The scientific method

Figure 1: Scientific method

A few self-evident objectives of empirical scientific inquiry

• Description, answering the question “what happens/happened?”
• explanation, answering the question “why does x happen?”
• prediction, answering the question “what will happen with x if …?”
• control, answering the question “how can x be influenced?”

But why use statistics for this?

• To describe, explain, and predict
  • objectively
  • precisely
  • comparably
  • concisely
• to cope with variability and to generalize: different samples even from the same population will yield different results
• thus, we need to be able to
  • quantify this variability
  • separate random from systematic/meaningful variability
• to assess the robustness of one’s generalizations

Three absolutely central notions

• Objectivity: independence of personal opinions
• reliability: precision (in the sense of ‘re-test reliability’)
• validity: one measures what one wants to measure; in a sense, this is probably the most important one

Two English verbs verb₁ and verb₂

• A published study discussed the complementation preferences verb₁ and verb₂ with regard to two grammatical patterns on the basis of the following data:

addmargins(example_1 <-matrix(c(295, 131, 104, 35), ncol=2,
   dimnames=list(VERB=1:2, PATTERN=1:2)))

     PATTERN
VERB    1   2 Sum
  1   295 104 399
  2   131  35 166
  Sum 426 139 565

    PATTERN
VERB    1    2
   1 0.74 0.26
   2 0.79 0.21

    Pearson's Chi-squared test

data:  example_1
X-squared = 1.5679, df = 1, p-value = 0.2105

Two English verbs verb₁ and verb₂

• Another study on two English verbs verb₁ and verb₂ discussed their complementation preferences with regard to 5 kinds of XPs on the basis of the following data:

addmargins(example_2 <- matrix(c(302,73, 8,0, 145,5, 19,3, 8,0), ncol=5,
   dimnames=list(VERB=1:2, PATTERN=c("NP", "PP", "VP", "AdjP", "AdvP"))))

     PATTERN
VERB   NP PP  VP AdjP AdvP Sum
  1   302  8 145   19    8 482
  2    73  0   5    3    0  81
  Sum 375  8 150   22    8 563

    PATTERN
VERB    NP    PP    VP  AdjP  AdvP
   1 -1.06  0.44  1.46  0.04  0.44
   2  2.59 -1.07 -3.57 -0.09 -1.07

Avoiding complete surprises 1

Figure 2: A correlation between two variables XX and YY

Avoiding complete surprises 2

Caveats: note, however

• Statistics don’t provide content –- the researcher does
• statistics are only useful to the extent that the researcher has been successful
  • in operationalizing his variables appropriately
  • eliciting/collecting the data correctly
  • choosing the right statistical technique
  • applying that right statistical technique the right way

The phases of an empirical study

• reconnaissance, leading to variables
• hypotheses (text and statistical forms)
• data collection ((operationalizations of) variables)
• evaluation of hypotheses given the data with
  • effect sizes
  • graphs
  • significance tests (p-values)

Phase 1 and 2: the notion of variables

• Variables
  • are measurable properties or characteristics of an item
  • vary across different items, where “items” are the individual measurements of the ‘thing of interest’ (see below); items can be people, events (words, utterances, …)
• non-linguistic examples:
  • annual income, number of children, IQ, …
  • party someone voted for in the last election, hair color, marital status, …
• linguistic examples:
  • reaction time to a word, length of a word, …
  • animacy of a subject noun phrase: human (Peter) vs. animate (the cat) vs. inanimate concrete object (the table) vs. abstract (time), …
• note: we can/need to decide on a resolution. For annual income:
  • numbers: the exact amount? the amount rounded to full US$?
  • ranked classes: ‘negative’, 0-30,000, 30,001-60,000, 60,001-100,000, 100,001-?
  • categories: none vs. any? Or above average vs. more than/above average?

Phase 1 and 2: variable types, part 1

• Variables can be distinguished in terms of their information value
  • categorical: different values → different properties
  • ordinal: categorical + different values → different ranks
  • numeric: categorical + ordinal + different values → sizes of differences
• here are results of a fictitious results of an Olympic 100m dash – what is the information level of each variable in a column?

