Looking at the OLS plot, we get the impression that we've identified certain text messages that outperform others. The Bayesian plots show us that this perception is incorrect. According to the Bayesian models, we have almost no idea which messages are better than others.

Imagine we randomly select one text message and put the data for that treatment in a locked box. What should our prior belief about the effect of this text message be? Empirical Bayes says, roughly, that our prior belief about the effect of the message we locked in the box should be the average effect of the other 18. We can also use the variability in the effects of the other 18 messages to tell us how confident we should be in our prior, giving us a prior distribution.

We can apply the same logic to the flu study. Imagine we randomly select one text message and put the data for that treatment in a locked box. What should our prior belief about the effect of this text message be? Empirical Bayes says, roughly, that our prior belief about the effect of the message we locked in the box should be the average effect of the other 18. We can also use the variability in the effects of the other 18 messages to tell us how confident we should be in our prior, giving us a prior distribution.

In Bayesian terms, we've constructed a prior belief about the next season's rookies' OBP from data about this season's rookies' OBP.

That's where empirical Bayes comes to the rescue. Empirical Bayes estimates a prior based on the data

The original data aren't yet available, but we can approximately reproduce the data given what we know about the study. We know that 47,306 participants were evenly assigned to one of 19 treatments or a control condition. The outcome was binary (did the patient get a vaccine or not), and we know the vaccination rate in each treatment from the PNAS publication.

You should use Bayesian analysis when comparing 4 or more "things."

declare_sampling(S = complete_rs(N, n = 50)) +

Similarly, we can compare two designs on the basis of their diagnoses:

compare_designs(planned_design, implemented_design)

diagnose_design(designs)

diagnose_design(designs)

Our simulation and diagnosis tools can take a list of designs and simulate all of them at once, creating a column called design to keep track. For example: diagnose_design(designs)

Warning messages:
1: The argument 'estimand = ' is deprecated. Please use 'inquiry = ' instead.

diagnose_design(designs)

A designer is a function that makes designs based on a few design parameters.

redesign(design, N = c(100, 200, 300, 400, 500))

diagnose_design(simulation_df, diagnosands = study_diagnosands)

Building a design from design steps

declare_inquiry(PATE = mean(Y_Z_1 - Y_Z_0)) +

draw_estimands(design)

difference_in_means

estimator <-

measurement <-

assignment <- declare_assignment(Z = complete_ra(N, prob = 0.5))

diagnose_design(design, diagnosands = study_diagnosands)

declare_sampling(S = complete_rs(N, n = 50))

declare_model( Y_Z_0 = U, Y_Z_1 = Y_Z_0 + 0.25)

draw_data(design)

When we run this population function, we will get a different 100-unit dataset each time, as shown here.

As an example of a two-level hierarchical data structure, here is a declaration for 100 households with a random number of individuals within each household.

As a bonus, the data also includes the probability that each unit is assigned to the condition in which it is in (Z_cond),

install.packages("tidyverse")

see also [39])

Although altruism is generally rewarded, several studies suggest that effective altruism is not [6,36,37].

d selected as leaders [33].

Parochialism also biases cost-benefit calculations that could lead to effective giving.

K.F. Law, D. Campbell, B. GaesserBiased benevolence: The perceived morality of effective altruism across social distancePersonal Soc Psychol Bull (2021), Article 014616722110027

approximately two thousand cases of trachoma—a bacterial infection of the eye that can lead to permanent blindness [4].

...: Optional model-function-specific parameters. As with args, these will be quosures and can be varying().

base_recipe %>% step_corr(all_of(stations), threshold = tune())

# remove any columns with a single unique value step_zv(all_predictors()) %>%

many numeric predictors that are highly correlated

Linear Models: the absolute value of the t-statistic for each model parameter is used.

Ways to contribute

Encourage more “here’s what we aren’t doing” posts (e.g. OpenPhil posting what they don’t fund)

Make EA easier for women with children

me sort of interpersonal conflict i

Number of people who listed that reason (without prompting)

I spoke with approximately 20 people who were recommended on the basis of their knowledge about EA retention. These were mostly non-University group organizers. There was moderate agreement about the reasons people leave and stay in EA.The major reasons why people leave EA are: inability to contribute, lack of cultural fit or interpersonal conflict, major life events (moving, having a child), burnout/mental health.

Given this, CEA is evaluating alternative metrics. Our current top choice is to focus on people who use our products, instead of those who are "engaged" with EA in a more subjective sense. This allows us to analyze larger populations, improving the power of our tests.

