effort.
\section{Model}\label{sectionmodel}
The broad features of our model are as follows.
A
population of students is assigned to different schools. At the end of school, a student's value to an employer can be either high or low.
A student's chances of becoming high value
are increasing in
ability and effort.
However, neither ability nor effort are observed by employers, who form beliefs about a student's productive value based solely on the school which this student attended and the grade which he obtained. Every student then receives a wage which reflects employers' perception about his productive value. The
details of the model are presented below.
There are $m \in \mathbb{N}$ students, labelled by $j \in \{1,\cdots, m\}$. The assignment of students to schools is
captured by a pair $(n, \phi)$, where
$n \in \mathbb{N}$
and
$\phi$
is a surjective mapping from
$\{1, \cdots, m\}$ to $\{1, \cdots, n\}$. We refer to such a pair $(n, \phi)$ as an \textit{assignment}; under this assignment, student $j \in \{1,\cdots, m\}$ is assigned to school $\phi(j)$.
Every student is characterized by his \textit{type}, which can be thought of as capturing his ability. The type of student $j$ is denoted by
$ \theta_j \in(0, \alpha)$, where $\alpha \in (0,1)$.
At the end of school,
each student's \textit{productive value}
is either 0 or 1.
The probability
that
student $j$'s productive value is $1$
equals the sum of this student's type $\theta_j$
and \textit{effort} $e_j \in [0, 1-\alpha]$.
The cost of exerting effort $e_j$ is given by
$c(e_j)$, where $c'>0$, $c''>0$, $c'''\geq 0$,
$c'(0)=0$,
and $c'(1-\alpha)>1$.
Letting $b$ denote a parameter in $(0,1)$, we refer to
a pair $G=(g^0, g^1) \in [0,1]^2$ such that
$ g^1> g^0$ and $g^1 -g^0\leq b$ as a
\textit{grading rule}.
The interpretation is as follows.
Every
school delivers a grade to each of its students, either
``pass" or ``fail".
A school's grading rule determines the probabilities of obtaining the grade pass conditional on one's productive value: under the grading rule $G$, a student with productive value 0 passes with probability $g^0$, while a student with productive value 1 passes with probability $g^1$.
The condition
$ g^1> g^0$ ensures that obtaining the grade pass conveys more favorable information about one's productive value than obtaining the grade fail;
the
condition $g^1 -g^0\leq b$
rules out perfectly informative grading rules. We
denote by $\mathbb{G}$ the set of all grading rules.
We refer to
a tuple $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ composed of an assignment $(n, \phi)$ and a grading rule $G_s$ for every school $s \in \{1, \cdots, n\}$ as a \textit{school
system}.
Given a school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$,
the number and average type of students attending school $s \in \{1, \cdots, n\}$ will be denoted by $m_{s}$ and $\Bar{\theta}_{s}$, respectively.
We assume that, for every $s \in \{1, \cdots, n\}$,
both $\Bar{\theta}_{s}$ and $G_s$
are public information. However, a
student's type, effort, and productive value
are all unobserved by employers.
Employers
only observe
the school which a student attended, and
the grade which he obtained. They form conjectures about effort, and ultimately pay each
student $j$ a competitive wage equal to their (Bayesian) posterior belief that this student's productive value is equal to 1.
\section{Equilibrium Effort}\label{section:eqeffort}
This section examines the effort exerted by different students in an arbitrary school system.
We show first
that each school system induces a unique effort profile. We then
show that
the grading rule
maximizing effort in a school depends on its average student type. Finally,
we show that
if two schools share the same grading rule then pooling their students in a single school increases total effort.
We start by defining the notion of equilibrium effort.
\begin{definition}
Given a school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$,
$e$
is an \emph{equilibrium effort} in school $s \in \{1, \cdots, n\}$ if choosing
effort $e$
maximizes the expected payoffs of the students attending school $s$ given employers' conjecture that students of $s$ exert effort $e$.
\end{definition}
Consider a school $s$
forming part of a given school system.
Let $q_s^+(\hat{e}_s)$
denote employers' posterior belief that a student of $s$ who obtained the grade pass has productive value 1, given
employers' conjecture that students of $s$ exert effort $\hat{e}_s$.
Let $q_s^-(\hat{e}_s)$ denote the corresponding belief for students of $s$ who obtained the grade fail.
Then
\[ q_s^+(\hat{e}_s)
= \frac{ (\Bar{\theta}_s + \hat{e}_s) g_s^1}{ (\Bar{\theta}_s + \hat{e}_s) g_s^1 + (1-\Bar{\theta}_s - \hat{e}_s) g_s^0}, \quad
q_s^-(\hat{e}_s)= \frac{ (\Bar{\theta}_s + \hat{e}_s) (1-g_s^1)}{ (\Bar{\theta}_s + \hat{e}_s) (1-g_s^1) + (1-\Bar{\theta}_s - \hat{e}_s) (1-g_s^0)}.
\]
A type-$\theta$
student
attending school $s$ and exerting effort $e $ obtains the grade
pass with probability $(\theta+e)g_s^1+ (1-\theta-e) g_s^0$, and the grade fail otherwise.
So letting $U_s(\theta, e, \hat{e}_s)$
denote the expected payoff of a
type-$\theta$
student
attending school $s$ and exerting effort $e$
(given
employers' conjecture that students of $s$ exert effort $\hat{e}_s$):
\[ U_s(\theta, e, \hat{e}_s)=
[(\theta+e)g_s^1+ (1-\theta-e) g_s^0] q_s^+(\hat{e}_s) +
[1-(\theta+e)g_s^1- (1-\theta-e) g_s^0] q_s^-(\hat{e}_s) -
c(e).\]
The unique $e$
maximizing
$ U_s(\theta, e, \hat{e}_s)$
satisfies
\[ c'(e)=
(g_s^1 - g_s^0) \big(q_s^+(\hat{e}_s) - q_s^-(\hat{e}_s)\big) = B(G_s, \Bar{\theta}_s+ \hat{e}_s ),
\]
where
\[
B(G,x):= (g^1 - g^0) \bigg( \frac{x g^1 }{ g^0 + x (g^1 - g^0) } - \frac{x(1-g^1)}{ 1- g^0 -x (g^1 - g^0) } \bigg), \quad \forall x\in (0,1).
\]
It follows that
$e$ is an equilibrium effort in school $s$ if and only if
\begin{equation}\label{eq:focbase}
c'(e)=B(G_s, \Bar{\theta}_s+ e ).
\end{equation}
The function $e \mapsto B(G_s, \Bar{\theta}_s+ e )- c'(e)$ is
strictly concave, positive at $e=0$, and negative at $e=1-\alpha$. So equation \eqref{eq:focbase} has a unique solution,
and this solution belongs to
$(0,1- \alpha)$; as it
is determined by $\bar{\theta}_s$ and $G_s$, we will denote it by $\xi(\Bar{\theta}_s, G_s)$. The next lemma summarizes the main takeaway.
\begin{lemma} To each school
$s$
forming part of a given school system
is associated a unique equilibrium effort $\xi(\Bar{\theta}_s, G_s)$. This equilibrium effort is the unique solution of equation \eqref{eq:focbase}.
\end{lemma}
Henceforth,
let
$G_{\star}$
denote the grading rule
$(1-b,1)$, and $G^{\star}$ the grading rule $(0, b)$.
The grading rule
$G_{\star}$ thus passes every student with productive value $1$;
the grading rule
$G^{\star}$, on the other hand, fails every student
with productive value $0$.
Moreover, note that $G_{\star}$ and $G^{\star}$ both belong to the set of
maximally ``informative" grading rules within $\mathbb{G}$.\footnote{Consider two grading rules, $G_A$ and $G_B$, such that either $g^1_A= g^1_B$ or $g^0_A= g^0_B$. Then $G_A$ is Blackwell-more-informative than $G_B$ if and only if
$g^1_A- g^0_A> g^1_B- g^0_B$. }
We can now state our first proposition.
\begin{proposition}\label{prop1}
For any $\Bar{\theta} \in (0, \alpha)$ and any grading rule $G \in \mathbb{G} \setminus \{G_{\star}, G^{\star}\}$,
either $\xi(\Bar{\theta}, G_{\star})>
\xi(\Bar{\theta}, G)$ or $\xi(\Bar{\theta}, G^{\star})>
\xi(\Bar{\theta}, G)$.
\end{proposition}
\begin{proof} Let $\Bar{\theta} \in (0, \alpha)$.
