When to average in the lab for indirect measurements? I am measuring distance and time to calculate the velocity. I repeat the experiment 10 times. What is better, to first calculate the average of distance and time and then the velocity, or to calculate the velocity ten times and then average the velocity?
I think the first option is better, but a colleague of mine usually does it the other way around.
EDIT: Some more information. The experiment is measuring the speed of sound. The students record a snap of their fingers and time how long it takes the echo to come back. We are using a 2m cardboard tube, so we only have one measurement done 10 times, and we cannot do a linear regression. 
 A: Calculate velocity ten times and then average it. If you were looking for average distance, then average the distance, if you were looking for average times, then average the times. 
What you are doing doesn't even give the same number as the average velocity:

A: Averaging rates is tricky.
To see why, let's use a different example

You're driving a car over a hill.  For the first mile, you're going up the hill and you use 1/20th of a gallon of gas.  On the second mile you coast down the hill using 1/100th of a gallon.  What is your average gas mileage?

For the first mile, you were traveling at 20 mpg, and for the second mile, you were traveling at 100 mpg.  So your average is 60 mpg, right?
Wrong!
Let's try a slightly different problem.

You're going up the same hill again.  Getting to the top requires another 1/20th of a gallon, but this time you turn off your car at the top and glide down to the bottom without using any gas at all.  What is your average mpg this time?

So let's see... add 1 mi / .05 gal + 1 mi / 0 gal, then divide by 2... good news! Your average gas mileage was infinite! You can travel anywhere in the world without ever running out of gas, and all you need to do is start on a hill!
Obviously, this result is ridiculous.
Summing first, and then dividing  ( 2 mi / .05 gal = 40 mpg ) gets us a far more useful answer.
To understand why, I'll leave you with two final problems.

Bob went for a drive.  He drove for 1 mile at 30 mph, then he drove 2 miles at 60 mph.  What was his average speed?
Alice went for a drive.  She drove for 1 hr at 30 mph, then she drove for 2 hrs at 60 mph.  What was her average speed?

Now consider: if you have two different trials that took different lengths of time, which one will contribute the most to your result? Is that what you want?
There's no such thing as a "right" way to summarize data.  The best you can do is to understand how your summarization formula affects the results.
A: Short answer: Depends.
Ideally you want to use the "minimum-variance unbiased estimator" given your data and assuming both systematic and random errors. In some cases it won't matter when you average, but generally the bias and the variance may change.
For example consider the formula:
$$ c = \sqrt{a^2 + b^2} $$
If you want to calculate $c$, given samples of $a$ and $b$ and being $a$ and $b$ fixed variables during the whole process. In this case is much better to average $a$ and $b$ before and then calculate $c$ to reduce the bias.
If we instead average different values of $c$, As the errors of $a$ and $b$ are squared they will be always positive, consider the case when a and b are zero and we will still get a non zero $c$ on average.
Now consider a system where we are measuring $c$ but $a$ and $b$ are uniformly distributed. Then averaging $a$ and $b$ before will be a terrible estimator as  $a$ and $b$ may approach zero when averaged. In this case we will reduce the bias averaging after calculating different values of $c$.
The truth
The way to calculate the best possible value of $c$ even more complicated.
What you can do is make a propagation of errors under certain assumptions and calculate the bias of your formula.
For instance assume normal distributions of the errors of a and b independence of the measurement of them. Then I can calculate the bias of calculating $c^2$ making the average after. The bias is the variance of $a$ and $b$ (don't forget Bessel's correction). Which lead to the following unbiased estimator of $c^2$
$$  <c^2> = <a^2 + b^2> - var(a) - var(b)^2 $$
Notice that this is as far as we can get with analytic closed expressions as we do have an unbiased expression for $c^2$, but not for $c$, for the same reason as we can calculate the unbiased variance but not the unbiased standard deviation, but you can still calculate $c$ mitigating the errors with some approximations.
When does this matter?
Fortunately for linear formulas and for the specific case of the question all this complexity can be drastically simplified, but for a general question as "When to average in the lab for indirect measurements?" then I think this is the correct answer.
As an additional note, the suggested procedure of making a "linear" regression for your case, or in general a regression constructed to minimize the squared mean error is often a very good strategy to find a good estimator for the dependent variable.
A: Averaging destroys information. Do it as late as is practical in your analysis.
A commenter points out that, if you are making many position/velocity measurements of the same object as it moves once, a simple linear regression is more robust as a velocity estimator than an ensemble of point-by-point analyses.
A: My original intuition was this: 

If you measured $x$ and $t$ once for each trial, then in your case $v$ is a function not a measurement. Every measurement generates its own pair of $x_i$ and $t_i$, which implies $v_i$. Thus you get $\hat v$ by averaging $v$. $x$ and $t$ depend on each other, as opposed to $x_i$-$x_j$, so it doesn't make sense to average $x_i$.
If you doubted your ability to precisely measure $x$ and $t$ (as well you should) then in each trial $i$ you would make $n$ measurements of $x_{ij}$. Now before you can do $v_i=x_i/t_i$, you must first have an estimate of $x_i$, which is given by $\frac{1}{n} \sum_j{x_{ij}}$. Thus you average $x$ for that trial, ditto $t$, get a $v_i$, repeat for each trial, and then average $v_i$.