For example, to detect a change in retention rate from 95% to 90%, we need a sample of 185 individuals.[2]

Number of people

otherwise engage with any of CEA's projects. 

remaining attendees who worked for an EA organization.

population, 50-70% of the individuals who engaged in 2020 also engaged in some way in 2021.

Results

Over the past six months, CEA has moved to unify our login systems. As of this writing, event applications, the EA Forum, and EA Funds/GWWC all use the same login system. This means that we are less likely to have issues with people using different emails.

Peter Wildeford has done the largest non-manual retention analysis I know, which looked at the percentage of people who answered the EA survey using the same email in multiple years. He found retention rates of around 27%, but cautioned that this was inaccurate due to people using different email addresses each year.

50-70% of people who engaged with CEA's projects in 2020 also engaged with one of our projects so far in 2021, using a naïve method of matching people (mostly looking at email addresses).

engagement with CEA's projects as a proxy.

29.8% is much closer to the annual retention estimate produced by Peter Wildeford based on the 2018 EA Survey

 * This is not a projection, but the retention rate observed 2019-2021. 

EAG Reattendees

Projected retentionover lifetime

We identified markers of retention that we could reliably analyze from year to year for the three cohorts. 

Employees at certain organizations connected to effective altruismRecipients of Community Building GrantsAttendees of EA Global

specify() allows you to specify the variable, or relationship between variables, that you’re interested in. hypothesize() allows you to declare the null hypothesis. generate() allows you to generate data reflecting the null hypothesis. calculate() allows you to calculate a distribution of statistics from the generated data to form the null distribution.

If this probability is below some pre-defined significance level 𝛼α\alpha, then we can reject our null hypothesis.

we start by assuming that the observed data came from some world where “nothing is going on” (i.e. the observed effect was simply due to random chance), and call this assumption our null hypothesis.

formalS4classes describe the data model and the conditional test procedures,consisting of multivariate linear statistics, univariate test statistics and a reference distribu-tion.

(−∞,−0.75)(−∞,−0.75)(-\infty,-0.75). T

To create a symbol, simply convert a string or a variable containing a string with the sym() function. Plug in these symbols anywhere dplyr expects a symbol or an expression of only one field name as a function argument using the !! operator.

To capture a quosure, simply capture an expression within the quo() function

# Capture a quosure
gear_filter <- quo(gear == 5)

# Filter our data with our quosure
mtcars %>%
  filter(!!gear_filter)

he R interpreter does not evaluate the expression before it is passed to the mutate function. Since the expression is passed in a raw format to the function, mutate can do whatever it wants with it and evaluate it however it likes

The funding overhang also created bottlenecks for people able to staff projects, and to work in supporting roles.

I’d typically prefer someone in these roles to an additional person donating $400,000–$4 million.

Personally, if given the choice between finding an extra person for one of these roles who’s a good fit or someone donating $X million per year, to think the two options were similarly valuable, X would typically need to be over three, and often over 10.

A big uncertainty here is what fraction of people will ever be interested in EA in the long term – it’s possible its appeal is very narrow, but happens to include an unusually large fraction of wealthy people. In that case, the overhang could persist much longer.

Working at an EA org is only one option, and a better estimate would aim to track the number of people ‘deployed’ in research, policy, earning to give, etc. as well.

Overall, my guess is that we’re only deploying 1–2% of the net present value of the labour of the current membership.

people only hit peak productivity around age 40–60

In terms of the level of ‘talent’ of new members, we don’t have great data.

The average age of community members is several years higher.

so most ways of measuring this growth will undercount it.

You can see some more statistics on what these people are like in the EA Survey.

GWWC members, EA Funds, Founders Pledge members, Longview Philanthropy, SFF, etc. have all grown significantly (i.e. more than doubling) in the last five years.

t could crash the price

I’ve

hen I think the stock is most relevant, since that determines how many resources will be deployed in the long term.

The CIs are not exactly identical, but very close.

the slope becomes r if x and y have identical standard deviations.

fit = lm(I(X2 * 0.5 + 0.4) ~ I(X1 * 0.5 + 0.2), D_correlation)

rbeta(1, alpha, alpha) - 1

(K - 2)/2

rlkj <- function (K, eta = 1) {

Account for a positive correlation between the growth premium and population growth, as people are more likely to move to a fast growing city

inary search is particularly useful for this. To do a binary search, you repeatedly remove half of the code until you find the bug. This is fast because, with each step, you reduce the amount of code to look through by half.

