You'll have data, and will try to find the distribution that the values are following. You'll try to find properties of your distribution, check if this is likely that the distribution is having such properties, and predict how your data will evolve.

This is a course where you will learn

  • (hard) about estimators, likelihood (vraisemblance in French), bias, ...
  • (soft) about population, variables (qualitative, quantitative), ...
  • how you could analyze a file of data
  • plots used in statistics
  • tests
  • linear regression

And everything will be coupled with examples in R, you may add examples in other languages. To be honest, I do not understand much about statistics so be sure to correct any mistakes, thanks! Be sure to cross-check your findings with a teacher/friend.


The main idea is that you got some data, and you need to find how this data got generated (What distribution function? What were the parameters?...). Once we established a model, then we can try to guess the future.

  • Population: data we are observing, a matrix
  • Variables: the columns of our matrix, can be quantitative or qualitative
  • Individuals: the rows of our matrix
  • Sample (échantillon): a part of our population
  • Model: characterization of the dataset
  • Population (paired): you took two times the data from the same individuals
  • Empiric: you are calculating values using the sample

Math (summary)

Before the next part, you need to know of those

  • mean (empiric/arithmetic): sum of the values divided by the number of values
  • mean (weighted=pondérée): E(X) = sum of { value * frequency }
  • median: 50% are greater and 50% are lesser that the median. It's F(x)=0.5. (if F is the empirical cumulative distribution function)
  • medial (Médiale): value that divide in two, the cumulative sum of values.
  • mode: most repeated value (wiki)
  • quantile: split the distribution in X part (4 = quartile, 10 = decile, 100 = centile).
  • covariance: if two values are independents, then cov=0.
  • correlation: if two variables are correlated then when one increase, the other may variate according to the correlation
  • moments: $\mathbb{E}(X)$,$V(X)$, Skewness, and Kurtosis

In R, you can use mean(x), median(x), quantiles(x)=fivenum(x), cov(x) or cor(x). In the library modeest, you can use mlv to get the mode value.

Analysis of a sample

This section is a summary of what you will do, but not how you will do it, as it'll be explained in other sections.

  1. Discovery
  2. Descriptive statistics
  3. Statistical inference

Follow this link to learn about distributions in R

Estimators and likelihood

We are using the notation $\theta$ (theta) for the vector of the parameters of a distribution. For instance, a binomial distribution $B(n,p)$ would become $B(\theta)$. Because it's convenient, in statistics, we are always using theta while you may not see it often in probabilities. We are calling $\Theta$ (big-theta), the space in which $\theta$ is defined.

Your goal is to estimate the vector "theta", like "what parameters seem to have generated this distribution?". We are calling estimator $\hat{\theta}$ (theta-hat), the value that is likely the value of $\theta$.

Finally, we are calling Likelihood function $L(\hat{\theta})$ (Fonction de vraisemblance) the function returning a value as to how much it seems $\hat{\theta}$ is equals to $\theta$. You will have to maximize this function to find the best $\hat{\theta}$, and that's called maximum likelihood estimation/value (maximum de vraisemblance).

According to the fitdistr function source code, I was able to learn more about estimators. You can find the source code here. I will use my newly earned knowledge a lot in the two next pages, but I don't know how did they learn that, unfortunately...

Other notes


In descriptive statistics, you will have to use plots and try to guess the distribution. You got all of them introduced in the R course.

  • plot (plot): simply plot your values
  • histogram (hist): used to see the proportion of the distribution (=frequency, if prob=T) or the number of individuals taking a value, for each value
  • contingency table (table): make a table of the unique values per the number of occurrences
  • pivot table (qhpvt): you can use this to observe a variable with more depth, especially observing the relation of your variable and others variables.
  • Bar chart: you can see the distribution of a quantitative variable split by group according to a qualitative variable
    • for instance, you can see the number of students per year
    • let d be a data.frame and d\$year be the year (ex: 2020) the student joined the school
    • then the call would be barplot(table(d$year))
    • you may add beside=T for xtabs barplot(xtabs(....), beside = TRUE)
  • Box plot: same as BarChart but you can see the quantiles, the min, and the max!
    • for instance, you can see the horsepower of a car per the number of cylinders
    • let d be a data.frame and d\$hp (horsepower), d\$cyl (cylinders) our variables
    • the call would be boxplot(d$hp ~ d$cyl)


We are using tests to check if it seems likely that a parameter has a value, the distribution has a property, etc. We are considering two hypotheses $H_0$ and $H_1$, and we are testing which one is more likely.

