Let $ \mathcal{L} $ be a language for first-order logic whose logical primitives are $ \neg$, $\vee$, $\wedge$, $\forall$, and $\exists$, with the usual formation rules. A sentence $ \sigma $ is then a (well-formed) formula with no free variables, and a positive sentence is a sentence that does not contain $ \neg $.
A reduced product $ \mathfrak{A} $ is defined as usual using an index set $ I $, a collection of models $ M_{i} $ indexed by $ I $, and a filter $ F $ on $ I $. If $ F $ is an ultrafilter, the product is an ultraproduct.
Łoś's Theorem concerns ultraproducts. One of its consequences is that for any sentence $\sigma$, $ \mathfrak{A} \models \sigma $ iff $ \lbrace i \: \epsilon \: I \: \vert \: M_{i} \models \sigma \rbrace \: \epsilon \: F $. This does not hold for reduced products in general
But the usual proof of Łoś's Theorem proceeds by induction on the construction of formulas. It makes use of the hypothesis that $F$ is an ultrafilter only in the clause that deals with $ \neg $. Since a positive sentence contains no negation symbols: For any reduced product $ \mathfrak{A} $ and any positive sentence $ \sigma $, $ \mathfrak{A} \models \sigma$ iff $ \lbrace i \: \epsilon \: I \: \vert \: M_{i} \models \sigma \rbrace \: \epsilon \: F $.
That's right, isn't it? Or did I miss something?
This is a surprisingly subtle issue. My previous answer was incorrect, and I've deleted it.
Let's start by fixing terminology.
A positive sentence is built up from (positive) atomic formulas using $\land$, $\lor$, $\forall$, and $\exists$ (but not $\lnot$).
A basic Horn formula is of the form $\psi_1\land\dots\land\psi_n \rightarrow \theta$, where the $\psi_i$ and $\theta$ are (positive) atomic formulas.
A Horn sentence is built up from basic Horn formulas using $\land$, $\forall$, and $\exists$ (but not $\lnot$ or $\lor$).
Let $I$ be an infinite set indexing a collection of structures $\langle A_i\rangle_{i\in I}$, $D$ a proper filter on $I$, and $A = \Pi_DA_i$ the reduced product.
Say a sentence $\phi$ is weakly preserved under reduced product if $\{i\,|\,A_i\models\phi\} = I$ implies $A\models \phi$.
Say a sentence $\phi$ is strongly preserved under reduced product if $\{i\,|\,A_i\models\phi\}\in D$ implies $A\models \phi$.
Say a sentence $\phi$ is preserved under reduced factors if $A\models\phi$ implies $\{i\,|\,A_i\models\phi\} \in D$.
Your question asked whether Los's theorem holds for positive sentences in reduced products, i.e. whether every positive sentence is strongly preserved under reduced product and preserved under reduced factors. In fact, positive sentences are preserved under reduced factors, but not necessarily under reduced product (strongly or weakly).
Surprisingly, the problem occurs with the $\lor$ case. Here's an example. Let $L = \{c, P,Q\}$, where $c$ is a constant and $P$ and $Q$ are both unary relation symbols. Define $\langle A_i\rangle_{i\in\mathbb{N}}$ as follows: all $A_i$ consist of just one element, $c$. If $i$ is even, $A_i\models P(c)\land \lnot Q(c)$, but if $i$ is odd, $A_i\models \lnot P(c)\land Q(c)$. Now let $D$ be the cofinite filter on $\mathbb{N}$. Then the reduced product $A$ consists of just one element, $c$, and $A\models \lnot P(c) \land \lnot Q(c)$. So the sentence $P(c)\lor Q(c)$ is false in $A$, despite holding in all of the $A_i$.
Okay, so $\lor$ isn't preserved under reduced products, but it turns out that the other operations are okay (in particular, both quantifiers are). In fact, we can expand our basic building blocks up from atomics to basic Horn sentences and still get the strong preservation under reduced product.
Here's a summary of what's known, as far as I know (most can be found in C+K Section 6.2, recently reprinted by Dover!)