Welcome

Citr package (addin) for RStudio

im %>% mutate(sims = invoke_map(f, params, n = 10)) #> # A tibble: 3 x 3

One restriction of summarise() is that it only works with summary functions that return a single value. That means that you can’t use it with functions like quantile() that return a vector of arbitrary length:

workflow for managing many models,

broom::glance)

by_country %>% mutate(glance = map(model, broom::glance)) %>% unnest(glance)

resids

resids = map2(data, model, add_residuals)

do(map_dfr(.$fit, tidy, .id = "dataset"))

mutate(fit = flatten(pmap(.l = list(.f = funcs, .formulas = models, data = dat), .f = modelr::fit_with)))

You can also use an integer to select elements by position: x <- list(list(1, 2, 3), list(4, 5, 6), list(7, 8, 9)) x %>% map_dbl(2) #> [1] 2 5 8

map_*() uses … ([dot dot dot]) to pass along additional arguments to .f each time it’s called:

The immediate consequence of the exogeneity assumption is that the errors have mean zero: E[ε] = 0, and that the regressors are uncorrelated with the errors: E[XTε] = 0.

OLS estimation can be viewed as a projection onto the linear space spanned by the regressors. (Here each of X 1 {\displaystyle X_{1}} and X 2 {\displaystyle X_{2}} refers to a column of the data matrix.)

applicable

This is illustrated at the right.

In other words, the gradient equations at the minimum can be written as: ( y − X β ^ ) T X = 0. {\displaystyle (\mathbf {y} -X{\hat {\boldsymbol {\beta }}})^{\rm {T}}X=0.}

when y is projected orthogonally onto the linear subspace spanned by the columns of X.

. Thus, the residual vector y − Xβ

ing from the moment conditions E [ x i ( y i − x i T β ) ] = 0. {\displaystyle \mathrm {E} {\big [}\,x_{i}(y_{i}-x_{i}^{T}\beta )\,{\big ]}=0.}

In particular, this assumption implies that for any vector-function ƒ, the moment condition E[ƒ(xi)·εi] = 0 will hold.

The OLS estimator β ^ {\displaystyle {\hat {\beta }}} in this case can be interpreted as the coefficients of vector decomposition of ^y = Py along the basis of X.

In statistics and signal processing, a minimum mean square error (MMSE) estimator is an estimation method which minimizes the mean square error (MSE), which is a common measure of estimator quality, of the fitted values of a dependent variable. In the Bayesian setting, the term MMSE more specifically refers to estimation with quadratic loss function. In such case, the MMSE estimator is given by the posterior mean of the parameter to be estimated. Since the posterior mean is cumbersome to calculate, the form of the MMSE estimator is usually constrained to be within a certain class of functions. Linear MMSE estimators are a popular choice since they are easy to use, easy to calculate, and very versatile. It has given rise to many popular estimators such as the Wiener–Kolmogorov filter and Kalman filter.

save persons’ factor scores, and

Principal axis factoring is not to be confused with principal components analysis (PCA), which strictly speaking is not a type of common factor analysis because it generates components rather than factors. Unlike factors, components include the unique variances of the observed variables. This is similar to how we calculate composite scores in classical test theory (though PCA finds the best possible solution to reducing variables, with their error, into composites). PCA is appropriate when we are not interested in assuming that there is an underlying latent variable such as a construct corresponding with the components. When we assume that the factors do represent latent traits, common factor analysis is more appropriate than PCA. Because we are interested in measuring latent traits or constructs, we do not use PCA in this chapter.

library(psych) fa(r = , nfactors = , n.obs = , fm = , rotate = ) The first argument, which can replace the r =, is for the data, whether it be a data frame of the raw data or a correlation or covariance matrix.

but in order to account for correlations, the current best-practice approach is to follow Katz, Kling and Liebman (2007) in calculating bootstrapped estimates of adjusted p-values using a modification of the free step-down algorithm of Westfall and Young (1993).

Do a bit more work. Re-check that your project is still in a functional state. Commit again but this time amend your previous commit. RStudio offers a check box for “Amend previous commit” or in the shell: git commit --amend --no-edit The --no-edit part retains the current commit message of “WIP”. Don’t push! Your history now looks like this: A -- B -- C -- WIP* but the changes associated with the WIP* commit now represent your last two commits, i.e. all the accumulated changes since state C. Keep going like this. Let’s say you’ve finally achieved

git checkout issue-5 git reset HEAD^

git commit --all -m "WIP" git checkout master Then when you come back to the branch and continue your work, you need to undo the temporary commit by resetting your state. Specifically, we want a mixed reset. This is “working directory safe”, i.e. it does not affect the state of any files. But it does peel off the temporary WIP commit. Below, the reference HEAD^ says to roll the commit state back to the parent of the current commit (HEAD). git checkout issue-5 git reset HEAD^

Check the remote was cloned successfully:

Unless you use the GitHub API, most of the GitHub bits really have to be done from the browser.

