National Security Whistleblowing

To understand how laws affect public beliefs, we will administer a survey experiment to the American public to analyze how the public's confidence in whistleblowers changes as the punishments whistleblowers face changes. The survey design is in Qualtrics and attached.

Sample frame and solicitation: We administer the survey to Mechanical Turk workers. We will close the survey once we receive 300 results, or after 7 days. The survey is open to American adults who have a 90\% payment/completion rate on Amazon Mechanical Turk.

Vignette and treatment
The vignette first presents all subjects with baseline information about whisteblowing. For a complete vignette see the Analysis plan.

We randomly present subjects with one of two treatments or a control. The treatments induce different beliefs about the costs whistleblowers face. To induce these beliefs, we exploit the fact that the public knows little about the technical details of laws governing national security classification and disclosure or the actual consequences for real-world whistleblowers but reasonably expects that national security professionals know about these things. Thus, we are able to induce in subjects different beliefs about what national security professionals expect will happen to them if they choose to blow the whistle. Critically, we only provide truthful information to subjects based on real-world cases, emphasizing different pieces of information across treatments.

To induce beliefs that whistleblowers face harsh punishments for coming forward, our high-cost treatment reads as follows:



\begin{quotation}
\noindent There is much we do not know about whistleblowers, yet one thing is for certain: the government has enormous powers to prosecute them. Whistleblowers face the prospect of life in prison for leaking classified information. The government consistently prosecutes those that leak classified information.

For example,
\begin{itemize}
\item Lawrence Franklin, a Department of Defense analyst, leaked classified information about U.S. foreign policy toward Iran to the American Israel Public Affairs Committee, a pro-Israel lobbying group. He was sentenced to 12-years in prison.
\item Jeffery Sterling, a former officer with the Central Intelligence Agency, leaked classified information about a secret U.S. operation to set back Iran's nuclear program. He was sentenced to three-and-a-half years in prison.
\end{itemize}

Although the public is largely unaware of these cases, national security professionals are acutely aware of them. They certainly reflect on these cases before they decide to blow the whistle.

\end{quotation}



To induce beliefs that whistleblowers face little punishments for coming forward, our low-cost treatment reads as follows:


\begin{quotation}
\noindent There is much we do not know about whistleblowers, yet one thing is for certain: it is incredibly difficult to prosecute them. Many prominent whistleblowers evade attempts at prosecution. Others have been pardoned or had their sentences commuted by subsequent administrations.

For example,

\begin{itemize}
\item Daniel Ellsberg leaked classified reports for the Secretary of Defense to blow the whistle on President Johnson's secret motives for continuing the Vietnam War. A Federal Judge dismissed the case against him.
\item Thomas Drake leaked NSA program information to the Baltimore Sun to blow the whistle on waste and mismanagement at the NSA. But prosecutors dropped felony charges because they feared that they could not secure a conviction.
\end{itemize}

Although the public is largely unaware of these cases, national security professionals are acutely aware of them. They certainly reflect on these cases before they decide to blow the whistle.

\end{quotation}


The control presents examples of whistleblowers through history but does not provide any information about what happened to them once they came forward.

The remainder of the survey measures subjects' beliefs and biographical information.




\subsection{Hypotheses and Measurement}

To increase confidence that our model captures the public's beliefs, we chose five hypotheses that address different aspects of our theory. Our main objective was to identify a relationship between the legal costs whistleblowers anticipated for coming forward and public trust in whistleblowers' when they do blow the whistle. We argued above that in a world where verification of whistleblowers claims is difficult, the public could only trust whistleblowers who came forward even in the face of criminal sanction. In section \ref{sec:incentive}, we showed that the public could trust whistleblowers who knew they would face harsh legal punishments for coming forward and not otherwise.\footnote{Using mechanism design, we demonstrated that these relationships held despite many different confounding variables.}

Main Hypotheses:
H1 American citizens are more likely to trust whistleblowers if they believe whistleblowers face harsh legal punishment for coming forward.

Our model also showed that increasing costs came at a price. When the costs were too high, it was not worth it for any whistelblowers, even those who notice a genuine abuse of power, to come forward. In section \ref{sec:incentive}, we showed that in a world where the whistleblower's personal gains for coming forward were known in advance for each case, it was always possible to identify a level of legal punishment that induced honest whistleblowers to come forward. But these costs may vary from case to case. It follows that in cases where even honest whistleblowers perceive low benefits to coming forward with allegations, the threat of legal punishment may force them to stay quiet.

H2: American citizens are more likely to believe that cases of government abuse where no whistleblower comes forward exist if they believe whistleblowers face harsh legal punishment for coming forward.


Secondary hypotheses:
H3: The differential effects of high-cost punishments described in hypothesis 1 is greatest for American citizens who believe the media is likely to verify stories that are false.

H4: The differential effects of high-cost punishments described in Hypothesis 1 is greatest for American citizens who believe that whistleblowers have little to gain from public attention.

