Week 10 - Design of experiments & causality Flashcards
Explain why selecting an SRS of single readings from n DIFFERENT cars may lead to a more precise estimation than selecting a simple random sample of n readings from a SINGLE car. [6m]
- values of y - where y is the variable of interest - for different units are NEGATIVELY CORRELATED in SRS without replacement scheme and this REDUCES the VARIANCE of an ESTIMATED MEAN compared to an SRS with replacement scheme. [3m]
> w/ replacement, where the y are independent for diff. units - repeatedly sampling the same car would introduce BIAS and a NON-INDEPENDENT SAMPLE. Any single car may NOT be REPRESENTATIVE of the car population as a whole. [3m]
Explain briefly why the stratified sampling scheme would lead to more precise estimation compared to SRS scheme. [3m]
The precision of stratified sampling does not depend on VARIATION BETWEEN STRATA, unlike the precision of simple random sampling.
{ref. to heterogeneity between]
Explain how you could modify the optimal allocation if you had information about the different costs of conducting the survey in households in different strata and precision is to be maximised subject to the cost not being increased beyond that incurred in the design considered in (d) (Neyman allocation). [7m, 2017]
- Assuming a linear cost function C = c0 +summation(ch*nh) the minimum variance design for a given cost allocates nh PROPORTIONAL to NhSh/√ch
- ie. Neyman allocation is adapted by DIVIDING by √ch.
- **So strata with higher costs will have their allocation reduced.
- The CONSTANT of PROPORTIONALITY is found by equating to
the COST INCURRED in the design in (d) (Neyman allocation).
Retrospective cross-sectional study vs Prospective longitudinal study
Retrospective cross-sectional study
- takes a sample now at a single point in time and ask about past things
- for descriptive analyses
Prospective longitudinal study
- selects a sample and follow the same COHORT into the future
- used to study CHANGES over TIME
- can be burdensome on both the researcher or the participant (members of longitudinal studies get tired)
Confounders
Variables that may influence the relationship between X and Y
eg. age
Randomisation (for experiments) + 2 ways to minimise risks of not getting randomisation
Experimental units are RANDOMLY ALLOCATED/{assigned} to treatment groups. [2011]
- confounders are (probabilistically) controlled for & and causal conclusions can be drawn from an experiment (with high probability)
- To make groups EQUIVALENT, on average, w.r.t. OBSERVED & Unobserved values // w.r.t. possible confounders Z which could affect outcome
- Ensures that, on average, there is NO BIAS due to the allocation of treatments
2 ways to minimise risks:
1. To control some of the factors & randomise others, thereby having a compromise between the 2 extremes
2. Large sample size, so that the chance of serious anomalies is negligible
5 types of experiments
eg. to study whether INTERVENTION X can increase OUTCOME VARIABLE Y
- Randomised controlled clinical trials (RCTs)
- Randomised blocks
- Multi-factorial experimental designs
- Cluster randomised trials
- Quasi-experiments
- no randomisation; observational study when expt.s are not possible
*Cannot say there is a CAUSAL EFFECT unless accounted for CONFOUNDERS
2 reasons why experimentation is not always possible
- Not feasible/UNETHICAL
- Sometimes don’t have the power to randomly assign (most of us can’t enact laws easily)
Laboratory experiment vs Field experiment with an example of each [3m, 2016]
Lab experiment in CONTROLLED setting, traditionally in psychology (also behavioural economics),
eg. given in lecture of effect of violent song lyrics on psychological measures.
Field experiment in REAL-LIFE setting with intention to IMPROVE EXTERNAL VALIDITY,
eg. study of effect of ’audit threats’ on measures of ’corruption’ in Indonesia. [1.5 marks each]
In no more than 200 words, explain the problem of confounding when assessing causal effects and how randomisation helps overcome this problem. [4m, 2016, 2018, 2019]
- Explain setting in terms of outcome Y, treatment X and confounders Z.
- Explain why confounding is a problem.
» The danger is that direct causation may be ascribed instead of indirect causation.
[2m] - Explain e.g. idea that randomisation breaks relation between treatment X and possible confounders Z so that correlation between outcome Y and X can be treated as measuring causal effect.
- Randomisation ensures that treatment groups are equivalent ’on average’ w.r.t. possible confounders Z which could affect outcome.
[2m]
Explain how, after an experiment has been conducted, you could check statistically to see whether randomisation has been implemented properly. [2m, 2015, 2018]
- Compare TREATMENT GROUPS w.r.t. variables which were defined prior to treatment,
eg. age and sex - Should find no ’significant differences’ between treatment groups on such variables.
Randomised blocking + why used in an experiment
[2015]
The arranging of experimental units in groups (blocks) that are SIMILAR to one another.
- This way, by applying randomisation within blocks we can increase ACCURACY
Multi-factorial experimental designs + why used in an experiment
[2015]
Experiments where several treatments are tested simultaneously
- to study interaction effects
- and reducing the costs of running the study.
- Placebo
- Placebo effect & 1 example
[2015, 2019]
- A simulated, inert treatment offered to the members of the CONTROL group of an experiment…
- designed to confuse the subject, who will not know if he or she is taking an active treatment or a dummy
- enable researchers to establish whether the active ingredient is in fact effective
- Placebo effect occurs when a participant believes he or she is benefiting from a treatment when any change is simply a PSYCHOLOGICAL EFFECT, rather than a physical effect resulting from the treatment being applied.
eg. patients may sleep better after ingesting a ‘sleeping pill’ which in fact is simply a sugar pill
Double blinding + why used in an experiment
[2015]
An experimental procedure in which neither the SUBJECTS of the expt. nor the persons ADMINISTERING the expt. know the critical aspects of the expt. / know which treatment was administered to which person
- To guard against both EXPERIMENTER BIAS and PLACEBO EFFECTS