. Bilder and Loughin (2007) introduceda flexible loglinear modeling approach that allows researchers to consider alternative associationstructures somewhere between SPMI and complete dependence. Within this framework, a modelunder SPMI is given aslog(μab(ij))=γij+ηWa(ij)+ηYb(ij)

asymptotic distribution is a linear combination ofindependentχ21random variables (Bilder and Loughin, 2004)

a test for simultaneous pairwise marginalindependence (SPMI), involves determining whether eachW1, . . . ,WIis pairwise independent of eachY1, . . . ,YJ. OurMI.test()function calculates their modified Pearson statistic

Examining all possible combinations of the positive/negative item responses between MRCVsis the preferred way to display and subsequently analyze MRCV data.

3 How EAs get Involved in EA

To deal with this, we organised all of the factors into six overarching categories, comprising three barriers and three facilitators: 1. Difficulties in accessing evidence (six studies) 2. Challenges in understanding the evidence (three studies) 3. Insufficient resources (six studies) 4. Knowledge sharing and ease of access (six studies) 5. Professional advisors and networks (three studies) 6. A broader definition of what counts as credible evidence and better standardisation of reporting (three studies).

hey run conjoint analysis: in which customers are offered goods with various combinations of characteristics and price – maybe a pink car with a stereo for £1,000, a pink car without a stereo for £800, a blue car for £1,100 and a blue car without a stereo for £950 – to identify how much customers value each characteristic.

et me tell you a story. Once upon a time, researcher Dean Karlan was investigating microloans to poor people in South Africa, and what encourages people to take them. He sent people flyers with various designs and offering loans at various rates and sizes. It turns out that giving consumers only one choice of loan size, rather than four, increased their take-up of loans as much as if the lender had reduced the interest rate by about 20 percent. And if the flyer features a picture of a woman, people will pay more for their loan – demand was as high as if the lender had reduced the interest rate by about a third. Nobody would say in a survey or interview that they would pay more if a flyer has a lady on it. But they do. Similarly, Nobel Laureate Daniel Kahneman reports that, empirically, people are more likely to be believe a statement if it is written in red than in green. But nobody would say that in a survey, not least because we don’t know it about ourselves.

do(cluster_summary = summary(.))

gives us the best segmentation possible.

Just like K-means and hierarchical algorithms go hand-in-hand with Euclidean distance, the Partitioning Around Medoids (PAM) algorithm goes along with the Gower distance.

The silhouette width is one of the very popular choices when it comes to selecting the optimal number of clusters. It measures the similarity of each point to its cluster, and compares that to the similarity of the point with the closest neighboring cluster. This metric ranges between -1 to 1, where a higher value implies better similarity of the points to their clusters.

library(cluster)gower_df <- daisy(german_credit_clean, metric = "gower" , type = list(logratio = 2))

  mutate_if(is.character, as.factor)

We find that the variable amount needs a log transformation due to the positive skew in its distribution.

e details about the mathematics of Gower distance are quite complicated and left out for another article.

Clustering datasets having both numerical and categorical variables

For each observation iii, calculate the average dissimilarity aiaia_i between iii and all other points of the cluster to which i belongs. For all other clusters CCC, to which i does not belong, calculate the average dissimilarity d(i,C)d(i,C)d(i, C) of iii to all observations of C. The smallest of these d(i,C)d(i,C)d(i,C) is defined as bi=minCd(i,C)bi=minCd(i,C)b_i= \min_C d(i,C). The value of bibib_i can be seen as the dissimilarity between iii and its “neighbor” cluster, i.e., the nearest one to which it does not belong. Finally the silhouette width of the observation iii is defined by the formula: Si=(bi−ai)/max(ai,bi)Si=(bi−ai)/max(ai,bi)S_i = (b_i - a_i)/max(a_i, b_i).

Average silhouette method

The total WSS measures the compactness of the clustering and we want it to be as small as possible.

To avoid distortions caused by excessive outliers, it’s possible to use PAM algorithm, which is less sensitive to outliers.