H5: The differential effects of high-cost punishments described in Hypothesis 2 is greatest for American citizens who believe that whistleblowers often come forward out of a sense of public duty.


Testing the direct effect of treatment on mechanism
We worry somewhat the treatment may not influence subjects' beliefs. To make sure that the treatment is strong enough we ask subjects the following question:

How much do you agree or disagree with the following statement: Whistleblowers face a high chance of jail-time if they come forward.


If the treatment is effective then:

\begin{equation}
\mu(E_H) < \mu(E_L).
\end{equation}

If we do not observe this difference, than we do not consider Null results from this study as informative.

Furthermore, we expect that this treatment will influence subjects overall calculation of costs and benefits. Thus, we ask subjects the following two questions:



\begin{enumerate}
\item [\textbf{F}:] When you add up all the benefits and costs that whistleblowers incur, do you think that those who come forward with \textbf{false but plausible} claims are ultimately better or worse off?
\item [\textbf{G}:] When you add up all the benefits and costs that whistleblowers incur, do you think that those who come forward with \textbf{true but plausible} claims are ultimately better or worse off?
\end{enumerate}

We measure using a 7-point scale from \textit{Much better off} to \textit{Much worse off}. We expect that those who observe the low-cost treatment will reply with answers between \textit{About the same} to \textit{Much better off} on average to both questions. However, those who observe the high-cost treatment will answer between \textit{About the same} to Much\textit{ Better off} to \textbf{G}, but \textit{About the same} to \textit{Much worse off} to question \textbf{G}.

2018-12-10

2019-03-24

To test H1 and H2 we ask:

\begin{enumerate}
\item [\textbf{A}:] In your opinion, what percentage of whistleblowers who come forward are telling the truth? The scale ranges from 0\% (no whistleblowers are telling the truth) to 100\% (all whistleblowers are telling the truth).

\item[\textbf{B}:] Do you believe that whistleblowers come forward in many or few of the total cases where the government abuses national security?
\end{enumerate}


Subjects respond to \textbf{A} using a percentage slider from 0-100 that moves in 5\% increments. Subjects respond to question \textbf{B} using a 7-point scale ranging from \textit{Almost all} to \textit{Almost none}.

If H1 correct then:

\begin{equation}
\mu(A_H) < \mu(A_L)
\end{equation}

where $\mu(A_H)$ reflects the mean of responses to question A for those in the high-cost treatment, and $A_L$ reflects the responses to question A for those in the low-cost treatment.


If H2 is correct then:

\begin{equation}
\mu(B_H) < \mu(B_L).
\end{equation}


We analyze results using a permutation test and confirm the difference is significant at the 95% confidence level.

To test this H3 we ask subjects:

\begin{enumerate}
\item [\textbf{C}:] On issues of national security, the media rushes to publication before they are confident that the story is accurate.
\end{enumerate}

We measure responses on a 5-point scale from \textit{Almost always} to \textit{Almost never}. To test it, we use linear regression that includes an interaction term. The unit of analysis is a respondent ($i$). Our model takes the form:

\begin{equation}
A_i = \beta_0 + \beta_1 T_i + \beta_2 C_i + \beta_3 C_i \times T_i + \epsilon
\end{equation}


where $T_i$ is an indicator function equal to 1 if respondent $i$ received the high cost treatment. Other letters represent answers to questions \textbf{A} and \textbf{C}. We expect that the marginal effects plot will show the effect of treatment on responses to \textbf{A} is largest for subjects that score high on question \textbf{C}.



To test this H4 we ask

\begin{enumerate}
\item [\textbf{D}:] How much do you agree or disagree with the following statements: Whistleblowers benefit from the public attention they receive.
\end{enumerate}

We measure responses on a 7-point scale from \emph{Strongly agree} to \emph{Strongly disagree}. To test it, we use linear regression:

\begin{equation}
A_i = \beta_0 + \beta_1 T_i + \beta_2 D_i + \beta_3 D_i \times T_i + \epsilon
\end{equation}

We expect that the marginal effects plot will show the effect of treatment is largest for subjects that score high on question \textbf{D}.


To test H5 we ask subjects:

\begin{enumerate}
\item [\textbf{E}:] How much do you agree or disagree with the following statements: Whistleblowers that come forward do so because they want to do the right thing.
\end{enumerate}

We measure responses on a 7-point scale from \emph{Strongly agree} to \emph{Strongly disagree}. To test it, we use linear regression:

\begin{equation}
B_i = \beta_0 + \beta_1 T_i + \beta_2 E_i + \beta_3 E_i \times T_i + \epsilon
\end{equation}

We expect that the marginal effects plot will show the effect of treatment on responses to \textbf{B} is largest for subjects that score high on question \textbf{E}.

See qualtrics printout

Use qualtrics in built randomization. There is just 1 treatment with two treated values and 1 control. We assign subjects equally to these groups.

individual randomization

No

We are not clustering. However, we expect 3 i.i.d treatment groups.

300 m-turk workers

300 m-turk workers