For any $x \in (0,1)$,
the function $g^1 \mapsto B\big( (g^0, g^1),x \big)$
is increasing, and
the function $g^0 \mapsto B\big( (g^0, g^1),x \big)$
decreasing. So either $g^1- g^0=b$ or we can find $\Tilde{G} \in \mathbb{G}$ such that
$B(\Tilde{G}, x)> B(G, x)$.
In particular, either
$g^1- g^0=b$ or we can find $\Tilde{G} \in \mathbb{G}$ such that
\begin{equation}\label{eqcupcaf2}
B\big(\Tilde{G}, \Bar{\theta} +\xi(\Bar{\theta}, G) \big)> B\big(G, \Bar{\theta} +\xi(\Bar{\theta}, G) \big)=c'\big( \xi(\Bar{\theta}, G)\big).
\end{equation}
The function $e \mapsto B(\Tilde{G}, \Bar{\theta}+ e )- c'(e)$ is
strictly concave, positive at $e=0$, and zero
at $e= \xi(\Bar{\theta}, \Tilde{G})$. So
\eqref{eqcupcaf2} implies $ \xi(\Bar{\theta}, G)<\xi(\Bar{\theta}, \Tilde{G})$.
In sum, either
$g^1- g^0=b$ or we can find $\Tilde{G} \in \mathbb{G}$ such that
$ \xi(\Bar{\theta}, \Tilde{G}) > \xi(\Bar{\theta}, G)$.
Next,
for any $x \in (0,1)$,
the
function $g^0 \mapsto B\big( (g^0, g^0+b),x \big)$
is strictly convex. So if
$G= (g^0, g^0+b)\in \mathbb{G} \setminus \{G_{\star}, G^{\star}\}$, we can find $\Tilde{G} \in \{G_{\star}, G^{\star}\}$ such that
$B(\Tilde{G}, x)> B(G, x)$. In particular,
if
$G= (g^0, g^0+b)\in \mathbb{G} \setminus \{G_{\star}, G^{\star}\}$, we can find $\Tilde{G} \in \{G_{\star}, G^{\star}\}$ such that \eqref{eqcupcaf2} holds.
Then reasoning as we did above yields
$ \xi(\Bar{\theta}, \Tilde{G}) > \xi(\Bar{\theta}, G)$.
\end{proof}
By
Proposition \ref{prop1}, if a grading
rule
$G$ maximizes effort in a school, either $G=G_{\star}$ or $G= G^{\star}$. The logic is as follows.
Any grading rule $G \notin \{G^{\star}, G_{\star}\}$
can be written as a strict convex combination of $G^{\star}$, $G_{\star}$, and some uninformative grading rule, $G^{\varnothing}$ say.\footnote{We here slightly abuse the terminology, since, the way we have defined them, every grading rule is informative. }
So using the grading rule $G$ is equivalent to randomizing between
$G^{\star}$, $G_{\star}$, and $G^{\varnothing}$. Yet, randomizing the grading rule is conducive to moral hazard. So there exists a grading rule in the set
$\{G^{\star}, G_{\star}, G^{\varnothing}\}$ such that adopting this grading rule
for sure induces more effort than $G$.\footnote{And, $G^{\varnothing}$ being uninformative, either
$G^{\star}$ or $G_{\star}$ must induce more effort than $G$.}
The next proposition tells us when
$G^{\star}$ induces more effort than $G_{\star}$,
or the other way around.
\begin{proposition}\label{prop2}
There exists $\Bar{\theta}^\dagger$ such that $\xi(\Bar{\theta}, G^{\star})> \xi(\Bar{\theta}, G_{\star})$ for $\Bar{\theta}<\Bar{\theta}^\dagger$, and
$\xi(\Bar{\theta}, G_{\star})> \xi(\Bar{\theta}, G^{\star})$ for $\Bar{\theta}>\Bar{\theta}^\dagger$.
\end{proposition}
\begin{proof}
The function $x \mapsto B( G^{\star},x)$ is decreasing.
Hence, for $0<\Bar{\theta}_A < \Bar{\theta}_B< \alpha$, we have
\begin{equation}\label{eqcupcaf3}
B \big(G^{\star}, \Bar{\theta}_A+ \xi(\Bar{\theta}_B, G^{\star}) \big) >
B \big(G^{\star}, \Bar{\theta}_B + \xi(\Bar{\theta}_B, G^{\star}) \big)=c' \big( \xi(\Bar{\theta}_B, G^{\star}) \big).
\end{equation}
The function $e \mapsto B(G^{\star}, \Bar{\theta}_A+ e )- c'(e)$ is
strictly concave, positive at $e=0$, and zero
at $e= \xi(\Bar{\theta}_A, G^{\star})$. So
\eqref{eqcupcaf3} implies $ \xi(\Bar{\theta}_A, G^{\star})> \xi(\Bar{\theta}_B, G^{\star})$. This shows that the function $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing.
The function $x \mapsto B( G_{\star},x)$ is increasing. Reasoning as we did above then shows that
the function $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ is increasing.
\end{proof}
By Proposition \ref{prop2}, which
grading rule maximizes effort in a school depends on the average type of students attending this school: if the average type is
greater than some cutoff ($\Bar{\theta}^\dagger$), effort is maximized by the grading rule
$G_{\star}$; if instead the average type is
less than this cutoff, effort is maximized by the grading rule
$G^{\star}$. The logic is as follows. A student's incentives to exert effort are increasing in the difference
between $q_s^+(\hat{e}_s)$
(employers' posterior belief that a student of $s$ who obtained the grade pass has productive value 1)
and $q_s^-(\hat{e}_s)$ (employers' posterior belief that a student of $s$ who obtained the grade fail has productive value 1).
When the average type of students assigned to school $s$ is high, then so
is $q_s^+(\hat{e}_s)$; in this case, maximizing $ q_s^+(\hat{e}_s) - q_s^-(\hat{e}_s)$ is achieved by minimizing $q_s^-(\hat{e}_s)$, which in turn implies passing every student with productive value 1. Symmetrically, when
the average type of students assigned to school $s$ is low, so
is $q_s^-(\hat{e}_s)$; maximizing $ q_s^+(\hat{e}_s) - q_s^-(\hat{e}_s)$ is then achieved by maximizing $q_s^+(\hat{e}_s)$, which in turn implies failing every student with productive value 0.
The final proposition of this section addresses the following question. Suppose that
two schools share the same grading rule:
if the goal is to maximize total effort,
should their students be pooled or kept separated?
\begin{proposition} \label{prop3}
For every $G \in \mathbb{G}$, the function
$ \bar{\theta} \mapsto \xi(\bar{\theta}, G)$ is strictly concave.
\end{proposition}
\begin{proof}
Let $G \in \mathbb{G}$,
$\bar{\theta}_A, \bar{\theta}_B \in (0, \alpha)$ with $\bar{\theta}_A\neq \bar{\theta}_B $, and $\lambda \in (0,1)$. The strict concavity of the function $x \mapsto B(G,x)$ together with the convexity of $c'$
imply
\begin{align*}
B\big(G, (1-\lambda)
(\bar{\theta}_A + \xi(\bar{\theta}_A, G)) +\lambda (\bar{\theta}_B + \xi(\bar{\theta}_B, G))\big)&> (1-\lambda)
B\big(G, \bar{\theta}_A + \xi(\bar{\theta}_A, G)\big) + \lambda B\big(G, \bar{\theta}_B + \xi(\bar{\theta}_B, G)\big) \\
&= (1-\lambda) c'\big( \xi(\bar{\theta}_A, G) \big) +
\lambda c'\big( \xi(\bar{\theta}_B, G) \big)
\\
&\geq c'\big( (1-\lambda) \xi(\bar{\theta}_A, G) + \lambda\xi(\bar{\theta}_B, G) \big).
\end{align*}
Finally,
the function $e \mapsto B(G, (1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B +e )- c'(e)$ is
strictly concave, positive at $e=0$, and zero
at $e= \xi\big((1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B, G\big)$. So the highlighted
expression
implies
\[ (1-\lambda) \xi(\bar{\theta}_A, G) + \lambda\xi(\bar{\theta}_B, G)< \xi\big((1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B, G\big).
\]
\end{proof}
By Proposition \ref{prop3}, if two schools share the same grading rule, then pooling their students in a single school increases total effort.
The basic intuition is as follows. Separating students provides employers with information about students' types.
This additional information weakens the informational content of the grades, and, consequently, also the benefit of exerting effort.
\section{Socially Optimal School Systems}\label{sec:mainresult}
We saw in the previous section
that each school system induces a unique effort profile: specifically,
students in a school $s$ exert effort $ \xi(\Bar{\theta}_s, G_s)$.
