1/49
Statistics Terms for MSc Conservation Courses
Numerical Data
Continuous (Measured)
Discrete (Counted in intervals) that can be analyzed statistically and represent quantities. Examples include height, weight, and age.
Categorical Variables
Nominal (Independent Categories)
Ordinal (Ranked in reference to another category)
Binary (Two options (Y/N, Present/Not Present)
Time-to-Event Variables
Latency or duration to an events (Time to fledge, time to hatch)
What does Descriptive Statistics tell us?
1) Informs us of central tendencies
2) Quanitifies variability
3) Visualize patterns/distributions
4) Detects Outliers
What are central tendencies? What do they tell us?
Mean → Gives the average/central values but is subject to outliers
Median → Gives the middle value in a ranked dataset and is more robust to outliers in skewed data
Mode → Most common value in the data
How does distribution change central tendencies?
Positive Skew (Lots of lower values, few high outliers): Mode < Median <Mean
Normal: Mode = Median =Mean
Negative Skew (Lots of higher values, few low outliers): Mean < Median <Mode
What terms quantify data spread? What do they tell us?
Range: Largest - Smallest, subject to outliers
Standard Deviation (SD): Average distance from the mean, subject to outliers, AND SD = sqrt(variance)
Variance: Average squared distance from the mean, subject to outliers. Squared SD → Population and Sample Variance
Interquartile Range (IQR): Measures data spread in quarters, making it robust to outliers
Population vs Sample Variance
Population Variance: Divide variance by population size
Sample Variance: Divide by n-1, OR Sample size -1
What is in the IQR?
Q1: The lower 25 % of the data
Q2: The median value
Q3: The lower 75 % of the data
IQR: Q3-Q1
How to detect outliers mathematically in a normal distribution? Where does data fall within SD in a nromal distribution?
Max Outlier Threshold: Q1 - 1.5(IQR)
Min Outlier Threshold: Q3 + 1.5(IQR)
Within 1 SD: 68% of data
Within 2 SD: 95% of data
Within 3 SD: 99% of data
What plots can you use to visualize data patterns?
Boxplots: IQR and Data range
Histograms: Probability/Density Curves
Bars in histogram show the probability that x is between Value A and Value B
Usually follows a bell curve when you add more data
What is Standardization?
A method to convert original units into units of SD from the mean
Gives a z-score: The amount of standard deviations an observation is from the mean
Variables used to analyze population and sample distribution
Population: N (Pop Size), u (Pop mean), o (Pop SD)
Sample: n (Sample Size), x (Sample mean), s (Sample SD)
As you include more samples to your sample size (n), this will start to resemble pop mean and SD
Standard Error
SE = Sample SD/sqrt (n)
How close your calculated sample mean is to the “true” pop mean (precision and reliability of estimate)
Decrease SE by 1) Decreasing SD/more data clustered around the mean or 2) Increasing n (taking more data points
When sample size (n) increases, standard error should ..
DECREASE (lower SD/higher n)
What are the two parts of the Confidence Interval?
1) Mean
2) +/1 Margin of error (ME)
Margin of Error = Tcrit * SE
What does CI tell us?
If the sampling procedure was repeated many times, about 95% of the resulting intervals would contain the TRUE POP mean.
Doesn’t tell us anything about our sample data, just the pop data
What is the Margin of Error made of?
ME = tcrit * SE
T-crit is the critical value of t-distribution for a given confidence level and (n-1) degrees of freedom
Normal “t” distribution
Assumes you know pop SD which you have to estimate with sample SD
T- distribution incorporates degrees of freedom (uncertainty) with a normal distribution
Low n - > Low df → High uncertainty
High n → High df → Greater certainty
Comparing 2 independant samples (t-test)
H0: No sig difference between the means of 2 populations (within-group variability; t=0)
HA: Sig difference between the means of 2 populations (t>0)
T = Difference in Means (X1-X2)/ SE of difference between X1-X2
T-Value
t= (X1-X2)/SE of (X1-X2)
t= signal/noise
T is likely >0 when:
There is a big difference between X1 and X2
There is a little bit of noise or low SE (high n, low sd, etc)

P-Value for t-test
Probability of observing a difference between X1 and X2 assuming populations are the same (t=0; H0)
T-Test Assumptions
1) Independence: Each observation is separate
2) Normality: Bell-shaped curve
3) Continuous data
4) Equal Variance
Non-parametric t-tests (Mann-Whitney/Wilcoxon)
Use ranked values rather than actual values
Compares ranked position of all observations pooled from smallest to largest, sums up ranks for each treatment, and askes whether one group is higher than the other
P-value: Probability of observing a U-value as extreme as the one observed assuming populations are the same (u=0; H0)
Paired t-test
-Comparing the SAME units before and after treatment (NON INDEPENDANT)
For continuous data only
When to use Chi-Squared test?
For discrete data, measuring whether some data are likely to fall into a certain category
Measures percentage or ratio
Steps of Chi-Squared test of indepedance
Step 1: 2×32 contingency table
Step 2: Created expected values using total values across the population
Step 3: (observed- expected)²/(expected) for each value.
Step 4: Sum up all values above to get an x² value
Step 5: Compare to p-value, or probability of getting a x² as extreme as the one observed assuming they come from the same distribution
Null Hypothesis
H0: There is no effect/difference/relationship between X and Y
Usually when p> 0.05 by convention
p-value measures “ If the null is true, what is the probability that I would observe a result as extreme just by chance”
Type 1 Error
“False Alarm”
Claiming there’s a difference when there isn’t (falsely rejecting H0)
Usually set at 5%
Increases with multiple comparisions
Type 2 Error
“Failed Detection”
There’s a difference but we accepted H0
Statical power or the ability to find an effect = 1-b
Multiple Comparision
Each additional category adds a 5% risk of making a Type 1 error
Family-wise error rate (FWER) - Risk of making a “wrong call”
Correct by:
Looking at all groups at once (ANOVA)
Correct for multiple comparisons
Analysis of Variance (ANOVA)
Measure continuous outcomes across all groups
H0: All population means are equal
HA: All least one population mean is different from the others
What does the ANOVA F-stat represent?
F-stat = Between Group Variance/ Within Group Variance (Signal/Noise)
Between group variance (Signal): How far apart are group means? Are group clusters far apart?
Within group variance (Noise): How much spread inside each group? Spread out data?
Significance: Large F → High signal/low noise
Non-significant: Small f → Low signal/high noise
What do post-hoc tests do?
Adjusts significance threshold so the total chance of making Type 1 stays at 5%
Shows you what groups significantly differ after ANOVA
Types of Post-hoc tests and when to use them?
Tukey’s: Tests all pairwise comparisons when variance is equal
Bonferroni: Tests a small, pre-planned subset of comparisons or mixes different test types.
Dunnett’s: Compares multiple treatments against control
Games-Howell: Compares groups when variance is unequal