However, I tried to simulate it and was surprised to find that it seems like there is hardly any difference. Below are sections of my R notebook along with results. Maybe I have a bug?

pacman::p_load(tidyverse, ggplot2)

Function to simulate an imprecise measurement:
# Relative measurement error
em = 0.01

measure = function(x, n) {
  # Attempt to get the value of a quantity x, and n measurements
  x_measured = mean(x*rnorm(n, 1, em))
  return(x_measured)
}

Let's test it with some simulated measurements:
df = expand.grid(case=1:3, measurement=1:5)
df$result = Vectorize(measure, vectorize.args = 'x')(replicate(nrow(df), 1), 1)

p = ggplot(df) + 
  geom_point(aes(y=result, x=as.factor(case))) +
  theme_classic()
show(p)


We expect repeated measurements to converge, but with diminishing returns:
df = expand.grid(case=1:1000, m=1:50)
df$result = Vectorize(measure, vectorize.args = 'n')(1, df$m)

df_sum = df %>% group_by(m) %>% summarise(mu = mean(result), sigma = sd(result), measurements = first(m))

p = ggplot(df_sum) + 
  # geom_boxplot(aes(y=result, x=measurements, group=measurements)) +
  geom_pointrange(aes(x=measurements, y=mu, ymin=mu-sigma, ymax=mu+sigma)) +
  theme_classic()
show(p)


We set up the ball experiment: The ball thrower has a velocity setting (the exact value of which we don't know) and the camera has a timer setting (the value of which we also don't know). So both v and t converge to one value, but have some experiment error. We try doing several experiments:
# Set ball roller to a certain energy level (value unknown to you)
v_expected = 5

# And the camera to photograph after a certain delay (value unknown to you)
t_expected = 20

# Relative apparatus error
ea = 0.001

# Number of trials
n = 10

# Number of experiments
k = 1000

df = expand.grid(experiment = 1:k, trial=1:n)

# Roll the ball for each trial, but the machine is slightly faster or slower sometimes
df$v_actual = v_expected*rnorm(nrow(df), 1, ea)

# The camera's timer isn't very consistent either
df$t_actual = t_expected*rnorm(nrow(df), 1, ea)

# We don't know the true distance yet, but nature does
df$x_actual = df$v_actual * df$t_actual

# Visualize
p = ggplot(df) + 
  geom_point(aes(x=x_actual, y=t_actual, color=v_actual), size=0.5) +
  theme_classic()
show(p)


Now we try measuring our experimental outcomes:
# You try to measure the distance, but your ruler isn't very accurate
df$x_measured = Vectorize(measure, vectorize.args = 'x')(df$x_actual, 1)

# You also tried to measure the time, but your stopwatch is not the best
df$t_measured = Vectorize(measure, vectorize.args = 'x')(df$t_actual, 1)

# Number of repeat measurements
m = 20

# What if you measured x multiple times?
df$x_hat = Vectorize(measure, vectorize.args = 'x')(df$x_actual, m)

# And had multiple assistants, each with their stopwatches?
df$t_hat = Vectorize(measure, vectorize.args = 'x')(df$t_actual, m)

Of course multiple measurements are much better:
df_sum = df %>% gather(key='method', value='measurement', c(6,8)) %>%
  mutate(error=abs(measurement-x_actual)/x_actual)

# Visualize
p = ggplot(df_sum) + 
  geom_boxplot(aes(x=method, y=error)) +
  theme_classic()
show(p)


But does it matter when we average?
df_sum = df %>% mutate(v_measured = x_measured/t_measured) %>% group_by(experiment) %>% 
  summarise(v_bar = mean(v_actual),
            t_bar = mean(t_actual),
            x_bar = mean(x_actual),
            v_bad = mean(x_measured)/mean(t_measured),
            v_good = mean(v_measured),
            v_best = mean(x_hat/t_hat),
            early = abs(v_bad-v_expected)/v_expected,
            late = abs(v_good-v_expected)/v_expected,
            multiple_measurements = abs(v_best-v_expected)/v_expected) %>% 
  gather(key='method', value='rel_error', 8:10)

df_sum$method = fct_inorder(df_sum$method)

# Visualize
p = ggplot(df_sum) +
  geom_boxplot(aes(x=method, y=rel_error)) +
  theme_classic()
show(p)


A: I assume you're talking about a straight line (so that by velocity you just mean speed)?
In all honesty I'm open to being wrong about this, but to me your method is what makes sense.
My reason is that repeatedly flying a distance 10 times is the same as flying 10 times the length of that distance only once, and hence I'd expect you to get the same average speed for both. Otherwise you're saying that the average speed over a longer distance can be a different value depending on how you choose to interpret chopping it up, which doesn't really make sense to me.
You can try to figure out which one has more error by working it out by the way... I think Markov or Chebyshev's inequalities could be helpful if you assume bounded moments etc.
But if you're like me and this doesn't sound too fun to you, just simulate a bunch of measurements using Gaussian noise with code and see which estimate ends up being more accurate. I tried to quickly whip up some code and the first one seemed to be more accurate.
A: Calculate your velocities and then report the following:


*

*Average (mean, median, or mode. Whatever makes sense for your application)

*Spread (Range, stdev, 95%CI, etc... Again which one depends on your application)

*Number of samples (in this case, n=10)


If you leave out any of the above, your number becomes less useful for the next person looking to build off of your work.
Since its only 10 measurements, it is easy enough to include them at the end in a supplement. That way if someone is unhappy with the values you reported, they can calculate their own charaters