The school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ therefore
generates social welfare
\[
W\big(n, \phi,
(G_{s})_{s=1}^n \big):=
\sum_{s=1}^n m_{s} \big[ \Bar{\theta}_{s} + \xi(\Bar{\theta}_s, G_s) - c\big( \xi(\Bar{\theta}_s, G_s) \big)\big].
\]
A school system is \textit{socially optimal} if it maximizes social welfare among all possible school systems.
Our goal in this section is to characterize the socially optimal school systems.
We will
say that a school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$
has a \textit{single-tier structure} if
$\bar{\theta}_{s_A}= \bar{\theta}_{s_B}$ for all $s_A,s_B \in \{1, \cdots, n\}$.
A school system
$\big(n, \phi,
(G_{s})_{s=1}^n \big)$
is said
to
have a
\textit{two-tier structure}
if there exist non-empty sets $S$ and
$S'$ partitioning $\{1,\cdots, n\}$
such that
\begin{itemize}
\item
$\bar{\theta}_{s_A}= \bar{\theta}_{s_B}< \bar{\theta}_{s'_A}= \bar{\theta}_{s'_B} $ for all $s_A, s_B \in S$ and
$s'_A, s'_B \in S'$;
\item
$\theta_j \leq \theta_{j'}$ whenever $\phi(j) \in S$ and $\phi(j') \in S'$.
\end{itemize}
We then refer to
$S$ as the bottom tier and to $S'$ as the top tier.
In words, a school system has a two-tier structure
if every school belongs to one of
two tiers, such that: (i)
all schools in a given tier have the same average student type, (ii) the
type of every student in a top-tier school is at least as a large as the type of any student in a bottom-tier school.
We can now state our first theorem.
\begin{theorem}\label{thm1}
Every socially optimal school system has either a
single-tier structure
or a two-tier structure. In the latter case, schools in the top tier have grading rule $G_{\star}$, and schools in the bottom tier have grading rule $G^{\star}$.
\end{theorem}
\begin{proof}
Any school $s$ forming part of a given school system is such that $c'\big( \xi(\Bar{\theta}_s, G_s)\big)=B\big(G_s, \Bar{\theta}_s+ \xi(\Bar{\theta}_s, G_s)\big)<1$. So $ \xi(\Bar{\theta}_s, G_s)<
\overline{\overline{e}}
$, where
$ \overline{\overline{e}}$ denotes the unique effort such that $c'( \overline{\overline{e}})=1$.
Yet
$e\mapsto e-c(e)$ is increasing on $[0, \overline{\overline{e}}]$.
It follows that any school
$s$ forming part of an efficient school system must be such that
\[
G_s \in \arg \max_{G \in \mathbb{G}}~ \xi(\bar{\theta}_s, G).
\] Applying
Proposition \ref{prop1} then shows that
any school
$s$ forming part of an efficient school system is such that
$G_s \in \{G_{\star}, G^{\star}\}$. Moreover, by Proposition \ref{prop2},
$G_s= G^{\star}$ if
$\Bar{\theta}_s<\Bar{\theta}^\dagger$, while
$G_s= G_{\star}$ if
$\Bar{\theta}_s>\Bar{\theta}^\dagger$.
Next,
consider $G \in \mathbb{G}$,
$\bar{\theta}_A, \bar{\theta}_B \in (0, \alpha)$ with $\bar{\theta}_A\neq \bar{\theta}_B $, and $\lambda \in (0,1)$.
We then have
\begin{align*}
(1-\lambda) \big[ \xi(\bar{\theta}_A, G) - & c\big( \xi(\bar{\theta}_A, G) \big) \big] + \lambda \big[ \xi(\bar{\theta}_B, G) - c\big( \xi(\bar{\theta}_B, G) \big) \big] \\ &\leq
(1-\lambda) \xi(\bar{\theta}_A, G) + \lambda \xi(\bar{\theta}_B, G) - c\big( (1-\lambda) \xi(\bar{\theta}_A, G) + \lambda \xi(\bar{\theta}_B, G) \big) \\
& < \xi\big((1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B, G\big) -c \big( \xi\big((1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B, G\big)\big),
\end{align*} where
the first inequality follows from the concavity of the function
$e\mapsto e-c(e)$, and the second inequality
follows from
the fact that this function is increasing on $[0, \overline{\overline{e}}]$
while, by Proposition \ref{prop3}, $ \xi\big((1-\lambda) \Bar{\theta}_A+ \lambda \Bar{\theta}_B, G\big)>(1-\lambda) \xi(\bar{\theta}_A, G) + \lambda\xi(\bar{\theta}_B, G)$.
The inequalities above imply that if
two schools with the same grading rule
form part of a socially optimal school system then their average student types
have to be equal.
Finally, we saw in the proof of Proposition \ref{prop2}
that $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing, and $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ increasing. Hence, if two
schools $s_A$ and $s_B$ forming part of a given school system
are such that
(i) $G_{s_A}=G^{\star}$, (ii) $G_{s_B}=G_{\star}$, (iii) there exist a student of $s_A$
and a student of $s_B$ such that the
type
of the former
is greater than the type of the latter, then swapping these students increases both equilibrium effort in school $s_A$ ($e^*_{s_A}$) and equilibrium effort in school $s_B$ ($e^*_{s_B}$), and therefore also
$m_{s_A} [ e^*_{s_A} - c(e^*_{s_A})] + m_{s_B} [ e^*_{s_B} - c(e^*_{s_B})]$.
We conclude that
if two
schools $s_A$ and $s_B$ forming part of an socially optimal school system
are such that (i) and (ii) hold, then the type of each student in $s_B$
must be at least as large as the type of any student in $s_A$.
\end{proof}
Theorem \ref{thm1} combines insights from Propositions
\ref{prop2} and \ref{prop3}.
Suppose that either all students
have a type
which is larger than the cutoff $\theta^\dagger$ in Proposition \ref{prop2}, or all students have a type
which is smaller than this cutoff. Then,
by the same proposition, every school forming part of a socially optimal school system must adopt the same grading rule (either $G_{\star}$ or $G^{\star}$).
Yet, by Proposition \ref{prop3}, if two schools share the same grading rule, then pooling their students in a single school increases total effort. It follows that, in this case, every socially optimal school system must essentially contain a single school. By contrast, if student types are very dispersed and low types are separated from high types then the efficient grading rule might vary across schools. In this case,
separating students in two groups may be optimal. We next illustrate Theorem \ref{thm1} in an example with two types.
\paragraph{Example.} Suppose that every student is of one of two types, $\mu- \sigma$ and $\mu + \sigma$, with $\mu$ representing the average student type. Then there exists
$\sigma^\dagger$
such that if $\sigma<\sigma^\dagger$ every socially optimal school system has a
single-tier structure, while if $\sigma>\sigma^\dagger$ every socially optimal school system has a
two-tier structure.\footnote{The proof is straightforward, and combines Theorem \ref{thm1} with the fact that $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing, and $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ increasing. }
\section{Incentive Compatibility}\label{sec:IC}
This section explores incentive compatibility.
Consider a designer (she) choosing the school system. As long as she can observe students' types, the designer can in principle implement the school system of her choice;
in particular, any socially optimal school system can be implemented. On the other hand, if the designer is either unable to observe students' types, or cannot freely assign students to schools, then the designer is forced to let students self-select into schools.
Subsection \ref{subsec:ICwithout} studies incentive compatibility without transfers, while Subsection \ref{subsec:ICwith} studies incentive compatibility with transfers.
\subsection{Incentive Compatibility without Transfers} \label{subsec:ICwithout}
To shorten notation, in this section let $e^*_s= \xi(\Bar{\theta}_s, G_s)$.
Given a school system
$\big(n, \phi,
(G_{s})_{s=1}^n \big)$, let
$U^*_s(\theta)$ denote a type-$\theta$ student's maximum expected payoff from attending school $s$, that is,
\[
U^*_s(\theta) :=U_s(\theta, e^*_s, e^*_s)=
[(\theta+e^*_s)g_s^1+ (1-\theta-e^*_s) g_s^0] q_s^+(e^*_s) +
[1-(\theta+e^*_s)g_s^1- (1-\theta-e^*_s) g_s^0] q_s^-(e^*_s) -
c(e^*_s).
\]
A school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is \textit{incentive compatible without transfers}
if every student $j$ weakly prefers attending school $\phi(j)$ than any other school, that is, if
for every $j\in \{1, \cdots, m\}$:
\[
U^*_{\phi(j)}(\theta_j) =\max_{s \in \{1, \cdots,n\} } ~ U^*_s(\theta_j).