Phase 1 and 2: variable types, part 2

• Variables can be distinguished in terms of their role in an investigation
  • response/dependent: the variable whose values/behavior/variation we want to explain
  • predictor/independent: often the assumed cause for the behavior of the response

• confounds (controlled, accounted for, or residualized out)
• moderators (accounted for by interactions w/ add. variables)
• colliders (accounted for differently)

Phase 1 and 2: variable types, practice

• In the following non-linguistic examples of text hypotheses, what is the response, what is the predictor, and what are the variables’ information values?
  • people with a univ degree are more intelligent than people without one
  • response: IQ (num) ~ predictor: HASUNIVDEGREE (cat): no vs. yes
  • men are better at parking than women
  • response: PARKINGSKILL (?) ~ predictor: SEX/GENDER (cat): female vs. male

Phase 1 and 2: variable types, practice

• In the following linguistic examples of text hypotheses, what is the response, what is the predictor, and what are the variables’ information values?
  • in essays, non-native speakers make more mistakes than native speakers
  • response: NUMBERofMISTAKES (num)~ predictor: SPEAKERTYPE (cat): nns vs. ns
  • subjects are shorter than objects
  • response: LENGTH (num) ~ predictor: GRAMREL (cat): object vs. subject

Phase 2: What are hypotheses?

• What are hypotheses? One definition:
  • universal statements (going beyond a singular event)
  • implicit structure of a conditional sentence
    • if [predictor] …, then [response] …
    • the more/less [predictor] …, the more/less [response] …
  • potentially falsifiable
  • empirically testable
• maybe most useful definition: a statement postulating a distribution of one or more response variables
• hypotheses come in different kinds

Phase 2: Kinds of hypotheses

• text hypotheses vs. statistical hypotheses (→ operationalization)
• alternative hypothesis H₁: a statement postulating
  • a particular distribution of a (response) variable (goodness-of-fit)
  • a relation between 1+ predictors & 1+ response variables (independence/difference(s))
    • stipulating some difference, but not its direction: non-directional/2-tailed
    • e.g., subjects and objects differ in their lengths
    • stipulating a difference and its direction: directional/1-tailed
    • e.g., subjects are shorter than objects
• null hypothesis H₀:
  • the logical counterpart to H₁: an alternative hypothesis with not in it

Phase 2: Operationalization 1

• Operationalization: the step from text hypotheses to statistical hypotheses
  • step 1: phrasing the variables in the text hypotheses such that they involve numbers
  • step 2: choose a statistical measure to be applied to those numbers
• non-linguistic examples
  • parking performance
  • physical fitness
  • financial wealth
• linguistic examples
  • the knowledge of a foreign language
  • the lengths of subjects and objects

Phase 2: Operationalization 2

• Operationalization: the step from text hypotheses to statistical hypotheses
  • step 1: phrasing the variables in the text hypotheses such that they involve numbers
  • step 2: choose a statistical measure to be applied to those numbers
• most frequent statistical measures:
  • counts/frequencies
  • averages/means
  • correlations
  • distributions and dispersions
• what statistic to use for the lengths of subjects & objects?
  • summed total of the above (i.e. counts/frequencies)?
  • means of the above (i.e. averages/means)?

Phase 2: an example

• Imagine you have the following alternative text hypothesis: “Across all clauses, subjects are shorter than objects in English”
  • what’s the corresponding null hypothesis?
  • “Across all clauses, subjects are not shorter than objects in English”
• what variables are involved?
  • response: LENGTH (numeric) ~ predictor: GRAMREL (binary/categorical)
• how to operationalize them?
  • LENGTH: let’s use length in words
  • GRAMREL:
    • object: the NP that is the ‘target’ of the action of a transitive verb and could become the subject of the sentence upon passivization
    • subject: the NP determining verbal morphology/agreement and that, prototypically, denotes the agent of the action denoted by the verb
• what statistic to use?
  • average length of all objects vs. average length of all subjects (not pairwise!)