Next, the wss (within sum of square) is drawn according to the number of clusters. The location of a bend (knee) in the plot is generally considered as an indicator of the appropriate number of clusters.

K-means algorithm can be summarized as follow: Specify the number of clusters (K) to be created (by the analyst) Select randomly k objects from the dataset as the initial cluster centers or means Assigns each observation to their closest centroid, based on the Euclidean distance between the object and the centroid For each of the k clusters update the cluster centroid by calculating the new mean values of all the data points in the cluster. The centoid of a Kth cluster is a vector of length p containing the means of all variables for the observations in the kth cluster; p is the number of variables. Iteratively minimize the total within sum of square. That is, iterate steps 3 and 4 until the cluster assignments stop changing or the maximum number of iterations is reached. By default, the R software uses 10 as the default value for the maximum number of iterations.

to use correlation distance, the data are input as z-scores.

A successful evaluation of discriminant validity shows that a test of a concept is not highly correlated with other tests designed to measure theoretically different concepts.

The remaining term, 1 / (1 − Rj2) is the VIF. It reflects all other factors that influence the uncertainty in the coefficient estimates. The VIF equals 1 when the vector Xj is orthogonal to each column of the design matrix for the regression of Xj on the other covariates. By contrast, the VIF is greater than 1 when the vector Xj is not orthogonal to all columns of the design matrix for the regression of Xj on the other covariates. Finally, note that the VIF is invariant to the scaling of the variables

It turns out that the square of this standard error, the estimated variance of the estimate of βj, can be equivalently expressed as:[3][4] var ^ ( β ^ j ) = s 2 ( n − 1 ) var ^ ( X j ) ⋅ 1 1 − R j 2 , {\displaystyle {\widehat {\operatorname {var} }}({\hat {\beta }}_{j})={\frac {s^{2}}{(n-1){\widehat {\operatorname {var} }}(X_{j})}}\cdot {\frac {1}{1-R_{j}^{2}}},} where Rj2 is the multiple R2 for the regression of Xj on the other covariates (a regression that does not involve the response variable Y). This identity separates the influences of several distinct factors on the variance of the coefficient estimate: s2: greater scatter in the data around the regression surface leads to proportionately more variance in the coefficient estimates n: greater sample size results in proportionately less variance in the coefficient estimates var ^ ( X j ) {\displaystyle {\widehat {\operatorname {var} }}(X_{j})} : greater variability in a particular covariate leads to proportionately less variance in the corresponding coefficient estimate The remaining term, 1 / (1 − Rj2) is the VIF. It reflects all other factors that influence the uncertainty in the coefficient estimates

Summary: Algorithms to Live By

When treatment assign-ment takes place in waves, it is natural to adapt Thompson sampling by assigning a non-random numberpdtNtof observations in wavetto treatmentd, in order to reduce ran-domness. The remainder of observations are assigned randomly so that expected sharesremain equal topdt.

Q = 12 n k ( k + 1 ) ∑ j = 1 k ( r ¯ ⋅ j − k + 1 2 ) 2 {\displaystyle Q={\frac {12n}{k(k+1)}}\sum _{j=1}^{k}\left({\bar {r}}_{\cdot j}-{\frac {k+1}{2}}\right)^{2}} . Note

is the rank of x i j {\displaystyle x_{ij}}

Find the values r ¯ ⋅ j = 1 n ∑ i = 1 n r i j {\displaystyle {\bar {r}}_{\cdot j}={\frac {1}{n}}\sum _{i=1}^{n}{r_{ij}}}

Definitions

Individual donors, governments and firms demonstrate substantial generosity (e.g., UK charity represents 0.5-1% of GDP, US charity around 2% of GDP).

how people react to the presentation of charity-effectiveness information.

Beem101: Project, discussion of research

pure’ altruism or ‘warm glow’ altruism (Andreoni 1990; Ashraf and Bandiera 2017)

The Global Priorities Institute’s vision and mission

Formula[edit] The Y-intercept of the SML is equal to the risk-free interest rate. The slope of the SML is equal to the market risk premium and reflects the risk return tradeoff at a given time: S M L : E ( R i ) = R f + β i [ E ( R M ) − R f ] {\displaystyle \mathrm {SML} :E(R_{i})=R_{f}+\beta _{i}[E(R_{M})-R_{f}]\,} where: E(Ri) is an expected return on security E(RM) is an expected return on market portfolio M β is a nondiversifiable or systematic risk RM is a market rate of return Rf is a risk-free rate

The Y-intercept of the SML is equal to the risk-free interest rate. The slope of the SML is equal to the market risk premium and reflects the risk return tradeoff at a given time: S M L : E ( R i ) = R f + β i [ E ( R M ) − R f ] {\displaystyle \mathrm {SML} :E(R_{i})=R_{f}+\beta _{i}[E(R_{M})-R_{f}]\,} where: E(Ri) is an expected return on security E(RM) is an expected return on market portfolio M β is a nondiversifiable or systematic risk RM is a market rate of return Rf is a risk-free rate

This abnormal extra return above the market's return at a given level of risk is what is called the alpha.