\]
\begin{theorem} \label{thm2}
Any
school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ that is incentive compatible without transfers
has a single-tier structure, and satisfies
$e^*_{s_A}=e^*_{s_B}$
for all $s_A,s_B \in \{1, \cdots, n\}$.
\end{theorem}
\begin{proof} Consider a school system
$\big(n, \phi,
(G_{s})_{s=1}^n \big)$. We have
\begin{align*}
U^*_s(\theta) &= q_s^-(e^*_s)+ \big[g_0^s + (\theta+e^*_s) (g_1^s- g_0^s) \big] \big[ q_s^+(e^*_s)- q_s^-(e^*_s)\big]- c(e^*_s)
\\
&= q_s^-(e^*_s)+ \big[g_0^s + (\theta+e^*_s) (g_1^s- g_0^s) \big] \left(\frac{c'(e^*_s)}{g_1^s- g_0^s} \right) - c(e^*_s)
\\
&= \zeta_s + \theta c'(e^*_s),
\end{align*}
where
\[
\zeta_s= q_s^-(e^*_s)+ \big[g_0^s + e^*_s (g_1^s- g_0^s) \big] \left(\frac{c'(e^*_s)}{g_1^s- g_0^s} \right) - c(e^*_s).
\]
So if $s_A$ and $s_B$ are two schools forming part of the school system considered, then
\begin{equation}\label{eq:singlecrossing}
U^*_{s_B}(\theta) -
U^*_{s_A}(\theta)= (\zeta_{s_B}- \zeta_{s_A}) + \theta\big(c'(e^*_{s_B})- c'(e^*_{s_A}) \big).
\end{equation}
Now reason by contradiction: suppose that some school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is incentive compatible without transfers, and we can find two schools $s_A$ and $s_B$ forming part of this system such that
$e^*_{s_B} >e^*_{s_A}$. Then,
by \eqref{eq:singlecrossing}, every student attending school $s_A$
must have a weakly lower type than all students attending school
$s_B$, whence
$\bar{\theta}_{s_A} \leq \bar{\theta}_{s_B}$.
We thus obtain
\begin{align*}
U^*_{s_B}(\bar{\theta}_{s_A}) = U_{s_B}(\bar{\theta}_{s_A}, &e^*_{s_B}, e^*_{s_B}) > U_{s_B}(\bar{\theta}_{s_A}, e^*_{s_A}, e^*_{s_B}) \\
&> U_{s_B}(\bar{\theta}_{s_A}, e^*_{s_A}, e^*_{s_A})
\geq
\bar{\theta}_{s_A} + e^*_{s_A} -c( e^*_{s_A})
=
U_{s_A}(\bar{\theta}_{s_A}, e^*_{s_A}, e^*_{s_A}) =U^*_{s_A}(\bar{\theta}_{s_A}).
\end{align*}
The first inequality captures the fact that effort $e^*_{s_B}$ is optimal in school $s_B$, while $e^*_{s_A}$ is not; the second inequality captures the benefit from increasing employers' conjecture regarding effort exerted in school $s_B$ (from $e^*_{s_A}$ to $e^*_{s_B}$); the third inequality captures the benefit from increasing employers' belief concerning the average student type in a school (from $\bar{\theta}_{s_A}$ to $\bar{\theta}_{s_B}$).
The last highlighted expression combined with
\eqref{eq:singlecrossing} gives
$ U^*_{s_B}(\theta)> U^*_{s_A}(\theta)$ for all $\theta \geq \bar{\theta}_{s_A}$.
Now, some student $j$ attending school $s_A$ has type $\theta_j \geq \bar{\theta}_{s_A}$, and so $ U^*_{s_B}(\theta_j)> U^*_{\phi(j)}(\theta_j)$, contradicting our assumption that $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is incentive compatible without transfers.
This shows that if
$\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is incentive compatible without transfers, then
$e^*_{s_A}=e^*_{s_B}$
for all $s_A,s_B \in \{1, \cdots, n\}$. Furthermore,
$\bar{\theta}_{s_A}= \bar{\theta}_{s_B}$, for the same reasoning as above shows that
$\bar{\theta}_{s_B}> \bar{\theta}_{s_A}$
implies
$ U^*_{s_B}(\theta)> U^*_{s_A}(\theta)$ for all $\theta \geq \bar{\theta}_{s_A}$.
\end{proof}
Theorem
\ref{thm2} tells us that if a designer is unable to observe students' types and cannot use transfers then she can do no better than placing all students in one school.
The basic idea is simple: if the average type of students in school A were greater than in school B, then all students would want to choose school A so as to benefit from a more favorable prior in the eyes of employers.
\subsection{Incentive Compatibility with Transfers} \label{subsec:ICwith}
Transfers, in the form of differentiated fees for example, can play an important role in the design of school systems. We show in this subsection that, with transfers, Theorem \ref{thm1} naturally extends, in the sense that every \emph{constrained} optimal school system has either a
single-tier structure
or a two-tier structure.
To streamline the exposition of our next result, we examine in this subsection the continuous version of our model.
There is a continuum of students, with types distributed according to the atomless
distribution $F$
such that $\text{supp}\; F= [0, \alpha]$. The assignment of students to schools is
captured by a pair $(n, \phi)$, where
$n \in \mathbb{N}$
and
$\phi$
is a measurable mapping from
$(0, \alpha)$ to $\{1, \cdots, n\}$
such that $m_s:=\int_{\phi^{-1}(s)} dF(\theta)>0$ for $s=1, \cdots, n$: we interpret $\phi$ as saying that type-$\theta$ students attend school $\phi(\theta)$.
The rest is as in the discrete model of Section \ref{sectionmodel}.
A school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is \textit{incentive compatible with transfers}
if
there exist school fees $t_1, \cdots, t_n$ such that, for every $\theta \in (0, \alpha)$,
type-$\theta$ students weakly prefer attending school $\phi(\theta)$ than any other school, that is,
\[
U^*_{\phi(\theta)}(\theta) - t_{\phi(\theta)}
=\max_{s \in \{1, \cdots,n\} } ~ U^*_s(\theta)-t_s, \qquad \forall\; \theta \in (0, \alpha).
\]
We say that a school system is \textit{constrained optimal} if it maximizes social welfare within the class of school systems that are incentive compatible with transfers.
\begin{theorem}\label{thm3}
Every constrained-optimal school system has either a
single-tier structure
or a two-tier structure.\footnote{In the continuous version of the model, a school system has a two-tier structure
if every school belongs to one of
two tiers, such that: (i)
all schools in a given tier have the same average student type, (ii) the
type of almost every student in a top-tier school is at least as a large as the type of almost any student in a bottom-tier school.
In other words, sets of students that have measure zero play no role. } In the latter case, schools in the top tier have grading rule $G_{\star}$, and schools in the bottom tier have grading rule $G^{\star}$.
\end{theorem}
The proof of Theorem \ref{thm3} is long and therefore relegated to the appendix.
We first show that
if
a school system is socially optimal and has either a single-tier structure,
or a two-tier structure such that
equilibrium effort in the top-tier schools is at least as large as equilibrium effort in the bottom-tier schools, then
this school system is also constrained optimal.
However, if
a school system is socially optimal and
has
a two-tier structure such that
equilibrium effort in the top-tier schools is less than equilibrium effort in the bottom-tier schools, then this school system is not incentive compatible with transfers.
The proof of Theorem \ref{thm3} shows that in this case a constrained-optimal school system is retrieved by
moving intermediate-type students from the top-tier schools to the bottom-tier schools. In other words,
the incentive compatibility constraints
distort the
socially optimal school system by
reducing the size of the top tier and increasing that of the bottom tier.
%The proof of Theorem \ref{thm3} combines various elements of the proofs of Theorem \ref{thm1} and \ref{thm2} presented in the text; we relegate it therefore to the appendix. Intuitively, if a constrained-optimal school system has a two-tier structure, incentive compatibility is ensured by making the fees of the bottom-tier schools smaller than those of the top-tier schools. High types willingly pay higher fees because they expect to get better grades than low types.
\section{An Example with Three Productive Values}\label{sec:morethantwo}
The grading rule $G_{\star}$ passes
every
student with productive value $1$ as well as a fraction $1-b$ of students with productive value $0$ and can thus be interpreted as a ``lenient" grading rule.
By contrast, the grading rules $G^{\star}$ fails every student
with productive value $0$ as well as a fraction $1-b$ of students with productive value $1$ and may thus be interpreted as a ``tough" grading rule.