Phase 3: Data storage rules

• Imagine you studied this using corpus data and collected the following data:
  • The younger bachelors ate the nice little cat
  • He was sleeping (note: no object!)
  • The quick brown fox hit the lazy dog
• rule: store the data in the so-called case-by-variable format:
  • each data point (i.e., measurement of the response variable) has a row on its own
  • every variable or every other characteristic of a data point has a column on its own
  • the very first row contains the names of all variables (header)
  • missing data are marked as NA –- do not use empty cells!
  • do not use numbers for categorical variables

Phase 3: Data storage (not great)

Phase 3: Data storage (better)

• Something like this would be the better format:

Phase 3: Data storage (excursus)

• If one did this study not across all clauses (i.e. including clauses that have no object), …
• … but only used clauses that have one subject and one object (i.e. each subject/object is associated with only one object/subject in a perfectly pairwise fashion), …
• then one could also use this format:

Phase 3: Data storage: direct comparison

The scientific method

• The logic of statistical testing is that of hypothesis falsification:
  • one does not prove that one’s own H₁ is correct
  • one ‘proves’ that the H₀ is wrong, which means one’s H₁ is right
• steps:
  • one defines a significance level p_critical, which quantifies how quickly one will reject H₀ / accept H₁
  • one computes the effect e observed in one’s data (using the statistic from the statistical hypothesis)
  • one computes the probability of error p how likely it is to find e if H₀ is correct
  • decision
    • if p<p_critical, one rejects H₀ and accepts H₁
    • if p≥p_critical, one must stick to H₀ and cannot accept H₁

Coin tossing, scenario 1

• You and I play a game, tossing a coin 100 times: heads: $1 for me; tails: $1 for you
• your hypotheses:
  • H₀: both players are honest: p_heads=p_tails=0.5
  • H₁: STG is not honest: p_heads>0.5 and p_tails<0.5
• the significance level is set to 0.05
• now, how often do you have to lose before you begin to accuse me of cheating a.k.a. accepting H₁?
  • when you lose 51 times?
  • when you lose 55 times?
  • when you lose 59 times?
• what are you doing? You’re looking at an effect e (the result STG: 59 vs. you: 41, i.e. your losses) and are determining when e becomes too unlikely to still believe in H₀

Tossing a coin just 3 times

• You set the significance level: p_critical=0.05
• we play, you lose 3 times out of 3: the effect e is 3:0

Tossing a coin more often

Coin tossing, scenario 2

• You and I play a game, tossing a coin 100 times: heads: $1 for me; tails: $1 for you
• an independent observer’s hypotheses:
  • H₀: both players are honest: p_heads=p_tails=0.5
  • H₁: at least one player is not honest: p_heads>0.5 or p_heads<0.5
• the significance level is set to 0.05
• now, how often does one of us have to lose before the independent observer begins to accuse the other of cheating a.k.a. accepting H₁?
  • when someone loses 51 times?
  • when someone loses 56 times?
  • when someone loses 61 times?
• what is the independent observer doing? Looking at an effect e (the results someone: 61 vs. someone else: 39, i.e. someone’s losses) and determining when e becomes too unlikely to still believe in H₀

Tossing a coin just 3 times

• An independent observer sets the significance level: p_critical=0.05
• we play, one of us (you) loses 3 times out of 3: the effect e is 3:0

Tossing a coin more often

Lessons to be learned, part 1

• Lesson 1 is about distributions and parametric testing:
• in this case of binomial trials, with increasing sample sizes,
  • we obtain a bell-shaped normal distribution
  • even if the ‘input probability’ is not normal
• thus, if the sample sizes are large enough and the distribution looks like one we can describe easily, then …
• we can use a parametric/asymptotic test – but only then!

Lessons to be learned, part 2

• Lesson 2 is about alternative hypotheses: there are
  • directional/one-tailed alternative hypotheses:
    • postulate an effect, a difference, a correlation,
    • and its direction (above, you)
  • non-directional/two-tailed alternative hypotheses:
    • postulate an effect, a difference, a correlation,
    • but not its direction (above: the independent observer)
• prior knowledge is rewarded: the former was/are easier to accept
• but where are those p-values coming from?