Capital asset pricing model

quantity beta (β)

systematic risk (beta) t

If the fraction q {\displaystyle q} of a one-unit (e.g. one-million-dollar) portfolio is placed in asset X and the fraction 1 − q {\displaystyle 1-q} is placed in Y, the stochastic portfolio return is q x + ( 1 − q ) y {\displaystyle qx+(1-q)y} . If x {\displaystyle x} and y {\displaystyle y} are uncorrelated, the variance of portfolio return is var ( q x + ( 1 − q ) y ) = q 2 σ x 2 + ( 1 − q ) 2 σ y 2 {\displaystyle {\text{var}}(qx+(1-q)y)=q^{2}\sigma _{x}^{2}+(1-q)^{2}\sigma _{y}^{2}} . The variance-minimizing value of q {\displaystyle q} is q = σ y 2 / [ σ x 2 + σ y 2 ] {\displaystyle q=\sigma _{y}^{2}/[\sigma _{x}^{2}+\sigma _{y}^{2}]} , which is strictly between 0 {\displaystyle 0} and 1 {\displaystyle 1} . Using this value of q {\displaystyle q} in the expression for the variance of portfolio return gives the latter as σ x 2 σ y 2 / [ σ x 2 + σ y 2 ] {\displaystyle \sigma _{x}^{2}\sigma _{y}^{2}/[\sigma _{x}^{2}+\sigma _{y}^{2}]} , which is less than what it would be at either of the undiversified values q = 1 {\displaystyle q=1} and q = 0 {\displaystyle q=0} (which respectively give portfolio return variance of σ x 2 {\displaystyle \sigma _{x}^{2}} and σ y 2 {\displaystyle \sigma _{y}^{2}} ). Note that the favorable effect of diversification on portfolio variance would be enhanced if x {\displaystyle x} and y {\displaystyle y} were negatively correlated but diminished (though not eliminated) if they were positively correlated.

Similarly, a 1985 book reported that most value from diversification comes from the first 15 or 20 different stocks in a portfolio.[6]

d(p)=(209000-130p)

CLV Formula

“Sue’s mother” RaRaR_a “Sue’s lecturer in the UK” →→\rightarrow false (so it’s not ‘transitive’)

intend

(Highly optional): Properties of binary relations - O-R problem 1a.

Students

We say that uu is ’a utility function for ≿\succsim.

Differentiating this wrt III yields Engel aggregation:

direct

Past Projects

That's because cause prioritization research is extremely difficult, not because no one has thought to do this.

4. It is difficult to find cause neutral funding.I think funders like to choose their cause and stick with it so there is a lack of cause neutral funding. 

Growth and the case against randomista development,

me that when reading the GPI research agenda, the economics parts read like it was written by philosophers.

(Also, I have never worked in academia so there may be theories of change in the academic space that others could identify.)

But for a new organisation to solely focus on doing the research that they believed would be most useful for improving the world it is unclear what the theory of change would be.

I think that people are hesitant to do something new if they think it is being done, and funders want to know why the new thing is different so the abundance of organisations that used to do cause prioritisation research or do research that is a subcategory of cause prioritisation research limits other organisations from starting up.

Theoretical cause selection beyond speculation. Evidence of how to reason well despite uncertainty and more comparisons of different causes.

more consideration of second order effect

Let me give just one example, if you look at best practice in risk assessment methodologies[5] it looks very different from the naive expected value calculations used in EA

theorists

e. From my point of view, I could save a human life for ~£3000. I don’t want to let kids die needlessly if I can stop it. I personally think that the future is really important but before I drop the ball on all the things I know will have an impact it would be nice to have:

(There could be experimental hits based giving.)

Now let’s get a bit more complicated and do some more research and find other interventions and consider long run effects and so on”. There could be research looking for strong empirical evidence into:the second order or long run effects of existing interventions.how to drive economic growth, policy change, structural changes, and so forth.