We saw in Proposition \ref{prop2} that whether the tough grading rule $G^{\star}$ is preferable to the lenient grading rule $G_{\star}$ depends on the type of students in the school considered:
when the average student type is high, $G_{\star}$ induces more effort than
$G^{\star}$, and vice versa when
the average student type is low. We now show through an example that the logic underlying Proposition \ref{prop2} extends beyond two productive
values.
Consider a modified version of the model of Section \ref{sectionmodel} with three productive values, $v_1<v_2<v_3$. Let $(\underline{\pi}_1, \underline{\pi}_2, \underline{\pi}_3)$ and $(\overline{\pi}_1, \overline{\pi}_2, \overline{\pi}_3)$ represent
two probability distributions over $\{v_1, v_2, v_3\}$, such that $(\overline{\pi}_1, \overline{\pi}_2, \overline{\pi}_3)$ first-order stochastically dominates $(\underline{\pi}_1, \underline{\pi}_2, \underline{\pi}_3)$.
At the end of school, each student's productive
value is either $v_1$, $v_2$, or $v_3$. We assume that the distribution of student $j$'s productive
value is a convex combination of $\underline{\pi}$
and $\overline{\pi}$:
\[
(1-\theta_j- e_j) \underline{\pi}+ (\theta_j+ e_j) \overline{\pi},
\]
where $\theta_j \in (0, \alpha)$ denotes student $j$'s type, and $e_j \in [0, 1-\alpha]$ his effort. Thus, effort increases a student's chances of acquiring a high productive value.
As in Section \ref{sectionmodel}, let $b$ denote a parameter in $(0,1)$, and
$G_{\star}$ and $G^{\star}$ be two grading rules, defined as follows. Under both grading rules,
a student with productive value $v_1$ obtains the grade $C$ with probability $1$, a student with productive value
$v_2$ obtains the grade $B$ with probability $b$,
and a student with productive value $v_3$ obtains the grade $A$ with probability $1$.
However, whereas under
$G_{\star}$ a student with productive value
$v_2$ obtains the grade $A$ with probability $1-b$, under
$G^{\star}$ a student with productive value
$v_2$ obtains the grade $C$ with probability $1-b$.
In words, $G_{\star}$ occasionally inflates $B$ to $A$, whereas $G^{\star}$ occasionally deflates $B$ to $C$.
In line with
Section \ref{sectionmodel}, the grading rules $G^{\star}$ and $G_{\star}$ may thus be interpreted as tough and lenient, respectively.
The following result mirrors
Proposition \ref{prop2}:
\begin{proposition}\label{propappendix}
There exists $\Bar{\theta}^\dagger$ such that $\xi(\Bar{\theta}, G^{\star})> \xi(\Bar{\theta}, G_{\star})$ for $\Bar{\theta}<\Bar{\theta}^\dagger$, and
$\xi(\Bar{\theta}, G_{\star})> \xi(\Bar{\theta}, G^{\star})$ for $\Bar{\theta}>\Bar{\theta}^\dagger$.
\end{proposition}
\begin{proof} For any school $s$ forming part of a given school system, let
$q_s^A(\hat{e}_s)$ denote employers' posterior expectation of the productive value of a student of $s$ who obtained the grade $A$ (given employers' conjecture that students of $s$ exert effort $\hat{e}_s$). Let $q_s^B(\hat{e}_s)$ and $q_s^C(\hat{e}_s)$ be similarly defined for grades $B$ and $C$, respectively.
If $G_s=G^{\star}$,
a student's marginal benefit of exerting effort
in school $s$ is given by
\[
B_s( G^{\star}, \hat{e}_s):= (\overline{\pi}_1- \underline{\pi}_1) q_s^C(\hat{e}_s) + (\overline{\pi}_2- \underline{\pi}_2) \big[b q_s^B(\hat{e}_s) + (1-b) q_s^C(\hat{e}_s) \big] + (\overline{\pi}_3- \underline{\pi}_3) q_s^A(\hat{e}_s);
\]
if $G_s=G_{\star}$,
the marginal benefit of effort
is given by
\[
B_s( G_{\star}, \hat{e}_s):= (\overline{\pi}_1- \underline{\pi}_1) q_s^C(\hat{e}_s) + (\overline{\pi}_2- \underline{\pi}_2) \big[b q_s^B(\hat{e}_s) + (1-b) q_s^A(\hat{e}_s) \big] + (\overline{\pi}_3- \underline{\pi}_3) q_s^A(\hat{e}_s).
\]
For any school $s$ forming part of a given school system, $e$ is an equilibrium effort in school $s$ if and only if
\[
c'(e)= B_s( G_s, \hat{e}_s).
\]
The function $e \mapsto B_s(G_s, + e )- c'(e)$ is
strictly concave, positive at $e=0$, and negative at $e=1-\alpha$. So the previous equation has a unique solution.
This solution is
determined by $\bar{\theta}_s$ and $G_s$, and so can be denoted by $\xi(\Bar{\theta}_s, G_s)$.
Now
consider a school $s$ such that $G_s=G^{\star}$. Then $q_s^A(\hat{e}_s)=v_3$, $q_s^B(\hat{e}_s)=v_2$, and
$q_s^C(\hat{e}_s)$ is increasing in $\hat{e}_s$. As
$\overline{\pi}_1- \underline{\pi}_1<0$ and either $\overline{\pi}_2- \underline{\pi}_2<0$ or $(1-b)|\overline{\pi}_2- \underline{\pi}_2|<|\overline{\pi}_1- \underline{\pi}_1|$, it follows that
the function $e \mapsto B_s( G^{\star}, e )$ is decreasing. This, in turn,
implies that the function $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing.
Next,
consider a school $s$ such that $G_s=G_{\star}$. Then $q_s^A(\hat{e}_s)$ is increasing in $\hat{e}_s$, $q_s^B(\hat{e}_s)=v_2$, and
$q_s^C(\hat{e}_s)=v_1$. As
$\overline{\pi}_3- \underline{\pi}_3>0$ and either
$\overline{\pi}_2- \underline{\pi}_2>0$ or
$(1-b)|\overline{\pi}_2- \underline{\pi}_2|<|\overline{\pi}_3- \underline{\pi}_3|$, it follows that
the function $e \mapsto B_s( G_{\star}, e )$ is increasing. This, in turn,
implies that the function $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ is increasing. Combining the previous remarks concludes the proof.
\end{proof}
While Proposition
\ref{propappendix} extends Proposition \ref{prop2} to an environment with three productive values, moving beyond binary productive values introduces various challenges. Firstly,
there appears to be no natural notion of
leniency ordering the set of all possible grading rules.
Secondly, beyond binary productive values,
leniency alone does not to capture all relevant aspects of a grading rule. To illustrate the latter point, consider the grading rules
$\Tilde{G}_{\star}$ and $\Tilde{G}^{\star}$ defined as follows. Under both grading rules,
a student with productive value $v_2$ obtains the grade $B$ with probability $1$.
Under
$\Tilde{G}_{\star}$,
a student with productive value
$v_3$ obtains the grade $A$ with probability $1$, while
a student with productive value
$v_1$ obtains the grade $C$ with probability $b$ and the grade $B$ with probability $1-b$. By contrast, under $\Tilde{G}^{\star}$,
a student with productive value
$v_1$ obtains the grade $C$ with probability $1$, while
a student with productive value
$v_3$ obtains the grade $A$ with probability $b$ and the grade $B$ with probability $1-b$.
In words, $\Tilde{G}_{\star}$ occasionally inflates $C$ to $B$, whereas $\Tilde{G}^{\star}$ occasionally deflates $A$ to $B$.
The grading rule $\Tilde{G}_{\star}$ may therefore, in a sense, be viewed as more lenient than $\Tilde{G}^{\star}$. Yet, whether
$\Tilde{G}_{\star}$ induces more effort than $\Tilde{G}^{\star}$ or the way around need not be monotonic in a school's average student type.
To see why, consider a school $s$ with $G_s=\Tilde{G}_{\star}$. Increasing $\bar{\theta}_s$ increases $q_s^B(\hat{e}_s)$, which raises the gain $q_s^B(\hat{e}_s)-q_s^C(\hat{e}_s)$ from obtaining grade $B$ instead of $C$, but reduces the gain $q_s^A(\hat{e}_s)-q_s^B(\hat{e}_s)$ from obtaining grade $A$ instead of $B$. Which of these effects dominates the other depends on the parameters.
\section{Conclusion}\label{sec:conc}
This paper proposes a framework to assess how to optimally sort and grade
students. The premise of our model is that an individual's productive value results from a combination of ability and effort. But potential employers observe neither productive value, nor ability, nor effort:
they have to make decisions based solely on the school which an individual attended and the grades that he obtained. The way in which students are sorted and graded therefore determines incentives to exert effort.