  • $H_0$, null Hypothesis: we believe this is true
  • $H_1$, alternate Hypothesis: what we want to demonstrate

A test could be

  • two-tailed test (test bilateral): $H_0 = \theta_0$, $H_1 \neq \theta_1$
  • left-tailed test (test unilateral gauche): $H_0 = \theta_0$, $H_1 \lt \theta_1$
  • right-tailed test (test unilateral droit): $H_0 = \theta_0$, $H_1 \gt \theta_1$
  • fitting test (test d'adequation): $H_0 \sim L_1$, $H_1 \sim L_2$ with $L_1$ and $L_2$ two distributions
  • ...

We may introduce two errors while picking

  • Type I error (risque de première espèce/seuil), $\alpha$
    • accepted $H_1$ but $H_0$ was true
    • $\mathbb{P}(reject\ H_0 | H_0\ true)$
  • Type II error (risque de seconde espèce), $\beta$
    • accepted $H_0$ but $H_1$ was true
    • $\mathbb{P}(reject\ H_1 | H_1\ true)$

We are calling the probabilities of the errors above

  • $1-\alpha$: confidence coefficient (Niveau de confiance)
  • $1-\beta$: power of a test (Puissance d'un test)

We are calling critical region or region of rejection $W$ (zone de rejet), the set of values for which the null hypothesis is rejected. Hence, the region covered by the null hypothesis is called the region of non-rejection. We are also calling it the critical region, because the bounds of the regions are called critical values. To find the critical region, you need to find a statistical test (test statistique) $T(x_1, ..., x_n) = T(x)$ which is a function taking a sample and returning a hypothesis. Of course, if you found $W$, then you don't have to look for $T$.

A test is successful if the result is in $W$. If that's the case, then $H_0$ is rejected. Otherwise, we do not accept $H_0$, and instead, we say that we can't reject $H_0$ at a significance level (seuil alpha).

@ R(x) = \frac{L(x|\theta_1)}{L(x|\theta_0)} @

@ W = \{ x; R(x) \gt k \} =^\text{some operation} W = \{ x; T(x) \gt c \} @

with $k$ a constant giving us

@ \mathbb{P}_{H_0}(W) = \mathbb{P}_{H_0}(T(x) \gt k) = a @

and $c$ a critical value. You can also see

@ W = \{ x; |T(x)| \gt c \} \quad or \quad W = \{ x; c_1 \lt T(x) \lt c_2 \} @

In R (or most of the time), computers are calculating a p-value,

  • according to Neyman–Pearson lemma
    • if $\text{p-value} < \alpha$: reject H0
    • else: accept
  • according to Fisher
    • the more $p$ is small, the more you can trust the result
    • if $\alpha=0.05$, with Neyman–Pearson $0.049$: H0 is rejected while $0.050$ is accepted, so Neyman–Pearson is less permissive than Fisher.

This is too complex for me, I made too many mistakes, so I removed everything. You may add back what you know (examples, well-known tests) or links.


The linear regression (Regression linéaire) is a line maximizing the distance between each point of the distribution and the line (minimizing the sum of the square vertical distances between our line and a point)

The linear regression equation is $Y = a + b X + residual$. If $Y \not\in \mathbb{R}$ but in $Y \in [0,1]$, then the residual is $0$, and you should look at logistic regression.


This is a list of all Wikipedia pages that you may want to check

Estimators and likelihood

other references