Introduction Flashcards
Definition of falsifiable
Can be proved to be false
Parsimonius
simple model to explain
Definition of theory
Set of principles that explain a topic on which a new hypothesis can be made
Descriptive statistics
Summarise a collection without inferences made
Inferential stats
Draws inferences about a population from estimation or hypothesis testing
Quantitative
Measured on interval/ratio scale or ordinal data
Qualitative
Assign objects into labelled groups without natural ordering
Interval variables
Equal intervals e.g. age
Ratio variables
Equal intervals with a clear 0
Binary
2 categories
Nominal
More than 2 categories
Average used for nominal data
Mode
Ordinal data
More than 2 categories with an order (e.g. 1st, 2:1, 2:2
Average used for ordinal data
Median
why does a variable type matter?
Alters tests that can be used
Measurement error
Discrepancy between actual number and one recorded
Systematic variance
Due to dependent variable
Random error
Random variance
What is validity
Measure of how well it measures what it’s supposed to measure
Problem with hypothesis testing
encourages all or nothing thinking, just because null hypothesis is rejected doesn’t mean it’s true
One-tailed when to reject H0
Reject null hypothesis if in extreme 5%
Two tailed when to reject H0
Reject is in either 2.5%
Type 1 error
Rejection of true null hypothesis, incorrectly preduct that variance is accounted for by the model, accepted p
Type 2 error
Fails to reject null hypothesis, incorrectly predict that too much variance is unaccounted for by the model, acceptable p=.2 at beta level
Benefit of one tailed (error)
Lower chance of type 2 error and more power BUT only in one direction have to be sure of result
Meaning of effect size
degree to whihc the mean of H1 differs from mean od H0 in terms of SD OR how much variance is explained
Ways to reduce type 1 error
Look at effect size (standardises results and not reliant on sample size)
Calculate cohen’s d
Mean1 - Mean2 / pooledSD
Calculate pearson’s r
Cov(xy) / SxSy
Pooled SD
Squarert ((SD12 + SD22)/2)
What does more power mean in terms of error
Reduces type 2 error, better chance of correctly rejecting null hypothesis with bigger sample
Ideal power
.8
Big r
.5
Big d
.8
Linearity
Variables are linearly related
Addivity
Several predictors combined effects best described with addition
Normality
Sampling distribution should be normal check with Kolmogorous-Smirnov (non-sign = normal) or Kurtosis (0=normal)
Homogenity of variance
Samples should have similar variances and outcome stable across predictor check (also referred to as heteroscedasticity) Levene’s test (non sig = variance assumed) or variance ratio (2 = assumed)
Independence
Errors should not be related
Checking for heteroscedasticity
Check scatterplots, want random pattern of standardised values
How to reduce bias
trimming data (or standardised), windowinising (replace outliers with highest value), bootstrapping or transforming (e.g. logs)