We show that if a planner can freely assign students to schools (as in, e.g., Austria, England, Germany and Switzerland) then any socially optimal school system either randomly assigns students to schools or separates students in just two ability groups.
If the variance in student ability is small, then
a random assignment is optimal; otherwise students whose ability is above a cutoff should be separated from those whose ability is below the cutoff. In the latter case,
bottom-tier students are evaluated through
a tough grading rule, and top-tier students through a lenient grading rule.
We then study the case in which students can choose the school they go to (as in, e.g., France and Italy). We show that
in the absence of transfers, a planner can do no better than randomly assigning students to schools. However, with transfers, any constrained-optimal school system has the same qualitative
features as when the
planner
can
freely assign students to schools.
\newpage
\section*{Appendix: Proof of Theorem
\ref{thm3}}
We start with a couple of definitions. We will say that
a school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ is \textit{$r$-regular} if there exist $r$ non-empty sets $S_1, \cdots, S_r$ partitioning $\{1,\cdots, n\}$, cutoffs $0= \hat{\theta}_0<\hat{\theta}_1< \cdots<\hat{\theta}_{r-1}< \hat{\theta}_r=\alpha$, and effort levels $e_1 \leq \cdots \leq e_r$
such that
\begin{itemize}
\item for every $k\in \{1, \cdots, r\}$,
$\phi(\theta) \in S_k$ for all $\theta \in ( \hat{\theta}_{k-1}, \hat{\theta}_k)$;
\item $\phi (\hat{\theta}_k) \in S_k \cup S_{k+1}$ for all $k\in \{0, \cdots, r-1\}$, and $\phi (\hat{\theta}_r) \in S_r$ ;
\item
$e^*_{s}= e_k $ for all school $s \in S_k$, $k\in \{1, \cdots, r\}$.
\end{itemize}
A school system is \textit{regular} if it is $r$-regular for some $r$.
We then refer to $\big( (\hat{\theta}_k)_{k=0}^r, (e_k)_{k=1}^r \big)$ as the \textit{signature} of the regular school system, and define
\[
M_k:= \int_{\phi^{-1}(S_k)} dF(\theta)= \int_{\hat{\theta}_{k-1}}^{\hat{\theta}_{k}} dF(\theta), \quad ~~ k \in \{1, \cdots, r\}.
\]
Notice that
the social welfare which
a regular school system generates is entirely determined by its signature:
\begin{equation*}
W\big(n, \phi,
(G_{s})_{s=1}^n \big):=
\sum_{k=1}^r M_k
\left[ \frac{ \int_{\phi^{-1}(S_k)} \theta dF(\theta)}{M_k} +e_k - c( e_k) \right].
\end{equation*}
Finally, given two regular school systems $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ and $\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)$ with
signatures $\big( (\hat{\theta}_k)_{k=0}^r, (e_k)_{k=1}^r \big)$ and $\big( (\hat{\theta}'_k)_{k=0}^{r'}, (e'_k)_{k=1}^{r'} \big)$,
we will say that $\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)$ is
a \textit{non-decreasing contraction} of $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ if
\begin{itemize}
\item $\{ \hat{\theta}'_k \}_{k=1}^{r'} \subseteq \{ \hat{\theta}_k \}_{k=1}^{r}$;
\item $e'_k \geq \frac{\sum_{l=k^{-}}^{k^+} M_le_l
}{\sum_{l=k^{-}}^{k^+} M_l}$, for all $k \in \{1, \cdots, r'\}$, where $k^-$ and $k^+$ are the indices such that $\hat{\theta}'_{k-1}= \hat{\theta}_{k^-}$ and $\hat{\theta}'_{k}= \hat{\theta}_{k^+}$.
\end{itemize}
By extension, we say that $\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)$ is
an \textit{increasing contraction} of $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ if it is a
non-decreasing contraction of the latter school system and $e'_k > \frac{\sum_{l=k^{-}}^{k^+} M_le_l
}{\sum_{l=k^{-}}^{k^+} M_l}$ for at least one $k \in \{1, \cdots, r'\}$.
The next two lemmas shed light on the above definitions.
\begin{lemma}\label{lemma:regss1}
A school system is incentive compatible with transfers if and only if it is regular.
\end{lemma}
\begin{proof}
We established in the proof of Theorem \ref{thm2} that, given a school system $\big(n, \phi,
(G_{s})_{s=1}^n \big)$,
\[
U^*_s(\theta) = \zeta_s + \theta c'(e^*_s),
\]
for some $\zeta_s$ independent of $\theta$. As $c'$ is increasing,
the lemma
follows from standard mechanism design arguments.
\end{proof}
\begin{lemma} \label{lemma:regss2}
If $\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)$ is
a non-decreasing contraction of $\big(n, \phi,
(G_{s})_{s=1}^n \big)$ then
\newline $W\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)\geq W\big(n, \phi,
(G_{s})_{s=1}^n \big)$.
If $\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)$ is
an increasing contraction of
$\big(n, \phi,
(G_{s})_{s=1}^n \big)$ then $W\big(n', \phi',
(G'_{s})_{s=1}^{n'} \big)> W\big(n, \phi,
(G_{s})_{s=1}^n \big)$.
\end{lemma}
\begin{proof}
We saw in the proof of Theorem \ref{thm1} that there exists $\overline{\overline{e}}$ such that
any school $s$ forming part of a given school system is such that $e^*_s<
\overline{\overline{e}}
$, and
$e\mapsto e-c(e)$ is increasing and concave on $[0, \overline{\overline{e}}]$.
\end{proof}
\quad
\begin{proof}[Proof of Theorem \ref{thm3}]
Pick a regular school system
with signature $\big( (\hat{\theta}_k^{(0)})_{k=0}^r, (e^{(0)}_k)_{k=1}^r \big)$. Suppose
that $\Sigma^{(0)}$ does not have a single-tier structure, nor a two-tier structure.
We will prove that $\Sigma^{(0)}$ is not constrained optimal.
In line with previous notation, let
$S^{(0)}_k$ denote the set of schools
such that $\phi(\theta) \in S^{(0)}_k$ for all $\theta \in ( \hat{\theta}^{(0)}_{k-1}, \hat{\theta}^{(0)}_{k})$.
As
$\Sigma^{(0)}$ has neither a single nor a two-tier structure, one of the following cases holds:
\begin{enumerate}
\item[Case 1.] $\Sigma^{(0)}$ is $r$-regular with $r\geq 3$;
\item [Case 2.] $\Sigma^{(0)}$ is $2$-regular, and one of $S_1^{(0)}$ and $S_2^{(0)}$ contains two schools whose average student types differ;
\item [Case 3.]
we can find two schools forming part of
$\Sigma^{(0)}$, $s_A$ and $s_B$ say, as well as two groups of students with positive mass, $A$ and $B$ say, such that (i) students in group $A$ attend school $s_A$, (ii) students in group $B$ attend school $s_B$, (iii)
$\bar{\theta}_{s_A}<\bar{\theta}_{s_B}$, (iv) the type of any students in group $A$ is larger than the type of every student in group $B$.
\end{enumerate}
We treat Case 1 below; the other cases are similar.
To streamline the exposition, suppose that
$\Sigma^{(0)}$
is $3$-regular, with signature
$\big((0, \hat{\theta}^{(0)}_1, \hat{\theta}^{(0)}_2, \alpha), (e^{(0)}_1, e^{(0)}_2, e^{(0)}_3 )\big)$, where $ \hat{\theta}^{(0)}_1<\bar{\theta}^\dagger < \hat{\theta}^{(0)}_2$ (with $\bar{\theta}^\dagger$ defined in Proposition \ref{prop2}).\footnote{The case in which $\bar{\theta}^\dagger \in \{\hat{\theta}^{(0)}_1, \hat{\theta}^{(0)}_2 \}$ is
simpler; we focus here on the more complicated case.}
\quad
\noindent \underline{ \textit{Step 1:}}
Let $\Sigma^{(1)}$ denote the school system obtained from $\Sigma^{(0)}$ by merging all schools in the set $S^{(0)}_3$ into one school with grading rule $G_{\star}$. As $\hat{\theta}^{(0)}_2> \bar{\theta}^\dagger$, all students in $S^{(0)}_3$ have type greater than $\bar{\theta}^\dagger$, whence, by Propositions
\ref{prop1}-\ref{prop3}: $e^{(1)}_3\geq e^{(0)}_3$.
It follows that $\Sigma^{(1)}$ is a non-decreasing contraction of $\Sigma^{(0)}$.