Choosing a method/test, part 1

• What kind of study is being conducted
  • descriptive, exploratory, hypothesis-generating
  • hypothesis-testing
• how many and what kinds of variables are involved?
  • 1 response (goodness-of-fit tests)
  • 1 response & 1 predictor (monofactorial test for independence or differences)
  • 1 response & 2+ predictors (multifactorial analyses)
  • 2 responses (multivariate analyses)
• are data points related such that you can associate them with each other in a meaningful principled way?
  • no: tests for independent samples
  • yes: tests for dependent samples
  • the latter are usually more powerful

Choosing a method/test, part 2

• What is the statistic of the dependent variable in the statistical hypothesis?
  • counts/frequencies, → often chi-squared tests
  • distributions → often Kolmogorov-Smirnov test
  • averages/means, → often t-tests
  • dispersions → often F-tests
  • correlations, → often r or ρ or τ
• What does the distribution of the data look like?
  • normal: often leads to parametric tests
  • non-normal: often leads to non-parametric, simulation, or exact tests
• how big are the samples to be collected?
  • <30: often a risk to normality assumptions
  • ≥30: often supporting normality assumptions

Significance testing (again)

• Your results section should usually include
  • the observed effect e
  • some significance results from some test(s)
  • how both these aspects of your results relate to your hypotheses
• but again: the p-value indicates how likely the observed result is given the H₀ – nothing else

Recall that the standard p-value required in the humanities and social sciences
is 0.05. [...] What does this statistical significance mean? It means that
there is at least a 95% chance that the null hypothesis is *incorrect*.

Effect sizes

• As mentioned above, your results should also include effect sizes
• effect sizes are correlated with p-values, but not deterministically so: often
  • strong effects will be significant, and
  • weak effects will be insignificant
• but,
  • given large sample sizes, even very weak effects can be significant
  • given large variability, even strong effects can be insignificant

   p-value odds ratio
    0.2320     1.7778 

   p-value odds ratio
    0.0002     1.7778 

For 20 nouns, you measured …

• a predictor IMAGEABILITY: whether one can imagine/visualize the referent of the noun (n: ‘no’ vs. y: ‘yes’)
• a response RT: a reaction time score ranging from 1 (fastest) to 20 (slowest)
• check out this nearly perfect correlation:

How do we determine whether that effect e is significant?

• The observed effect e (n-y) is 14 minus 7 = 7 but H₀ hypothesizes an effect of 0
• how about we generate relevant H₀ data and check how the observed effect e compares to those H₀ data?
• relevant H₀ data
  • have the same IMAGEABILITY frequencies of n and y (10 each), &
  • have the same values of RT, but
  • are somehow random and, thus, compatible with H₀ – how?
• simple: we destroy the association of RT ~ IMAGEABILITY (n & slow / y & fast) by randomly reordering the values of the predictor IMAGEABILITY!

RT ~ IMAGEABILITY (randomized 1)

set.seed(1); d_rand <- data.frame(RT=d$RT, IMAGEABILITY=sample(d$IMAGEABILITY))

RT ~ IMAGEABILITY (randomized 2)

d_rand <- d_rand <- data.frame(RT=d$RT, IMAGEABILITY=sample(d$IMAGEABILITY))

RT ~ IMAGEABILITY (randomized 3)

d_rand <- d_rand <- data.frame(RT=d$RT, IMAGEABILITY=sample(d$IMAGEABILITY))

We need this much more often …

• Let’s generate not 3, but 100,000 random H₀ distribution, i.e. 100,000 effects e, …

But how do we evaluate this?

• We can represent all the H₀ effects e_1-100,000, e.g. in a histogram
• we can add a vertical line to the histogram that represents the actual observed effect e of 7
• we can count how often we get a value of 7 or higher in the H₀ data and …
• … express that as a percentage – that is p
• here, p is 0.0034 – the observed difference of 7 between objects and subjects is significant(ly different from 0)

How well does this work?

• Reminder: the p-value from the simulation is 0.0034
• the ’gold standard p-value from an exact (!) t-test for independent samples is 0.00342, …
• … which means that the simulation approach scores a nearly perfect result
• what about the parametric t-test (according to Welch)? Its p-value is 0.00232, which is also pretty close (but worse than the simulation!)
• what about the parametric t-test (according to Student)? Its p-value is 0.00227, which is also close (but worse than the simulation!)
• simulation-based approaches are very versatile and useful – they can often help when few other things can!