\quad
\noindent \underline{ \textit{Step 2:}}
Let $S^{(1)+}_2$ denote the subset of $S^{(1)}_2$ comprising the schools whose average student type is at least as large as
$\bar{\theta}^\dagger$, and $S^{(1)-}_2$ the subset of $S^{(1)}_2$ comprising the schools whose average student type is less than
$\bar{\theta}^\dagger$.
Now
let $\Sigma^{(2)}$ denote the school system obtained from $\Sigma^{(1)}$ by (i) merging all schools in the set $S^{(1)-}_2$ into one school, $s_A$ say, with grading rule $G^{\star}$, and (ii) merging all schools in the set $S^{(1)+}_2$ into one school, $s_B$ say, with grading rule $G_{\star}$. By Propositions
\ref{prop1}-\ref{prop3}:
\begin{equation}\label{eqtrainnevy1}
\min \{e^*_{s_A}, e^*_{s_B} \}\geq e^{(1)}_2.
\end{equation}
\quad
\noindent \underline{ \textit{Step 3:}}
Let $\Sigma^{(3)}$ denote the school system obtained from $\Sigma^{(2)}$ by swapping students between $s_A$ and $s_B$
so as to create two schools, $s_A'$ and $s_B'$ say, such that (i) $m_{s_A'}=m_{s_A}$, (ii) $m_{s_B'}=m_{s_B}$, and (iii) the type of each student in $s_B'$ is greater than the type of any student
in $s_A'$. We saw in the proof of Proposition \ref{prop2}
that $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing, and $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ increasing. Therefore: $e^*_{s_A'} \geq e^*_{s_A}$ and $e^*_{s_B'} \geq e^*_{s_B}$. Then \eqref{eqtrainnevy1} yields
\begin{equation}\label{eqtrainnevy2}
\min \{e^*_{s_A'}, e^*_{s_B'} \}\geq e^{(1)}_2.
\end{equation}
\quad
\noindent \underline{ \textit{Step 4:}}
Let $\Sigma^{(4)}$ denote the school system obtained from $\Sigma^{(3)}$ as follows. If $e^*_{s_B'}<e^*_{s_A'}$, replace the grading rule of
$s_A'$
by the grading rule $(0,\Tilde{b})$, with $\Tilde{b}<b$ chosen so as to
ensure $e^*_{s_B'}=e^*_{s_A'}$. If instead $e^*_{s_A'}<e^*_{s_B'}$, replace the grading rule of
$s_B'$
by the grading rule $(1-\Tilde{b},1)$, with $\Tilde{b}<b$ chosen so as to
ensure $e^*_{s_B'}=e^*_{s_A'}$. Then, by \eqref{eqtrainnevy2}, $\Sigma^{(4)}$ is a non-decreasing contraction of $\Sigma^{(1)}$.
\quad
\noindent \underline{ \textit{Step 5:}}
Let $S^{(4)+}$ denote the subset of schools in $\Sigma^{(4)}$
whose average student type is at least as large as
$\bar{\theta}^\dagger$, and $S^{(4)-}$
the subset of schools in $\Sigma^{(4)}$
whose average student type is less than
$\bar{\theta}^\dagger$. Now
let $\Sigma^{(5)}$ denote the school system obtained from $\Sigma^{(4)}$
by (i) merging all schools in $S^{(4)-}$ into one school, $s''_A$ say, with grading rule $G^{\star}$, and (ii) merging all schools in $S^{(4)+}$ into one school, $s_B''$ say, with grading rule $G_{\star}$. By Propositions
\ref{prop1}-\ref{prop3}:
\begin{equation}\label{eqmonbard}
e^*_{s_A''}\geq \frac{\sum_{s \in S^{(4)-}} m_s e^*_s}{ \sum_{s \in S^{(4)-}} m_s}
, \quad
e^*_{s_B''}\geq \frac{\sum_{s \in S^{(4)+}} m_s e^*_s}{ \sum_{s \in S^{(4)+}} m_s},
\end{equation}
and one of those inequalities must be strict.
\quad
\noindent \underline{ \textit{Step 6:}}
Let $\Sigma^{(6)}$ denote the school system obtained from $\Sigma^{(5)}$ as follows. If
$ e^*_{s_B''}\geq e^*_{s_A''}$, keep $\Sigma^{(5)}$ unchanged. If $ e^*_{s_B''}< e^*_{s_A''}$, increase the cutoff type separating students between the two schools so as to ensure $ e^*_{s_B''} =e^*_{s_A''}$ (recall,
$\Bar{\theta} \mapsto \xi(\Bar{\theta}, G^{\star})$ is decreasing, and $\Bar{\theta} \mapsto \xi(\Bar{\theta}, G_{\star})$ increasing).
Then, by \eqref{eqmonbard},
$\Sigma^{(6)}$ is an increasing contraction of
$\Sigma^{(4)}$.
\quad
As $\Sigma^{(1)}$ is a non-decreasing contraction of $\Sigma^{(0)}$,
$\Sigma^{(4)}$ is a non-decreasing contraction of $\Sigma^{(1)}$, and
$\Sigma^{(6)}$ is an increasing contraction of
$\Sigma^{(4)}$, we conclude that $\Sigma^{(6)}$ is an increasing contraction of
$\Sigma^{(0)}$. By Lemma \ref{lemma:regss2}, the social welfare generated by $\Sigma^{(6)}$ is thus strictly larger than the social welfare generated by $\Sigma^{(0)}$.
Yet, by Lemma \ref{lemma:regss1}, $\Sigma^{(6)}$ is incentive compatible. This shows that $\Sigma^{(0)}$ is not constrained optimal.
\end{proof}
\newpage
%We are unable to characterize the socially optimal school systems in environments with more than two productive values, but conjecture that the broad ideas underlying Theorem \ref{thm1} are very general. If student types are very dispersed, then because separating students enables the grading rule to be tailored to their types, creating a tiered school system could be socially optimal.
\bibliographystyle{apecon}
\bibliography{library}
\end{document}
\begin{thebibliography}{42}
\providecommand{\natexlab}[1]{#1}
\bibitem[{Arnott and Rowse(1987)}]{arnott1987peer}
Arnott, R. and Rowse, J. (1987) Peer group effects and educational attainment,
\textit{Journal of Public Economics}, \textbf{32}, 287--305.
\bibitem[{Arrow(1973)}]{arrow1973}
Arrow, K. (1973) The theory of discrimination, in \textit{Discrimination in
Labor Markets}, Princeton University Press, vol.~3, pp. 3--33.
\bibitem[{Becker and Rosen(1992)}]{becker1992learning}
Becker, W.~E. and Rosen, S. (1992) {The Learning Effect of Assessment and
Evaluation in High School}, \textit{Economics of Education Review},
\textbf{11}, 107--118.
\bibitem[{Bergemann \textit{et~al.}(2018)Bergemann, Bonatti and
Smolin}]{bergemann2018design}
Bergemann, D., Bonatti, A. and Smolin, A. (2018) {The Design and Price of
Information}, \textit{American Economic Review}, \textbf{108}, 1--48.
\bibitem[{Betts(1998)}]{Betts98}
Betts, J. (1998) {The Impact of Educational Standards on the Level and
Distribution of Earnings}, \textit{American Economic Review}, \textbf{88},
266--275.
\bibitem[{Betts(2011)}]{Betts2011Review}
Betts, J.~R. (2011) {The Economics of Tracking in Education}, in
\textit{{Handbook of the Economics of Education}} (Eds.) E.~Hanushek,
S.~Machin and L.~Woessmann, Elsevier, vol.~3 of \textit{Handbook of the
Economics of Education}, chap.~7, pp. 341--381.
\bibitem[{Board(2009)}]{board2009monopolistic}
Board, S. (2009) {Monopolistic Group Design with Peer Effects},
\textit{Theoretical Economics}, \textbf{4}, 89--125.
\bibitem[{Boleslavsky and Cotton(2015)}]{BoleslavskyCotton15}
Boleslavsky, R. and Cotton, C. (2015) {Grading Standards and Education
Quality}, \textit{American Economic Journal: Microeconomics}, \textbf{7},
248--279.
\bibitem[{Chan \textit{et~al.}(2007)Chan, Hao and Suen}]{chan2007signaling}
Chan, W., Hao, L. and Suen, W. (2007) {A Signaling Theory of Grade Inflation},
\textit{International Economic Review}, \textbf{48}, 1065--1090.
\bibitem[{Coate and Loury(1993)}]{coate1993will}
Coate, S. and Loury, G.~C. (1993) Will affirmative-action policies eliminate
negative stereotypes?, \textit{The American Economic Review}, pp. 1220--1240.
\bibitem[{Cook(1924)}]{cook1924study}
Cook, R. (1924) A study of the results of homogeneous grouping of abilities in
high-school classes, \textit{the Education of Gifted Children, National
Society of the Study of Education Yearbook}, \textbf{23}.
\bibitem[{Costrell(1994)}]{costrell1994simple}
Costrell, R.~M. (1994) {A Simple Model of Educational Standards},
\textit{American Economic Review}, pp. 956--971.
\bibitem[{Dubey and Geanakoplos(2010)}]{DubeyGeanakoplos10}
Dubey, P. and Geanakoplos, J. (2010) {Grading Exams: 100, 99, 98 ... or A, B,
C?}, \textit{Games and Economic Behavior}, \textbf{69}, 72--94.
\bibitem[{Duflo \textit{et~al.}(2011)Duflo, Dupas and Kremer}]{Duflo2011Kenya}
Duflo, E., Dupas, P. and Kremer, M. (2011) {Peer Effects, Teacher Incentives,
and the Impact of Tracking: Evidence from a Randomized Evaluation in Kenya},
\textit{American Economic Review}, \textbf{101}, 1739–1774.
\bibitem[{Ehlers and Schwager(2016)}]{ehlers2016honest}
Ehlers, T. and Schwager, R. (2016) Honest grading, grade inflation, and
reputation, \textit{CESifo Economic Studies}, \textbf{62}, 506--521.
\bibitem[{Es{\H{o}} and Szentes(2007)}]{esHo2007optimal}
Es{\H{o}}, P. and Szentes, B. (2007) {Optimal Information Disclosure in
Auctions and the Handicap Auction}, \textit{The Review of Economic Studies},
\textbf{74}, 705--731.
\bibitem[{Galindo-Rueda and Vignoles(2007)}]{galindo2007heterogeneous}
Galindo-Rueda, F. and Vignoles, A. (2007) The heterogeneous effect of selection
in uk secondary schools, \textit{Schools and the equal opportunity problem},
pp. 103--128.
\bibitem[{Guo and Shmaya(2019)}]{guo2019interval}
Guo, Y. and Shmaya, E. (2019) {The Interval Structure of Optimal Disclosure},
\textit{Econometrica}, \textbf{87}, 653--675.
\bibitem[{Himmler and Schwager(2013)}]{himmler2013double}
Himmler, O. and Schwager, R. (2013) Double standards in educational
standards--do schools with a disadvantaged student body grade more
leniently?, \textit{German Economic Review}, \textbf{14}, 166--189.
\bibitem[{Jehiel and Moldovanu(2001)}]{jehiel2001efficient}
Jehiel, P. and Moldovanu, B. (2001) {Efficient Design with Interdependent
Valuations}, \textit{Econometrica}, \textbf{69}, 1237--1259.
\bibitem[{Jehiel \textit{et~al.}(1996)Jehiel, Moldovanu and
Stacchetti}]{jehiel1996not}
Jehiel, P., Moldovanu, B. and Stacchetti, E. (1996) {How (Not) to Sell Nuclear
Weapons}, \textit{The American Economic Review}, pp. 814--829.
\bibitem[{Jehiel \textit{et~al.}(1999)Jehiel, Moldovanu and
Stacchetti}]{jehiel1999multidimensional}
Jehiel, P., Moldovanu, B. and Stacchetti, E. (1999) {Multidimensional Mechanism
Design for Auctions with Externalities}, \textit{Journal of Economic Theory},
\textbf{85}, 258--293.
\bibitem[{Kerr \textit{et~al.}(2013)Kerr, Pekkarinen and
Uusitalo}]{Kerr2013Cognitive}
Kerr, S.~P., Pekkarinen, T. and Uusitalo, R. (2013) {School Tracking and
Development of Cognitive Skills}, \textit{Journal of Labor Economics},
\textbf{31}, 577--602.
\bibitem[{Kolotilin \textit{et~al.}(2017)Kolotilin, Mylovanov, Zapechelnyuk and
Li}]{kolotilin2017persuasion}
Kolotilin, A., Mylovanov, T., Zapechelnyuk, A. and Li, M. (2017) {Persuasion of
a Privately Informed Receiver}, \textit{Econometrica}, \textbf{85},
1949--1964.
\bibitem[{Li and Shi(2017)}]{li2017discriminatory}
Li, H. and Shi, X. (2017) {Discriminatory Information Disclosure},
\textit{American Economic Review}, \textbf{107}, 3363--85.
\bibitem[{Matthewes(2021)}]{matthewes2021better}
Matthewes, S.~H. (2021) Better together? heterogeneous effects of tracking on
student achievement, \textit{The Economic Journal}, \textbf{131}, 1269--1307.
\bibitem[{Onuchic(2022)}]{onuchic22}
Onuchic, P. (2022) Recent contributions to theories of discrimination,
\textit{arXiv preprint arXiv:2205.05994}.
\bibitem[{Ostrovsky and Schwarz(2010)}]{OstrovskySchwarz10}
Ostrovsky, M. and Schwarz, M. (2010) {Information Disclosure and Unraveling in
Matching Markets}, \textit{American Economic Journal: Microeconomics},
\textbf{2}, 34--63.
\bibitem[{Rayo(2013)}]{rayo2013monopolistic}
Rayo, L. (2013) {Monopolistic Signal Provision}, \textit{The BE Journal of
Theoretical Economics}, \textbf{13}, 27--58.
\bibitem[{Rothschild and Scheuer(2013)}]{rothschild2013redistributive}
Rothschild, C. and Scheuer, F. (2013) {Redistributive Taxation in the Roy
Model}, \textit{The Quarterly Journal of Economics}, \textbf{128}, 623--668.
\bibitem[{Rothschild and Scheuer(2016)}]{rothschild2016optimal}
Rothschild, C. and Scheuer, F. (2016) {Optimal Taxation with Rent-seeking},
\textit{The Review of Economic Studies}, \textbf{83}, 1225--1262.
\bibitem[{Slavin(1990)}]{slavin1990achievement}
Slavin, R.~E. (1990) Achievement effects of ability grouping in secondary
schools: A best-evidence synthesis, \textit{Review of educational research},
\textbf{60}, 471--499.
\bibitem[{Spence(1973)}]{spence1973press}
Spence, M. (1973) Job market signaling, \textit{The Quarterly Journal of
Economics}, \textbf{87}, 355--374.
\bibitem[{Tampieri(2019)}]{tampieri2019students}
Tampieri, A. (2019) Students’ social origins and targeted grading,
\textit{The BE Journal of Theoretical Economics}, \textbf{20}, 20170146.
\bibitem[{Wei and Green(2020)}]{wei2020reverse}
Wei, D. and Green, B. (2020) {(Reverse) Price Discrimination with Information
Design}, \textit{Available at SSRN 3263898}.
\bibitem[{Whipple(1936)}]{whipple1936grouping}
Whipple, G.~M. (1936) The grouping of pupils, national society for the study of
education yearbook vol. 35, 1, \textit{Bloomington III: Public School
Publishing}.
\bibitem[{Whitty and Power(2015)}]{WhittyPower2015}
Whitty, G. and Power, S. (2015) Selective, comprehensive and diversified
secondary schooling in england: A brief history, in \textit{"The Ins and Outs
of Selective Secondary Schools: A Debate"} (Ed.) A.~de~Waal, London: Civitas,
pp. 1--8.
\bibitem[{W\"{o}{\ss}mann(2009)}]{Woessmann2009Evidence}
W\"{o}{\ss}mann, L. (2009) {International Evidence on School Tracking: A
Review}, \textit{CESifo DICE Report}, \textbf{1}, 26--34.
\bibitem[{Yamashita and Sarkisian(2021)}]{yamashita2021large}
Yamashita, T. and Sarkisian, R. (2021) {Large Mechanism Design with
Moment-based Allocation Externality}, \textit{TSE Working Papers 21-1241}.
\bibitem[{Yamashita and Zhu(2021)}]{yamashita2021type}
Yamashita, T. and Zhu, S. (2021) {Type-contingent Information Disclosure},
\textit{TSE Working Paper - 1242}.
\bibitem[{Yang(2022)}]{yang2022selling}
Yang, K.~H. (2022) {Selling Consumer Data for Profit: Optimal
Market-Segmentation Design and Its Consequences}, \textit{American Economic
Review}, \textbf{112}, 1364--93.
\bibitem[{Zubrickas(2015)}]{Zubrickas15}
Zubrickas, R. (2015) {Optimal Grading}, \textit{International Economic Review},
\textbf{56}, 751--776.
\end{thebibliography}